Download as pdf or txt
Download as pdf or txt
You are on page 1of 18

Remote Sensing of Environment 290 (2023) 113529

Contents lists available at ScienceDirect

Remote Sensing of Environment


journal homepage: www.elsevier.com/locate/rse

Estimating and mapping forest age across Canada’s forested ecosystems


James C. Maltman a, *, Txomin Hermosilla b, Michael A. Wulder b, Nicholas C. Coops a, Joanne
C. White b
a
Department of Forest Resources Management, Faculty of Forestry, University of British Columbia, 2424 Main Mall, Vancouver, BC V6T 1Z4, Canada
b
Canadian Forest Service (Pacific Forestry Centre), Natural Resources Canada, 506 West Burnside Road, Victoria, BC V8Z 1M5, Canada

A R T I C L E I N F O A B S T R A C T

Edited by Marie Weiss Forest age is an important variable for assessments of biodiversity and habitat, sustainable forest and land
management, as well as forest carbon science and modeling. Tree and stand age are typically measured directly
Keywords: on site, or estimated through visual photo interpretation, with spatially explicit maps of forest age not often
Forest age produced over large areas. Remote sensing enables the generation of wall-to wall, spatially explicit maps of
Forest inventory
disturbance events within the satellite record; however, as disturbance is relatively rare on the landscape in a
Land cover
given year, additional means of determining forest age are required. As reviewed herein, the estimation of forest
Landsat
Time since disturbance age using optical Earth observation data is challenging due to the limited spectral link to the attribute of interest,
especially as forests get older. The temporally dictated multi-method approach to forest age estimation outlined
herein acknowledges these limitations, by applying the approach that is best suited to the quality of the infor­
mation available, depending on the epoch of interest. In this research, we combine three approaches to estimate
forest age at a 30-m spatial resolution using Landsat data. The first approach uses change detection protocols to
detect disturbance from 1985 to 2019, with time since disturbance used as a proxy for forest age. The second
approach uses Landsat surface reflectance composites to identify pixels exhibiting evidence of recovery from a
disturbance that occurred within the twenty years prior to 1985, allowing for the extension of forest age esti­
mates to 1965. Finally, given an understanding of the linkage between forest age and canopy height, inverted
allometric equations are coupled with maps of forest structure and productivity metrics to model forest age for
those pixels that show no evidence of disturbance or recovery to a maximum of 150 years, acknowledging that
uncertainty in age estimate increases with increasing age. Combining these three approaches, forest age esti­
mates are made for every treed pixel found within the 650 Mha forested ecosystems of Canada. Nationwide,
mean estimated forest age for forests ≤150 years old (representing 94.1% of treed area) was 70 years (standard
deviation = 32.1 years). For confidence building, forest age estimates were compared to reported forest age in
the National Forest Inventory (NFI) both spatially and aspatially. Nationally, 5.9% of the forested area was
estimated to be older than 150 years, while 9.5% of area within in the NFI sample was recorded as older than
150 years. The median estimated forest age for forested pixels ≤150 years old was 68 years while median forest
age reported in the NFI was 73 years, with regional variability matching expectations related to disturbance
regimes and productivity. Spatially explicit maps of forest age provide important information for understanding
forest ecosystems and can be used to inform a wide range of policy, science, and management needs.

1. Introduction et al., 2011a), ecosystem services (Costanza et al., 1997), structure


(Franklin et al., 2018), and wildlife habitat (Conner and Dickson, 1997;
Accurate knowledge of forest age is critical for informing effective Ecke et al., 2002; Russo et al., 2010). It has additionally been linked to
forest management (Franklin et al., 2018), regional or national reporting both wood quality (Wylie et al., 2019) and volume (Scott and Voorhis,
(Gillis et al., 2005), and as a variable for predicting a variety of forest 1986), and is informative for fire management (Boulanger et al., 2017)
attributes (e.g., site index, mean annual increment). Forest age has been and fire behavior modeling (Finney, 1993). Further, knowledge of lo­
shown to be an important predictor of forest carbon sequestration (Pan cations with older forests allows for targeted investigations or policies,

* Corresponding author.
E-mail address: jmaltman@student.ubc.ca (J.C. Maltman).

https://doi.org/10.1016/j.rse.2023.113529
Received 21 November 2022; Received in revised form 7 February 2023; Accepted 27 February 2023
Available online 15 March 2023
0034-4257/Crown Copyright © 2023 Published by Elsevier Inc. This is an open access article under the CC BY-NC license (http://creativecommons.org/licenses/by-
nc/4.0/).
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

including those related to ecosystem integrity and risk (Rogers et al., airborne laser scanning (ALS) forest structural metrics have also been
2022). used to infer forest age due to strong relationships between forest age
Regional to national characterizations of attributes such as stand age and stand structural information captured by ALS derived point clouds
are commonly included within jurisdictional forest inventory programs, (Sanchez-Lopez et al., 2019; Schumacher et al., 2020; Wylie et al.,
either as maps or samples depending on the nature and scope of the 2019). When interpreting the literature summary presented in Table 1,
program (Kangas and Maltamo, 2006). From an initial characterization recall that the studies represent different spatial extents, diversity of
perspective, forest age estimation poses a number of difficulties. forest maturity and structure, number, type, and distribution of plots, as
Ground-based methods such as tree coring require manual collection of well as differences in methods and quality assurance protocols. To
samples, limiting the total area over which they can be implemented further aid in interpreting the summary, we note that tree cores are
(Metsaranta, 2020). While generally regarded as an accurate measure of understood to be the most accurate method of estimating tree age, yet
tree age, coring can be subject to uncertainty from a number of sources even this method is known to have precision ranging from 10 to 20 years
including the time it takes for a tree to reach breast height (which varies across a range of forest environments (Fraver et al., 2011; Koch et al.,
by species), missing information from cores that do not strike pith (often 2008; Metsaranta, 2020). In western Canadian boreal forests for
in irregularly shaped trees, or those with significant heartwood rot), and example, Metsaranta (2020) found that standard ring counts used in
the miscounting of rings caused by false, missing, or hard to see rings forest inventories typically underestimated age by an average of 9.3
(Metsaranta, 2020). Critically, ring counts are known to underestimate years.
both tree and stand age, as well as time since the most recent stand Overall, as shown in Table 1, estimation of forest age over large areas
replacing disturbance (Gutsell and Johnson, 2002; Metsaranta, 2020; has been typically undertaken at coarser spatial resolutions (e.g. 1 km),
Vasiliauskas and Chen, 2002). To reduce logistical and cost issues, with finer resolutions (i.e. <20 m) reserved for smaller study areas. ALS
photo-interpretation is also used to estimate forest attributes, including and satellite imagery are the two most common data sources for age
stand age, over relatively larger spatial extents, especially in a Canadian estimation, and the most common modeling approaches include linear
context (Gillis et al., 2005); however, photo-interpretation is qualitative regression and Random Forests.
in nature, and relies on the subjective experience and skill of individual Identified challenges in estimating forest age with optical Earth
photo-interpreters, and can therefore be prone to bias and error. Indeed, observation data include availability of cloud-free imagery, limited
Magnussen and Russo (2012) identified age and species composition as spectral variation in undisturbed areas, as well as the disparity between
the two forest inventory attributes in Canada’s National Forest In­ time since disturbance and age (Bradford et al., 2008). The integration
ventory (NFI) that were most sensitive to photo-interpretation uncer­ of remotely sensed spectral information and derived products offers
tainty. Nationally consistent information on forest age in Canada and opportunities to overcome some of these considerations and to proto­
elsewhere (Kangas and Maltamo, 2006), is therefore typically limited in type fine grained estimates of forest age over large spatial extents. In this
spatial detail and extent, and not of a uniform vintage or consistency. research, we aimed to map forest age at a 30-m spatial resolution over
Satellite-based remote sensing technologies provide a cost-effective the forest-dominated ecozones of Canada. To achieve this goal, we
method to obtain information regarding the Earth’s surface with a availed upon a number of national datasets derived from annual Landsat
relatively fine level of detail (Lechner et al., 2020) that is informative of surface-reflectance best available pixel composites, including estimates
human activities (e.g., forest harvesting, urbanization) or natural pro­ of canopy height and forest cover (Matasci et al., 2018a), disturbance
cesses such as wildfire (Wulder et al., 2022). Free and open access to the information (Hermosilla et al., 2016, 2017), and dominant tree species
Landsat archive (Woodcock et al., 2008), coupled with increases in (Hermosilla et al., 2022a). Given an understanding of the data available
computing power, has led to marked advances in land cover classifica­ and age inference and estimation approaches, we partition the forested
tion (Wulder et al., 2020), change detection (Hansen and Loveland, area of Canada and apply three approaches to either directly estimate,
2012; Schroeder et al., 2017), and forest structure mapping (Matasci infer, or model, forest age. Recognizing that forest age can be defined in
et al., 2018a, 2018b). These advances can be leveraged in conjunction a number of ways, including age from stand establishment, age from
with other environmental data, and forest plot information to provide seed, breast− height age, and time since disturbance (stand replacing),
estimates of forest age over a range of scales and quality. Notably, me­ herein we define age as time since disturbance when that is known.
dium spatial resolution (10 to 100-m pixel) Earth observation data, such Otherwise, we define age as analogous to that of forest inventory: the
as from Landsat or Sentinel-2, offers the potential to generate manage­ mean age of trees over a given unit area. Time since disturbance rep­
ment and policy relevant forest age estimates over large areas in a resents the amount of time since an area underwent a stand replacing
transparent and systematic manner. disturbance. As a result of differing definitions, some variability can be
A variety of approaches have previously been used to predict and found in the estimation of forest age. Given the inherent difficulties in
map forest age, including change detection, regression, machine the estimation of forest age, we link the methods used to the results
learning, and allometric equations (Table 1). Spatially explicit mapping obtained and offer implications regarding methods, definitions, and the
of forest disturbance is enabled by the use of calibrated surface reflec­ quality of the estimated forest age outcomes.
tance values from satellite-based datasets. Temporal sequences of
spectral values (i.e. spectral trajectories) can be used to accurately detect 2. Study area
the occurrence of forest disturbances within the satellite record, and
thus time since disturbance, often used as a proxy for forest age (Diao Canada’s forested ecosystems (~650 Mha in area) are comprised of a
et al., 2020; Gu et al., 2016; Hermosilla et al., 2015a). However, age mixture of trees, wetlands, shrubs, and lakes (Wulder et al., 2008).
estimation of forests that are not disturbed within the satellite record is Forest composition varies as a result of topographic, climatic, and
more challenging, with a number of approaches being proposed across a disturbance drivers, among others. To capture this diversity, Canada’s
range of spatial scales utilizing various remote sensing datasets. Such forested ecosystems are subdivided into 12 ecozones (Fig. 1). Ecozones
approaches involve using spectral indices derived from satellite imagery are broad, generalized ecological units which contain similar geo­
(Diao et al., 2020; Li et al., 2014; Zhang et al., 2014) as input variables to morphology, soils, vegetation, and climate (Ecological Stratification
predictive models generated via machine learning (Besnard et al., 2021; Working Group, 1995). Average forest age varies between ecozones due
Reyes-Palomeque et al., 2021; Sanchez-Lopez et al., 2019), or para­ to biological and disturbance related factors (Natural Resources Canada,
metric regression approaches (Kayitakire et al., 2006; Maltamo et al., 2021). Wet, more productive, southwestern coastal forests are home to
2020; Schumacher et al., 2020) to predict forest age. Pre-existing forest the majority of Canada’s oldest forests, which are typically coniferous.
disturbance records have also often been used as an input variable for Drier forests tend to have more frequent wildfire, resulting in younger
forest age prediction (Gu et al., 2016; Pan et al., 2011b). More recently, forests comprised of broadleaf and mixedwood stands (Natural

2
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

Table 1
Summary of selected studies using remotely sensed data to estimate forest age. Acronyms used in this table: AGB: above ground biomass, ALS: airborne laser scanning,
FIA: forest inventory and analysis, kNN: k nearest-neighbors, LEDAPS: Landsat ecosystem disturbance adaptive processing system, MTBS: monitoring trends in burn
severity, NFI: national forest inventory, RMSD: root mean square deviation, RMSE: root mean square error, rf-KNN: random forests-k nearest-neighbors, VCT:
vegetation change tracker.
Study Location Study area size Spatial Age Data source Modeling approach Validation data Validation
resolution range and approach results
(years)

