Download as pdf or txt
Download as pdf or txt
You are on page 1of 20

ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Contents lists available at ScienceDirect

ISPRS Journal of Photogrammetry and Remote Sensing


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

Field-level crop yield estimation with PRISMA and Sentinel-2


Michael Marshall a, *, Mariana Belgiu a, Mirco Boschetti b, Monica Pepe b, Alfred Stein a,
Andy Nelson a
a
Faculty of Geo-Information Science and Earth Observation, University of Twente, Hengelosestraat 99, 7514 AE Enschede, the Netherlands
b
Institute for Electromagnetic Sensing of the Environment, Italian National Research Council, Via Bassini 15, Milan 20133, Italy

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

Keywords: Satellite image data deliver consistent and frequent information for crop yield estimation over large areas.
Agriculture Hyperspectral narrowbands are more sensitive spectrally to changes in crop growth than multispectral broad­
Imaging spectroscopy bands but few studies quantified the gains in the former over the later. The PRecursore IperSpettrale della
Hyperspectral
Missione Applicativa (PRISMA) mission offers narrow (≤10 nm) band capability across the full optical range. The
Remote sensing
multispectral broadband Sentinel-2 mission carries four experimental red-edge and near infrared (NIR) hyper­
Machine learning
Random forest spectral narrow (≤20 nm) bands. We compared the performance of PRISMA and Sentinel-2 spectral bands at
important phases of crop development (vegetative, reproductive, maturity) in estimating field-level biomass and
yield for corn, rice, soybean, and wheat. We selected three data-driven methods: two-band vegetation indices
(TBVIs), partial least squares regression (PLSR), and random forest (RF). The PRISMA and Sentinel-2 models on
average explained approximately 20% more variability in biomass and yield with RF than TBVIs and PLSR. The
mean RMSE of the PRISMA RF models was 0.42 and 0.17 kg m− 2, which was lower than the Sentinel-2 RF models
(0.48 and 0.18 kg m− 2). Multidate image (seasonal) model performance was generally higher than single-date
image model performance. PRISMA shortwave infrared narrowbands and Sentinel-2 red-edge and near
infrared bands were among the top-performing spectral regions. The results highlight potential complementarity
between the PRISMA and Sentinel-2 missions for predicting crop biomass and yield. The results also show the
benefits, limitations, and pitfalls of hyperspectral imaging in agricultural monitoring, which is important for
upcoming operational hyperspectral missions, such as ESA CHIME and NASA Surface Biology and Geology
(formerly HysPIRI).

1. Introduction Machine learning and other data-driven methods are increasingly


employed for crop modeling because of the challenges of process-based
Crop yield is the marketable fraction of aboveground biomass methods to account for the complex interactions governing yield (Leng
accumulated over a growing season for food, feed, fiber or oil (Hay, and Hall, 2020). At the same time, modern processing cycles substitute
1995). It varies widely over space and time because of differences in or integrate multispectral broadband optical data with satellite image
cultivars (Genetic – G), environmental conditions (Environment – E), data capturing extra spectral features related to crop growth and
and agronomic practices (Management – M): G × E × M (Ben-Ari and development (Weiss et al., 2020).
Makowski, 2016; Licker et al., 2010; Ray et al., 2015). Crop models Early data-driven yield models consisted of simple regression re­
typically account for the cumulative effects of G × E × M with seasonal lationships with vegetation indices (e.g., normalized difference index­
meteorological and other geospatial information. Satellite image data —NDVI) or biophysical features (e.g., leaf area index—LAI) that track
are an important input to the models because they capture eco- the evolution of crop canopy spectral reflectance patterns over the
physiological conditions on the ground consistently and frequently, growing season (Moulin et al., 1998). The relationships were based on
and they offer the possibility to do so over large areas (Doraiswamy the principle of light use efficiency: biomass accumulation is propor­
et al., 2003). Studies that combine data-driven crop models and tional to the amount of solar radiation absorbed by canopies during
emerging EO technology are a promising line of research and inform photosynthesis (Monteith, 1969). Contemporary data-driven methods
agricultural monitoring system development (Jones et al., 2017). combine spectral features with other geospatial information

* Corresponding author.

https://doi.org/10.1016/j.isprsjprs.2022.03.008
Received 21 October 2021; Received in revised form 17 February 2022; Accepted 11 March 2022
Available online 18 March 2022
0924-2716/© 2022 The Author(s). Published by Elsevier B.V. on behalf of International Society for Photogrammetry and Remote Sensing, Inc. (ISPRS). This is an
open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Fig. 1. Technical flow chart describing the methodology adopted in this study. TBVs: Two-band vegetation indices; PLSR: Partial least squares regression; RF:
Random Forest.

legacy (Wulder et al., 2012). The European Space Agency (ESA)


Sentinel-2 optical sensor is increasingly used for global agricultural
monitoring because it has a higher spatial resolution (10–20 m) and
return frequency (5 days in a two-satellite constellation) than Landsat,
as well as four red-edge and near infrared (NIR) hyperspectral narrow
(≤20 nm) bands (Defourny et al., 2019). The narrowbands are more
important predictors of yield than Sentinel-2 broadbands and environ­
mental variables (Hunt et al., 2019) or Sentinel-1 radar backscatter
measures (Jin et al., 2019) because they are particularly sensitive to
changes in the chlorophyll content and the structural properties of
canopies (Frampton et al., 2013). The promising results of the Sentinel-2
experimental hyperspectral narrowbands for estimating yield and other
environmental indicators motivated the NASA Harmonized Landsat and
Sentinel-2 (HLS) project (Claverie et al., 2018) and the Italian Space
Agency PRecursore IperSpettrale della Missione Applicativa (PRISMA)
(Cogliati et al., 2021). PRISMA ushers in a new era of spaceborne
monitoring missions with narrow (≤10 nm) band capability across the
full optical range (Rast and Painter, 2019). The missions scheduled or
planned for 2022 and beyond include the German Space Agency EnMAP,
Israeli-Italian Space Agency SHALOM, ESA CHIME, and NASA Surface
Biology and Geology (formerly HysPIRI).
Studies characterize many biochemical processes and biophysical
properties related to yield with hyperspectral narrowbands (Thenkabail
et al., 2013). Illustrative examples include the relative reflectance of
chlorophyll and accessory pigments associated with photosynthetic ac­
tivity in the visible spectrum (i.e., green peak) (Gitelson and Solo­
vchenko, 2017); absorption of chlorophyll in the visible red (Thenkabail
et al., 2000); blue-shift in the red-edge inflection point caused by stress
(Schlemmer et al., 2013); reflectance in the NIR due to canopy structure;
absorption of canopy moisture across the NIR and short-wave infrared
(SWIR) (Casas et al., 2014); and absorption of non-photosynthetically
Fig. 2. The location of the elementary sampling units in the Bonifiche Ferraresi active protein and carbon base constituents in the SWIR (Féret et al.,
study area overlaying a true color Sentinel-2 composite captured on August 2021).
11, 2020. The estimation of yield from space with fluorescence, microwave,
and thermal emission features has received comparatively more atten­
characterizing climate, human influences, soil, and topography to esti­ tion by the remote sensing community than full-optical-range narrow­
mate yield. They use powerful data mining methods to overcome the band absorption and reflectance features (Guan et al., 2017). Only a few
limitations of linear regression: artificial neural networks (Feng et al., studies were carried out to the authors’ knowledge. They showed 5–25%
2021), Bayesian neural networks (Ma et al., 2021), boosted regression gains in crop biomass and yield estimation performance with single date
(Wang et al., 2020); convolutional neural networks (Sagan et al., 2021), Hyperion narrowbands and narrowband vegetation indices over Land­
partial least squares regression (PLSR) (Kowalik et al., 2014), random sat, MODIS, and commercially available high-resolution broadbands
forest (RF) (Hunt et al., 2019), and support vector machine (Kamir et al., and broadband vegetation indices (Mariotto et al., 2013; Marshall and
2020). Thenkabail, 2015a). The lack of studies could be due to the fact that
Yield can be estimated from space with features across the optical there were only two spaceborne sensors in the past (Chris from ESA and
range (400–2500 nm) (Hank et al., 2019). Landsat images were the Hyperion from NASA)—neither of which were designed for monitoring
workhorse given their free accessibility, global coverage, and long (Rast and Painter, 2019). Further, hyperspectral data typically require

192
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Fig. 3. Satellite image acquisition dates (●) according to the three main growth stages: vegetative ( ), reproductive ( ), and maturity ( ).

Table 1
Summary of PRISMA and Sentinel-2 image attributes excluding the panchromatic bands.
Sensor Designation Spectral Range (nm) # Bands / Band # Band Width (nm) Spatial Resolution (m) Swath (km)

PRISMA VNIR 400–1000 67 9 30 30


SWIR1 1000–1700 66 11
SWIR2 1700–2500 98 8
Sentinel-2 Coastal aerosol 433–453 1 20 60 100
Blue 458–523 2 65 10
Green 543–578 3 35 10
Red 650–680 4 30 10
Red-edge 1 698–713 5 15 20
Red-edge 2 733–748 6 15 20
Red-edge 3 773–793 7 20 20
NIR 785–900 8 115 10
Narrow NIR 855–875 8a 20 20
Water vapor 935–955 9 20 60
SWIR-Cirrus 1360–1390 10 30 60
SWIR1 1565–1655 11 90 20
SWIR2 2100–2280 12 180 20

sophisticated data mining and filtering techniques given the large field-level assessment.
number of bands and low signal-to-noise ratio (Broge and Leblanc,
2001). 2.1. Study area
The overall goal of our study was to estimate end-of-season above­
ground biomass (hereafter referred to as biomass) and yield from We performed the analysis in the year 2020 over fields with a total
PRISMA and Sentinel-2 spectral bands. PRISMA and Sentinel-2 captured area of 3800 ha in the Bonifiche Ferraresi farm near the city of Ferrara in
spectral changes according to the main phases of plant development Emilia-Romagna, Italy (Fig. 2). The farm, like the surrounding area,
(vegetative, reproductive, maturity) for four globally important food primarily grows food and feed crops, such as alfalfa, barley, corn, rice,
crops (corn, rice, soybean, wheat). Two-band vegetation indices soybean, and wheat. The farm consists mainly of silty-clay and silty-clay
(TBVIs), Partial least squares regression (PLSR), and RF characterized loam soils and is highly productive since it is reclaimed land and rich in
the spectral response to field-level estimates of biomass and yield. The mineral deposits. The climate is classified as humid subtropical (Köppen
results of the three models were used to quantify the relative gains in classification: Cfa) with average annual daily temperatures and monthly
performance of a full-optical-range narrowband sensor over a multi­ precipitation of 13.6 ◦ C and 52.9 mm. Most crops are grown in the
spectral broadband spaceborne sensor with hyperspectral capability. summer when it is warm (daily mean temperature = 23.1 ◦ C) and moist
They also uncovered potential complementarity between PRISMA and (average monthly precipitation = 54.6 mm). Wheat is planted in the fall
Sentinel-2, which is important for remote sensing data integration and and harvested at the beginning of summer.
fusion.
2.2. Field data
2. Material and methods
Crop cuttings were randomly harvested at the end of the growing
The following sub-sections expand on the methods implemented in season (June for wheat and September for corn, soybean, rice) within 60
our study (Fig. 1). They describe the study area over which the com­ × 60 m2 elementary sample units (ESUs). Each ESU was homogenous in
parison was made, field-level biomass and yield data handling, PRISMA terms of crop variety and soil type. The ESUs were randomly stratified
and Sentinel-2 data handling, analytical methods (TBVIs, PLSR, RF), and by crop variety and soil type however to maximize the spatial

