1 s2.0 S1574954122001984 Main

You might also like

Download as pdf or txt
Download as pdf or txt
You are on page 1of 13

Ecological Informatics 70 (2022) 101748

Contents lists available at ScienceDirect

Ecological Informatics
journal homepage: www.elsevier.com/locate/ecolinf

Using multi-temporal tree inventory data in eucalypt forestry to benchmark


global high-resolution canopy height models. A showcase in Mato
Grosso, Brazil
Adrián Pascual a, *, Frederico Tupinambá-Simões b, c, Tiago de Conto a
a
Department of Geographical Sciences, University of Maryland, College Park, MD 20742, United States of America
b
Universidad de Valladolid, UVA University Institute for Research in Sustainable Forest Management, Palencia 34004, Spain
c
Sustainable Forest Management Research Institute UVa-INIA, Avda. Madrid 50, 34071 Palencia, Spain

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

Keywords: The global monitoring of forest structure worldwide is increasingly being supported by refined and enhanced
Tree height satellite mission datasets. Forest canopy height is a global metric to characterise and monitor dynamics in forest
Remote sensing ecosystems worldwide. Satellite mapping missions as NASA’s Global Ecosystem Dynamics Investigation (GEDI)
Calibration
are creating opportunities to refine global forest canopy height models adding forest structural information to
Forest mapping
Tropical forestry
time-series satellite imagery. A recent global canopy height model presented by Lang et al., (2022) using GEDI
and 10-m Sentinel-2 and the map from Potapov et al., (2020) using GEDI and Landsat are both tested in this study
using multi-temporal tree-level data collected over eucalypt plantations in Brazil. Our results at plot-level
showed Lang et al., (2022)’s estimates of canopy height came short compared to 2020 maximum and mean
tree height records in the plots, 7.6 and 3.6 m, respectively, but adding CHM standard deviation improves the
agreement of ground records for maximum tree height. Higher errors were computed for the plots in 2019 using
the Potapov’s 30-m CHM: 14.2 and 9.5 m, respectively. Averaged stand values were more similar between the
three sources tested. We report improvement from the 30-m CHM to the 10-m, but still height saturation
problems were observed when accounting for height differences in tall eucalypt trees. As more global products
for forest height and biomass are becoming available to users, more validation exercises as presented in this study
are needed to assess the suitability of CHM products to forestry needs, and facilitate the uptake and actionability
of the next generation of global height and biomass products. We provide recommendations and insights on the
use of GEDI laser data for global mapping and on the potential of commercial forestry areas to benchmark the
accuracy of satellite mapping missions focusing on tree height estimation in the tropics.

1. Introduction or combined with satellite imagery (e.g., Asner et al., 2013; Pardini
et al., 2019; Qi et al., 2019). One the most discussed solutions in CHM
Measuring forest canopy height using remote sensing observations is during the last years is the 2019 map from Potapov et al. (2020) using
a timely and global-scale research topic constantly evolving from the Landsat-8 time-series and NASA’s GEDI full waveform three-
assimilation of new sources of remote sensing data (Asner et al., 2013; dimensional laser data to produce a global, 30-m resolution CHM. The
Duncanson et al., 2019). Time series of satellite imagery and ground spaceborne lidar technology from GEDI represents a major step forward
observations have been used to map forest canopy height across the in the characterization of forest ecosystems as the mission is specifically
globe (e.g., Hansen et al., 2013). The top-oh-height of forest vegetation designed to collect 3D data for global canopy height estimation and
is usually represented in two-dimensions as a gridded product called aboveground biomass density (Dubayah et al., 2020b; Duncanson et al.,
canopy height model (CHM) where pixel/raster cell estimates represent 2022). Relative height metrics from GEDI full waveforms can be gridded
the average top-of-tree height for each pixel. Remote sensing data to conform global CHM using different techniques means and sources of
collected from active sensors such as LiDAR or SAR are solid approaches satellite imagery as showcased by Potapov et al. (2020).
on which to rely the estimation on 3D forest attributes globally – alone The product from Potapov et al. (2020) used a limited sampling

* Corresponding author.
E-mail address: apascual@umd.edu (A. Pascual).

https://doi.org/10.1016/j.ecoinf.2022.101748
Received 27 May 2022; Received in revised form 23 June 2022; Accepted 9 July 2022
Available online 18 July 2022
1574-9541/© 2022 Elsevier B.V. All rights reserved.
A. Pascual et al. Ecological Informatics 70 (2022) 101748