Kayitakire Plantations in 7900 ha 1m 0–110 IKONOS-2 imagery (1 m Linear regression Leave-one-out RMSE: 0.164
et al., 2006 Eastern resolution) Field based age cross-validation (log
Belgium Stand level map measurements transformed
denoting plantation date age)
Schumacher Norway 18.2 Mha 16 m 0–250 Sentinel 2 imagery Independent linear Comparison to RMSE: 19–56
et al., 2020 (10–20 m resolution) regression models NFI age (n = 63) years
Plot level age data from
NFI, determined via tree
cores (n = 4829)
ALS
Predicted Site Index map
Tree species map
Wylie et al., Ontario, 1.2 Mha 20 m 0–120 ALS rf-kNN Reported model RMSD: 15
2019 Canada Sample plots with age performance on years, R2:62%
determined by tree cores training data
(10 per plot, n 134)
Frate et al., Central Italy 128,000 ha 20 m 0–200 IRS LISS III imagery (20 kNN, Comparison to RMSE: 15.8
2016 m resolution) Inverted allometric manually years
Plot-level forest equations measured tree age r: 0.93
inventory data, age
defined by time since last
harvest
Véga and St- Quebec, 2025 ha 20 m 0–250 ALS/Stereo photography Inverted allometric Withheld plots RMSE: 7 years
Onge, 2009 Canada Manually sampled plots equations in single
3 trees per plot species stands
destructively sampled
and growth rings
counted (n = 23)
Sanchez- Idaho, USA 52,000 ha 30 m 0–142 ALS Random Forests Comparison to RMSE:17.5
Lopez et al., Timber harvest records stands with years
2019 (polygon based) known age from Bias: 0.08
Burned area dataset (fire harvest/ fire years
atlas records, aerial
photography)
Diao et al., Lishui, China 1.8 Mha 30 m 0–70 Landsat imagery (30 m VCT algorithm Comparison to RMSE:
2020 resolution) Random Forests sample plots 4.85–5.55
Plot-level Forest withheld as years in single
inventory age data based validation data plantations
on diameter (n = 1223)
Gu et al., Washington Entirety of 30 m 0–200 North American Forest Identification of Aspatial No summary
2016 and Oregon, Washington and Dynamics disturbance disturbed pixels comparison of metrics
USA Oregon polygons, Inverted allometric age distribution reported
Monitoring Trends in equations to FIA records and
Burn Severity (MTBS) Pan et al., 2011b.
polygons,
Aerial Detection Survey
Polygons,
National Biomass and
Carbon Dataset biomass
map
Li et al., 2014 Daxinganling 240,000 ha 30 m 0–170 Landsat imagery (30 m) Linear/ nonlinear Comparison to RMSE
Forest, China Stand level forest regression NFI 10.1–21.3
inventory data referring Neural Network years in single
to disturbance history species stands
R2 0.5–0.6
Besnard et al., Global Global 1 km 0–300 Plot-level forest age and Random Forests Cross-validation RMSE: 47.64
2021 AGB from various years
sources (n = 25,000), NRMSE:
Landsat tree cover and 51.42%
disturbance (Hansen
et al., 2013), Climate
data (WorldClim)
Zhang et al., China Entirety of China 1 km 0–300 Published plot-level field Inverted allometric Comparison to RMSE: 11–21
2014 inventory data from equations NFI years
number of different R2: 0.16–0.32
studies (n = 3543)
1 km resolution species
map
Global 1-km canopy
height map (Simard
et al., 2011)
(continued on next page)
3
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

Table 1 (continued )
Study Location Study area size Spatial Age Data source Modeling approach Validation data Validation
resolution range and approach results
(years)

Pan et al., Canada and Entirety of 1 km 0–125 Canadian and American Disturbance No validation No validation
2011b USA Canada and USA national forest inventory detection, recorded reported reported
data disturbance,
SPOT-VEGETATION and extrapolation from
NASA LEDAPS forest inventory
Canadian large fire age
database and American
MTBS polygons
Reyes- Yucatan 70,300 ha Object-based Four age SPOT-5 multispectral Random Forests 50% of sample Overall
Palomeque Peninsula, MX classes imagery (10 m sites withheld accuracy:
et al., 2021 to 50 resolution) 88.4–91%
years Purpose collected plot-
level forest succession
class and structure
measurements (n = 362)
Xu et al., New Zealand 51,237 ha Object-based 9–30 Plot level inventory data kNN, 10-fold cross RMSE:
2018 RapidEye imagery (5-m Random Forests, validation 2.05–8.73
spatial resolution) Multiple linear years
ALS regression,
Seemingly
unrelated
regression
Maltamo Finland Model designed N/A, only a 0–100 ALS Linear Regression Leave-one-out RMSE: 14
et al., 2020 to be applicable model was Stand-level national cross validation years
to all of Finland, generated, no forest inventory data
10 survey areas spatially including soil type, site
actually studied explicit output fertility, land use

Fig. 1. Map of Canada’s forest-dominated ecozones and treed area as of 2019 (Hermosilla et al., 2022b). Forest-dominated ecozones: Atlantic Maritime (AM), Boreal
Cordillera (BC), Boreal Plains (BP), Boreal Shield East (BSE), Boreal Shield West (BSW), Hudson Plains (HP), Montane Cordillera (MC), Pacific Maritime (PM), Taiga
Cordillera (TC), Taiga Plains (TP), Taiga Shield East (TSE), and Taiga Shield West (TSW).

4
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

Resources Canada, 2022). Natural disturbance return intervals vary target temporal window for the years 1984 to 2019, in conjunction with
widely across ecozones, from fire adapted forests in areas of the Boreal pixel scoring functions, are used to create best available pixel surface
Shield West ecozone having fire return intervals on the order of 50–100 reflectance image composites. The pixel scoring functions serve to rank
years, to cool, wet, forests in the Pacific Maritime ecozone having return and enable selection of the optimum annual pixel observation to be used
intervals on the order of thousands of years (Coops et al., 2018). The in an image composite based on sensor type, day of acquisition, distance
history of fire suppression in many of these areas however, can often to cloud or cloud shadow, and atmospheric opacity (White et al., 2014).
result in a higher average forest age (Franklin et al., 2018). In managed As a result of varying data availability at different latitudes and
forests, rotation age (i.e. the age at which forests are typically harvested) persistent clouds in certain regions, data gaps are present in the annual
can differ depending on the productivity of the site, but forest man­ surface reflectance image composite produced. The composites are
agement activities are often predicated on a 100 year rotation (Bergeron further refined through the removal of anomalous values (e.g., wildfire
et al., 2017). Nationally, black spruce dominates Canada’s forests, smoke, haze; Kennedy et al., 2010) and by the infilling of data gaps with
making up 57.3% of forested area. Trembling aspen, lodgepole pine, and proxy values assigned through a spectral trend analysis of the Normal­
jack pine are the next three most common species, comprising 9.8%, ized Burn Ratio (NBR) index. A time series of NBR is calculated for each
5.9%, and 4.2% of forested area respectively (Hermosilla et al., 2022a). pixel from the annual Landsat surface reflectance composites. Break­
point detection is applied to the pixel series to identify change events,
3. Materials and methods with a further contextual analysis in the spatial domain to reduce spatial
or temporal inconsistencies possibly resulting from the pixel based
3.1. Data compositing process (Hermosilla et al., 2015b). A Random Forests
modeling approach is then applied to label the detected changes by type
3.1.1. National Forest Inventory (e.g., wildfire, harvest) that incorporates spectral as well as object-level
To support model development and subsequent forest age validation (i.e. disturbance patch) information (Hermosilla et al., 2015a). Hermo­
we utilized Canada’s National Forest Inventory (NFI), a program that silla et al. (2016) informed on the accuracy of the C2C approach
surveys 1% of Canada’s landmass. The NFI is a systematic survey spatially (overall accuracy = 90%) and temporally (89% of the changes
wherein square 2 × 2-km permanent sampling units referred to as photo detected the correct occurrence year and 98% within ±1 year). The
plots are distributed on a 20 × 20-km grid across the country. This disturbance agent for stand replacing disturbances, such as fires and
survey utilizes aerial photography and very high spatial resolution sat­ harvest were attributed with 87% accuracy.
ellite imagery to enable manual interpretation of a number of forest Canada-wide, annual land cover layers were generated to represent
characteristics including species, age, and disturbance history. Prior to each year from 1985 to 2019 using the Virtual Land Cover Engine
interpretation, the photointerpreter will delineate homogenous spatial (VLCE) methodological framework, which avails upon a time series of
units (e.g., forest stands or polygons), representing areas of relatively Landsat annual surface-reflectance image composites to enable change-
uniform forest characteristics (Gillis et al., 2005). Attributes such as age informed annual land cover mapping across Canada’s forested ecosys­
are then interpreted for each polygon. tems (Hermosilla et al., 2018). Hermosilla et al. (2022b) extended and
Forest age of each NFI polygon is derived from one of four main enhanced the framework through the refinement of calibration data, the
sources: manual interpretation of aerial photography (interpreted); re­ integration of new predictor variables and the optimization of regional
cords of a recent disturbance event (updated); field plot data such as tree model development and implementation. The 30-m spatial resolution
cores (ground survey); and growth and yield models (modelled), annual land cover maps for 1984 to 2019 comprise 12 land cover classes
whereby stand attributes are grown forward to represent current con­ of which four represent vegetated treed classes: broadleaf, coniferous,
ditions from estimated time of stand establishment or last measurement. mixedwood and wetland-treed. Reported overall accuracy of the land
Manual interpretation is the most common method for age estimation cover product was 77.9% ± 1.4% (95%-confidence interval).
nationwide in the NFI and, as noted earlier, age is among the NFI at­ Annual, wall-to-wall, 30-m forest structure metrics (i.e., 95th
tributes most sensitive to photo-interpretation uncertainty (Magnussen percentile of canopy height, percent canopy cover, total aboveground
and Russo, 2012). We highlight here that these four sources do not biomass, and elevation covariance) were also derived from the annual
provide the same level of accuracy or precision in the estimation of stand Landsat surface reflectance composites using the method described in
age, with modelled ages not considering disturbances that may have Matasci et al. (2018a, 2018b). This method uses ALS and field plot data
occurred after the original data were collected. Therefore, to ensure data to estimate forest structure metrics from topographic and Landsat
integrity, NFI polygons where stand age was derived from a modelled spectral predictors, using a k-nearest neighbor imputation approach.
information source were discarded in all ecozones where they made up Models were trained based on acquired ALS and field data, and then
<90% of total polygons (i.e., all ecozones but the Pacific Maritime and used to estimate structural metrics using annual Landsat surface
Montane Cordillera). Additionally, age within the NFI is defined in one reflectance composites. Estimates of 95th percentile of canopy height
of three ways (i.e., age from stand establishment, age from seed, and age had an RMSE of 3.56 m, while estimates of percent canopy cover had an
at breast height), and is denoted by polygon, with age from establish­ RMSE of 18.41%.
ment being the most common. These differing definitions of age can lead To determine tree species, the dominant tree species map for 2019
to significant variability in recorded age (Wong and Lertzman, 2001). produced following the methodology described in Hermosilla et al.
(2022a) was used. This map represents 37 tree species at 30-m spatial
3.1.2. Landsat-derived datasets resolution over the forested ecosystems of Canada and was generated
Annual Landsat image composites, produced following the Compo­ using a series of regional Random Forests models utilizing a 2019
site2Change (C2C) approach (Hermosilla et al., 2016), were used as a Landsat surface reflectance composite, as well as climatic, phenological,
primary data source for forest age estimation. The C2C approach utilizes topographic, and geographic data. An accuracy assessment was con­
the free and open Landsat archive (White et al., 2014) to produce ducted using independent validation data, reporting an overall accuracy
annual, gap-free, surface-reflectance image composites with a 30-m of 93.1% ± 0.1% (95%-confidence interval).
spatial resolution (Hermosilla et al., 2015b) as well as to allow for the
detection and characterization of forest disturbance information (Her­ 3.1.3. MODIS GPP data
mosilla et al., 2015a). The target date for image compositing is set to The MODIS 500 m Gross Primary Production (GPP) data product
August 1st ± 30 days for centrality to the growing season for most of (Running et al., 2015) was used as a surrogate for site productivity with
Canada’s forested ecosystems. All atmospherically corrected Landsat the underlying algorithm described in detail by Running et al. (2004).
images (Masek et al., 2006; Schmidt et al., 2013) acquired within this The algorithm relies on the light-use efficiency approach, relating GPP