193
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Fig. 4. Additional PRISMA Level-2c processing: geometric (a) and spectral (b) corrections. ( ), ( ), and ( ) are the farm boundaries, uncorrected PRISMA
spectral signature, and corrected PRISMA spectral signature with smoothing splines.

heterogeneity among ESUs and therefore increase the robustness of the phase, images for our study area became available upon request in 2020.
results. Forty ESUs were demarcated using a handheld global posi­ We acquired nine images over the 2020 growing season. Three images in
tioning system (horizonal position error ≤ 5 m) for silage corn, rice, May (day of year = 125, 131, 137) and one image in July (day of year =
soybean, and winter wheat varieties (160 ESUs in total). The dimensions 195) were discarded because of high-cloud coverage. In the end, we had
of the ESUs were double the spatial resolution of the PRISMA sensor images for the (i) vegetative and reproductive stages of corn and soy­
following Marshall and Thenkabail (2015b) to account for potential geo- bean; (ii) vegetative and maturity stages of rice; and (iii) vegetative,
positioning errors. Except for silage corn, ten 0.25 × 0.25 m2 randomly reproductive, and maturity stages of wheat (Fig. 3).
distributed clear-cut samples were taken within each ESU (total = 0.625 Images also corresponded to the early vegetative phases for corn
m2). Corn samples were harvested over 3 m2 instead because it is (jointing) and rice (tillering) when the canopies were not fully closed.
considerably larger and wider spaced than the other crops. The samples The sensor characteristics of PRISMA are shown in Table 1. The PRISMA
within an ESU were well-mixed and weighed in the field to obtain system consists of: (i) two hyperspectral sensors in the spectral range
aboveground “wet” biomass. The grain was immediately separated from 400–2500 nm with overlapping narrowbands between 920 and 1010
the plant material and weighed (wet yield). Wet-weight measurements nm, a spatial resolution of 30 m, and an average spectral resolution < 10
are inconsistent because the moisture content of harvested material nm; (ii) a panchromatic camera working in the spectral range 400–700
varies according to soil moisture and other environmental conditions nm with a spatial resolution of 5 m. We downloaded PRISMA narrow­
(Barrs and Weatherley, 1962). We therefore dried the samples consis­ bands as Level-2D (atmospherically- and geometrically-corrected). The
tently in a forced air oven at 105˚C to estimate the final dry-weight data for each acquisition date were integrated into one data cube with
biomass and yield in kg m− 2. the “PRISMAread” package in R (Busetto, 2021). The cube consisted of
231 narrowbands ranging from 400 to 2500 nm with no overlaps.
Thirty-nine bands from 1338 to 1449 nm, 1793 to 1993 nm, and 2356 to
2.3. PRISMA data 2393 nm were omitted from the analysis because the signals within
these ranges are degraded due to strong absorption by atmospheric CO2,
PRISMA was launched on March 22, 2019. After the commissioning

194
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Fig. 5. The statistical distribution of biomass (a) and yield (b) samples collected during the field campaign (N = 40 for each crop).

Table 2
Summary of the top-performing TBVI PRISMA model according to crop type and the day of year (DOY) of image acquisition. N is the number of samples, m and b are
the slope and intercept of the linear model. The top-performing growth phase is in bold.
Crop Growth Phase DOY λ1 λ2 N m b R2 RMSE Visible 400–680

Biomass Red-edge 680–730


Corn Vegetative 143 1656 749 39 − 0.0445 0.2602 0.54 0.03 NIR 730–1000
Vegetative 177 1047 739 40 0.0202 0.0770 0.53 0.01 SWIR1 1000–1700
Reproductive 212 579 507 37 − 0.0075 0.1679 0.45 0.01 SWIR2 1700–2500
Mean 1109 729 40 0.0239 0.0975 0.43 0.02
Rice Vegetative 177 1152 1109 40 − 0.0367 0.0577 0.46 0.03
Vegetative 212 1554 1316 37 − 0.0272 − 0.2988 0.43 0.02
Maturity 259 2035 2027 39 0.0084 0.0247 0.35 0.01
Mean 1196 918 40 ¡0.0338 0.1378 0.55 0.02
Soybean Vegetative 177 2407 2061 26 0.0496 − 0.6707 0.26 0.01
Reproductive 212 897 844 26 ¡0.0154 ¡0.0555 0.42 0.00
Mean 1088 749 26 − 0.0228 0.1451 0.32 0.00
Wheat Vegetative 97 1774 698 38 0.1016 0.1718 0.58 0.07
Reproductive 143 1284 708 39 0.0705 0.3109 0.73 0.03
Maturity 177 507 448 38 0.0346 0.2242 0.57 0.02
Mean 759 739 39 0.0207 0.0198 0.57 0.01
Yield
Corn Vegetative 143 1163 770 39 − 0.0328 0.1327 0.39 0.01
Vegetative 177 978 729 40 0.0792 0.0770 0.54 0.02
Reproductive 212 570 514 37 − 0.0154 0.1290 0.43 0.01
Mean 978 729 40 0.0455 0.1047 0.41 0.02
Rice Vegetative 177 1131 1120 40 − 0.0273 0.0125 0.24 0.01
Vegetative 212 622 514 37 − 0.0479 0.0678 0.25 0.02
Maturity 259 908 679 39 0.1458 0.4938 0.43 0.05
Mean 1141 918 40 − 0.0663 0.0690 0.33 0.03
Soybean Vegetative 177 650 463 26 0.1730 0.2887 0.44 0.02
Reproductive 212 2222 470 26 0.3395 0.4380 0.59 0.02
Mean 587 419 26 0.2890 0.3786 0.61 0.02
Wheat Vegetative 97 1784 698 38 0.2416 0.1739 0.47 0.07
Reproductive 143 1595 1543 39 0.0634 0.0337 0.68 0.01
Maturity 177 822 749 38 0.0143 0.0709 0.51 0.00
Mean 739 728 39 0.0448 0.0248 0.47 0.01

water vapor, and other atmospheric constituents (Marshall and Then­ ground control point network at the time of data acquisition. Second, we
kabail, 2014). removed unusual anomalies in the NIR region of the spectra with the
We took additional steps after the standard atmospheric correction findpeaks function (“pracma” package) in R (Kuhn, 2021) following
and orthorectification pre-processing to remove spectral and geometric (Belgiu et al., 2021; Pepe et al., 2020). Finally, we used a smoothing-
inconsistencies (Fig. 4). First, we manually shifted (without spatial spline filter to fill in data gaps.
transformation) image corners so the images aligned with the study area
field boundaries. The geocoding error was due to a nonoperational

195
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Fig. 6. Lambda-lambda plots highlighting regions of high and low correlations between PRISMA TBVIs and biomass for corn (a), rice (b), soybean (c), and wheat (d)
during the top-performing growth phase.

2.4. Sentinel-2 data 2.5. Analytical methods

Approximately fifteen Sentinel-2 images were available over the 2.5.1. Two-band hyperspectral narrowband vegetation indices (TBVIs)
same period as PRISMA. The images were visually inspected online The most common EO input to data-driven crop models is NDVI and
(Copernicus Open Access Hub:). Images were downloaded for analysis if other vegetation indices (VIs) derived from two or more spectral bands
the study area was not obscured by clouds or haze. This selection cri­ (Huete et al., 2002). VIs reduce the noise in individual bands because of
terion ensured the spectral properties were not substantially different varying soil backgrounds, solar geometries, and terrain. Further, they
between PRISMA and Sentinel-2. Eight images met these criteria. are easy to apply and interpret. A downside of VIs is that they discard
Sentinel-2 includes thirteen multispectral broadbands that span a com­ spectral information that could otherwise increase the signal-to-noise
parable range to PRISMA (Fig. A1.1.). We excluded bands 1, 9, and 10 ratio (Atzberger et al., 2010). TBVIs take the normalized difference
from the analysis because these bands are designated for coastal/at­ form of NDVI (Tucker, 1979):.
mospheric studies. Band 2 was also not included because this channel is
λ2 − λ1
severely impacted by atmospheric scattering. Sentinel-2 images were TBVI = (1)
λ2 − λ1
geometrically and atmospherically corrected (designated Level-2A).
Sentinel-2 was resampled from 10 to 20 m to 30 m resolution using where λ1 and λ2 is the surface reflectance captured by two different
nearest neighbor for comparison purposes with PRISMA. spectral bands. We used λ-λ plots (Thenkabail et al., 2000) to identify
hotspots where TBVIs derived from PRISMA and Sentinel-2 spectral
bands during the main phases of crop development and across the

196
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Fig. 7. Lambda-lambda plots highlighting regions of high and low correlations between PRISMA TBVIs and yield for corn (a), rice (b), soybean (c), and wheat (d)
during the top-performing growth phase.