catalog of the GEDI mission for year 2019 using version 001 of GEDI The purpose of this study is to FILLIN ""evaluate the suitability of
observations, FILLIN ""whose mean ground geo-location error could relevant global CHM products, Potapov et al. (2020)’s and Lang et al.
exceed 10 m (Dubayah et al., 2020b). Since then and until mission is (2022a, 2022b), to estimate forest top-of-canopy height. The study area
completed by the end of 2023, the geo-location of GEDI footprints is for this assessment is in Mato Grosso, Brazil, over 1400 ha of commercial
being improved along successive releases of GEDI mission products. eucalypt forestry previously used to monitor the effects of climate
Estimates of forest canopy height and aboveground carbon density change on tree mortality (Pascual et al., 2022; Tupinambá-Simões et al.,
(AGBD) for >10 billion footprints are to be collected along the lifespan 2022). For this study, >10,000 tree height measurements between 2018
of the GEDI mission (Dubayah et al., 2020b). These laser footprints and 2020 were used to assess the suitability of the latest 10-m CHM
collecting forest structure data can be then gridded, considering product and to measure the improvement from the earlier 30-m CHM
geographic regions and plant functional types (PFT), to compose global product. The exercise was carried out in challenging conditions: fast-
CHM for the case of forest canopy height. As the current GEDI version growing, tall trees in the tropics. The refinement of global CHM prod­
002 is publicly available and it reduces the geo-location error substan­ ucts is gaining momentum towards the inclusion of 3D structural data,
tially (below 5 m) from data version 001, further refinements have and this has important synergies towards sustainability and conserva­
margin to improve the quality of earlier CHM products (Roy et al., tion (Cordell et al., 2017; Csillik et al., 2019; Laclau et al., 2013), the
2021). Depending on the scale of geo-location errors, these could mapping of forest AGBD based on tree structural variables (Duncanson
severely impact the statistical relationships between laser footprint data et al., 2022) and ongoing satellite mapping missions such as Norway’s
and both satellite imagery and ground reference data. New solutions in International Climate & Forests Initiative (NICFI) Planet tropical mo­
CHM and AGBD modelling should then integrate GEDI laser data under saics - releasing monthly image mosaics to enforce land-use monitoring
version 002 to limit the influence of co-registration errors. Climate in tropical forest ecosystems.
conditions or plant functional types are two other factors that can
impact the quality of CHM as documented in Potapov et al. (2020) and 2. Material and methods
other previous studies, especially for the case of tall trees – under­
estimated when using optical remote sensing data in the statistical 2.1. Study area and multi-annual mortality surveys
inference. The problem, referred to as height saturation in the field of
forest mapping, is a concern when assessing carbon stocks in tropical The pilot area is Tangará da Serra, a rural area Mato Grosso state in
forests as large, tall trees yielding high AGBD stocks might be under- Brazil (Fig. 1). The showcase research area is a 1400-ha mosaic of
estimated (e.g., Mitchard et al., 2014; Urbazaev et al., 2015). Several planted eucalypt stands (57) showing 13 annual classes. Spacing,
studies have reported on this limitation (Lang et al., 2022a; Silva et al., stocking and clone selection varies across the site: E. urograndis, E.
2021) and hence, new refinements on forest canopy height estimation urophylla x E. grandis, E. urophylla and E. camaldulensis x E. grandis are the
have been proposed (Healey et al., 2020; Lang et al., 2022b). four clone selections. The site follows an intense management and a
The study from Lang et al. (2022b) is one of the latest and most so­ close monitoring: The area was severely affected by a drought in 2019
phisticated refinements on global CHM and it tackles the afore-said affecting growth conditions and lower tree survival rates as observed for
needs unresolved in earlier attempts (e.g., Potapov et al., 2020). The the 2018–2020 campaigns (Tupinambá-Simões et al., 2022). We used
product from Lang et al. (2022b) uses GEDI laser data and satellite image ground data from the 2019–2021 to cover the temporal range defined by
spectroscopy data but at a fine-grained 10-m resolution using Sentinel-2 Lang et al. (2022a, 2022b)’s CHM build at 10-m resolution later
(S2) FILLIN ""imagery as support. The expected improvements in the 10- described. Field campaigns began on April 1st and June 1st for 2019 and
m CHM product are the result of using a larger portfolio of GEDI data 2020, respectively, and they were both completed in a month (see
collected under enhanced geo-location conditions, and giving more Supplementary for an extended footage of the field data collection
weight to tall trees when designing an approach based on convolutional including aerial overviews collected with a commercial-grade drone).
neural network (Fayad et al., 2021a: Lang et al., 2022b). The statistical The dataset included a total of 13,347 tree height measurements and
relationship between metrics from GEDI full waveforms metric and S1 120 plots measured in 2019 and 2020 located within 57 stands. Trees
10-m pixels was based on quality filters and on the proximity between whose height was not measured in the fields (19,738) were predicted
GEDI 25-m footprint center locations and pixels from S1 imagery (Lang from a local-calibrated model using diameter at breast height (dbh, cm)
et al., 2022b). The temporal scope of the 10-m CHM product is set by the as predictor (Tupinambá-Simões et al., 2022). Plots (18-m radius)
authors to years 2019 and 2020. For the purpose of auditing and testing approximately accounts for 10 raster cells of Lang et al. (2022a, 2022b)
different CHM products, independent and temporally coincident data global CHM. Tall-tree conditions were represented in most of the large
sources such as tree-level measurements are useful for benchmarking stands for which, on average, comprised 4–5 sample plots. A grid of 18-
(Andersen et al., 2006), and to confirm the hypothesized improvement m side was created for the plantation and a 20-m buffer was used to filter
in the estimation of height for tall trees and specifically over regions out sampling locations over and close edges between stands. From the
where moisture-saturated conditions are common such as in the tropics set of possible sampling locations within the boundaries of each stand,
(Healey et al., 2020). random sampling was used to select the minimum number of plots to be
Forest plantations in the tropics bring then opportunities for forest measured as determine by private forestry guidelines for the site and to
height mapping by providing fast-growing conditions from which to meet the desirable sampling intensity.
model growth dynamics or effects of climate change episodes (Pascual
et al., 2022). In terms of geolocation accuracy, homogenous and well- 2.2. Global 30-m canopy height modelling
defined annual plantation units reduce geolocation uncertainty when
using GEDI laser footprints (Fayad et al., 2021a). Regarding data peri­ Landsat multi-temporal metrics and the 95th relative height from
odicity and the overlap with satellite missions’ data, fast-growing spe­ GEDI full waveforms were used to compose a 30-m CHM globally
cies can help to validate simultaneous satellite missions taking place (Potapov et al., 2020). The team at the University of Maryland has made
such as IceSat-2, NASA’s GEDI or the coming NISAR (Duncanson et al., the FILLIN ""map available here. Selected GEDI orbits ranged from
2019; Guerra-Hernández and Pascual, 2021). Fast-growing plantations (April–October 2019), same starting date as the 2019 ground campaign.
can help to refine tree and forest canopy height mapping at different From these FILLIN ""GEDI observations, metrics on surface topography,
scales from tree- and plot-level estimates to homogenous stand defined canopy height and cover, and vertical structure can be extracted.
by annual units where estimates of AGBD stocking are accurate and Textural metrics from satellite images are good indicators of land uses
management operations traceable (Fayad et al., 2021a; Guerra- while GEDI L2A metrics add forest structural information. The map has
Hernández and Pascual, 2021; Milenković et al., 2022). known issues regarding height saturation and authors report the value of