5
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

to the fraction of absorbed photosynthetically active radiation (fPAR) Table 2


through a radiation use conversion efficiency term ε (with the units kg C Data layers utilized for age estimation methods.
MJ− 1) (Monteith, 1972; Prince and Goward, 1995), which is dependent Data used Description Age approach Source
on vegetation type (as defined by land cover classification of MODIS used in
imagery) and is reduced by two climate-based multipliers. Of the mul­ Landsat spectral Annual best available Recovery White et al.,
tipliers, one reduces the conversion efficiency when cold temperatures composites pixel composites 2014
limit plant function, while the second reduces the maximum conversion Hermosilla
efficiency when the vapour pressure deficit is sufficiently high to inhibit et al., 2015b
Hermosilla
photosynthesis. The algorithm uses three sources for input data. Firstly, et al., 2016
biome specific parameters such as ε, are assigned based on an eight-class Disturbance time Annual map of Disturbance Hermosilla
land cover classification (Hansen et al., 2000; Running et al., 2004). series (Landsat) disturbance events Recovery et al., 2015a
Secondly, incoming solar radiation is obtained along with air tempera­ Hermosilla
et al., 2016
ture and relative humidity from global scale meteorology (Zhao et al.,
Tree species 2019 Leading species map Allometric Hermosilla
2005). Lastly, daily fPAR is used, which is derived from atmospherically (Landsat) et al., 2022a
corrected surface reflectance values at from MODIS spectral bands, land Forest structure Annual maps of forest Recovery Matasci et al.,
cover, as well as information on viewing and illumination angles time series structural metrics Allometric 2018a, 2018b
(Myneni et al., 2002). Heinsch et al. (2006) provide a detailed discussion (Landsat) including height and
percent cover
of the accuracy of the MODIS GPP product over a range of ecosystems, Land cover time Annual land cover maps Disturbance Hermosilla
and the potential for errors introduced by a variety of input data sources. series (Landsat) of Canada Recovery et al., 2018
Cubic spline interpolation was used to resample the 500–m MODIS GPP Allometric Hermosilla
product to 30-m pixels to match Landsat’s spatial resolution. Mean et al., 2022b
MODIS 8-day 500-m resolution gross Allometric Running et al.,
maximum annual GPP is then calculated from the resampled product.
cumulative Gross primary productivity 2015
Relative productivity of a pixel was determined by partitioning Canada’s Primary measure
forested ecozones into a grid of 388, 150 × 150-km tiles (as per Her­ Productivity
mosilla et al., 2022b). Productivity was divided into three classes (i.e., Canadian National Polygons denoting Recovery Canadian
high, medium, and low). For each tile, classes were evenly determined Fire Database recorded historical fires Allometric Forest Service,
Polygon Data across Canada 2020
based on the quantiles 33% and 66% of the GPP distribution within the
National Forest Polygon-based survey of Agreement National Forest
5 × 5-neighborhood of the tile. Inventory Canadian forests assessment Inventory,
2021
3.1.4. Canadian National Fire Database
The Canadian National Fire Database (CNFDB) collects historic and
spatially and temporally, with reported overall accuracy of 90% in the
current forest fire data from a number of fire management agencies
change detection and 98% of changes correctly dated within ±1 year.
including provincial and territorial agencies, as well as Parks Canada. It
The disturbance date provides a valuable baseline of consistent infor­
contains polygon perimeters of recorded fires from the early 1900s to
mation across the forested ecosystems of Canada. As a result, forest age
present (Canadian Forest Service, 2020). The exact date of the earliest
for disturbed pixels between 1985 and 2019 was derived by computing
recorded fires in the database depends on the province in which the fire
the time since disturbance from the change year indicated in the
occurred. In British Columbia, fires are recorded as far back as 1917,
Landsat-derived forest change layers.
while in New Brunswick and Newfoundland and Labrador, fires are only
recorded from 1980 onward. Fire perimeter information within the
3.2.2. Recovery approach
database is not always precise, error free, or inclusive of all possible fires
Pixels showing indirect evidence of disturbance were identified by
in a given (historic) year. Especially in earlier recorded fires, fire pe­
examining their spectral trajectories from 1985 to 1995 for evidence of
rimeters are approximations of fire location, rather than exact spatial
recovery from disturbance immediately prior to 1985. Time series of
records and do not capture unburned areas within larger fire perimeters
NBR derived from the Landsat annual surface-reflectance image com­
(White et al., 2017). However, the database provides a vital tool for
posites were used to examine spectral recovery trends. After a distur­
evaluating fire activity and location nationwide over long-time horizons.
bance event, the NBR value of a pixel will decrease with the loss of
Table 2 summarises the datasets used in this research.
vegetation. Assuming no change in land use, NBR is expected to slowly
recover to a value similar to its pre-disturbance level in the years
3.2. Forest age estimation
following the disturbance (White et al., 2017). In a study of forest re­
covery following disturbance across Canada, White et al. (2022) used
Forest age was estimated for the year 2019 following three distinct
measures of forest canopy height and cover derived from ALS data to
approaches (Fig. 2), which are based on the identification of: (1) pixels
validate post-disturbance measures of spectral recovery. The authors
with direct evidence of stand replacing disturbances within the period of
found that a majority (87–97%) of pixels that were considered to be
the Landsat record (disturbance approach); (2) pixels with evidence of
spectrally recovered (using an NBR-based metric with recovery indi­
recovery from disturbance in the period directly preceding the Landsat
cated by a return to an NBR value of 80% of pre-disturbance NBR), also
record (recovery approach); and (3) pixels with no apparent disturbance
achieved at least one ALS-derived benchmark indicative of recovery).
or recovery within the Landsat record (allometric approach). Pixels with
Therefore, by analyzing the NBR trajectories of pixels following known
an estimated age that is >150 years were grouped as an “old” class, as
stand replacing disturbance within the Landsat record, the anticipated
uncertainty in forest age estimation increases with age due to saturation
rate of spectral recovery associated with disturbances is determined. The
in spectral signals (Sader et al., 1989; Zhang et al., 2004, 2014). The
spectral recovery rate is then used as baseline for comparison with each
detailed methodological steps followed are outlined in Fig. 3.
pixel’s spectral trajectory from 1985 to 1995 to determine if the tra­
jectory is indicative that the pixel is recovering from a stand replacing
3.2.1. Disturbance approach
disturbance that occurred in the twenty years prior from 1965 to 1985.
Pixels showing direct evidence of disturbance were identified using
Based on analysis reported in White et al. (2022), most areas in
the Landsat-derived forest change layers produced by Hermosilla et al.
Canada’s forested ecosystems experiencing stand replacing disturbance
(2016), which enabled forest age estimation from 1985 to 2019. The
caused by timber harvest or wildfire were found to have spectrally
forest change layers have been shown to be highly accurate, both

6
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

Fig. 2. Illustration of the three forest age estimation approaches applied.

Fig. 3. Workflow of the framework to estimate forest age. Acronyms used in this figure: BAP: Best Available Pixel, GPP: Gross Primary Productivity, CNFDB: Ca­
nadian National Fire Database, NFI: National Forest Inventory.

recovered within 20 years. Similarly, a synthesis of post-disturbance previous approaches used to assess post-disturbance recovery (White
field-based studies also found that forests achieve benchmarks of can­ et al., 2017, 2022).
opy cover and height within 20 years (Bartels et al., 2016). Additionally, The spectral recovery baselines indicating the distribution of NBR
less data are available as baseline quartile values increase in years-since- values for time since disturbance were derived from a pool of pixels with
disturbance. Only disturbances which occurred in 1985 could be used to known disturbances (described in Matasci et al., 2018a,b) and defined
generate the theoretically oldest baseline quartile for TSD = 34, mean­ by ecozone in order to acknowledge that the variation of spectral re­
ing it would by highly subject to influence from the annual variations of covery rates between differing geographic, climatic, and biological
that year. By limiting the recovery approach to 20 years, we ensured that conditions (White et al., 2022). From the spectral recovery baseline of
we had at least five years of data to derive our oldest baseline quartile at each ecozone, interquartile range (IQR) values were calculated per years
TSD = 30. This five-year window post-disturbance is also consistent with since disturbance, from 0 to 30 years. This resulted in a range of

7
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

potential NBR values for which a pixel at each year in its recovery is height, age, and site condition of sample plots are recorded, either
expected to fall. Then, for each pixel, NBR values from 1985 to 1995 following stands through time, or by comparing multiple stands of
were compared to the IQR of the baseline at each time-since-disturbance different ages (Hanson et al., 2002). These data are then used to
step to determine age. parameterize equations characterizing the expected growth of a tree
Fig. 4 shows an example of the approach to estimate forest age using over time, given the productivity of the site where the tree is located.
the recovery approach for a given pixel. The line segment represents the To determine which species to select site index equations for, the
time series of NBR values of the pixel for the period 1985–1995. To most common dominant species as recorded in the NFI were calculated
determine if the time series indicates that the pixel is recovering from a for all tiles within the 150 × 150-km tiling grid. Site index equations for
previous disturbance, values are compared against the recovery baseline all species which were dominant in at least five tiles were selected. 12
of their respective ecozone. The time series values are compared for each conifer species and six deciduous species met this criterion, and equa­
year across the recovery baseline for time since disturbance (TSD) 0–20. tions were compiled from various sources within the literature (Table 3).
Any location where the entire time series falls within the IQR baseline For a given pixel, forest age is computed from the inverted site index
values is recorded as a potential TSD (TSD = 10 in Fig. 4), and all lo­ curve for its leading tree species, canopy height, and representative site
cations where the time series has values outside of the IQR baseline index value. On a pixel level, species was determined using the tree
(TSD = 0, TSD = 20 in Fig. 4) are discarded. To determine a singular age species layer generated by Hermosilla et al. (2022a). If the pixel had a
value, the mean of all potential TSD values is utilized. If a pixel has no species for which no site-index equation was present, the most common
valid TSD values, then it is considered to not show evidence of recovery tree species in the ecozone of the pixel’s forest type (i.e., coniferous,
from disturbance, and is not given an age in this approach. deciduous) was utilized.
As an additional line of evidence, once TSD has been estimated, Canopy height values were derived from forest structure layers
pixels also had to display increases in their canopy cover and NBR values representing the year 1985 (Matasci et al., 2018a). The representative
between 1985 and 1995 to be considered as recovering from a distur­ site index value was based on MODIS GPP productivity class (i.e., low,
bance. This NBR increase requirement is to prevent uncommon in­ medium, high) and leading tree species. Pixels with a high productivity
stances of pixels with slight NBR decreases over time which may still fit class, based on MODIS GPP, were assigned a site index value of 75% of
within the IQR bands. Pixels showing decreases in either value were the maximum reported site index for its tree species; pixels with a me­
excluded from the recovery approach. As final step, a minimum map­ dium productivity class were assigned a site index value of 50% of the
ping unit (MMU) was applied, so disturbed pixels identified with the maximum; and pixels with a low productivity class were assigned a site
recovery approach had to be adjacent to other disturbed pixels, repre­ index value of 25% of the maximum. We aimed to assign a site index that
senting a minimum area of 0.5 ha. was representative of the broader productivity class, with these selected
values representing the midpoint of site index values for each class.
3.2.3. Allometric approach The inverted forms of the site index equations used in this research
For those pixels not showing evidence of disturbance or recovery are indicated below.
within the Landsat record, forest age was estimated by utilizing the ( ( ) )
inverted forms of pre-existing site index equations. Site index equations ln 1 − H 1
d⋅S e

are a critical component of conventional forest inventories, providing


a⋅Sb
A= (1)
foresters with estimates of anticipated height growth at a given age. c
These equations are derived from field-based measurements, where the

Fig. 4. Comparison of Normalized Burn Ratio (NBR) time series to interquartile range (IQR) recovery baseline to estimate forest age. Locations where all values in
the time series fall within the IQR bands are recorded to determine the potential age for that pixel. Locations where values do not fall within the IQR bands are
discarded. TSD: Time Since Disturbance.