growing season, correlated well with biomass and yield. A λ-λ plot is Y, followed by the second, third and so on, until the maximum number
essentially a cross-correlation plot in which the horizontal and vertical of components (n). The X scores (t) predict the Y-scores, which are then
coordinates are defined by the bands in a TBVI. The value at the inter­ used to predict Y with standard multivariate linear regression (Equation
section of λ1, λ2 indicates the coefficient of determination (R2) for a 1). The loadings (w) are weights that indicate the relative contribution
linear model between the TBVI and biomass or yield.. of features to explaining Y. Variables with more influence have higher
absolute loadings. Positive loadings have an opposite influence to
2.5.2. Partial least squares regression (PLSR) negative loadings. Each variable (i) receives a score and loading. ε is the
PLSR models (Wold, 1980) unlike TBVIs, account for all available error term.
spectral information. PLSR uncovers underlying structures by projecting ∑∑
features into lower dimensional space. The projections are less prone to Y⋯ = ε + wn,i tn,i (1)
n i
outliers and non-linearity than TBVIs and linear regression models. PLSR
projects the features to maximize their co-variance with the response PLSR estimated biomass and yield from PRISMA and Sentinal-2
variable (Atzberger et al., 2010). This means the features and the spectral bands retrieved during the main phases of crop development
response variable in a PLSR model are interrelated. Feature loadings are and across the growing season. We performed 100 iterations of 10-fold
a key characteristic of PLSR. They reveal the underlying relationship of cross-validation to determine the optimal number of PLSR components
features to the response variable relative to other features. for each biomass or yield model (maximum components ≤ 5) with the R
PLSR reduces a matrix of features (X) to an additive series of factors “caret” package (Borchers, 2021).
or components that explain a unique proportion of the variation in the
response variable (Y). The first component explains the most variation in

197
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Table 3
Summary of the top-performing TBVI Sentinel-2 model according to crop type and the day of year (DOY) of image acquisition. N is the number of samples, m and b are
the slope and intercept of the linear model. The top-performing growth phase is represented in bold.
Crop Growth Phase DOY λ1 λ2 N m b R2 RMSE Blue 2

Biomass Green 3
Corn Vegetative 114 11 6 40 − 0.0197 0.2654 0.23 0.02 Red 4
Vegetative 144 12 7 40 − 0.0711 0.2810 0.54 0.05 Red-edge 1 5
Vegetative 174 8 6 40 0.0358 0.0311 0.57 0.02 Red-edge 2 6
Vegetative 194 8 6 40 0.0324 0.0745 0.45 0.03 Red-edge 3 7
Reproductive 224 8a 6 40 0.0223 0.1213 0.40 0.02 NIR 8
Reproductive 234 11 3 40 0.0257 0.4336 0.40 0.02 Narrow NIR 8a
Mean 8a 6 40 0.0486 0.0801 0.48 0.01 SWIR1 11
Rice Vegetative 174 11 6 40 ¡0.1186 0.3062 0.21 0.17 SWIR2 12
Vegetative 194 12 11 40 − 0.0160 − 0.2677 0.14 0.03
Vegetative 224 7 6 40 0.0142 0.1089 0.13 0.03
Reproductive 234 12 5 40 − 0.0277 − 0.0838 0.11 0.06
Mean 11 7 40 − 0.0410 − 0.2225 0.21 0.06
Soybean Vegetative 174 12 3 40 − 0.0798 0.3903 0.18 0.02
Vegetative 194 8 6 40 − 0.0766 0.1992 0.07 0.03
Reproductive 224 8a 7 40 − 0.0259 0.0537 0.23 0.01
Reproductive 234 8a 7 40 − 0.0477 0.0803 0.25 0.01
Mean 8a 7 40 ¡0.0151 0.0351 0.26 0.00
Wheat Vegetative 94 4 2 40 − 0.0830 0.1455 0.51 0.06
Vegetative 114 4 2 40 − 0.0892 0.1605 0.67 0.05
Reproductive 144 7 6 40 0.0481 0.0361 0.76 0.02
Maturity 174 5 2 40 0.0382 0.2919 0.62 0.02
Mean 11 7 40 − 0.0920 − 0.0562 0.67 0.05
Yield
Corn Vegetative 114 3 2 40 0.0208 0.0592 0.21 0.01
Vegetative 144 12 7 40 − 0.1342 0.2463 0.32 0.06
Vegetative 174 8a 6 40 0.0778 0.0434 0.51 0.02
Vegetative 194 8a 6 40 0.0844 0.0663 0.57 0.02
Reproductive 224 8a 6 40 0.0612 0.1029 0.49 0.02
Reproductive 234 11 3 40 0.0759 0.4043 0.58 0.02
Mean 8a 6 40 0.0486 0.0801 0.48 0.02
Rice Vegetative 174 8a 6 40 − 0.0314 0.0828 0.07 0.03
Vegetative 194 5 3 40 − 0.0524 0.2454 0.12 0.04
Vegetative 224 11 7 40 − 0.0938 − 0.3538 0.27 0.04
Reproductive 234 11 8 40 ¡0.1529 ¡0.2560 0.32 0.05
Mean 7 6 40 0.0289 0.0802 0.16 0.02
Soybean Vegetative 174 12 3 40 − 0.0798 0.3903 0.18 0.02
Vegetative 194 8 6 40 − 0.0766 0.1992 0.07 0.03
Reproductive 224 8a 7 40 − 0.0259 0.0537 0.23 0.01
Reproductive 234 8a 7 40 − 0.0477 0.0803 0.25 0.01
Mean 11 5 40 0.3239 0.3155 0.43 0.03
Wheat Vegetative 94 11 5 40 0.1135 0.1337 0.42 0.04
Vegetative 114 4 2 40 − 0.2342 0.1681 0.65 0.05
Reproductive 144 8a 5 40 0.3897 0.2298 0.74 0.06
Maturity 174 5 2 40 0.0912 0.2971 0.50 0.02
Mean 6 5 40 0.1859 0.2142 0.58 0.05

2.5.3. Random forest (RF) mean error difference before and after permutation is calculated over all
RF regression models (Breiman, 2001) are widely used to analyze decision trees. RFE applies a 10-fold cross-validation strategy to itera­
hyperspectral data because of its insusceptibility to outliers, non- tively discard the least important predictors.
linearity, serial autocorrelation, and high dimensionality (Wójtowicz We developed models for each crop based on single date and multi-
et al., 2021). RF is a non-parametric machine learning algorithm that date (i.e., seasonal) images using ≤ 5 predictors as determined through
consists of an ensemble of unpruned decision trees. These decision trees RFE. This conforms to Thenkabail et al. (2000) who suggested the ratio
are trained on samples randomly selected through replacement. Roughly of predictors to number of samples should be between 0.1 and 0.2. The
two-thirds of the samples are used for training. The remaining out-of- predictors were first-derivative transformed to increase the signal-to-
bag (OOB) samples are used for error assessment. Each tree consists of noise ratio given the small number of predictors (Tsai and Philpot,
a subset of the available predictors defined by the “mtry” parameter. In 1998).
our study, we set mtry to p/3 (where p is the number of predictors). One
hundred ensembles consisting of 1000 decision trees were generated for 2.6. Field-level biomass and yield assessment
each predictive model to increase the robustness of the predictions and
reduce the risk of over-fitting (Belgiu and Drăguţ, 2016). We present the We assessed the performance of TBVIs, PLSR, and RF using the co­
results of average predictions across all iterations. efficient of determination (R2) and root mean square error (RMSE).
We used recursive feature elimination (RFE) (Guyon et al., 2002) to These metrics were calculated with 10-fold cross-validation because the
remove redundant predictors. RFE is a wrapper feature selection method sample size of the biomass and yield data was small. R2 and RMSE were
that relies on the internal capability of RF to estimate the predictors’ generated from linear models driven by the top-performing TBVI in each
importance. RF takes advantage of the OOB samples for this purpose. λ-λ plot because of the large number of TBVIs. Cross-validation was
Thus, in the first step, the Mean Squared Error (MSE) for each decision performed over several iterations for PLSR and RF since errors can vary
tree is computed using OOB samples. In the next step, each predictor’s substantially based on the seed conditions and feature ranking.
values are permuted and the MSE is calculated again. In the end, the

198
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Fig. 8. Lambda-lambda plots highlighting regions of high and low correlations between Sentinel-2 TBVIs and biomass for corn (a), rice (b), soybean (c), and wheat
(d) during the top-performing growth phase.

3. Results 32–57% of the variability in biomass (Table 2). The average RMSE was
0.02 kg m− 2. The relationships were strongest for wheat and weakest for
3.1. Field data summary soybean. The majority of TBVIs included narrowbands from the short­
wave infrared-1 (SWIR-1: 1000–1700 nm), followed by the NIR
The biomass and yield samples were generally skewed, but contained (730–1000 nm), visible (400–730 nm) and shortwave infrared-2 (SWIR-
few outliers (Fig. 5). Corn on average produced more biomass and yield 2: 1700–2500 nm), and red-edge (680–730 nm). The performance of the
than the other crops, followed by rice, wheat, and soybean. Yield on models was highest for the vegetative phase for corn, reproductive phase
average accounted for nearly half (41–45%) of the end-of-season for soybean and wheat, and seasonal mean for rice. The corn TBVIs were
biomass. The interquartile range of biomass and yield between crops concentrated in a narrow region of the NIR at 739 nm (Fig. 6a). The rice
with the exception of soybean were similar. Biomass distributions TBVIs were centered on a narrow region shifted to longer wavelengths in
however had larger interquartile ranges and were more skewed than the NIR at 918 nm (Fig. 6b). The soybean TBVIs clustered around the
yield. NIR at 850 nm and SWIR-2 at 2300 nm (Fig. 6c). The performance of
wheat TBVIs was generally high across the available narrowbands
(Fig. 6d). Particularly strong correlations were centered on a narrow
3.2. Estimation of biomass and yield with PRISMA and Sentinel-2 using region at 708 nm.
TBVs The PRISMA TBVIs tended to perform better for yield estimation
(33–62% of the variability). The average RMSE for yield was similar to
The top-performing PRISMA TBVIs on average accounted for

199
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Fig. 9. Lambda-lambda plots highlighting regions of high and low correlations between Sentinel-2 TBVIs and yield for corn (a), rice (b), soybean (c), and wheat (d)
during the top-performing growth phase.