2
A. Pascual et al. Ecological Informatics 70 (2022) 101748

Fig. 1. Location of the study area in Mato Grosso (a) showing the stand boundaries and Planet NICFI tropical basemaps (May–June 2020 mosaic is represented) using
false-colour vegetation to highlight vegetation area outside the study area within stand boundaries.

30-m as height-break for which the estimation is biased towards tall each CHM estimate to account and inform user of the reliability of
trees (Potapov et al., 2020). We included the 2019 CHM from Potapov predicted height estimates. We used these two layers, publicly available,
et al. (2020) into the assessment to benchmark the expected improve­ to compare real measured data collected from eucalypt forestry assets at
ment in the newest global CHM map: the product from Lang et al. two scales: plot and stand levels.
(2022a) which provides estimates of CHM and the associated uncer­
tainty (Fig. 2), expressed by the standard deviation product, to all 10-m
2.4. Assessment of global canopy height estimates
raster cells globally (Fig. 3)

Sample plot locations and stand boundaries were used to extract


2.3. Global 10-m canopy height modelling FILLIN ""mean and maximum height values for the two 10-m CHM raster
available from Lang et al. (2022b) - model estimate and model uncer­
The global CHM for Lang et al. (2022a, 2022b) builds upon previous tainty expressed as standard deviation (SD) - and the CHM estimate from
global forest height mapping guidelines (Hansen et al., 2013; Potapov Potapov et al. (2020). The spatial layout of the two products over the
et al., 2020) and combines GEDI L2A relative height 98 (RH98) of forest boundaries and inner divisions of the test area are presented in Fig. 3 for
canopy height computed for track of 25-m ground footprints with raw the 10-m resolution products as showcases. The maximum measured
satellite images from S2 constellation from the European Space Agency tree height (hmax) and the mean tree height (hmean) were compared for
(ESA). The temporal scope of the product ranged from April to August in each year to the maximum (Hmax) and mean (Hmean) value of the CHM
2019 and 2020. The spatial alignment between discrete GEDI footprints raster estimate on each plot for year 2019 using Potapov et al. (2020)
and S2 mosaics was done for 600 million samples, and those samples and for year 2020 using Lang et al. (2022b). The average values of hmax
were defined as follows: “the sparse GEDI data is rasterized to the and hmean w FILLIN ""ere computed for each individual stand and for
Sentinel-2 10-meter grid by setting the pixel corresponding to the center each annual rotation class. For the case of Hmax and Hmean, boundaries
of each GEDI footprint to the associated footprint height” (Lang et al., were used to extract summary statistics for stands and annual rotation
2022b). The statistical inference was based on deep learning methods by units.
means of a convolutional approach (CNN). The 2020 CHM build upon To compare prediction uncertainty, we compared only at plot level
previous studies on global mapping using sophisticated machine using the SD of all measured tree heights in the plots (hstd) and the mean
learning techniques (Fayad et al., 2021b; Lang et al., 2022a). The raster value of SD for each plot (Hstd) from the 10-m product from Lang
emphasis in this recent product was placed on addressing the saturation et al. (2022b). Preliminary tests showed low co-variance of CHM esti­
effect for canopy heights above 25 m (Healey et al., 2020) – here tested mates and SD values within 18-m-plot boundaries as expected in forest
with tree-level data as it was claimed by the authors as a main plantation conditions. Measurements on tree height were modelled
improvement from the 30-m CHM model tested here indeed. An esti­ using mean and maximum values from both CHM products tested.
mation of uncertainty for each 10-m raster cell (Fig. 3) accompanies We used simple linear regression modelling to explain the

3
A. Pascual et al. Ecological Informatics 70 (2022) 101748

Fig. 2. Representation of the 30-m canopy height model (CHM) from Potapov et al. (2020) over the showcase area showing stand boundaries and permanent sample
locations. Ground campaign from 2019 was used for the assessment (a). Same for the 10-m CHM from Lang et al. (2022b) showing more clearly the locations of the
sample plots (b).

relationships between measured values of mean and maximum tree • Models 1–2 compute model error fixing the intercept and slope to
height at plot level. At stand level, we aggregated plot values to estimate explain measured mean and maximum tree height, hmax and hmean,
mean and maximum tree height in the stand. The explanation of the respectively, using as predictor variables the corresponding mean
models presented in Table 1 is as follows: and maximum value on each two CHM rasters here compared.

4
A. Pascual et al. Ecological Informatics 70 (2022) 101748

Fig. 3. Spatial layout and distribution of estimates of canopy height model (CHM) from Lang et al. (2022b) and the associated standard deviation (SD) for each 10-m
CHM pixels.