8
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

Table 3 ( 1
)
Canada’s most common tree species and site index equations utilized to estimate Hc
ln 1 −
forest age. a⋅S (3)
A=
Common Scientific Equation Equation Source for site b
name name used Source index values
where A is forest age for black spruce, tamarack, jack pine, eastern white
Conifer
Black spruce Picea mariana Eq. (3) Dolid and Dolid and pine, trembling aspen and paper birch (Dolid and Lundgren, 1970), H is
Lundgren, Lundgren, canopy height, S is site index, and a, b and c are constants as defined in
1970 1970 Table 5.
Balsam fir Abies balsamea Eq. (1) Payandeh, Carmean and ( )
1974 Hahn, 1981 1
ln
Douglas fir Pseudotsuga Eq. (1) Payandeh, McArdle et al., C (4)
menziesii 1974 1949 A=
0.03298
Eastern Pinus strobus Eq. (3) Dolid and Dolid and
white pine Lundgren, Lundgren,
1970 1970
where A is forest age for largetooth aspen (Scott and Voorhis, 1986), H is
( )
Jack pine Pinus Eq. (3) Dolid and Dolid and H S0.54414
canopy height, S is site index, and C = 1 − 17.608+1.118⋅S
14.6962 .
banksiana Lundgren, Lundgren,
1970 1970 ( )
Lodgepole Pinus contorta Eq. (8) Thrower et al., Smithers, 1
ln
pine 1994 1961 C (5)
Red spruce Picea rubens Eq. (1) Payandeh, Burns and A=
0.02134
1974 Honkla, 1990
Subalpine fir Abies Eq. (6) Chen and Chen and
where A is forest age for sugar maple (Scott and Voorhis, 1986), H is
lasiocarpa Klinka, 2000 Klinka, 2000
Tamarack Larix laricina Eq. (3) Dolid and Dolid and canopy height, S is site index, and
Lundgren, Lundgren, ( )
1970 1970
H 1
C = 1− 1.1734⋅S0.07876 .
Western Tsuga Eq. (1) Payandeh, Meyer, 1937 − 5.152 + 2.057⋅S
hemlock heterophylla 1974
( )
Western Thuja plicata Eq. (2) Nigh, 2000 Thrower et al., 1
1 − KP
redcedar 1994 A = ln (6)
White spruce Picea glauca Eq. (1) Dolid and Dolid and − 0.01798
Lundgren, Lundgren,
1970 1970 where A age is forest age for subalpine fir (Chen and Klinka, 2000), H is
Deciduous
Largetooth Populus Eq. (4) Scott and Scott and
canopy height, S is site index, K = 5.88585H−(S−1.31.3)0.59774
and P = −
( )
aspen grandidentata Voorhis, 1986 Voorhis, 1986
(S− 1.3)0.0683
Red alder Alnus rubrum Eq. (7) Nigh and Nigh and 1.91387*ln 5.88585 .
Courtin, 1998 Courtin, 1998
Red maple Acer rubrum Eq. (1) Carmean and Carmean and
(7)
K
Hahn, 1981 Hahn, 1981 A = e1.24 + 0.1789S − 4.994
Sugar maple Acer Eq. (5) Scott and Scott and
saccharinum Voorhis, 1986 Voorhis, 1986 where A is forest age for red alder (Nigh and Courtin, 1998), H is canopy
Trembling Populus Eq. (3) Dolid and Dolid and ( )
aspen tremuloides Lundgren, Lundgren, height, S is site index, and K = 3.6 − ln 1.693(S−
H− 1.3
1.3)
− 1 .
1970 1970
White birch Betula Eq. (3) Dolid and Dolid and
42.64
papyrifera Lundgren, Lundgren, (8)
7.815− 1.007ln(S− 1.3)− ln(K− 1)
A=e 1.285 + 5.6 +
1970 1970 S

where A is forest age for lodgepole pine (Thrower et al., 1994), H is


(S− 1.3)*(1+e(2.788− 1.007ln(S− 1.3) )
where A is forest age for balsam fir, white spruce, Douglas-fir, red canopy height, S is site index and K= H− 1.3 .
spruce, western hemlock and red maple (Carmean and Hahn, 1981; The curves produced from site index equations are generally
Dolid and Lundgren, 1970; Payandeh, 1974), H is canopy height, S is site asymptotic and this asymptote represents the maximum height a tree is
index, and a, b, c, d and e are constants as defined in Table 4. expected to grow (Nigh et al., 2016). Consequently, site index equations
(( ) ) may result in a non-real solution for input canopy heights taller than
− ln
(S− 1.3)⋅c
H− 13
− 1 +1.244⋅ln(S− 1.3)− 9.474
(2) certain values (Smith, 1984). Pixels with canopy height taller than the
A=e 1.34 + 0.5 canopy height that the site index curve can estimate within a produc­
tivity class, are assumed to be more productive than estimated, and
where A is forest age for western redcedar (Nigh, 2000), H is canopy therefore the site index equation is iteratively solved by assigning the
height, S is site index, and c = 1 + e(4.245− 1.244⋅ln(S− 1.4)). next higher productivity class. Those pixels with a high productivity
class that return a non-real age are assumed to be taller—and thus

Table 4 Table 5
Constants for species covered by Eq. (1). Parameterized for imperial units. Constants for species covered by Eq. (3). Parameterized for imperial units.
Species a b c d e Species a b c

Balsam fir 1.276 1.0096 − 0.0401 1.9605 0.0182 Black spruce 1.762 ¡0.02011 1.2307
Douglas fir 1.1585 1.0011 − 0.023 1.2956 0.0134 Eastern white pine 1.966 − 0.02399 1.8942
Red maple 2.9435 0.9132 − 0.0141 1.658 − 0.1095 Jack pine 1.633 − 0.02233 1.2419
Red spruce 2.3106 0.8856 − 0.0561 8.163 − 0.0813 Paper birch 1.598 − 0.01938 0.9824
Western hemlock 1.5469 1.0018 − 0.0114 1.0883 0.0072 Tamarack 1.547 − 0.02246 1.1129
White spruce 1.8939 0.9591 − 0.023 1.2765 0.005 Trembling aspen 1.48 − 0.0214 0.9377

9
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

older—than expected, and accordingly labelled as old class. ferences in scale and resolution, as well as potential temporal mis­
matches (Wulder et al., 2006). As such, a number of data filtering steps
3.2.4. Canadian National Fire Database integration were implemented prior to spatial comparison. To avoid geolocation
Forest age estimates for pixels that were not disturbed within the discrepancies, pixels located on NFI polygon boundaries were excluded
period of the Landsat record (i.e., forest ages determined by recovery or from the analysis. To avoid temporal mismatches, only NFI polygons
allometric approaches), were adjusted by comparing them against the with a data collection date more recent than the estimated forest age
wildfire records in CNFDB (Canadian Forest Service, 2020).The CNFDB were retained. A total of 10,000 NFI polygons were selected for the
was used to determine forest age as it contains records of wildfire events spatial assessment via stratified random sampling, with strata defined by
over 100 years. As the spatial information regarding fire extents may be ecozone and by ten-year age class. Agreement was calculated by
coarser and inaccurate for older fires, our age estimates are corroborated comparing the median estimated forest age within an NFI polygon to the
with the CNFDB records to determine if they both represent similar ages. NFI recorded forest age for that polygon. Measures of agreement
Given that the recovery approach was found to be more accurate than included mean absolute deviation (MAD), mean deviation (MD), Pear­
the allometric approach (related to more limited range of years and son’s correlation coefficient, and the percentage of polygons with me­
direct spectral corroboration), pixels with forest age defined by the re­ dian estimated forest age falling within ±1 standard deviation of NFI
covery approach within 10% of the age of a fire event in the CNFDB were recorded forest age.
assumed to be disturbed by the recorded fire, and the year of disturbance ∑
|AgeNFI − AgeEstimated |
in the CNFDB was used to derive forest age. In contrast, due to the lower MAD = (9)
n
accuracy of the allometric approach, pixels with age defined by the
allometric approach were assumed to be disturbed by fire if their forest ∑
(AgeNFI − AgeEstimated )
age was within 20% of the age of a fire event in the CNFDB, and the year MD = (10)
n
of disturbance of the fire was then utilized.
where AgeNFI is NFI recorded forest age in a polygon, AgeEstimated is me­
3.3. Comparison to NFI forest age dian estimated forest age within the same polygon, and n is the number
of sample units in that stratum (i.e., ecozone and age class).
Agreement between the estimated forest age obtained using the
approach presented in this research and the forest age recorded in the 4. Results
NFI polygons was assessed. While the NFI represents the best nationally
consistent information on forest age, the stand ages reported therein are 4.1. Summary of age estimation approaches
also estimates and are subject to varying levels of uncertainty, accuracy,
and precision. Thus, any comparison to these data must be interpreted Nationally, a large majority of forested area had age estimated by the
accordingly and seen as a confidence building exercise rather than a allometric approach, accounting for 82% of all pixels (Table 6; Fig. 5).
validation against an absolute truth. Forest age estimates obtained with The Taiga Shield West had the largest percentage of area with age
the recovery and allometric approaches were assessed for overall estimated by the disturbance approach, accounting for 30.8% of it’s
agreement and agreement within their respective eras (i.e., 1985–1965 forested area, while the Pacific Maritime had the smallest percentage of
for the recovery approach, prior to 1965 for the allometric approach). area with age estimated by this approach at 8.1%. The recovery
Pixels with forest age estimated by the disturbance approach were approach covered only 2.8% of forested area nationwide. The ecozone
not assessed for agreement with the NFI, as the forest change product with the largest percentage of age estimated by the recovery approach
utilized in this approach was previously validated and shown to be was the Boreal Cordillera, accounting for 13.1% of its area, while the
highly accurate spatially (overall accuracy of 89%) and temporally (97% area with the smallest percentage of age estimated by the recovery
of change detected within ±1 year; see Hermosilla et al., 2016). Accu­ approach was the Taiga Plains, accounting for 0.2% of its area. It is
racy assessment in Hermosilla et al. (2016) was conducted using and important to note that each approach does not account for an equal span
independent set of photointerpreted point validation samples following of time, with the disturbance approach estimating age over 34 years, the
the practices recommended in Wulder et al. (2007) and Wulder et al. recovery approach covering 20 years, and the allometric approach
(2008). The validation samples used in Hermosilla et al. (2016) were spanning 116 years of estimated age, plus the old class (>150 years).
points as opposed to NFI records which are polygons. The use of polygon
samples entails summarizing the pixel values within the polygons via
statistical descriptors (i.e., mean, median). The point samples used are of
high quality and enable one-to-one comparisons with the raster values,
hence providing a better depiction of the accuracy of the age derived Table 6
using the disturbance approach. Percentage of total forested area by ecozone with age classified by each age
Agreement was assessed both aspatially and spatially. Aspatial estimation approach.
agreement was conducted by comparing estimated and NFI recorded Ecozone Disturbance Recovery Allometric
forest age distributions by ecozone. Area weighted summary statistics (i. approach approach approach
e., mean, median, IQR) were computed from estimated forest age and Atlantic Maritime 17.5% 5.7% 76.8%
from the NFI recorded forest age for polygons that were ≤150 years old. Boreal Cordillera 8.8% 13.1% 78.1%
The percent of forest area belonging to the old class (over 150 years old Boreal Plains 14.9% 0.6% 84.5%
Boreal Shield East 12.6% 3.1% 84.3%
in 2019) was also reported. Boreal Shield West 23.8% 1.8% 74.5%
The spatial assessment compared spatially concurrent estimated Hudson Plains 8.2% 1.6% 90.2%
forest age (i.e. pixel-based) with NFI recorded forest age (polygon-based Montane 13.7% 2.6% 83.7%
assessment). In the absence of other independent data for age, such a Cordillera
Pacific Maritime 8.1% 1.4% 90.4%
spatial comparison provides some indication of model performance but
Taiga Cordillera 9.7% 0.5% 89.8%
should not be considered an assessment of absolute pixel-level accuracy, Taiga Plains 14.6% 0.2% 85.2%
particularly given the uncertainties associated with methods used for Taiga Shield East 8.8% 2.1% 89.1%
age estimation within the NFI. In addition, the spatial comparison of Taiga Shield West 30.8% 2.2% 67.0%
polygon-based with pixel-based age estimates presents a number of All Forested 15.2% 2.8% 82.0%
Ecozones
challenges including potential discrepancies in geolocation due to dif­

10
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

Fig. 5. Spatial representation of approaches used to estimate forest age in Canada’s forest-dominated ecozones. Legend applicable to treed pixels.