biomass (0.02 kg m− 2). The relationships were strongest for soybean, TBVIs tended to perform better for corn (ΔR2 = 0.05) and wheat (ΔR2 =
followed by wheat, corn, and rice. The highest correlations were ach­ 0.10) than the PRISMA TBVIs. PRISMA TBVIs performed better for rice
ieved for corn, rice, and wheat during the vegetative and reproductive (ΔR2 = 0.34) and soybean (ΔR2 = 0.07). The bulk of the Sentinel-2
phases. Soybean yield was best estimated with a seasonal mean TBVI. TBVIs for biomass included the red-edge, followed by the visible and
The TBVIs mainly consisted of visible and NIR narrowbands. SWIR-1, NIR, and SWIR. The strongest correlations occurred during the vegeta­
SWIR-2, and red-edge narrowbands were less abundant. As with corn tive phase for corn and rice, the seasonal mean for soybean, and the
biomass, the top-performing TBVIs included a narrow region in the NIR reproductive phase for wheat. The highest performing TBVIs for corn
centered at 729 nm (Table 7a). The rice TBVIs were centered on two consisted of band 6 and band 8 (Fig. 8a). The various combinations of
narrow regions of the visible and NIR centered at 679 nm and 908 nm Sentinel-2 bands were weakly correlated with rice biomass (Fig. 8b).
(Fig. 7b). A second region of high correlations occurred in the SWIR-2 Band 11 was the most abundant in the top-performing TBVIs. Similarly,
around 2000 nm. The soybean TBVIs clustered in the visible and soybean TBVIs tended to perform poorly across Sentinel-2 bands
SWIR-2 (Fig. 7c). As with wheat biomass, TBVIs for wheat yield corre­ (Fig. 8c). Like PRISMA, several Sentinel-2 bands performed well for
lated well across the available narrowbands (Fig. 7d). The highest cor­ wheat biomass (Fig. 8d). The red-edge (band 6, 7) and NIR produced the
relations however were shifted to longer wavelengths in the SWIR-1 highest correlations.
(around 1500 nm). The performance of Sentinel-2 TBVIs was generally lower for yield
Sentinel-2 TBVIs on average explained 21–67% of the variability in estimation (16–58% of the variability). Like biomass, Sentinel TBVIs
biomass (Table 3). The average RMSE was 0.03 kg m− 2. The Sentinel-2 performed better for corn (ΔR2 = 0.07) and wheat (ΔR2 = 0.11) and

200
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Table 4
Summary of the PLSR decomposition of PRISMA spectral data at a given day of year (DOY) of image acquisition. The percentage of total explained variance is shown for
each component alongside model performance.
Crop Growth Phase DOY N R2 RMSE Comp.1 Comp.2 Comp.3 Comp.4 Comp.5

Biomass
Corn Vegetative 143 39 0.55 0.58 22.78 71.78
Vegetative 177 40 0.42 0.62 74.78
Reproductive 212 37 0.47 0.67 50.92 18.08 15.83 2.06
Mean Mean 40 0.48 0.65 47.62 44.11 1.38 1.14
Rice Vegetative 177 40 0.51 0.65 75.66
Vegetative 212 37 0.45 0.67 19.76 62.55
Maturity 259 39 0.58 0.59 31.49 39.48 12.04 8.60
Mean Mean 40 0.56 0.61 62.84 28.39 4.14
Soybean Vegetative 177 26 0.24 0.14 69.16
Reproductive 212 26 0.23 0.14 71.22
Mean Mean 26 0.24 0.13 71.26
Wheat Vegetative 97 38 0.57 0.59 75.92 14.81
Reproductive 143 39 0.67 0.47 68.06 22.51 2.45 1.05 0.74
Maturity 177 38 0.70 0.52 71.60 7.40 12.47
Mean Mean 39 0.62 0.50 50.46 41.03
Yield
Corn Vegetative 143 39 0.46 0.27 56.41 38.15
Vegetative 177 40 0.48 0.25 74.52 16.38
Reproductive 212 37 0.50 0.28 52.92 18.01 14.02 1.59
Mean Mean 40 0.43 0.28 42.43 49.36
Rice Vegetative 177 40 0.39 0.26 75.06
Vegetative 212 37 0.38 0.26 41.84 38.88
Maturity 259 39 0.56 0.22 37.08 38.39 8.00
Mean Mean 40 0.38 0.25 62.60
Soybean Vegetative 177 26 0.49 0.06 52.07 44.83 0.46 0.92 0.60
Reproductive 212 26 0.47 0.07 67.94 24.71
Mean Mean 26 0.38 0.07 57.13 38.88
Wheat Vegetative 97 38 0.49 0.25 76.17 8.58
Reproductive 143 39 0.62 0.19 67.96 22.63 2.40 1.12 1.17
Maturity 177 38 0.66 0.21 71.15 11.38 9.18
Mean Mean 39 0.55 0.21 52.82 38.51

worse for rice (ΔR2 = -0.17) and soybean (ΔR2 = -0.19) than PRISMA tended to be lower for yield than biomass. Performance was highest for
TBVIs. Most top-performing Sentinel-2 TBVIs included bands in the red- wheat, followed by corn, soybean, and rice. The first component of the
edge, followed by the NIR, visible, SWIR-1, and SWIR-2. Band 11 PLSR models generally explained more variability (37–76%) in yield
together with bands 3 and 5, produced the strongest correlations with than the other components. The loadings for yield were similar to
corn yield (Fig. 9a). Band 11 stood out for rice and soybean as well, but biomass for corn, soybean, and wheat (Fig. A1.3). Rice showed a clear
the correlations were considerably lower (Fig. 9b-c). Sentinel-2 bands muted response in the NIR and SWIR during all growth phases.
correlated well for most combinations. The highest correlations were The PLSR models for Sentinel-2 explained 15–78% and 19–73% of
occurred in the NIR (Fig. 9d). the variability in biomass and yield, respectively (Table 5). The average
RMSE was 0.50 and 0.44 kg m− 2. As with the PLSR models for PRISMA,
performance was lower for yield than biomass estimation. It was also
3.3. Estimation of biomass and yield with PRISMA and Sentinel-2 using highest for wheat, followed by corn, soybean, and rice. Performance was
PLSR consistently better in the earlier phases of crop development. The first
component of the Sentinel-2 models tended to explain more variability
The PLSR models explained 23–70% of the variability in biomass (upwards of 90%) than the PRISMA models. The Sentinel-2 models
(Table 4). The mean RMSE was 0.50 kg m− 2. The explanatory power of (except for soybean) performed poorer than the PRISMA models. The
the models was comparable to the TBVIs, but the errors were much PRISMA models greatly out-performed the Sentinel-2 models for rice
higher. As with the TBVIs, the PLSR models performed best for wheat (ΔR2 = 0.27). The performance was comparable for the two sensors for
and worst for soybean. The first component of the PLSR models tended yield estimation. The ΔR2 for rice was lower (0.18).
to explain the most variability (20–76%) in biomass. The reproductive
and maturity phases included more components reflecting the
complexity of spectra in these phases. The narrowbands in the PLSR 3.4. Prediction of biomass and yield with Random forest (RF)
models consistently showed strong positive loadings in the NIR and
SWIR-1 during the vegetative phases (Fig. A1.2). The loadings in the The RF-based PRISMA models generally explained more variability
visible were positive in the early vegetative phase when the canopies in biomass and yield (10% and 7%) with considerably fewer (≤5) nar­
were not completely closed. The loadings became negative later in the rowbands than PLSR (Table 6). The best performing models for biomass
vegetative phase when the canopies were completely closed. Loadings in were obtained when all PRISMA images were used as input predictors:
the SWIR-2 were low. A similar pattern to the spectral response in the corn (R2 = 0.64 and RMSE = 0.50 kg m− 2), rice (R2 = 0.74 and RMSE =
advanced vegetative phase was seen during the reproductive phase, but 0.47 kg m− 2), soybean (R2 = 0.81 and RMSE = 0.09 kg m− 2), and wheat
the loadings were typically lower. During the maturity phase when (R2 = 0.82 and RMSE = 0.36 kg m− 2). The best yield predictions were
canopies lose most of their chlorophyll, loadings became consistently obtained using all PRISMA data for rice (R2 = 0.63 and RMSE = 0.20 kg
positive across the spectral range. m− 2), soybean (R2 = 0.86 and RMSE = 0.04 kg m− 2), and wheat (R2 =
The PLSR models explained 38–66% of the variability in yield 0.80 and RMSE = 0.15).
(Table 4). The mean RMSE was 0.21 kg m− 2, which was quite a bit When single-date images were used as predictors, the best prediction
higher than TBVI RMSE. Contrary to the TBVIs, model performance for the corn biomass was obtained at the early vegetative stage.