• Model 3: the difference between estimated and measured maximum


Table 1
height in the plots was explained by using the latter as a categorical
Models to test Potapov et al. (2020) 30-m and Lang et al. (2022b)’s 10-m canopy
predictor rather than as a continuous variable. We used 5-m intervals
height models.
to model the deviation in height between measurements and esti­
Model Formula Parameters mates from the CHM products (h5m max). The coefficients from the
1 hmax~β0 + β1Hmax + ε β0 = 0 β1 = 1 model fitting were used to interpret differences between height
2 hmean~β0 + β1Hmean + ε β0 = 0 β1 = 1 classes and test whether saturation persists for tall trees or not.
3 hmax~β0 + β1 h5m β0 = 0 f (OLS)
max + ε
• Model 4 was designed for Lang et al. (2022a, 2022b) product as the
β1 =
4 hmax~β0 + β1 (Hmax + Hstd) + ε β0 = 0 β1 = f (OLS)
model combines as a single predictor the sum of the average CHM
*β1 coefficients can be fixed or optimized using ordinary least squares (OLS) in uncertainty for each plot and the maximum height value for each
regression modelling. plot. We outlined Model 4 aiming to maximize the usability of the
existing SD product that informs on uncertainty for each 10-m cell.

5
A. Pascual et al. Ecological Informatics 70 (2022) 101748

The model fitting results for Model 4 can confirm whether or not this 3.2. Plot-level assessment of 30-m CHM estimates
sub-product can be useful to better align estimated heights to
measured maximum tree height (hmax). The mean difference between measured maximum tree height and
the maximum value of the 30-CHM product for the 120 plots measured
3. Results in year 2019 was 14.12 m. For mean tree height, the value was 9.51 m.
In terms of RMSE for Model 1 and Model 2, the values reached 15.34 and
3.1. Field data on tree height and stand canopy height 11.27 m, respectively, which represent 50.0% and 34.7% when
expressed as relative RMSE. Trends for both variables are quite consis­
The distribution of tree height measurements for the three surveys tent when comparing absolute values and when accounting for differ­
showed similar tendencies for mean and maximum tree height. Most of ences (Fig. 5). Model coefficients in Table 2 showed an error increasing
the maximum tree height values were above the 25–30 m threshold in systematic difference between field observations and 30-m CHM values
the plots- meeting the conditions for testing the expected improvement for tall canopy height values.
in height saturation to assess CHM products (Fig. 4).

Fig. 4. Distribution of measured maximum (grey) and mean tree height (blue) in the 2019–2021 annual surveys (displayed from left to right). Tree measurements
were computed for plots, and these were re-scaled to stand values. (For interpretation of the references to colour in this figure legend, the reader is referred to the web
version of this article.)

6
A. Pascual et al. Ecological Informatics 70 (2022) 101748

Fig. 5. Height estimate and differences at plot-level between measurements in the 2019 campaign (y-axis) and canopy height model (CHM) estimates in the x-axis
from Potapov et al. (2020). A set of 120 plots measured in 2019 were used.

Table 2
Model 3 fitting statistics and coefficient values for the maximum height 5-m classes.
Height class 30-m Canopy Height Model 10-m Canopy Height Model

Potapov et al. (2020) Lang et al. (2022b)

Height class Estimate Std. t Pr (>|t|) Estimate Std. t Pr (>|t|)

β2 Error value β2 Error value

10–15 1.025 2.580 0.397 0.692 4.734 1.869 2.534 0.01248*


15–20 9.525 1.490 6.394 3.59e-09 *** 4.586 1.999 2.294 0.02339*
20–25 14.603 0.860 16.979 <2e-16*** 4.01 1.413 2.840 0.00524**
25–30 14.328 0.737 19.437 <2e-16*** 11.365 1.037 10.959 <2e-16***
30–35 18.316 1.184 15.471 <2e-16*** 5.647 0.741 7.626 4.66e-12***
35–40 – – – – 10.346 1.037 9.976 <2e-16***
40–45 – – – – 14.200 2.644 5.370 3.54e-07***
*
Significance level at p < 0.01.
**
Significance level at p < 0.005.
***
Significance level at p < 0.001.

3.3. Plot-level assessment of 10-m CHM estimates The results for Model 4 showed the improvement in the relationship
between measured and estimated values for maximum tree height in the
The 2019–2021 field surveys showed the effect of temporal plots when accounting for estimates of uncertainty (Fig. 7). Adding the
mismatch in the 10-m CHM designed for year 2020: some plots were available uncertainty of each pixel to the prediction decreased the RMSE
harvested in 2019 and 2021 (Fig. 6). Hence, we narrowed the assess­ from 9.52 m in Model 1 to 6.4 m for Model 4. Indeed, the mean differ­
ment to the 2020 ground campaign as dependent predictor in the set of ence between measured hmax and (Hmax + Hstd) computed from Lang
models. The RMSE was 9.52 m for Model 1 (hmax) and 7.21 m for Model et al. (2022b)’s CHM was zero in the plot training data. In this regard,
2 (hmean). In terms of model error, the values for relative RMSE were some plots around the 25–30 m height class showed low estimated
32.2% and 28.5%, respectively. Model fitting results for Model 3 showed values of both Hmax and Hmean comparing to the corresponding measured
the difference between measured hmax and estimated Hmax was the variable, hmax and hmean, respectively. Those plots showed low values -
largest for height classes 25–30 m and for the two classes between 35 and sometimes zero - for 10-m estimates of uncertainty(Hstd), lowering
and 45 m as derived from Model 3 coefficients and their significance the overall good performance of Model 4 and mitigating the improve­
(Table 2). ment for the entire set of plots when compared to Models 1–2 in terms of

7
A. Pascual et al. Ecological Informatics 70 (2022) 101748

Fig. 6. Relationship between measured maximum tree height in the samples for 2019, 2020 and 2021 (x-axis) and the maximum value of the 10-m canopy height
model from Lang et al. (2022a, 2022b) (y-axis).