4.2. Mapping forest age 4.3. Agreement with NFI

The result of the 30-m spatially explicit mapping of forest age esti­ 4.3.1. Aspatial comparison
mates for Canada’s forested ecosystems is shown in Fig. 6A. Detailed The aspatial comparison of estimated and NFI recorded age for for­
insets over selected areas, in support of the nation-wide map, inform on ests ≤150 years old indicated similar forest age distributions, overall
forest age diversity in the Pacific Maritime (Fig. 6B), Boreal Shield East and by ecozone (Fig. 8). Across Canada’s forest-dominated ecozones,
(Fig. 6C), and Atlantic Maritime (Fig. 6D) ecozones. Forest age distri­ median estimated forest age in forests ≤150 years, was 66 years, and
butions by ecozone are summarized in Fig. 7. median NFI recorded forest age was 73 years. Similarly, mean estimated
Nationally, the mean estimated forest age for pixels ≤150 years old forest age was 68 years, and mean NFI recorded forest age was 74 years.
was 70 years (standard deviation = 32.1 years) and the median forest The Taiga Shield West had the lowest difference in median forest age
age was 68 years. Across Canada’s forested ecosystems, 22.7% of (<1 year). Conversely, the largest difference in median forest age was
forested area ranged between 0 and 50 years old, 56.6% of forested area found in the Montane Cordillera (~26 years).
was estimated to be between 50 and 100 years old, 14.7% was between The comparison of percent forested area belonging to the old class
100 and 150 years old, and 6.1% was over 150 years old. The most (>150 years) showed varying levels of agreement between the NFI
common age range nationwide was between 60 and 70 years old, ac­ recorded and estimated forest ages (Table 7). Across Canada’s forest-
counting for 14.4% of forested area. The least common age class dominated ecozones, 9.5% of forested area covered by the NFI was
nationwide was 140–150 years, accounting for 1.4% of forested area. identified as being within the old class, in contrast to the 5.9% of the
By ecozone, the Boreal Cordillera, Boreal Plains, and Taiga Plains area of estimated forest age. The Taiga Shield West had the highest
had the youngest median forest age (60 years), whereas the Hudson agreement in old class area, differing by only 0.7% between estimated
Plains had the oldest median forest age (80 years). The Taiga Shield and NFI recorded forest age, while the Pacific Maritime had the largest
West had the largest proportion of forest between zero and ten years old disagreement, at 61% less old class than the NFI.
with 9.5% of its area being this class. Conversely, the Taiga Shield East
had 0.9% of its forested area fall within that age range. The Montane 4.3.2. Spatial comparison
Cordillera had the largest proportion of old forest (15.4%), while old The stand-level spatial comparison of the estimated age to the NFI
forest was <5% of the forested area for the Atlantic Maritime, Boreal recorded age resulted in a MAD of 38.2 years nationally (Table 8; Fig. 9).
Cordillera, Taiga Cordillera, and Taiga Shield East and West. Estimated forest age in the recovery approach era of 1965–1985 was
Overall, Canada’s forests have ages concentrated in the 50–100 year more in agreement with the NFI recorded forest age (MAD = 31.8 years)
range, accounting for 56.6% of all forested area. Ecozones with longer than estimated forest age within the allometric approach era, prior to
disturbance return intervals such as the Montane Cordillera tend to show 1965 (MAD = 39.3 years).
higher percentages of old forest when compared to more frequently By ecozone, the strongest agreement was in the Taiga Cordillera with
disturbed ecozones such as the Taiga Cordillera. It is worth noting that an overall MAD of 28.3 years, a MAD for the recovery approach era of
despite the relatively long nature of its natural disturbance return in­ 19.4 years, and a MAD for the allometric era of 30 years. The weakest
terval, the Pacific Maritime ecozone has a modest amount of forest area agreement was found in the Pacific Maritime with an overall MAD of
that is older than 150 years, at 5.5% of forested area. 153.2 years.
Nationally, the MD was 7.8 years, indicating a relatively small bias,
with an underestimation of age relative to the NFI; however, at the
ecozone level, MD indicated both over and underestimation. The
Atlantic Maritime and Taiga Shield West ecozones had an MD less than

11
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

Fig. 6. Predicted forest age in (A) Canada’s forest-dominated ecozones. Insets display areas in (B) Pacific Maritime, (C) Boreal Shield East, and (D) Atlantic
Maritime ecozone.

or equal to − 10 years, and the Boreal Cordillera, Montane Cordillera, source. By combining disturbance-, recovery-, and allometric-based
Taiga Shield East and Pacific Maritime having an MD greater than or approaches to estimating forest age, we were able to take advantage
equal to 10 years. of best available information within a given epoch.
Across all of Canada’s forested ecosystems, 79% of NFI polygons had While a means to estimate forest age for pixels without more direct
an estimated age that fell within one standard deviation of the NFI evidence, the use of inverted site index curves also poses challenges. At
recorded age. The Pacific Maritime had the lowest percentage of poly­ older ages, the rate of height growth declines. For this reason, site index
gons with estimated age falling within one standard deviation, at 34%, curves are asymptotic, reaching a maximum tree height beyond which,
while the Taiga Shield East had the highest at 78%. predicted height will remain the same, regardless of the age input. When
inverted, these equations therefore become very sensitive to changes in
5. Discussion height as the measured height approaches the asymptote of the equa­
tion, and small changes in height can result in large changes in predicted
Forest age is a fundamentally important attribute in forest in­ age. Similarly, for a given site index value, height inputs beyond the
ventories and for understanding forest ecosystems. Herein, we proto­ asymptote of the equation result in non-real solutions, which indicate
typed an approach to estimate forest age for Canada’s forested ecozones. that the pixel is old, but give no concrete estimate of age. By setting an
Landsat-derived data products were combined with other supplemen­ upper age threshold of 150 years, inaccurate estimates of very old ages
tary layers to allow for a fine-scale, consistent, and spatially explicit due to small changes in height are avoided.
estimation of age that is derived from a nationally consistent data Availability of appropriate and representative data is also a challenge

12
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

Fig. 7. Distribution of estimated forest age across Canada’s forest-dominated ecozones, by 10-year age class.

for the allometric method. Both input height and site index rely on data truth. Our results indicated a closer agreement between the NFI recor­
products with associated uncertainty. The reported RMSE of height es­ ded and estimated forest ages when compared aspatially than when
timates used in this study was 3.5 m (Matasci et al., 2018b), which has compared spatially. Disagreement between forest age products may be
the potential to affect the quality of the age estimates made. Appropriate caused by errors in either dataset, and all methods of forest age esti­
site index curves were assigned based on dominant species information mation have uncertainty associated with them, whether that be tree
derived from a national Landsat-derived species map, which also has cores (Metsaranta, 2020), photointerpretation (Magnussen and Russo,
some associated uncertainty (Hermosilla et al., 2022a). Moreover, it 2012), or time since disturbance as a proxy for age (Bradford et al.,
must be noted that the site index curves themselves are typically derived 2008). Indeed, forest age may be one of the most challenging forest
from a limited sample of managed, even-aged stands. Site index values attributes to estimate and map with remotely sensed data due to the
used herein were derived from GPP measurements based upon 500–m uncertainties associated with all sources of potential calibration and
spatial resolution MODIS observations that are resampled to 30-m. Site validation data. Moreover, age is considered one of the NFI attributes
index values are assigned one of three (high, medium, low) productivity that is most sensitive to photo-interpretation uncertainty (Magnussen
classes. The availability of consistent, national productivity measure­ and Russo, 2012).
ments with a finer spatial resolution would provide an enhanced char­ Our results also highlighted the challenges in the spatial comparison
acterization of local variation in productivity, and hence enable finer of polygon-based forest age records versus pixel-based estimates. While
groupings of site index that in turn may strengthen the age estimates. the forest age of a stand is recorded as a single value in a polygon, there
When compared aspatially, estimated ages had a similar statistical is likely a finer scale variation in the ages of trees that make up that
distribution to the NFI for forested areas ≤150 years. However, in the polygon, especially in older, uneven-aged stands, which may be
Pacific Maritime and Montane Cordillera, the percent area estimated to captured in a pixel-based approach, further exacerbating disagreement
be older than 150 years was markedly lower than percent area recorded (Wulder et al., 2006). Moreover, characterizing spatial agreement by
in the NFI (61% lower and 21.7% lower respectively). It is notable here taking the median of our pixel-level estimated ages within the NFI
that these were the only two ecozones in which modelled NFI polygons polygon does not provide a pixel-level assessment of our estimated age
were included in the spatial comparison between estimate age and NFI map.
age (to ensure a sufficient sample size). Recall that modelled ages in the The availability of reference data to adequately train and validate
NFI are derived from growth and yield models and have greater asso­ models of forest age is a challenge common to many published studies
ciated uncertainties. (Table 1). Unsurprisingly, the spatial comparison between estimated
As noted earlier, comparison to the NFI should be considered as a and NFI recorded forest age reported herein indicated a lower level of
confidence building exercise rather than a validation against absolute agreement than reached in studies that focused on smaller study areas

13
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

Fig. 8. Boxplots represent the mean, median, interquartile range, and extreme values of estimated and National Forest Inventory (NFI) recorded forest age
≤150 years.

using ALS data as the main predictive driver (Maltamo et al., 2020; 2021; Zhang et al., 2014).
Wylie et al., 2019; Xu et al., 2018). Additionally, many studies utilized Our three approaches to estimating age can be construed as a con­
NFI data to both train and test their age estimation methodologies, tinuum of data confidence and reliability. Direct observation of changes
whereas herein we have used NFI data exclusively for assessment pur­ in the forest that determine age (i.e. stand replacing disturbance) pro­
poses. Our results were commensurate to studies that covered similar or vide a spatially explicit and transparent assessment of forest age with
larger study areas at a lower (1 km) spatial resolution (Besnard et al., high confidence and reliability in the disturbance approach that relies on
the Landsat time series to provide accurate dating of disturbance events.
Table 7 The recovery approach involves inference based on the capacity of
Percentage of forested area belonging to the old class (i.e. >150 years), by observed forest characteristics to relay the influence of disturbance
ecozone. NFI: National Forest inventory. outside the satellite record and thereby may have greater uncertainty
Ecozone Estimated old class area NFI recorded old class area
associated with it. The allometric approach is undoubtedly the least
reliable, whilst also representing the approach applied for the majority
Atlantic Maritime 4.2% 0.1%
of the forest area in Canada. Thus, while improvements to the meth­
Boreal Cordillera 3.9% 4.8%
Boreal Plains 5.1% 1.8% odology applied herein accrue automatically with the lengthening of the
Boreal Shield East 6.4% 1.1% Landsat record, improvements in the estimation of age for those areas of
Boreal Shield West 5.3% 1.1% forest that have not been disturbed in the past 50 years will be more
Hudson Plains 6.4% 5.8%
challenging to achieve.
Montane Cordillera 15.4% 37.1%
Pacific Maritime 5% 66.0%
Given the high accuracy of disturbance detection (Hermosilla et al.,
Taiga Cordillera 4% 2.4% 2016), as time progresses and more annual disturbance data become
Taiga Plains 5.3% 3.2% available in the satellite record, the relative need for other age estima­
Taiga Shield East 4% 14.8% tion methodologies will decrease somewhat, as age estimation for pro­
Taiga Shield West 2.2% 2.9%
gressively older stands will be achievable through disturbance
All Forested Ecozones 5.9% 9.5%

14
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

Table 8
Agreement of age predictions by ecozone, including Mean Absolute Deviation (MAD), Pearson’s correlation coefficient, percentage of polygons with estimated age
falling within one standard deviation of National Forest Inventory (NFI) recorded age (±1SD%), and Mean Deviation (MD).
Ecozone Recovery approach era MAD (years) Allometric approach era MAD (years) Overall MAD (years) r ±1SD% MD