201
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Table 5
Summary of the PLSR decomposition of Sentinel-2 spectral data at a given day of year (DOY) of image acquisition. The percentage of total explained variance is shown
for each component alongside model performance.
Crop Growth Phase DOY N R2 RMSE Comp.1 Comp.2 Comp.3 Comp.4 Comp.5

Biomass
Corn Vegetative 114 40 0.29 0.67 90.03 NA NA NA NA
Vegetative 144 40 0.57 0.51 63.51 34.95 NA NA NA
Vegetative 174 40 0.55 0.53 81.71 15.80 1.39 0.83 NA
Vegetative 194 40 0.41 0.61 54.72 16.23 26.66 NA NA
Reproductive 224 40 0.30 0.66 46.96 18.20 32.46 NA NA
Reproductive 234 40 0.36 0.63 49.75 11.54 36.56 NA NA
Mean 40 0.43 0.60 44.35 47.13 6.64 1.39 NA
Rice Vegetative 174 40 0.36 0.68 48.16 20.10 31.24 0.36 0.06
Vegetative 194 40 0.15 0.79 37.85 58.77 NA NA NA
Vegetative 224 40 0.19 0.77 16.05 70.53 NA NA NA
Reproductive 234 40 0.34 0.70 62.03 30.52 2.63 3.22 1.43
Mean 40 0.22 0.76 35.06 26.28 NA NA NA
Soybean Vegetative 174 26 0.44 0.11 59.51 40.12 0.23 NA NA
Vegetative 194 26 0.23 0.13 85.97 NA NA NA NA
Reproductive 224 26 0.25 0.14 29.75 NA NA NA NA
Reproductive 234 26 0.28 0.12 80.42 NA NA NA NA
Mean 26 0.39 0.12 59.80 37.22 0.40 2.41 NA
Wheat Vegetative 94 40 0.53 0.56 75.85 21.34 1.86 NA NA
Vegetative 114 40 0.64 0.48 65.64 NA NA NA NA
Reproductive 144 40 0.78 0.37 67.26 30.51 1.20 NA NA
Maturity 174 40 0.60 0.52 85.83 5.09 7.84 0.93 NA
Mean 40 0.65 0.46 61.26 NA NA NA NA
Yield
Corn Vegetative 114 40 0.37 0.26 90.20 NA NA NA NA
Vegetative 144 40 0.41 0.25 78.20 20.26 0.89 0.43 0.13
Vegetative 174 40 0.61 0.20 81.70 15.54 1.74 0.73 0.12
Vegetative 194 40 0.56 0.22 52.53 24.59 20.34 1.51 NA
Reproductive 224 40 0.44 0.24 47.16 15.47 35.27 NA NA
Reproductive 234 40 0.54 0.22 49.30 16.05 32.54 NA NA
Mean 40 0.53 0.22 42.99 42.68 12.26 1.58 NA
Rice Vegetative 174 40 0.23 0.27 53.41 41.23 2.95 2.28 0.05
Vegetative 194 40 0.19 0.28 60.62 33.95 3.44 1.16 NA
Vegetative 224 40 0.27 0.26 50.18 46.39 1.62 0.71 NA
Reproductive 234 40 0.33 0.24 20.19 72.97 NA NA NA
Mean 40 0.23 0.27 46.94 46.74 3.14 2.38 0.63
Soybean Vegetative 174 26 0.46 0.06 51.71 47.95 0.16 0.10 NA
Vegetative 194 26 0.27 0.07 85.34 NA NA NA NA
Reproductive 224 26 0.45 0.06 68.35 22.10 7.90 NA NA
Reproductive 234 26 0.39 0.07 79.90 NA NA NA NA
Mean 26 0.48 0.06 67.90 28.87 2.83 NA NA
Wheat Vegetative 94 40 0.43 0.24 75.64 21.90 1.26 0.90 NA
Vegetative 114 40 0.56 0.20 65.50 31.85 1.09 0.29 NA
Reproductive 144 40 0.73 0.15 67.42 NA NA NA NA
Maturity 174 40 0.58 0.20 85.79 7.25 5.72 0.90 0.23
Mean 40 0.55 0.20 61.01 NA NA NA NA

Narrowbands with RF were able to reasonably predict rice and soybean images (R2 = 0.72 and RMSE = 0.17 kg m− 2). As with PRISMA, pre­
biomass at the early vegetative stages as well. Contrary to corn, rice and dictions were highest early in the season when single-date Sentinel-2
soybean, what biomass was best predicted with PRISMA images images were used. Sentinel-2 performed particularly well for wheat
collected at reproductive and maturity stages. PRISMA images collected biomass and yield when all images collected along the entire growing
at the reproductive stage obtained the best prediction for all crops with season were used as predictors. Sentinel-2 red-edge band 5 and NIR band
the exception of rice, which was best predicted at the vegetative stage. 8 were the most important explanatory variables across crops and
There was much variation in the most important predictors among the growth stages. Contrary to PRISMA, the Sentinel-2 SWIR bands were
RF models, but some important patterns emerged. Narrowbands in the generally less important.
NIR and SWIR were more important than in the visible or red-edge
across crops and growth stages. Narrowbands in the NIR and SWIR1 4. Discussion
(SWIR1 and SWIR2) were important for predicting yield (biomass) at the
maturity phase. Narrowbands in the SWIR1 were also important for 4.1. Potential of PRISMA and Sentinel-2 images for yield and biomass
predicting biomass at the reproductive stage. SWIR2 was more impor­ predictions
tant at the vegetative stage for biomass. The most important predictors
mostly linked to regions of the spectrum with the highest positive and Our study revealed that both PRISMA and Sentinel-2 spectral data
negative loadings, but this was not always the case. For example, rice can accurately predict the biomass and yield of corn, soybean, rice and
biomass exhibited strong negative loadings in the visible at the vege­ wheat.
tative stage, which did not appear in the final RF model. This conclusion confirmed the results reported in the previous
Sentinel-2 broadbands and red-edge narrowbands on average studies assessing the potential of hyperspectral and multispectral mis­
explained 6% less variability in end-of-season biomass with higher error sions (Mariotto et al, 2013, Marshall and Thenkabail, 2015a), UAV-
(ΔRMSE = 0.05 kg m− 2) (Table 7). based red–green-blue and hyperspectral imaging (Li et al., 2020),
The best prediction for corn yield was obtained with Sentinel-2 Sentinel-2 (Hunt et al., 2019) for yield and biomass predictions of

202
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Table 6
Summary of Random Forest predictive models at a given day of year (DOY) PRISMA acquisition. The predictors are expressed as wavelengths from left to right in order
of importance.
Wavelength (nm)

Crop DOY Stage N R2 RMSE 1 2 3 4 5

Biomass
Corn 143 Vegetative1 39 0.63 0.52 729 739 2127 1099 708
177 Vegetative 40 0.60 0.52 770 492 538 2143 759
212 Reproductive 37 0.54 0.57 650 641 538 530 1501
All 37 0.64 0.50 143–729 212–650 143–739 177–770 143–708
Rice 177 Vegetative1 40 0.64 0.54 2086 1131 854 1141 1028
212 Vegetative 37 0.66 0.51 770 2143 780 739 2175
259 Maturity 39 0.66 0.52 587 1120 579 2035 2027
All 37 0.74 0.47 259–1120 259–1240 259–1273 259–2035 259–579
Soybean 177 Vegetative 26 0.81 0.09 507 1480
212 Reproductive 26 0.79 0.09 833 1306
All 26 0.81 0.09 177–506 212–833 212–2320 212–1306
Wheat 97 Vegetative 38 0.59 0.58 708 587 570 562
143 Reproductive 39 0.74 0.41 470 1141 1131 538 1008
177 Maturity 38 0.74 0.48 2229 660 822 2222 1606
All 38 0.82 0.36 143–470 143–1141 177–2229 143–1131 177–660
Yield
Corn 143 Vegetative1 39 0.52 0.25 1784 1765 708 729 1295
177 Vegetative 40 0.52 0.24 770 780 759 2143 2414
212 Reproductive 37 0.63 0.21 538 650 641
All 37 0.71 0.19 143–1784 212–538 143–1765 177–770 212–729
Rice 177 Vegetative1 40 0.57 0.21 854 669 641
212 Vegetative 37 0.55 0.21 1078 1163 876 570 2206
259 Maturity 39 0.51 0.23 759 1109 1037 1284 1273
All 37 0.63 0.20 259–1120 259–759 259–579 259–1037 259–1273
Soybean 177 Vegetative 26 0.75 0.06 530 2283 2290 978 522
212 Reproductive 26 0.86 0.04 1616 739 886 1606 1595
All 26 0.86 0.04 212–1616 212–739 212–1606 177–530 212–886
Wheat 97 Vegetative 38 0.48 0.23 1993 2462 2496 587 554
143 Reproductive 39 0.71 0.17 1131 470 1141 1295 1008
177 Maturity 38 0.69 0.19 812 801 822 770 1606
All 38 0.80 0.15 177–812 177–822 177–801 143–470 143–1131
1
Early vegetative stage.

various crops. Mariotto et al. (2013) used stepwise regression to build who stated that, in general, yield is more difficult to predict than
models between Hyperion hyperspectral narrowbands and wet- and dry- biomass because of its sensitivity to G × E × M. In previous studies (Li
weight biomass and yield on a smaller sample size of field data (8 and 23 et al., 2020; Montesinos-López et al., 2017), prediction models devel­
samples respectively). Higher accuracies in that study were produced oped using VIs obtained lower accuracy than PLSR models that account
(wet-weight corn biomass R2 = 0.99 and dry-weight yield for wheat R2 for the full optical range. In our work PLSR, performed better than TBVI
= 0.98) with four and eight predictors, respectively. In our work, we except for soybean biomass and yield, wheat biomass and yield, and
reduced the number of input predictors to five for 40 samples following corn yield prediction obtained using PRISMA images.
the recommendation of Thenkabail et al. (2000) to prevent over-fitting. The RF-based models developed using PRISMA performed better
Marshall and Thenkabail, (2015a) analyzed more samples and used with lower error and considerably fewer predictors (≤5) than the TBVI
fewer predictors with stepwise regression. Their model outperformed and PLSR models for both yield and biomass. These results confirmed
ours for rice (N = 20, R2 = 0.92) and corn (N = 20, R2 = 0.93) biomass previous studies that showed that RF-based prediction models out­
using only two predictors. Wet-weight biomass samples were collected performed other methods including PLSR (Li et al., 2020) or support
throughout the growing season, especially during the vegetative and vector regression and artificial neural network (Wang et al., 2016). RF
reproductive stages. Their models therefore captured more variability performed better than the other two methods because it exploited the
and were more sensitive to environmental stressors that may otherwise explanatory power of first-derivative transformed PRISMA as predictors.
be obscured in end-of-season biomass. Like our study, spectra were first- The first-derivative transformation adopted in our study was a way to
derivative transformed. accentuate important wavelengths, while diminishing the importance of
Hunt et al. (2019) evaluated the performance of Sentinel-2 broad­ adjacent narrowbands that explain essentially the same variance (i.e.,
bands and red-edge and NIR narrowbands for predicting wheat yields. eliminated redundant information). Further, RF accounted only for the
Yield data were collected from combined harvesters. This resulted in a five (or less) most relevant features selected using RFE feature selection
large sample size (N = 8000). Peak performance using RF and normal­ method.
ized two-band ratios occurred at the maturity stage, but the models
explained 15% less variability than our study. Better results were re­
4.3. Importance of the multi-temporal images for yield and biomass
ported by Li et al. (2020) for potato yield (R2 > 0.90) and biomass (R2 >
predictions
0.80) estimated with PLSR and RF models and UAV images.