Fig. 7. Relationship between 2020 Lang et al. (2022b) canopy height model (CHM) at plot-level and measured tree maximum and mean height in the plots (Model 1,
left). One standard deviation (SD) was added to the CHM estimate for each plot to improve model fitting (Model 4, right).

8
A. Pascual et al. Ecological Informatics 70 (2022) 101748

RMSE reduction. 4. Discussion

3.4. Standard deviation of tree canopy estimates The assessment presented here evaluates the suitability of global
CHM for the purpose of operational needs in commercial forestry.
The SD of tree height measurement was constant along the three Despite the challenging fast-growing conditions and the broad global
years: the mean values for hstd were 2.12 m for 2019 campaign, 2.29 m scope of both CHM tested here, the agreement on forest canopy height is
for 2020 and 2.32 m for year 2021. These field-based measurements of remarkable under Model 4 when using the 10-m CHM product while
SD represent a third of the SD estimated using the 10-m CHM product adding estimates of SD. Under those conditions, we observed a good
using samples: mean standard deviation (Hstd) was 6.84 m, on average, relationship between field data collected in year 2020 and Lang et al.
with a maximum of 7.11 m and a low standard deviation of 21 cm. (2022b) global CHM scoped for year 2020. Beyond the precision of CHM
Hence, the estimated variability of tree canopy heights was a third of estimates for the training area and the benchmarking numbers in this
what Lang et al. (2022a, 2022b) predicted at plot-level. The variability assessment, it is important to express some considerations regarding the
of SD estimates within stand boundaries was 1.01 m on average for the statistical methods behind CHM estimates for both CHM products here
57 stands, ranging from 0.4 m to a maximum of 3.7 m in areas coloured benchmarked.
in dark green (Fig. 3). Low SD values for the 10-m product were expected The product from Lang et al. (2022b) is a major and outstanding
as preliminary suggested by plot-level values and considering the even- contribution to forest mapping science. Authors were able to fuse GEDI
aged conditions of stands in the study. FILLIN ""We identified the full waveforms metric relative height 98 (RH98) with satellite imagery to
occurrence of edge effects between forest and non-forest areas by produce a fine-grained map of forest canopy height through the asso­
checking NICFI tropical basemaps along the contours of stand bound­ ciation of GEDI discrete observations (25-m footprints) to adjacent
aries (Fig. 8). pixels of 10-m S1 satellite image mosaics. The novel 10-m CHM product
and the approach for connecting GEDI shots and imagery is tested here
using eucalypt plantation data overlapping in time and space, and
further compared to the 30-m res baseline CHM product – which we

Fig. 8. Standard deviation map of CHM estimates from Lang et al. (2022b) showing plots location and stand boundaries, the two scales assessed (a). Zoomed area in
the transition between standing forest and harvested stands using a false-colour composite from Planet NICFI basemaps (May–June 2020) (b). Same zoomed area
showing the standard deviation map of CHM estimates from Lang et al. (2022b) to highlight the edge effect as shown by yellow pixels between forest and non-forest
areas (c). (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of this article.)

9
A. Pascual et al. Ecological Informatics 70 (2022) 101748

Fig. 8. (continued).

further validated using data for year 2019. Lang’s study produced a means of Model 4, it can help towards data calibration, substantially
trained model based on neural networks (NN) which can be specifically improving the agreement between CHM estimates to ground-truth
refined for the purpose of commercial forest assets to produce local height estimates – this is essential for monitoring (Mitchard et al., 2014).
CHMs at narrower time intervals by simply using S2 images with the Our results for Model 4 showed height saturation is an issue that still
lowest time lag with field measurements. Commercial plantations persists in the product according to our data within the tested bound­
represent a proper environment where to verify height estimates from aries. The addition of pixel estimates of uncertainty improved the
spaceborne and airborne sensors as homogenous compartments can help alignment to ground measured height, both for tree maxima and for
to further mitigate the geolocation error associated to GEDI footprints mean tree values, paving the way to correct the bias reported in Models
(Guerra-Hernández and Pascual, 2021; Milenković et al., 2022). These 1–2 for maximum height. These results highlight the need of computing
favorable conditions can provide valuable benchmarks to assess the uncertainty estimates for model predictions to further inform on the
expected correction in height saturation implemented in the 10-m CHM robustness of the estimation but also to correct biases as shown with
product. Model 4. The benefit of aggregating uncertainty estimates to CHM es­
The 10-m CHM from Lang et al. (2022b) had the upper bound at 33 m timates was less evident for mean tree height for which CHM estimates
for the boundaries of the study area. For Potapov’s 2019 map, the value already showed better agreement to plot data. Some plots below 30 m
was 26 m. Our field records showed multiple plots above 40 m in (see Fig. 6b) had no increment in height when adding plot estimates of
maximum height and some homogenous stands showed uniform tall tree SD as these values were close to zero. The 3-year inventory series sup­
conditions above the 26 and 33-m height-breaks thresholds. The dif­ porting the model calibration allowed us to identify temporal changes as
ferences between ground truth and CHM data for our training plots some plots were harvested in 2019 and 2021, but not in 2020 - this
decreased from 14 m for Potapov’s to 7.6 m in Lang’s when it comes to helped us to reduce the uncertainty at interpreting the analyses and to
maximum tree height. The improvement from the 30- to the 10-m better align the temporal scope of the field data and CHM products.
product is notorious in maxima but also for mean tree height: average The showcased products by Potapov et al. (2020) and Lang et al.
difference decreased from 9.5 to 3.6 m. The refinement in Lang’s CHM is (2022b) set refence thresholds for ongoing and coming global studies on
a remarkable step forward in forest satellite mapping and we want to tree height and biomass mapping. With this assessment we have iden­
acknowledge the major contribution this product brings to the forest tified limitations but also solutions based on the data provided by the
remote sensing community. Not only resolution improved to 10 m, au­ authors, publicly available and easy to access. The improvement in
thors also added the prediction of uncertainty which, as shown here by Model 4 solutions outlines practitioners of commercial forestry and