Atlantic Maritime 21.9 28.1 27.1 0.20 59% − 13


Boreal Cordillera 32.9 33.1 33 0.19 66% 11.2
Boreal Plains 27.2 29.9 29.4 0.20 67% 4
Boreal Shield East 25.1 28.5 28 0.24 67% − 1.9
Boreal Shield West 21.4 30.1 28.8 0.23 62% − 8.1
Hudson Plains 19.4 31.6 30.3 0.15 62% − 9.9
Montane Cordillera 57.1 69.2 67.4 0.25 61% 57.1
Pacific Maritime 150.3 153.7 153.2 0.06 34% 144.3
Taiga Cordillera 19.4 30 28.3 0.11 56% − 7.6
Taiga Plains 25.1 31.4 30.4 0.19 67% 2.2
Taiga Shield East 37.7 47 45.8 − 0.02 78% 13
Taiga Shield West 17.2 31.5 29.4 0.12 56% − 16.1
All Forested Ecozones 31.8 39.3 38.2 0.11 79% 7.8

whereby the data are grown forward using growth and yield programs.
Issues with spectral saturation post canopy closure (Pan et al., 2011b; Xu
et al., 2018) coupled with structural heterogeneity and changing
structural growth patterns (Wylie et al., 2019; Zhang et al., 2014) lead to
significant difficulties in accurately determining age as forests grow
older. Future research could combine spectral and ALS data to charac­
terize the heterogeneity present in older forests (Sanchez-Lopez et al.,
2019). By tracking the development of heterogeneity, both spectrally
and structurally in a stand, and using it as an indicator of age, it may be
possible to increase the accuracy of age estimates in older stands.
Studies in the literature that have used remote sensing data to esti­
mate forest age have pursued a variety of approaches, using a wide range
of data sources (Table 1). Few studies have leveraged time series Landsat
data for this purpose. The recovery approach outlined herein is rela­
tively novel, with none of the reviewed studies having utilized spectral
trajectories to identify pixels recovering from disturbances that
happened prior to the Landsat record, thereby extending age estimation
outside of the satellite record. Age estimation methods in the literature
have varied greatly, as have validation approaches. Assessing the ac­
curacy of estimated age is very difficult given a lack of spatially explicit
reference data and indeed some studies have not reported validation
Fig. 9. Relationship of estimated forest age to forest age recorded in forest
stands in the National Forest Inventory (NFI) photo plots. Hexagonal bins results.
represent 10-year steps, and the dashed line represents the 1:1 line. ±1SD% = Using different data sources or methods offer abundant future op­
percent of pixels with estimated age within one standard deviation of recorded portunities for the large-area estimation of forest age. ALS-based struc­
NFI forest age, MAD = Mean Absolute Deviation in years, r = Pearson’s cor­ tural metrics such as those used in Schumacher et al. (2020) and
relation coefficient, MD = Prediction bias in years. Maltamo et al. (2020) have been shown to be effective predictor vari­
ables for forest age in boreal forests of Finland and Norway over smaller
detection. However, stand replacing disturbance is known to impact spatial extents with less variability in forest structure and complexity.
<0.5% of Canada’s net forested land area annually (White et al., 2017). Fine-scale, high-accuracy characterization of forest structural attributes
With advances in Canada-wide species classification, the recovery may allow for more precise age estimation using either allometric
approach presented herein could potentially be improved by developing equations or machine learning techniques, however it is costly and
recovery baselines based on tree species rather than ecozone. While difficult to obtain these data over very large areas. Herein we have used
currently only for 2019, if applied over the entire Landsat record, the height and canopy cover metrics imputed from ALS data (Matasci et al.,
Canada-wide species map (Hermosilla et al., 2022a), could be used to 2018a) introducing additional uncertainty in estimation.
determine recovery baselines using tree species, rather than ecozone, However, while different data sources with alternate sensing modes
potentially allowing for a more refined estimate of age for those stands or capabilities (e.g., ALS or stereo photogrammetry) or higher spatial
disturbed in the era just preceding the Landsat record. resolutions do reduce some sources of error, they are often reliant on
There are substantial challenges associated with the estimate of age high quality supplementary data, and cannot be considered a be-all and
for older forests (i.e. older than 100–200 years). A recurring problem in end-all solution for accuracy improvement. For example, Schumacher
remotely sensed age estimation is the prediction of age in older stands, et al. (2020) combined optical and ALS data to estimate forest age using
with many studies limiting their estimations to 100 to 200 years site index as a key variable. They found that for some tree species, error
(Table 1; Frate et al., 2016; Gu et al., 2016; Maltamo et al., 2020; Reyes- in age estimation nearly doubled when switching from utilizing site
Palomeque et al., 2021; Sanchez-Lopez et al., 2019). Overall, our age index values recorded in the NFI, to a modelled site index value based on
estimation approach performed most poorly in areas with relatively remotely sensed data (as would be required for any large-scale
larger amounts of older forests in the reference data. In the Pacific application).
Maritime, our MAD was 153.2 years, with much of the reference NFI All methods used to estimate forest age have associated un­
data having ages that were beyond our prediction range (i.e. >150 certainties, and consequently these uncertainties should be quantified
years). Recall that the NFI data we used as reference had ages that were and acknowledged in models used to subsequently derive other attri­
determined through a variety of approaches including modelled ages, butes of interest (Metsaranta, 2020). Uncertainty will vary according to

15
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

the approach used for estimation, with greater uncertainties associated CRediT authorship contribution statement
with the allometric approach. Information of forest age is critical for a
wide range of applications beyond forest inventory, including for carbon James C. Maltman: Conceptualization, Methodology, Investigation,
modeling and the development of greenhouse gas budgets, where Visualization, Software, Writing – review & editing. Txomin Hermo­
knowledge of forest age allows for the understanding of future carbon silla: Conceptualization, Methodology, Visualization, Investigation,
cycling and variability (Pan et al., 2011a; Vilén et al., 2012). For carbon Software, Writing – review & editing. Michael A. Wulder: Conceptu­
modeling, precision in age estimation may be most relevant for younger alization, Methodology, Investigation, Writing – review & editing.
forests (<40 years; Zhou et al., 2015), which may be well captured Nicholas C. Coops: Conceptualization, Methodology, Investigation,
within the Landsat record. For unmanaged forest areas, the absence of Writing – review & editing. Joanne C. White: Conceptualization,
forest inventory data creates a paucity of information concerning forest Methodology, Investigation, Writing – review & editing.
age. Nationwide forest age estimates allow for a more complete under­
standing of forest age spatial distribution.
By way of context for the results presented herein, the mean age of Declaration of Competing Interest
Canada’s forests, as reported by the NFI, is 90 years, with regional
variation present. When stratified to forests ≤150 years old, the agree­ The authors declare that they have no known competing financial
ment (MAD) between our estimation approach and the NFI is 38.2 years. interests or personal relationships that could have appeared to influence
Our age estimation approach results in spatially explicit outcomes, with the work reported in this paper.
all 30-m treed pixels found in Canada’s 650 Mha forested ecosystems
receiving an age estimate from one of the three methods presented Data availability
herein. Our estimated mean age of 74 years only considers forests that
are ≤150 years, which represents 94.1% of the treed area. Including The forest age map described herein for Canada’s forested ecosys­
forests older than 150 years, which represent 5.9% of the treed area, tems is open access and is available at https://opendata.nfis.
would result in a mean age >70 years. org/mapserver/nfis-change_eng.html.

6. Conclusion Acknowledgments

In this paper we prototyped a methodology for the estimation of This research was undertaken as part of ongoing efforts by the Ca­
forest age across Canada’s forested ecozones at a 30-m spatial resolu­ nadian Forest Service (CFS) of Natural Resources Canada (NRCan) to
tion. Remotely sensed data are utilized to determine age where distur­ develop a National Terrestrial Ecosystem Monitoring (NTEMS) for
bance can be identified (disturbance approach) or inferred (recovery Canada. Coops was supported in part by a NSERC grant (RGPIN-2018-
approach), and inverted allometric equations are utilized to model age 03851).
where there is no evidence of disturbance (allometric approach). The Geordie Hobart, of the Canadian Forest Service, is thanked for his
disturbance approach is based directly upon satellite data and is the assistance in acquiring the Landsat data and preparing the annual best
most accurate. The recovery approach also avails upon satellite data and available pixel image composites. This research was enabled in part by
logic regarding forest succession, with an accuracy that is greater than the computational support provided by the Digital Research Alliance of
pure modeling. Given the lack of widespread recent disturbance over Canada (www.alliancecan.ca). Open access supported by the Govern­
Canada’s forests, the allometric approach is required over the greatest ment of Canada.
area (82%). Using information regarding realized heights and growth
and yield modeling, ages are estimated where none are otherwise References
possible. As such, we mapped trees of all ages, with older trees put into a
>150 years category, representing 5.9% of national forest area. Given Bartels, S.F., Chen, H.Y.H., Wulder, M.A., White, J.C., 2016. Trends in post-disturbance
recovery rates of Canada’s forests following wildfire and harvest. For. Ecol. Manag.
the spatially explicit nature of the mapping implemented, those inter­
361, 194–207. https://doi.org/10.1016/j.foreco.2015.11.015.
ested in older trees can stratify or investigate further as required for a Bergeron, Y., Vijayakumar, D.B., Ouzennou, H., Raulier, F., Leduc, A., Gauthier, S., 2017.
given application or information need. The estimated ages can be used Projections of future forest age class structure under the influence of fire and
as is or placed into categories resembling those found in jurisdictional harvesting: implications for forest management in the boreal forest of eastern
Canada. For. Int. J. For. Res. 90, 485–495. https://doi.org/10.1093/forestry/
forest inventories. Creation of maturity classes (e.g., young, intermedi­ cpx022.
ate, old) to offer information or enable inference regarding independent Besnard, S., Koirala, S., Santoro, M., Weber, U., Nelson, J., Gütter, J., Herault, B.,
ecological response variables is also possible. Kassi, J., N’Guessan, A., Neigh, C., Poulter, B., Zhang, T., Carvalhais, N., 2021.
Mapping global forest age from forest inventories, biomass and climate data. Earth
Herein, different approaches to forest age estimation were applied Syst. Sci. Data 13, 4881–4896. https://doi.org/10.5194/essd-13-4881-2021.
depending on data availability and presence of stand replacing distur­ Boulanger, Y., Girardin, M., Bernier, P.Y., Gauthier, S., Beaudoin, A., Guindon, L., 2017.
bance on the landscape. Uncertainty in age estimation increased with Changes in mean forest age in Canada’s forests could limit future increases in area
burned but compromise potential harvestable conifer volumes. Can. J. For. Res. 47,
increasing age, as demonstrated by both aspatial and spatial compari­ 755–764. https://doi.org/10.1139/cjfr-2016-0445.
sons to NFI data. Our results point to the importance of continuity of Bradford, J.B., Birdsey, R.A., Joyce, L.A., Ryan, M.G., 2008. Tree age, disturbance
Earth observation programs such as Landsat, which provide valuable history, and carbon stocks and fluxes in subalpine Rocky Mountain forests. Glob.
Change Biol. 14 (2882–2897), 14. https://doi.org/10.1111/j.1365-
observations of disturbance and regrowth that can inform age
2486.2008.01686.x.
estimation. Burns, R.M., Honkla, 1990. Silvics of North America. U.S. Government Printing Office.
The age estimates generated in this research provide a heretofore- Canadian Forest Service, 2020. Canadian National Fire Database.
Carmean, W.H., Hahn, J.T., 1981. Revised Site Index Curves for Balsam Fir and White
unavailable, 30-m resolution, spatially explicit map of forest age
Spruce in the Lake States, No. NC-RN-269. U.S. Department of Agriculture, Forest
across Canada. The methods shared in this paper are portable to other Service, North Central Forest Experiment Station, St. Paul, MN. https://doi.org/
study locations where similar data are available and are likely to be most 10.2737/NC-RN-269.
effective when applied to boreal and temperate forests where distur­ Chen, H.Y.H., Klinka, K., 2000. Height growth curves and site index tables for subalpine
fir, engelmann spruce, and lodgepole pine in the ESSF zone of B.C. West. J. Appl. For.
bance strongly drives the forest age class distribution. Spatially explicit 62–69.
estimates of forest age at a medium spatial resolution, such as those Conner, R.N., Dickson, J.G., 1997. Relationships between bird communities and Forest
presented herein, are an important attribute for monitoring and age, structure, species composition and fragmentation in the west gulf coastal plain.
Tex. J. Sci. 493 (Suppl), 123–138.
modeling purposes in forestry, conservation, and scientific inquiries. Coops, N.C., Hermosilla, T., Wulder, M.A., White, J.C., Bolton, D.K., 2018. A thirty year,
fine-scale, characterization of area burned in Canadian forests shows evidence of