The best prediction results were obtained by RF when all available


4.2. Performance of TBVI, PLSR and RF-based prediction models images were considered as input predictors. These results confirmed the
study of Hunt et al. (2019) who showed that incorporating multiple
Our results showed that TBVI, PLSR, and RF-based prediction acquisitions of remote sensing spectra over the growing season led to
models, with the exception of soybean and corn, performed better for considerably higher model performance.
biomass than yield. A similar conclusion was drawn by Li et al (2020) Previous studies (e.g., Johnson, 2016) identified the highest

203
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Table 7
Summary of Random Forest predictive models at a given day of year (DOY) Sentinel-2 acquisition. The predictors are expressed as band numbers from left to right in
order of importance. N = 40 for corn, rice and wheat and N = 26 for soybean.
Band #

Crop Stage DOY R2 RMSE 1 2 3 4 5

Biomass
Corn Vegetative1 114 0.50 0.63 5 7
Vegetative1 144 0.55 0.60 7 6 8a 8 4
Vegetative 174 0.46 0.64 5 7 8a 8 3
Vegetative 194 0.51 0.60 3 6 5 8a 7
Reproductive 224 0.52 0.59 5 6 3 11 8
Reproductive 234 0.50 0.64 3 5 4 7 8
All 0.57 0.45 234–3 144–6 144–7 224–11 114–5
Rice Vegetative1 174 0.52 0.65 11 4 8a
Vegetative 194 0.44 0.72 12 4 11 7 5
Vegetative 224 0.45 0.71 8 12 8a 5 6
Reproductive 234 0.51 0.62 12 11 7 8a 8
All 0.47 0.69 174–11 174–12 234–8 234–7 174–4
Soybean Vegetative 174 0.69 0.13 5 3 7 6 8a
Vegetative 194 0.73 0.12 3 4
Reproductive 224 0.70 0.12 4 11 5 12 3
Reproductive 234 0.73 0.11 7 8
All 0.71 0.11 174–5 224–11 174–3 194–3 234–7
Wheat Vegetative 94 0.65 0.50 3 5 12 4 6
Vegetative 103 0.69 0.49 4 12 5 3 11
Reproductive 144 0.81 0.37 4 12 5
Maturity 174 0.55 0.60 5 8a 8 7 11
All 0.81 0.37 4 12 5
Yield
Corn Vegetative1 114 0.57 0.22 4 5
Vegetative1 144 0.39 0.29 7 8
Vegetative 174 0.42 0.27 5 8 8a 7 3
Vegetative 194 0.67 0.20 6 8a 3 7 8
Reproductive 224 0.62 0.22 5 3 4 6 8a
Reproductive 234 0.58 0.23 3 5 6 8a
All 0.72 0.17 234–3 224–5 224–3 194-8a 194–8
Rice Vegetative1 174 0.31 0.28 8a 4 8 7 5
Vegetative 194 0.40 0.26 12 11
Vegetative 224 0.48 0.24 4 8 7 8a 5
Reproductive 234 0.56 0.21 8 7 12 8a 5
All 0.46 0.22 234–7 234–8 234-8a 224–4 224–5
Soybean Vegetative 174 0.78 0.06 12 8 5 4
Vegetative 194 0.82 0.05 11 6 3 12
Reproductive 224 0.77 0.06 3 6 8 8a 11
Reproductive 234 0.76 0.06 6 8 8a 12 4
All 0.84 0.04 194–11 234–6 194–3 194–6 234–12
Wheat Vegetative 94 0.55 0.21 5 3 4 6 12
Vegetative 103 0.60 0.21 4 12 5 7 3
Reproductive 144 0.77 0.15 4 5 12 3 7
Maturity 174 0.59 0.22 12 11 7 6 5
All 0.77 0.15 144–4 144–12 144–5 144–3 144–7
1
Early vegetative stage.

correlation between yield and spectral indicators at the onset of the The strong performance of spectra at the early vegetative phase for
reproductive stage (i.e., anthesis). This is the point in crop development corn and rice for yield was surprising. Biomass accumulation is less
when the canopy reaches peak productivity and maximum chlorophyll critical to grain-filling pre-anthesis than post-anthesis, but early crop
absorption. This phase is also the most complex because plants transi­ development does have an impact on end-of-season biomass and yield
tion from the development of vegetative to reproductive organs (i.e., the (Kemanian et al., 2007). At the same time, corn and rice spectra at the
grain-filling process triggers translocation of assimilate from plant to early vegetative stage when canopies are open, are significantly
yield). Photosynthesis slows during the maturity phase, but chlorophyll impacted by soil and water background presence, respectively. Soils
can still impact the spectral response. Longer wavelengths associated tend to scatter more light with increasing wavelength from the visible to
with non-photosynthetic vegetation (NPV) crop components however, NIR. Scatter from soil remains high, but is lower in the SWIR when soil is
as displayed by the high PLSR loadings in the SWIR1 and SWIR2, moist. Rice fields are flooded or saturated, so the rice signal is impacted
become more prominent. We hypothesize this is due to the depletion of more by water presence than by soil characteristics. Water, contrary to
canopy moisture during maturity. Absorption due to canopy moisture soil, absorb most incoming radiation. Soil conditions (e.g., soil fertility)
tends to obscure the NPV absorption signal during earlier growth stages. at these early stages could improve predictions, but we suspect soil in
The NIR region, which during maturity is associated with scatter from general suppresses the vegetation signal. In order to prove this, we
NPV canopy structure was more important for yield than biomass. The would need a spectral library consisting of crop, soil, and water end­
yield is completely covered by a spikelet for rice and wheat during members to sperate the signal through unmixing.
maturity. Perhaps the spikelet scatters NIR without impacting absorp­ The transition from the reproductive to maturity stage is best illus­
tion features of other plant material. A larger sample size, more so­ trated by rice and wheat spectra. Rice narrowbands were acquired at the
phisticated data mining techniques, and additional field data collected beginning of the maturity phase, while wheat narrowbands were ac­
through the growing season are required to test these hypotheses. quired at the end of the maturity phase. The rice signal showed a weaker

204
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

but clear response to chlorophyll in this stage than at earlier stages other satellite images such as Sentinel-2 either via data fusion (Acito
where background influence was still dominant. The wheat signal on the et al., 2022) or integration.
other hand was devoid of the chlorophyll signal and instead resembled Machine learning algorithms, in general, require a large number of
soil spectra. samples collected across large areas and several growing seasons. To
Soybean is faster growing than the other crops, so it was rotated out address the challenges related to sample availability, transfer learning
earlier with more variable timing. Fourteen soybean samples had to be approaches similar to those recently used for Nitrogen estimation (Wan
discarded because their growth cycles did not correspond well to the et al., 2022) using hyperspectral measurements should be considered.
available image acquisition dates. The smaller size likely impacted
soybean model performance. 5. Conclusion
The PLSR loadings at the later vegetative, reproductive, and maturity
stages showed similar patterns in the visible, NIR, and SWIR across crop Within-season forecasting of agricultural production with space­
types. Negative loadings in the visible during the vegetative and borne hyperspectral remote sensing remains in its infancy. We evaluated
reproductive phase are influenced mainly by chlorophyll in the visible the performance of narrowbands as detected by PRISMA spaceborne
red and less so by absorption of chlorophyll and accessory pigments platform against Sentinel-2 using end of season clear-cut dry-weight
outside the visible red. This absorption outweighs the influence of crop biomass and yield samples acquired during the primary crop
scatter near the green peak. Negative loadings in the SWIR2 could be growth phases (vegetative, reproductive, maturity). The assessment was
due to absorption by water content in the canopy (main effect) and made for four global food crops (corn, rice, soybean, wheat) using TBVI,
proteins or carbon base constituents (secondary effect). The strong PLSR, and RF methods. Our study showed that both PRISMA and
positive loadings in the NIR and SWIR1 during the vegetative and Sentinel-2 images are promising data sources for predicting crop yield
reproductive phase are affected mostly by scatter from canopy structure. and biomass. RF obtained better prediction accuracy than TBVI and
The results of our analysis confirm Mariotto et al. (2013) and PLSR for all crops with one exception. Corn biomass was similarly pre­
Marshall and Thenkabail, (2015a), which highlighted the importance of dicted with all three methods when Sentinel-2 images were used as input
NIR and SWIR over visible and red-edge narrowbands for predicting predictors.
crop biomass and yield. The SWIR region was more important for RF yielded the best prediction accuracy when multiple acquisitions
biomass estimation and PRISMA outperformed Sentinel-2. Several of remote sensing spectra over the growing season were used as input
important features in the SWIR region are related to biomass accumu­ predictors. The prediction phase also revealed the importance of
lation, which are indistinguishable by Sentinel-2 SWIR broadbands. On PRISMA NIR and SWIR and Sentinel-2 red-edge narrowbands in pre­
the other hand, the NIR region was more important for yield estimation dicting crop biomass and yield. PRISMA slightly underperformed for
and PRISMA slightly underperformed. PRISMA had large data anomalies crop yield because of data anomalies in the NIR.
(e.g., spikes) in the NIR. The largest spike corresponded to lower PLSR Future work should expand the sample size of our analysis, exam the
loadings near the red-edge inflection point (730 nm) and the first- reliability of early season crop yield predictions based on mixed
derivative transformation could have augmented it. We tested other vegetation-soil signals, and evaluate the performance of PRISMA/
smoothing techniques but this did not impact the outcome of the anal­ Sentinel-2 data fusion/integration.
ysis. The data spike issue in the standard PRISMA product is currently
being addressed by two of the co-authors. Sentinel-2 red-edge narrow­
bands at the vegetative and reproductive stages with their sensitivity to Declaration of Competing Interest
chlorophyll content, clearly boost the predictive power of NIR broad­
bands. This led to better Sentinel-2 performance for yield estimation The authors declare that they have no known competing financial
when data from active vegetation were analyzed. Sentinel-2 performed interests or personal relationships that could have appeared to influence
particularly well for wheat at the reproductive stage. Timing clearly the work reported in this paper.
played an important role. Sentinel-2 data are available continuously,
whereas PRISMA is available on-demand. As a result, more Sentinel-2 Acknowledgments
images were acquired at each growth stage, which increased the likeli­
hood of capturing a biophysical property or biochemical process related Our research was supported by the European Space Agency (ESA) –
to biomass and yield. EO Science for society (EOEO-5 Block 4, 4000130277/20/I-DT) under
the projected titled, “HyNutri: Sensing ‘Hidden Hunger’ with Sentinel-2
4.4. Future work and Hyperspectral (PRISMA)” (http://www.hynutri.nl). The authors
gratefully acknowledge the PRISCAV project (ASI Contract 2019-5-HH-
The study showed the potential of hyperspectral imagery and ma­ 0) for the acquisition of PRISMA imagery and field spectral measure­
chine learning techniques for crop yield estimation but future work ments to assess data quality and pre-processing. We would like to extend
should address the limitations of such data and methods given the re­ our gratitude to Dr. Donato Cillis and Gabriele Dottori of IBF-Servizi
quirements of agricultural monitoring systems. who collected and processed the field samples. Finally, we would like
The narrow swath (30 km) and temporal characteristics of PRISMA to thank Dr. Caroline Lievens and Kathrin Zweers from ITC’s Geoscience
are not ideal for operational monitoring. At-nadir acquisitions are Laboratory who coordinated the receipt, processing, and analysis of field
theoretically available over the same target every 29 days. PRISMA samples shipped by IBF-Servizi.
imagery is currently available only on demand, which means off-nadir
acquisitions are frequently made. This hinders periodic and systematic Appendix A
acquisition over the same target. Further, the spatial resolution of these
images makes it difficult to use them in smallholder farming systems. Annexes.
One way of addressing these challenges is to combine PRISMA with See Figs. A1.1-A1.3.