10
A. Pascual et al. Ecological Informatics 70 (2022) 101748

tropical environments a solution to correct CHM estimates if needed and The 30-m CHM from Potapov et al. (2020) used GEDI data from
shows the actionability of uncertainty maps to imp. We identified the orbits between April–October 2019 while the GEDI catalog in Lang’s
need of correction CHM estimates for plot-level estimates of maximum et al., (2022b) included April and August both in 2019 and 2020.
tree height. Mean tree height, by definition, is smoothed by the aver­ Phenology in the winter hemisphere might have led authors to filter
aging considering the number of trees in the samples, which was high GEDI data only over short-term periods to produce global CHM, but for
and varies across the experiment. The upscaling of results at stand-level the case of tropical ecosystems this filtering leaves out many observa­
showed the same trends: FILLIN "" maximum tree height remained tions that could have improved the accuracy of global CHM models.
clearly on top of 10-m CHM estimates for Model 1–2 without correcting Indeed, there are GEDI orbits in 2019 and 2021 over the 1400-ha study
with estimates of SD. The homogenous environment commercial site, some of them outside of the filtered range by the authors (Fig. 9).
forestry stands represents is a helpful source for global product valida­ The neural network model structure behind the 10-m product can be
tion. The assessment at plot-level showed bias in estimates of SD: the SD further calibrated with enhanced image selection while using many
of tree height measurements was a third of the value compute from the more GEDI observations under enhanced geolocation uncertainty. The
10-m CHM product. version of GEDI data used for training models must be clearly specified
The 10-m CHM is valid for year 2020 although authors used GEDI every now and then in global CHM products as GEDI orbit data pro­
and S2 observations for year 2019 and 2020. The fusion of GEDI RH98 cessed under version 001 can present geolocation errors that could make
metrics at footprint level and image mosaics deserves some consider­ the association between 10-m images and footprints invalid (Roy et al.,
ations. Authors filter GEDI shots ranging between April and August in 2021). Next developments on global CHM must broad the temporal
both 2019 and 2020 and they assigned height to pixels by associating scope of the product to fully maximize the global coverage of satellite
GEDI L2A height (RH98) to the image cell ranging within the center of missions. The NASA GEDI mission is still collecting data and it will be
the 25-m GEDI footprint, exceeding the scale of S1 imagery. The reso­ doing so until later 2023 - producing refined data releases of height and
lution of 10-m for the 2020 CHM can be considered as nominal as the 10- biomass products (e.g., Dubayah et al., 2020a).
m threshold is not capable to mask out edge effect as with stand The issue of product validation and calibration will become even
boundaries in transition areas where forest structure abruptly varies. more pressing and relevant as the number of global mapping applica­
The impact of these edge effects should be negligible when computing tions is expanding boosted by satellite mapping mission such as GEDI or
average values for stands, but it highlights the mismatch in the scale NISAR among others. Same as presented here for the global forest height
between GEDI 25-m footprints where canopy height is retrieved and the maps, forest biomass is another niche (Duncanson et al., 2019; Mitchard
S2 imagery at 10-m used as auxiliary, wall-to-wall support for the CHM et al., 2014; Yu et al., 2022) on which coming research articles could
product. The selection of GEDI observations to train the statistical follow the steps in the assessment presented here: the use of temporally
models under these CHM solutions is another important factor. coincident, high-quality ground data to evaluate model performance

Fig. 9. Availability of GEDI L2A shots over the study area. Orbits along year 2019 (blue) and along 2021 (red) intersecting the study area are represented. Dots
represent 25-m footprint of GEDI laser full waveforms. (For interpretation of the references to colour in this figure legend, the reader is referred to the web version of
this article.)