16
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

regionally increasing trends in the last decade. PLOS ONE 13, e0197218. https://doi. forest monitoring. Int. J. Digit. Earth 9, 1035–1054. https://doi.org/10.1080/
org/10.1371/journal.pone.0197218. 17538947.2016.1187673.
Costanza, R., d’Arge, R., de Groot, R., Farber, S., Grasso, M., Hannon, B., Limburg, K., Kangas, A., Maltamo, M. (Eds.), 2006. Forest Inventory: Methodology and Applications,
Naeem, S., O’Neill, R.V., Paruelo, J., Raskin, R.G., Sutton, P., van den Belt, M., 1997. Managing Forest Ecosystems. Springer, Dordrecht.
The value of the world’s ecosystem services and natural capital. Nature 387, Kayitakire, F., Hamel, C., Defourny, P., 2006. Retrieving forest structure variables based
253–260. https://doi.org/10.1038/387253a0. on image texture analysis and IKONOS-2 imagery. Remote Sens. Environ. 102,
Diao, J., Feng, T., Li, M., Zhu, Z., Liu, J., Biging, G., Zheng, G., Shen, W., Wang, H., 390–401. https://doi.org/10.1016/j.rse.2006.02.022.
Wang, J., Ji, B., 2020. Use of vegetation change tracker, spatial analysis, and random Kennedy, R.E., Yang, Z., Cohen, W.B., 2010. Detecting trends in forest disturbance and
forest regression to assess the evolution of plantation stand age in Southeast China. recovery using yearly landsat time series: 1. LandTrendr — temporal segmentation
Ann. For. Sci. 77, 27. https://doi.org/10.1007/s13595-020-0924-x. algorithms. Remote Sens. Environ. 114, 2897–2910. https://doi.org/10.1016/j.
Dolid, W., Lundgren, A., 1970. Biological growth functions describe published site index rse.2010.07.008.
curves for Lake States timber Species. Koch, A.J., Driscoll, D.A., Kirkpatrick, J.B., 2008. Estimating the accuracy of tree ageing
Ecke, F., Löfgren, O., Sörlin, D., 2002. Population dynamics of small mammals in relation methods in mature Eucalyptus obliqua forest, Tasmania. Aust. For. 71, 147–159.
to Forest age and structural habitat factors in northern Sweden. J. Appl. Ecol. 39, https://doi.org/10.1080/00049158.2008.10676281.
781–792. https://doi.org/10.1046/j.1365-2664.2002.00759.x. Lechner, A.M., Foody, G.M., Boyd, D.S., 2020. Applications in remote sensing to Forest
Ecological Stratification Working Group, 1995. A National Ecological Framework for ecology and management. One Earth 2, 405–412. https://doi.org/10.1016/j.
Canada. Centre for Land and Biological Resources Research, Research Branch, oneear.2020.05.001.
Agriculture and Agri-Food Canada. Li, D., Ju, W., Fan, W., Gu, Z., 2014. Estimating the age of deciduous forests in Northeast
Finney, M., 1993. Modeling the spread and behavior of prescribed natural fires, China with enhanced thematic mapper plus data acquired in different phenological
presented at the 12th conference on fire and Forest meteorology, (Society of seasons. J. Appl. Remote. Sens. 8, 083670 https://doi.org/10.1117/1.JRS.8.083670.
American Foresters. In: Jekyll Island, GA, pp. 138–144. Magnussen, S., Russo, G., 2012. Uncertainty in photo-interpreted forest inventory
Franklin, J.F., Johnson, K.N., Johnson, D.L., 2018. Ecological forest management. variables and effects on estimates of error in Canada’s National Forest Inventory.
Waveland Press Inc, Long Grove, Illinois. For. Chron. 88, 439–447. https://doi.org/10.5558/tfc2012-080.
Frate, L., Carranza, M., Garfì, V., Febbraro, M., Tonti, D., Marchetti, M., Ottaviano, M., Maltamo, M., Kinnunen, H., Kangas, A., Korhonen, L., 2020. Predicting stand age in
Santopuoli, G., Chirici, G., 2016. Spatially explicit estimation of forest age by managed forests using National Forest Inventory field data and airborne laser
integrating remotely sensed data and inverse yield modeling techniques. IForest - scanning. For. Ecosyst. 7, 44. https://doi.org/10.1186/s40663-020-00254-z.
Biogeosci. For. 9, 63–71. https://doi.org/10.3832/ifor1529-008. Masek, J.G., Vermote, E.F., Saleous, N.E., Wolfe, R., Hall, F.G., Huemmrich, K.F., Gao, F.,
Fraver, S., Bradford, J.B., Palik, B.J., 2011. Improving tree age estimates derived from Kutler, J., Lim, T.-K., 2006. A landsat surface reflectance dataset for North America,
increment cores: a case study of red pine. For. Sci. 572 164–170 57, 164–170. 1990–2000. IEEE Geosci. Remote Sens. Lett. 3, 68–72. https://doi.org/10.1109/
https://doi.org/10.1093/forestscience/57.2.164. LGRS.2005.857030.
Gillis, M.D., Omule, A.Y., Brierley, T., 2005. Monitoring Canada’s forests: the National Matasci, G., Hermosilla, T., Wulder, M.A., White, J.C., Coops, N.C., Hobart, G.W.,
Forest Inventory. For. Chron. 81, 214–221. https://doi.org/10.5558/tfc81214-2. Bolton, D.K., Tompalski, P., Bater, C.W., 2018a. Three decades of forest structural
Gu, H., Williams, C.A., Ghimire, B., Zhao, F., Huang, C., 2016. High-resolution mapping dynamics over Canada’s forested ecosystems using landsat time-series and lidar
of time since disturbance and forest carbon flux from remote sensing and inventory plots. Remote Sens. Environ. 216, 697–714. https://doi.org/10.1016/j.
data to assess harvest, fire, and beetle disturbance legacies in the Pacific northwest. rse.2018.07.024.
Biogeosciences 13, 6321–6337. https://doi.org/10.5194/bg-13-6321-2016. Matasci, G., Hermosilla, T., Wulder, M.A., White, J.C., Coops, N.C., Hobart, G.W.,
Gutsell, S.L., Johnson, E.A., 2002. Accurately ageing trees and examining their height- Zald, H.S.J., 2018b. Large-area mapping of Canadian boreal forest cover, height,
growth rates: implications for interpreting forest dynamics. J. Ecol. 90, 153–166. biomass and other structural attributes using landsat composites and lidar plots.
https://doi.org/10.1046/j.0022-0477.2001.00646.x. Remote Sens. Environ. 209, 90–106. https://doi.org/10.1016/j.rse.2017.12.020.
Hansen, M.C., Defries, R.S., Townshend, J.R.G., Sohlberg, R., 2000. Global land cover McArdle, R.E., Meyer, W.H., Bruce, D., 1949. The Yield of Douglas Fir in the Pacific
classification at 1 km spatial resolution using a classification tree approach. Int. J. Northwest (No. Technical Bulletin No. 201). US Department of Agriculture,
Remote Sens. 21, 1331–1364. https://doi.org/10.1080/014311600210209. Washington DC.
Hansen, M.C., Loveland, T.R., 2012. A review of large area monitoring of land cover Metsaranta, J.M., 2020. Dendrochronological procedures improve the precision and
change using landsat data. Remote Sens. Environ. Landsat Legacy Spec. Issue 122, accuracy of tree and stand age estimates in the western Canadian boreal forest. For.
66–74. https://doi.org/10.1016/j.rse.2011.08.024. Ecol. Manag. 457, 117657 https://doi.org/10.1016/j.foreco.2019.117657.
Hansen, M.C., Potapov, P., Moore, R., Hancher, M., Turubanova, S., Tyukavina, A., Meyer, W.H., 1937. YIELD OF EVEN-AGED STANDS OF SITKA SPRUCE AND WESTERN
Thau, D., Stehman, S., Goetz, S., Loveland, T., Kommareddy, A., Egorov, A., HEMLOCK, No. Bulletin No. 544. US Department of Agriculture.
Chini, L., Justice, C.O., Townshend, J., 2013. High-resolution global maps of 21st- Monteith, J.L., 1972. Solar radiation and productivity in tropical ecosystems. J. Appl.
century Forest cover change. Science 342, 850–853. https://doi.org/10.1126/ Ecol. 9, 747. https://doi.org/10.2307/2401901.
science.1244693. Myneni, R.B., Hoffman, S., Knyazikhin, Y., Privette, J.L., Glassy, J., Tian, Y., Wang, Y.,
Hanson, E.J., Azuma, D.L., Hiserote, B.A., 2002. In: Site Index Equations and Mean Song, X., Zhang, Y., Smith, G.R., Lotsch, A., Friedl, M., Morisette, J.T., Votava, P.,
Annual Increment Equations for Pacific Northwest Research Station Forest Inventory Nemani, R.R., Running, S.W., 2002. Global products of vegetation leaf area and
and Analysis Inventories, 1985–2001, p. 26. fraction absorbed PAR from year one of MODIS data. Remote Sens. Environ. 83,
Heinsch, F.A., Maosheng Zhao, S.W., Running, J.S., Kimball, R.R., Nemani, K.J., Davis, P. 214–231. https://doi.org/10.1016/S0034-4257(02)00074-3.
V., Bolstad, B.D., Cook, A.R., Desai, D.M., Ricciuto, B.E., Law, W.C., Oechel, Hyojung National Forest Inventory, 2021. Canada’s National Forest Inventory – first
Kwon, Hongyan Luo, S.C., Wofsy, A.L., Dunn, J.W., Munger, D.D., remeasurement (2007-2017) photo-plot data, version 1.0. https://nfi.nfis.org.
Baldocchi, Liukang Xu, Hollinger, D.Y., Richardson, A.D., Stoy, P.C., Siqueira, M.B. Natural Resources Canada, 2022. The State of Canada’s Forests: Annual Report 2022.
S., Monson, R.K., Burns, S.P., Flanagan, L.B., 2006. Evaluation of remote sensing https://natural-resources.canada.ca/sites/nrcan/files/forest/sof2022/SoF_Annual2
based terrestrial productivity from MODIS using regional tower eddy flux network 022_EN_access.pdf.
observations. IEEE Trans. Geosci. Remote Sens. 44, 1908–1925. https://doi.org/ Nigh, G.D., 2000. Western redcedar site index models for the interior of British
10.1109/TGRS.2005.853936. Columbia. British Columbia, Ministry of Forests Research Program, Victoria, BC.
Hermosilla, T., Bastyr, A., Coops, N.C., White, J.C., Wulder, M.A., 2022a. Mapping the Nigh, G.D., Columbia, British, Ministry of Forests, 2016. British Columbia, Forest
presence and distribution of tree species in Canada’s forested ecosystems. Remote Analysis and Inventory Branch. Revised site index models for Western Redcedar for
Sens. Environ. https://doi.org/10.1016/j.rse.2022.113276. Coastal British Columbia.
Hermosilla, T., Wulder, M.A., White, J.C., Coops, N.C., 2022b. Land cover classification Nigh, G.D., Courtin, P.J., 1998. Height models for red Alder (Alnus rubra bong.) in
in an era of big and open data: optimizing localized implementation and training British Columbia. New For. 16, 59–70.
data selection to improve mapping outcomes. Remote Sens. Environ. 268, 112780 Pan, Y., Birdsey, R.A., Fang, J., Houghton, R., Kauppi, P.E., Kurz, W.A., Phillips, O.L.,
https://doi.org/10.1016/j.rse.2021.112780. Shvidenko, A., Lewis, S.L., Canadell, J.G., Ciais, P., Jackson, R.B., Pacala, S.W.,
Hermosilla, T., Wulder, M.A., White, J.C., Coops, N.C., Hobart, G.W., 2018. Disturbance- McGuire, A.D., Piao, S., Rautiainen, A., Sitch, S., Hayes, D., 2011a. A large and
informed annual land cover classification maps of Canada’s forested ecosystems for a persistent carbon sink in the World’s forests. Science 333, 988–993. https://doi.org/
29-year landsat time series. Can. J. Remote. Sens. 44, 67–87. https://doi.org/ 10.1126/science.1201609.
10.1080/07038992.2018.1437719. Pan, Y., Chen, J.M., Birdsey, R., McCullough, K., He, L., Deng, F., 2011b. Age structure
Hermosilla, T., Wulder, M.A., White, J.C., Coops, N.C., Hobart, G.W., 2017. Updating and disturbance legacy of north american forests. Biogeosciences 8, 715–732.
landsat time series of surface-reflectance composites and forest change products with https://doi.org/10.5194/bg-8-715-2011.
new observations. Int. J. Appl. Earth Obs. Geoinformation 63, 104–111. https://doi. Payandeh, B., 1974. Nonlinear Site Index Equations for Several Major Canadian Timber
org/10.1016/j.jag.2017.07.013. Species.
Hermosilla, T., Wulder, M.A., White, J.C., Coops, N.C., Hobart, G.W., 2015a. Regional Prince, S.D., Goward, S.N., 1995. Global primary production: a remote sensing approach.
detection, characterization, and attribution of annual forest change from 1984 to J. Biogeogr. 22, 815–835. https://doi.org/10.2307/2845983.
2012 using landsat-derived time-series metrics. Remote Sens. Environ. 170, Reyes-Palomeque, G., Dupuy, J.M., Portillo-Quintero, C.A., Andrade, J.L., Tun-Dzul, F.J.,
121–132. https://doi.org/10.1016/j.rse.2015.09.004. Hernández-Stefanoni, J.L., 2021. Mapping forest age and characterizing vegetation
Hermosilla, T., Wulder, M.A., White, J.C., Coops, N.C., Hobart, G.W., 2015b. An structure and species composition in tropical dry forests. Ecol. Indic. 120, 106955
integrated landsat time series protocol for change detection and generation of annual https://doi.org/10.1016/j.ecolind.2020.106955.
gap-free surface reflectance composites. Remote Sens. Environ. 158, 220–234. Rogers, B.M., Mackey, B., Shestakova, T.A., Keith, H., Young, V., Kormos, C.F.,
https://doi.org/10.1016/j.rse.2014.11.005. DellaSala, D.A., Dean, J., Birdsey, R., Bush, G., Houghton, R.A., Moomaw, W.R.,
Hermosilla, T., Wulder, M.A., White, J.C., Coops, N.C., Hobart, G.W., Campbell, L.B., 2022. Using ecosystem integrity to maximize climate mitigation and minimize risk in
2016. Mass data processing of time series landsat imagery: pixels to data products for international forest policy. Front. For. Glob. Change 5.

