Science of Remote Sensing 12 (2025) 100267

Contents lists available at ScienceDirect

Science of Remote Sensing

journal homepage: [Link]/journal/science-of-remote-sensing

Machine learning algorithms for maize yield prediction with multispectral

imagery: Assessing robustness across varied growing environments
Bala Ram Sapkota a,b , Gurjinder S. Baath a,* , K. Colton Flynn c , Kabindra Adhikari c ,
Chad Hajda c, Douglas R. Smith c
a
Blackland Research & Extension Center, Texas A&M AgriLife Research, 720 E. Blackland Rd, Temple, TX, 76502, USA
b
Department of Soil and Crop Sciences, Texas A&M University, College Station, TX, 77843, USA
c
USDA-ARS Grassland Soil and Water Research Laboratory, 808 E. Blackland Road, Temple, TX, 76502, USA

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

Keywords: Multispectral imagery acquired via Unmanned Aircraft Systems (UAS) can provide an on-demand and cost-
Corn effective approach to crop yield estimation. Traditional methods relying solely on vegetation indices (VIs) for
Remote sensing yield prediction face challenges such as saturation in dense canopies, like those of maize (Zea Mays L.), and
Canopy reflectance
inconsistencies due to environmental variability. This study investigates the potential use of canopy spectral
Vegetation indices
reflectance, directly applied in conjunction with machine learning (ML) to address these limitations and produce
Extra trees regressor
more consistent and reliable yield predictions. The performance of reflectance, VI and hybrid (utilizing both
reflectance and VIs) based ML models was evaluated for maize yield predictions across various crop growth
phases. The research utilizes multispectral imagery and maize yield data from diverse growing environments,
comprising seven maize planting dates tested across three field locations over two years. Among five ML algo­
rithms tested, the Extra Trees Regressor (ETR) showed superior performance at predicting maize yield across
most maize growth phases. Reflectance-based models consistently outperformed VI-based models at mid-
vegetative (R2 = 0.71 vs. 0.50), late vegetative (R2 = 0.85 vs. 0.80), flowering (R2 = 0.83 vs. 0.60), and mid-
reproductive (R2 = 0.83 vs. 0.74) growth phases. Hybrid models combining spectral bands and VIs achieved
the greatest prediction accuracy but required more extensive data pre-processing compared to models using
reflectance or VIs alone. Our findings indicate that direct incorporation of canopy reflectance bands into ML
models, bypassing the complexities associated with VIs, can lead to more consistent, timely and reliable maize
yield predictions across varied growing environments.

1. Introduction Yield prediction plays a crucial role in helping producers optimize

resource allocation throughout the growing period, which includes
Maize (Zea mays L., also known as corn) is one of the dominant cereal adjusting the practices like irrigation, fertilizer application and pest
crops cultivated extensively worldwide. In the United States alone, control to maximize production. Moreover, it allows producers to make
maize for grain was planted across ~35 million hectares, producing informed decisions regarding pricing strategies, storage needs, and po­
~390 million metric tons yield in 2023 (USDA-NASS, 2024). However, tential export opportunities, thereby increasing economic benefits (Zhu
maize yield fluctuates from year to year due to the variation in factors et al., 2019; Ma et al., 2021; Darra et al., 2023; Grabowski and Self,
like weather, soil conditions, agronomic management, or economic as­ 2022).
pects related to fertilizer, seeds, and equipment. For instance, many Traditional methods of yield prediction often rely on manual labor
maize producers adjust their planting time due to low soil moisture and through ground-based assessments, possible mostly in the later stages of
low soil temperature at the time of planting, which affects the crop the maize (Cen and Wan, 2023; Attri et al., 2023). For example, Yield
growth and overall production (Nielsen et al., 2002; Sah et al., 2020). Component Method (YCM) is popular among producers for estimating
The ability to accurately predict crop yield in such changed situations is the yield production that requires taking measurements like number of
of paramount importance (Lobell and Asner, 2003). cobs on plants, average number of kernels per ear, and kernel weight at

* Corresponding author.
E-mail address: [Link]@[Link] (G.S. Baath).

[Link]
Received 7 February 2025; Received in revised form 29 July 2025; Accepted 4 August 2025
Available online 5 August 2025
2666-0172/© 2025 The Authors. Published by Elsevier B.V. This is an open access article under the CC BY license ([Link]
B.R. Sapkota et al. Science of Remote Sensing 12 (2025) 100267