205
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Fig. A1.1. Spectral signatures for wheat recorded in our study area during the early vegetative ( ), reproductive ( ), and maturity ( ) phases by (a) PRISMA
and (b) Sentinel-2. The black lines denote the average of all phases.

206
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Fig. A1.2. Relative contribution of PRISMA spectral features to biomass estimation for corn (a), rice (b), soybean (c), and wheat (d) at the three main growth stages:
early vegetative ( ), vegetative ( ), reproductive ( ), maturity ( ), and mean ( ).

207
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

Fig. A1.3. Relative contribution of PRISMA spectral features to yield estimation for corn (a), rice (b), soybean (c), and wheat (d) at the three main growth stages:
early vegetative ( ), vegetative ( ), reproductive ( ), maturity ( ), and mean ( ).

208
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

References Huete, A., Didan, K., Miura, T., Rodriguez, E.P., Gao, X., Ferreira, L.G., 2002. Overview
of the radiometric and biophysical performance of the MODIS vegetation indices.
Remote Sens. Environ. 83 (1-2), 195–213. https://doi.org/10.1016/S0034-4257(02)
Acito, N., Diani, M., Corsini, G., 2022. PRISMA spatial resolution enhancement by fusion
00096-2.
with Sentinel-2 data. IEEE J Sel. Top. Appl. Earth Obs. Remote Sens. 15, 62–79.
Hunt, M.L., Blackburn, G.A., Carrasco, L., Redhead, J.W., Rowland, C.S., 2019. High
https://doi.org/10.1109/JSTARS.2021.3132135.
resolution wheat yield mapping using Sentinel-2. Remote Sens. Environ. 233,
Atzberger, C., Guérif, M., Baret, F., Werner, W., 2010. Comparative analysis of three
111410. https://doi.org/10.1016/j.rse.2019.111410.
chemometric techniques for the spectroradiometric assessment of canopy
Jin, Z., Azzari, G., You, C., Di Tommaso, S., Aston, S., Burke, M., Lobell, D.B., 2019.
chlorophyll content in winter wheat. Comput. Electron. Agric. 73 (2), 165–173.
Smallholder maize area and yield mapping at national scales with Google Earth
https://doi.org/10.1016/j.compag.2010.05.006.
Engine. Remote Sens. Environ. 228, 115–128. https://doi.org/10.1016/j.
Barrs, H.D., Weatherley, P.E., 1962. A re-examination of the relative turgidity technique
rse.2019.04.016.
for estimating water deficits in leaves. Aust. J. Biol. Sci. 15, 413–428. https://doi.
Johnson, D.M., 2016. A comprehensive assessment of the correlations between field crop
org/10.1071/bi9620413.
yields and commonly used MODIS products. Int. J. Appl. Earth Obs. Geoinformation
Belgiu, M., Drăguţ, L., 2016. Random forest in remote sensing: A review of applications
52, 65–81. https://doi.org/10.1016/j.jag.2016.05.010.
and future directions. ISPRS J. Photogramm. Remote Sens. 114, 24–31. https://doi.
Jones, J.W., Antle, J.M., Basso, B., Boote, K.J., Conant, R.T., Foster, I., Godfray, H.C.J.,
org/10.1016/j.isprsjprs.2016.01.011.
Herrero, M., Howitt, R.E., Janssen, S., Keating, B.A., Munoz-Carpena, R., Porter, C.
Belgiu, M., Marshall, M., Boschetti, M., Pepe, M., Stein, A., Lievens, C., 2021. HyNutri:
H., Rosenzweig, C., Wheeler, T.R., 2017. Toward a new generation of agricultural
Estimating the Nutritional Composition of Wheat from Multi-Temporal Prisma Data,
system data, models, and knowledge products: State of agricultural systems science.
in: 2021 IEEE International Geoscience and Remote Sensing Symposium IGARSS.
Agric. Syst. 155, 269–288. https://doi.org/10.1016/j.agsy.2016.09.021.
Presented at the 2021 IEEE International Geoscience and Remote Sensing
Kamir, E., Waldner, F., Hochman, Z., 2020. Estimating wheat yields in Australia using
Symposium IGARSS, pp. 471–474. https://doi.org/10.1109/
climate records, satellite image time series and machine learning methods. ISPRS J.
IGARSS47720.2021.9553614.
Photogramm. Remote Sens. 160, 124–135. https://doi.org/10.1016/j.
Ben-Ari, T., Makowski, D., 2016. Analysis of the trade-off between high crop yield and
isprsjprs.2019.11.008.
low yield instability at the global scale. Environ. Res. Lett. 11 (10), 104005. https://
Kemanian, A.R., Stöckle, C.O., Huggins, D.R., Viega, L.M., 2007. A simple method to
doi.org/10.1088/1748-9326/11/10/104005.
estimate harvest index in grain crops. Field Crops Res. 103 (3), 208–216. https://doi.
Borchers, H.W., 2021. caret: Classification and Regression Training [WWW Document].
org/10.1016/j.fcr.2007.06.007.
URL https://CRAN.R-project.org/package=caret.
Kowalik, W., Dabrowska-Zielinska, K., Meroni, M., Raczka, T.U., de Wit, A., 2014. Yield
Breiman, L., 2001. Random Forests. Mach. Learn. 45, 5–32. https://doi.org/10.1023/A:
estimation using SPOT-VEGETATION products: A case study of wheat in European
1010933404324.
countries. Int. J. Appl. Earth Obs. Geoinformation 32, 228–239. https://doi.org/
Broge, N.H., Leblanc, E., 2001. Comparing prediction power and stability of broadband
10.1016/j.jag.2014.03.011.
and hyperspectral vegetation indices for estimation of green leaf area index and
Kuhn, M., 2021. pracma: Practical Numerical Math Functions [WWW Document]. URL
canopy chlorophyll density. Remote Sens. Environ. 76 (2), 156–172. https://doi.
https://cran.r-project.org/package=pracma.
org/10.1016/S0034-4257(00)00197-8.
Leng, G., Hall, J.W., 2020. Predicting spatial and temporal variability in crop yields: an
Busetto, L., 2021. prismaread: A tool for facilitating access and analysis of PRISMA L1/L2
inter-comparison of machine learning, regression and process-based models.
hyperspectral imagery [WWW Document]. URL https://ranghetti.github.io/
Environ. Res. Lett. 15 (4), 044027. https://doi.org/10.1088/1748-9326/ab7b24.
prismaread/.
Li, B., Xu, X., Zhang, L., Han, J., Bian, C., Li, G., Liu, J., Jin, L., 2020. Above-ground
Casas, A., Riaño, D., Ustin, S.L., Dennison, P., Salas, J., 2014. Estimation of water-related
biomass estimation and yield prediction in potato by using UAV-based RGB and
biochemical and biophysical vegetation properties using multitemporal airborne
hyperspectral imaging. ISPRS J. Photogramm. Remote Sens. 162, 161–172. https://
hyperspectral data and its comparison to MODIS spectral response. Remote Sens.
doi.org/10.1016/j.isprsjprs.2020.02.013.
Environ. 148, 28–41. https://doi.org/10.1016/j.rse.2014.03.011.
Licker, R., Johnston, M., Foley, J.A., Barford, C., Kucharik, C.J., Monfreda, C.,
Claverie, M., Ju, J., Masek, J.G., Dungan, J.L., Vermote, E.F., Roger, J.-C., Skakun, S.V.,
Ramankutty, N., 2010. Mind the gap: how do climate and agricultural management
Justice, C., 2018. The Harmonized Landsat and Sentinel-2 surface reflectance data
explain the ‘yield gap’ of croplands around the world? Glob. Ecol. Biogeogr. 19,
set. Remote Sens. Environ. 219, 145–161. https://doi.org/10.1016/j.
769–782. https://doi.org/10.1111/j.1466-8238.2010.00563.x.
rse.2018.09.002.
Ma, Y., Zhang, Z., Kang, Y., Özdoğan, M., 2021. Corn yield prediction and uncertainty
Cogliati, S., Sarti, F., Chiarantini, L., Cosi, M., Lorusso, R., Lopinto, E., Miglietta, F.,
analysis based on remotely sensed variables using a Bayesian neural network
Genesio, L., Guanter, L., Damm, A., Pérez-López, S., Scheffler, D., Tagliabue, G.,
approach. Remote Sens. Environ. 259, 112408. https://doi.org/10.1016/j.
Panigada, C., Rascher, U., Dowling, T.P.F., Giardino, C., Colombo, R., 2021. The
rse.2021.112408.
PRISMA imaging spectroscopy mission: overview and first performance analysis.
Mariotto, I., Thenkabail, P.S., Huete, A., Slonecker, E.T., Platonov, A., 2013.
Remote Sens. Environ. 262, 112499. https://doi.org/10.1016/j.rse.2021.112499.
Hyperspectral versus multispectral crop-productivity modeling and type
Defourny, P., Bontemps, S., Bellemans, N., Cara, C., Dedieu, G., Guzzonato, E.,
discrimination for the HyspIRI mission. Remote Sens. Environ. 139, 291–305.
Hagolle, O., Inglada, J., Nicola, L., Rabaute, T., Savinaud, M., Udroiu, C., Valero, S.,
https://doi.org/10.1016/j.rse.2013.08.002.
Bégué, A., Dejoux, J.-F., El Harti, A., Ezzahar, J., Kussul, N., Labbassi, K.,
Marshall, M., Thenkabail, P., 2015a. Advantage of hyperspectral EO-1 Hyperion over
Lebourgeois, V., Miao, Z., Newby, T., Nyamugama, A., Salh, N., Shelestov, A.,
multispectral IKONOS, GeoEye-1, WorldView-2, Landsat ETM+, and MODIS
Simonneaux, V., Traore, P.S., Traore, S.S., Koetz, B., 2019. Near real-time agriculture
vegetation indices in crop biomass estimation. ISPRS J. Photogramm. Remote Sens.
monitoring at national scale at parcel resolution: Performance assessment of the
108, 205–218. https://doi.org/10.1016/j.isprsjprs.2015.08.001.
Sen2-Agri automated system in various cropping systems around the world. Remote
Marshall, M., Thenkabail, P., 2015b. Developing in situ Non-Destructive Estimates of
Sens. Environ. 221, 551–568. https://doi.org/10.1016/j.rse.2018.11.007.
Crop Biomass to Address Issues of Scale in. Remote Sens. 7, 808–835. https://doi.
Doraiswamy, P.C., Moulin, S., Cook, P.W., Stern, A., 2003. Crop yield assessment from
org/10.3390/rs70100808.
remote sensing. Photogramm. Eng. Remote Sens. 69 (6), 665–674.
Marshall, M., Thenkabail, P., 2014. Biomass modeling of four leading world crops using
Feng, L., Wang, Y., Zhang, Z., Du, Q., 2021. Geographically and temporally weighted
hyperspectral narrowbands in support of HyspIRI Mission. Photogramm Eng Remote
neural network for winter wheat yield prediction. Remote Sens. Environ. 262,
Sensing 80 (8), 757–772.
112514. https://doi.org/10.1016/j.rse.2021.112514.
Monteith, J., 1969. Light Interception and Radiative Exchange in Crop Stands. Agron.
Féret, J.-B., Berger, K., de Boissieu, F., Malenovský, Z., 2021. PROSPECT-PRO for
Hortic. – Fac. Publ.
estimating content of nitrogen-containing leaf proteins and other carbon-based
Montesinos-López, O.A., Montesinos-López, A., Crossa, J., de los Campos, G.,
constituents. Remote Sens. Environ. 252, 112173. https://doi.org/10.1016/j.
Alvarado, G., Suchismita, M., Rutkoski, J., González-Pérez, L., Burgueño, J., 2017.
rse.2020.112173.
Predicting grain yield using canopy hyperspectral reflectance in wheat breeding
Frampton, W.J., Dash, J., Watmough, G., Milton, E.J., 2013. Evaluating the capabilities
data. Plant Methods 13 (1). https://doi.org/10.1186/s13007-016-0154-2.
of Sentinel-2 for quantitative estimation of biophysical variables in vegetation.
Moulin, S., Bondeau, A., Delecolle, R., 1998. Combining agricultural crop models and
ISPRS J. Photogramm. Remote Sens. 82, 83–92. https://doi.org/10.1016/j.
satellite observations: From field to regional scales. Int. J. Remote Sens. 19 (6),
isprsjprs.2013.04.007.
1021–1036. https://doi.org/10.1080/014311698215586.
Gitelson, A., Solovchenko, A., 2017. Generic Algorithms for Estimating Foliar Pigment
Pepe, M., Pompilio, L., Gioli, B., Busetto, L., Boschetti, M., 2020. Detection and
Content. Geophys. Res. Lett. 44 (18), 9293–9298. https://doi.org/10.1002/
classification of non-photosynthetic vegetation from PRISMA hyperspectral data in
2017GL074799.
croplands. Remote Sens. 12, 3903. https://doi.org/10.3390/rs12233903.
Guan, K., Wu, J., Kimball, J.S., Anderson, M.C., Frolking, S., Li, B., Hain, C.R., Lobell, D.
Rast, M., Painter, T.H., 2019. Earth observation imaging spectroscopy for terrestrial
B., 2017. The shared and unique values of optical, fluorescence, thermal and
systems: an overview of its history, techniques, and applications of its missions. Surv.
microwave satellite data for estimating large-scale crop yields. Remote Sens.
Geophys. 40 (3), 303–331. https://doi.org/10.1007/s10712-019-09517-z.
Environ. 199, 333–349. https://doi.org/10.1016/j.rse.2017.06.043.
Ray, D.K., Gerber, J.S., MacDonald, G.K., West, P.C., 2015. Climate variation explains a
Guyon, I., Weston, J., Barnhill, S., Vapnik, V., 2002. Gene Selection for Cancer
third of global crop yield variability. Nat. Commun. 6 (1) https://doi.org/10.1038/
Classification using Support Vector Machines. Mach. Learn. 46, 389–422. https://
ncomms6989.
doi.org/10.1023/A:1012487302797.
Sagan, V., Maimaitijiang, M., Bhadra, S., Maimaitiyiming, M., Brown, D.R., Sidike, P.,
Hank, T.B., Berger, K., Bach, H., Clevers, J.G.P.W., Gitelson, A., Zarco-Tejada, P.,
Fritschi, F.B., 2021. Field-scale crop yield prediction using multi-temporal
Mauser, W., 2019. Spaceborne Imaging Spectroscopy for Sustainable Agriculture:
WorldView-3 and PlanetScope satellite data and deep learning. ISPRS J.
Contributions and Challenges. Surv. Geophys. 40 (3), 515–551. https://doi.org/
Photogramm. Remote Sens. 174, 265–281. https://doi.org/10.1016/j.
10.1007/s10712-018-9492-0.
isprsjprs.2021.02.008.
Hay, R.K.M., 1995. Harvest index: a review of its use in plant breeding and crop
Schlemmer, M., Gitelson, A., Schepers, J., Ferguson, R., Peng, Y., Shanahan, J.,
physiology. Ann. Appl. Biol. 126 (1), 197–216. https://doi.org/10.1111/j.1744-
Rundquist, D., 2013. Remote estimation of nitrogen and chlorophyll contents in
7348.1995.tb05015.x.