17
J.C. Maltman et al. Remote Sensing of Environment 290 (2023) 113529

Running, S., Mu, Q., Zhao, M., 2015. MOD17A2H MODIS/Terra Gross Primary White, J.C., Wulder, M.A., Hobart, G.W., Luther, J.E., Hermosilla, T., Griffiths, P.,
Productivity 8-Day L4 Global 500m SIN Grid V006. https://doi.org/10.5067/ Coops, N.C., Hall, R.J., Hostert, P., Dyk, A., Guindon, L., 2014. Pixel-based image
MODIS/MOD17A2H.006. compositing for large-area dense time series applications and science. Can. J.
Running, S.W., Nemani, R.R., Heinsch, F.A., Zhao, M., Reeves, M., Hashimoto, H., 2004. Remote. Sens. 40, 192–212. https://doi.org/10.1080/07038992.2014.945827.
A continuous satellite-derived measure of global terrestrial primary production. Wong, C.M., Lertzman, K.P., 2001. Errors in estimating tree age: implications for studies
Bioscience 54, 547. https://doi.org/10.1641/0006-3568(2004)054[0547:ACSMOG] of stand dynamics. Can. J. For. Res. 31, 1262–1271. https://doi.org/10.1139/x01-
2.0.CO;2. 060.
Russo, D., Cistrone, L., Garonna, A., Jones, G., 2010. Reconsidering the importance of Woodcock, C.E., Allen, R., Anderson, M., Belward, A., Bindschadler, R., Cohen, W.,
harvested forests for the conservation of tree-dwelling bats. Biodivers. Conserv. 19, Gao, F., Goward, S.N., Helder, D., Helmer, E., Nemani, R., Oreopoulos, L., Schott, J.,
2501–2515. https://doi.org/10.1007/s10531-010-9856-3. Thenkabail, P.S., Vermote, E.F., Vogelmann, J., Wulder, M.A., Wynne, R., 2008. Free
Sader, S.A., Waide, R.B., Lawrence, W.T., Joyce, A.T., 1989. Tropical forest biomass and access to landsat imagery. Science 320, 1011. https://doi.org/10.1126/
successional age class relationships to a vegetation index derived from landsat TM science.320.5879.1011a.
data. Remote Sens. Environ. 28, 143–198. https://doi.org/10.1016/0034-4257(89) Wulder, M.A., Hermosilla, T., Stinson, G., Gougeon, F.A., White, J.C., Hill, D.A.,
90112-0. Smiley, B.P., 2020. Satellite-based time series land cover and change information to
Sanchez-Lopez, N., Boschetti, L., Hudak, A.T., 2019. Reconstruction of the disturbance map forest area consistent with national and international reporting requirements.
history of a temperate coniferous forest through stand-level analysis of airborne For. Int. J. For. Res. 93, 331–343. https://doi.org/10.1093/forestry/cpaa006.
LiDAR data. For. Int. J. For. Res. cpz048 https://doi.org/10.1093/forestry/cpz048. Wulder, M.A., Roy, D.P., Radeloff, V.C., Loveland, T.R., Anderson, M.C., Johnson, D.M.,
Schmidt, G.L., Jenkerson, C.B., Masek, J., Vermote, E., Gao, F., 2013. Landsat ecosystem Healey, S., Zhu, Z., Scambos, T.A., Pahlevan, N., Hansen, M., Gorelick, N.,
disturbance adaptive processing system (LEDAPS) algorithm description. US Geol Crawford, C.J., Masek, J.G., Hermosilla, T., White, J.C., Belward, A.S., Schaaf, C.,
Surv Open File Rep 1057. Woodcock, C.E., Huntington, J.L., Lymburner, L., Hostert, P., Gao, F., Lyapustin, A.,
Schroeder, T., Schleeweis, K., Moisen, G., Toney, C., Cohen, W., Freeman, E., Yang, Z., Pekel, J.-F., Strobl, P., Cook, B.D., 2022. Fifty years of landsat science and impacts.
Huang, C., 2017. Testing a landsat-based approach for mapping disturbance Remote Sens. Environ. 280, 113195 https://doi.org/10.1016/j.rse.2022.113195.
causality in U.S. Forests. Remote Sens. Environ. 195, 230–243. https://doi.org/ Wulder, M.A., White, J.C., Cranny, M., Hall, R.J., Luther, J.E., Beaudoin, A.,
10.1016/j.rse.2017.03.033. Goodenough, D.G., Dechka, J.A., 2008. Monitoring Canada’s forests. Part 1:
Schumacher, J., Hauglin, M., Astrup, R., Breidenbach, J., 2020. Mapping forest age using completion of the EOSD land cover project. Can. J. Remote. Sens. 34, 549–562.
National Forest Inventory, airborne laser scanning, and Sentinel-2 data. For. Ecosyst. https://doi.org/10.5589/m08-066.
7, 60. https://doi.org/10.1186/s40663-020-00274-9. Wulder, M.A., White, J.C., Luther, J.E., Strickland, G., Remmel, T.K., Mitchell, S.W.,
Scott, C.T., Voorhis, N.G., 1986. Northeastern Forest survey site index equations and site 2006. Use of vector polygons for the accuracy assessment of pixel-based land cover
productivity classes. North. J. Appl. For. 3, 144–148. https://doi.org/10.1093/njaf/ maps. Can. J. Remote. Sens. 32, 268–279. https://doi.org/10.5589/m06-023.
3.4.144. Wulder, M.A., White, J.C., Magnussen, S., McDonald, S., 2007. Validation of a large area
Simard, M., Pinto, N., Fisher, J.B., Baccini, A., 2011. Mapping forest canopy height land cover product using purpose-acquired airborne video. Remote Sens. Environ.
globally with spaceborne lidar. J. Geophys. Res. Biogeosciences 116. https://doi. 106, 480–491. https://doi.org/10.1016/j.rse.2006.09.012.
org/10.1029/2011JG001708. Wylie, R.R.M., Woods, M.E., Dech, J.P., 2019. Estimating stand age from airborne laser
Smith, V.G., 1984. Asymptotic site-index curves, fact or Artifact? For. Chron. 60, scanning data to improve models of black spruce wood density in the boreal Forest of
150–156. https://doi.org/10.5558/tfc60150-3. Ontario. Remote Sens. 11, 2022. https://doi.org/10.3390/rs11172022.
Smithers, L., 1961. Lodgepole Pine in Alberta, No. Bulletin 127. Canada Department of Xu, C., Manley, B., Morgenroth, J., 2018. Evaluation of modelling approaches in
Forestry. predicting forest volume and stand age for small-scale plantation forests in New
Thrower, J., Nussbaum, A., Di Lucca, C., 1994. Site index curves and tables for British Zealand with RapidEye and LiDAR. Int. J. Appl. Earth Obs. Geoinform. 73, 386–396.
Columbia: Interior species. BC Ministry of Forests, Victoria, BC. https://doi.org/10.1016/j.jag.2018.06.021.
Vasiliauskas, S., Chen, H.Y., 2002. How long do trees take to reach breast height after fire Zhang, C., Ju, W., Chen, J.M., Li, D., Wang, X., Fan, W., Li, M., Zan, M., 2014. Mapping
in northeastern Ontario? Can. J. For. Res. 32, 1889–1892. https://doi.org/10.1139/ forest stand age in China using remotely sensed forest height and observation data:
x02-104. CHINA’S FOREST STAND AGE MAPPING. J. Geophys. Res. Biogeosci. 119,
Véga, C., St-Onge, B., 2009. Mapping site index and age by linking a time series of 1163–1179. https://doi.org/10.1002/2013JG002515.
canopy height models with growth curves. For. Ecol. Manag. 257, 951–959. https:// Zhang, Q., Pavlic, G., Chen, W., Latifovic, R., Fraser, R., Cihlar, J., 2004. Deriving stand
doi.org/10.1016/j.foreco.2008.10.029. age distribution in boreal forests using SPOT VEGETATION and NOAA AVHRR
Vilén, T., Gunia, K., Verkerk, P.J., Seidl, R., Schelhaas, M.-J., Lindner, M., Bellassen, V., imagery. Remote Sens. Environ. 91, 405–418. https://doi.org/10.1016/j.
2012. Reconstructed forest age structure in Europe 1950–2010. For. Ecol. Manag. rse.2004.04.004.
286, 203–218. https://doi.org/10.1016/j.foreco.2012.08.048. Zhao, M., Heinsch, F.A., Nemani, R.R., Running, S.W., 2005. Improvements of the
White, J.C., Wulder, M.A., Hermosilla, T., Coops, N.C., Hobart, G.W., 2017. A nationwide MODIS terrestrial gross and net primary production global data set. Remote Sens.
annual characterization of 25 years of forest disturbance and recovery for Canada Environ. 95, 164–176. https://doi.org/10.1016/j.rse.2004.12.011.
using landsat time series. Remote Sens. Environ. 194, 303–321. https://doi.org/ Zhou, T., Shi, P., Jia, G., Dai, Y., Zhao, X., Shangguan, W., Du, L., Wu, H., Luo, Y., 2015.
10.1016/j.rse.2017.03.035. Age-dependent forest carbon sink: estimation via inverse modeling. J. Geophys. Res.
White, J.C., Hermosilla, T., Wulder, M.A., Coops, N.C., 2022. Mapping, validating, and Biogeosciences 120, 2473–2492. https://doi.org/10.1002/2015JG002943.
interpreting spatio-temporal trends in post-disturbance forest recovery. Remote
Sens. Environ. 271, 112904 https://doi.org/10.1016/j.rse.2022.112904.

18

You might also like