11
A. Pascual et al. Ecological Informatics 70 (2022) 101748

and help in the continuous calibration of estimates of forest structure Duncanson, L., Armston, J., Disney, M., Avitabile, V., Barbier, N., Calders, K., Carter, S.,
Chave, J., Herold, M., Crowther, T.W., Falkowski, M., Kellner, J.R., Labrière, N.,
from satellite remote sensing.
Lucas, R., MacBean, N., McRoberts, R.E., Meyer, V., Næsset, E., Nickeson, J.E.,
The approach outlines by the authors and the geographical scope of Williams, M., 2019. The importance of consistent global Forest aboveground
the training area is under the radar of novel remote sensing programmes. biomass product validation. Surv. Geophys. 40 (4), 979–999. https://doi.org/
One example are Planet tropical mosaics funded by THE Norway’s In­ 10.1007/s10712-019-09538-8.
Duncanson, L., Kellner, J.R., Armston, J., Dubayah, R., Minor, D.M., Hancock, S.,
ternational Climate & Forests Initiative (NICFI), which is delivering Healey, S.P., Patterson, P.L., Saarela, S., Marselis, S., Silva, C.E., Bruening, J.,
continuous monitoring in tropical ecosystems using almost daily ob­ Goetz, S.J., Tang, H., Hofton, M., Blair, B., Luthcke, S., Fatoyinbo, L., Abernethy, K.,
servations to produce annual and monthly mosaics over tropical forests. Zgraggen, C., 2022. Aboveground biomass density models for NASA ’ s global
ecosystem dynamics investigation (GEDI) lidar mission. Remote Sens. Environ. 270,
The data fusion approach based applied in Lang et al. (2022a, 2022b) 112845 https://doi.org/10.1016/j.rse.2021.112845.
and the relevance of tropical forest for climate change and conservation Fayad, I., Baghdadi, N.N., Alvares, C.A., Stape, J.L., Bailly, J.S., Scolforo, H.F., Zribi, M.,
policies might advocate for a tropical tree height map model where the Maire, G.Le., 2021a. Assessment of GEDI’s LiDAR data for the estimation of Canopy
Heights and wood volume of Eucalyptus plantations in Brazil. IEEE JSTARS 14,
specifics of moisture saturation, fast-growing conditions can be more 7095–7110. https://doi.org/10.1109/JSTARS.2021.3092836.
effectively addressed while downscaling the resolution of CHM esti­ Fayad, I., Ienco, D., Baghdadi, N., Gaetano, R., Alvares, C.A., Stape, J.L., Ferraço
mates using the 4.77-m resolution of Planet NICFI data, for example. Scolforo, H., Le Maire, G., 2021b. A CNN-based approach for the estimation of
canopy heights and wood volume from GEDI waveforms. Remote Sens. Environ. 265,
112652 https://doi.org/10.1016/j.rse.2021.112652.
5. Conclusions Guerra-Hernández, J., Pascual, A., 2021. Using GEDI lidar data and airborne laser
scanning to assess height growth dynamics in fast-growing species: a showcase in
Spain. Forest Ecosystems 8 (14), 1–17. https://doi.org/10.1186/s40663-021-00291-
The assessment presented here evaluates the suitability of global
2.
CHM for the monitoring of forest landscape characterized by tall trees. Hansen, M.C., Potapov, P.V., Moore, R., Hancher, M., Turubanova, S.A., Tyukavina, A.,
Despite the challenging fast-growing conditions, the agreement on forest Thau, D., Stehman, S.V., Goetz, S.J., Love-Land, T.R., et al., 2013. High-resolution
canopy height between field measurements and the 10-m CHM tested global maps of 21st- century forest cover change. Science 342, 850–853. https://doi.
org/10.1126/science.1244693.
here is remarkable when accounting for prediction uncertainty. The Healey, S.P., Yang, Z., Gorelick, N., Ilyushchenko, S., 2020. Highly local model
improvement in the tested 10-m CHM is notorious in the study area calibration with a new GEDI LiDAR asset on google earth engine reduces landsat
when adding uncertainty to calibrate field measurements compared to forest height signal saturation. Remote Sens. 12 (17), 1–10. https://doi.org/
10.3390/rs12172840.
the performance of the 30-m CHM previously available. Based on our Laclau, J.P., Gonçalves, J.L., Stape, J.L., 2013. Perspectives for the management of
results and the elaborated insights here, future refinements on global eucalypt plantations under biotic and abiotic stresses. For. Ecol. Manag. 301, 1–5.
canopy height model need to better explore the possibilities and con­ https://doi.org/10.1016/j.foreco.2013.03.007.
Lang, N., Kalischek, N., Armston, J., Schindler, K., Dubayah, R., Wegner, J.D., 2022a.
straints when using data from the still in -orbit NASA GEDI mission – an Global canopy height regression and uncertainty estimation from GEDI LIDAR
actionable, flagship tool for global forest monitoring. waveforms with deep ensembles. Remote Sens. Environ. 268, 112760 https://doi.
org/10.1016/j.rse.2021.112760.
Lang, N., Walter, J., Schindler, K., Wegner, J.D., 2022b. A high-resolution canopy height
Declaration of Competing Interest model of the. Earth. arXiv https://doi.org/10.48550/arxiv.2204.08322.
Milenković, M., Reiche, J., Armston, J., Neuenschwander, A., De Keersmaecker, W.,
The authors declare that they have no known competing financial Herold, M., Verbesselt, J., 2022. Assessing amazon rainforest regrowth with GEDI
and ICESat-2 data. Science of Remote Sensing 100051. https://doi.org/10.1016/j.
interests or personal relationships that could have appeared to influence