209
M. Marshall et al. ISPRS Journal of Photogrammetry and Remote Sensing 187 (2022) 191–210

maize at leaf and canopy levels. Int. J. Appl. Earth Obs. Geoinformation 25, 47–54. Wang, Li’ai, Zhou, X., Zhu, X., Dong, Z., Guo, W., 2016. Estimation of biomass in wheat
https://doi.org/10.1016/j.jag.2013.04.003. using random forest regression algorithm and remote sensing data. Crop J. 4 (3),
Thenkabail, P.S., Mariotto, I., Gumma, M.K., Middleton, E.M., Landis, D.R., 212–219. https://doi.org/10.1016/j.cj.2016.01.008.
Huemmrich, K.F., 2013. Selection of Hyperspectral Narrowbands (HNBs) and Wang, Y., Zhang, Z., Feng, L., Du, Q., Runge, T., 2020. Combining multi-source data and
Composition of Hyperspectral Twoband Vegetation Indices (HVIs) for Biophysical machine learning approaches to predict winter wheat yield in the conterminous
Characterization and Discrimination of Crop Types Using Field Reflectance and United States. Remote Sens. 12, 1232. https://doi.org/10.3390/rs12081232.
Hyperion/EO-1 Data. IEEE J Sel. Top. Appl. Earth Obs. Remote Sens. 6 (2), 427–439. Weiss, M., Jacob, F., Duveiller, G., 2020. Remote sensing for agricultural applications: a
https://doi.org/10.1109/JSTARS.2013.2252601. meta-review. Remote Sens. Environ. 236, 111402. https://doi.org/10.1016/j.
Thenkabail, P.S., Smith, R.B., De Pauw, E., 2000. Hyperspectral vegetation indices and rse.2019.111402.
their relationships with agricultural crop characteristics. Remote Sens. Environ. 71 Wójtowicz, A., Piekarczyk, J., Czernecki, B., Ratajkiewicz, H., 2021. A random forest
(2), 158–182. https://doi.org/10.1016/S0034-4257(99)00067-X. model for the classification of wheat and rye leaf rust symptoms based on pure
Tsai, F., Philpot, W., 1998. Derivative analysis of hyperspectral data. Remote Sens. spectra at leaf scale. J. Photochem. Photobiol. B 223, 112278. https://doi.org/
Environ. 66 (1), 41–51. https://doi.org/10.1016/S0034-4257(98)00032-7. 10.1016/j.jphotobiol.2021.112278.
Tucker, C.J., 1979. Red and photographic infrared linear combinations for monitoring Wold, H., 1980. Model Construction and Evaluation When Theoretical Knowledge Is
vegetation. Remote Sens. Environ. 8 (2), 127–150. https://doi.org/10.1016/0034- Scarce: Theory and Application of Partial Least Squares. In: Kmenta, J., Ramsey, J.B.
4257(79)90013-0. (Eds.), Evaluation of Econometric Models. Academic Press, pp. 47–74. https://doi.
Wan, L., Zhou, W., He, Y., Wanger, T.C., Cen, H., 2022. Combining transfer learning and org/10.1016/B978-0-12-416550-2.50007-8.
hyperspectral reflectance analysis to assess leaf nitrogen concentration across Wulder, M.A., Masek, J.G., Cohen, W.B., Loveland, T.R., Woodcock, C.E., 2012. Opening
different plant species datasets. Remote Sens. Environ. 269, 112826. https://doi. the archive: how free data has enabled the science and monitoring promise of
org/10.1016/j.rse.2021.112826. Landsat. Remote Sens. Environ Landsat Legacy Special Issue 122, 2–10. https://doi.
org/10.1016/j.rse.2012.01.010.

210

You might also like