srs.2022.100051.
the work reported in this paper. Mitchard, E.T., Feldpausch, T.R., Brienen, R.J., Lopez-Gonzalez, G., Monteagudo, A.,
Baker, T.R., Lewis, S.L., Lloyd, J., Quesada, C.A., Gloor, M., et al., 2014. Markedly
Data availability divergent estimates of Amazon forest carbon density from ground plots and
satellites. Glob. Ecol. Biogeogr. 23, 935–946. https://doi.org/10.1111/geb.12168.
Pardini, M., Armston, J., Qi, W., Lee, S.K., Tello, M., Cazcarra Bes, V., Choi, C.,
Data will be made available on request. Papathanassiou, K.P., Dubayah, R.O., Fatoyinbo, L.E., 2019. Early lessons on
combining Lidar and multi-baseline SAR measurements for Forest structure
characterization. Surv. Geophys. 40 (4), 803–837. https://doi.org/10.1007/s10712-
Appendix A. Supplementary data 019-09553-9.
Pascual, A., Tupinambá-Simões, F., Guerra-Hernández, J., Bravo, F., 2022. High-
Supplementary data to this article can be found online at https://doi. resolution planet satellite imagery and multi-temporal surveys to predict risk of tree
mortality in tropical eucalypt forestry. J. Environ. Manag. 310, 114804 https://doi.
org/10.1016/j.ecoinf.2022.101748. org/10.1016/j.jenvman.2022.114804.
Potapov, P., Li, X., Hernandez-Serna, A., Tyukavina, A., Hansen, M.C., Kommareddy, A.,
References Pickens, A., Turubanova, S., Tang, H., Silva, C.E., et al., 2020. Mapping global forest
canopy height through integration of GEDI and Landsat data. Remote Sens. Environ.
253, 112165 https://doi.org/10.1016/j.rse.2020.112165.
Andersen, H.E., Reutebuch, S.E., McGaughey, R.J., 2006. A rigorous assessment of tree
Qi, W., Lee, S.K., Hancock, S., Luthcke, S., Tang, H., Armston, J., Dubayah, R., 2019.
height measurements obtained using airborne lidar and conventional field methods.
Improved forest height estimation by fusion of simulated GEDI Lidar data and
Can. J. Remote. Sens. https://doi.org/10.5589/m06-030.
TanDEM-X InSAR data. Remote Sens. Environ. 221 (11), 621–634. https://doi.org/
Asner, G.P., Mascaro, J., Anderson, C., Knapp, D.E., Martin, R.E., Kennedy-bowdoin, T.,
10.1016/j.rse.2018.11.035.
Van Breugel, M., Davies, S., Hall, J.S., Muller-landau, H.C., Potvin, C., Sousa, W.,
Roy, D.P., Kashongwe, H.B., Armston, J., 2021. The impact of geolocation uncertainty on
Wright, J., Bermingham, E., 2013. High-fidelity national carbon mapping for
GEDI tropical forest canopy height estimation and change monitoring. Science of
resource management and REDD +. Carbon Balance Manag. 8 (1), 1. https://doi.
Remote Sensing 4, 100024. https://doi.org/10.1016/j.srs.2021.100024.
org/10.1186/1750-0680-8-7.
Silva, C.A., Duncanson, L., Hancock, S., Neuenschwander, A., Thomas, N., Hofton, M.,
Cordell, S., Questad, E.J., Asner, G.P., Kinney, K.M., Thaxton, J.M., Uowolo, A.,
Fatoyinbo, L., Simard, M., Marshak, C.Z., Armston, J., Lutchke, S., Dubayah, R.,
Brooks, S., Chynoweth, M.W., 2017. Remote sensing for restoration planning: how
2021. Fusing simulated GEDI, ICESat-2 and NISAR data for regional aboveground
the big picture can inform stakeholders. Restor. Ecol. 25 (12), 147–154. https://doi.
biomass mapping. Remote Sens. Environ. 253, 112234 https://doi.org/10.1016/j.
org/10.1111/rec.12448.
rse.2020.112234.
Csillik, O., Kumar, P., Mascaro, J., Shea, T.O., Asner, G.P., 2019. Monitoring tropical
Tupinambá-Simões, F., Guerra-Hernández, J., Bravo, F., Pascual, A., 2022. Assessment of
forest carbon stocks and emissions using planet satellite data. Sci. Rep. 9 (17831),
drought effects on survival and growth dynamics in eucalypt commercial forestry
1–12. https://doi.org/10.1038/s41598-019-54386-6.
using remote sensing photogrammetry. A showcase in Mato Grosso, Brazil. For. Ecol.
Dubayah, R., Luthcke, S., Blair, J.B., Hofton, M., Armston, J., Tang, H., 2020a. GEDI L1B
Manag. 505, 119930 https://doi.org/10.1016/j.foreco.2021.119930.
Geolocated Waveform Data Global Footprint Level V001. NASA EOSDIS Land
Urbazaev, M., Thiel, C., Mathieu, R., Naidoo, L., Levick, S.R., Smit, I.P.J., Asner, G.P.,
Processes DAAC. Accessed 2022-05-26 from. https://doi.org/10.5067/GEDI/
Schmullius, C., 2015. Assessment of the mapping of fractional woody cover in
GEDI01_B.001.
southern African savannas using multi-temporal and polarimetric ALOS PALSAR L-
Dubayah, R., Blair, J.B., Goetz, S., Fatoyinbo, L., Hansen, M., Healey, S., Hofton, M.,
band images. Remote Sens. Environ. 166, 138–153. https://doi.org/10.1016/j.
Hurtt, G., Kellner, J., Luthcke, S., Armston, J., Tang, H., Duncanson, L., Hancock, S.,
rse.2015.06.013.
Jantz, P., Marselis, S., Patterson, P.L., Qi, W., Silva, C., 2020b. The global ecosystem
Yu, K., Ciais, P., Seneviratne, S.I., Liu, Z., Chen, H.Y.H., Barichivich, J., Allen, C.D.,
dynamics investigation: high-resolution laser ranging of the Earth’s forests and
Yang, H., Huang, Y., Ballantyne, A.P., 2022. Field-based tree mortality constraint
topography. Science of Remote Sensing 1 (1), 100002. https://doi.org/10.1016/j.
srs.2020.100002.

12
A. Pascual et al. Ecological Informatics 70 (2022) 101748

reduces estimates of model-projected forest carbon sinks. Nat. Commun. 13 (1)


https://doi.org/10.1038/s41467-022-29619-4.

13

You might also like