milk stage (R3) of kernel development (Nielsen, 2021; Blumenthal and to spectral reflectance data could enhance robustness and predictive
Thompson, 2009). Another method is to select the ear or ears that accuracy of yield prediction, particularly when dealing with higher
represent the average ear size in the field, followed by counting the variability. ML algorithms, such as Random Forest (RF), K-Nearest
number of kernels and multiplying by 0.300, to obtain a rough yield Neighbors algorithm (KNN), Support Vector Regressor (SVR), and
estimate (Lee and Herbek, 2005). However, these manual methods are Extra-tree Regressor (ETR), offer the flexibility to capture nonlinear
often time-consuming and labor-intensive, especially in large-scale relationships and interactions within complex datasets. These tech­
agricultural production settings. niques help eliminate redundancy, reduce the risk of overfitting, and
Technological advancements, particularly in the realm of remote generate robust models, thereby improving the quality of yield pre­
sensing, have opened new avenues for crop management, yield predic­ dictions (Darra et al., 2023; Sarker, 2021; Kamilaris and Pre­
tion, and planning. For instance, the U.S. Department of Agriculture - nafeta-Boldú, 2018).
National Agricultural Statistics Service (USDA-NASS) has been utilizing Therefore, we hypothesize that incorporating multispectral reflec­
the satellite imagery to monitor crop conditions across the United States tance bands directly into ML models may lead to more accurate and
(Johnson et al., 2021). Similarly, the Group on Earth Observation Global consistent yield predictions for maize compared to models based solely
Agricultural Monitoring (GEOGLAM) provides international crop yield on VIs. This research aims to assess five ML algorithms for estimating
estimates using remote sensing data (Becker-Reshef et al., 2023). In maize yield at various growth stages, considering the diverse environ­
addition to satellite imagery, the integration of unmanned aerial systems mental conditions encountered by maize across different geographic
(UAS) with multispectral cameras offers a non-invasive and efficient field locations, weather patterns, and variations in planting times. The
means of capturing high-resolution agricultural imagery (Colomina and objectives of this research were to: (1) assess the performance of ML
Molina, 2014). It allows acquisition of reflectance data across multiple models that are solely reflectance-dependent, solely VIs dependent, and
spectral bands, which can be radiometrically calibrated to reduce in­ a combination of reflectance and VIs for maize yield estimation, and (2)
consistencies due to atmospheric conditions and flight time (Poncet investigate the effectiveness of these approaches across different crop
et al., 2019). growth stages. The research leverages an extensive and diverse multi­
Spectral information can be transformed into various mathematical spectral imagery and yield dataset obtained from experimental maize
forms known as vegetation indices (VIs), which can correlate with the plots planted at seven different times across three distinct field locations
crop physiological parameters and provide valuable insights into crop during two years (2022 and 2023), allowing a thorough and compre­
health, development, and ultimately yield potential (Camenzind and Yu, hensive comparison.
2024; Barzin et al., 2020; Laliberte et al., 2011). One of the most used VI
is the normalized difference vegetation index (NDVI) that shows strong 2. Materials and methods
association with leaf and dry biomass (Kross et al., 2015; Liu et al., 2012;
Rouse et al., 1974), leaf area index (LAI) (Kross et al., 2015), and yield 2.1. Field experiment sites and agronomic treatments
(Maresma et al., 2020; Hassan et al., 2019). Similarly, several other VIs,
including normalized difference red edge index (NDRE), green The study was conducted during the maize growing seasons (Feb­
normalized difference vegetation index (GNDVI), plant senescence ruary–August) of 2022 and 2023 at the U.S. Department of Agriculture –
reflectance index (PSRI), visible atmospherically resistant index (VARI), Agricultural Research Service (USDA-ARS) Grassland Soil and Water
enhanced vegetation index (EVI), soil adjusted vegetation index (SAVI), Research Laboratory and Texas A&M AgriLife Blackland Research &
modified soil adjusted vegetation index (MSAVI), and triangular vege­ Extension Center, situated in Temple, TX, USA. The dominant soil types
tation index (TVI), are reported to exhibit strong relationship with on the research site are Vertisols mapped as Houston Black Clay (Fine,
chlorophyll, biomass, and are utilized for yield predictions in different smectitic, thermic, udic Haplusterts) (Soil Survey Staff and Keys to Soil
crops (Wang et al., 2019; Gitelson et al., 2002a; Broge and Leblanc, Taxonomy, 2010). The climate conditions at Temple are characterized
2001; Huete et al., 2010). by mild winters with January as the coldest month (Average high 16 ◦ C,
The relationship between VI and crop yield across different growth low 3 ◦ C, precipitation 57 mm), warmer February (Average high 17 ◦ C,
stages has been explored in previous studies using linear models, which low 6 ◦ C, precipitation 51 mm), followed by wet and hot springs
have shown generally inconsistent trends (Maresma et al., 2020; Teal (March–May, average high 25 ◦ C, low 14 ◦ C, precipitation 276 mm), and
et al., 2006; Shrestha et al., 2023). The inconsistency can be attributed drier summers (June–August, average high 35 ◦ C, low 23 ◦ C, precipi­
to environmental variability encountered by crops across different tation 182 mm) (TWDB and TexMesonet, 2024; USDA-ARS, 2024).
geographical locations, inter-annual and intra-annual weather patterns, Producers often begin maize planting from around the last week of
or shifts in planting times. There are some instances like early increase in February as temperature rises above 15 ◦ C.
soil temperature, good moisture, which make producers plant early, In 2022, the study was conducted at a single field location (Field 1)
while extended freezing weather and/or low soil moisture could delay using mid-maturing maize hybrids (NK 1444, A185755, A1457VT2P,
the planting. To improve the accuracy of models accounting for re­ A1687VT2P, A1548DGVT2P, and 57RR) planted on seven different
lationships between spectral information and crop yield across different dates: March 4, March 10, March 18, March 24, April 1, April 8, and
growth stages, it is essential to consider environmental variations April 18. Each planting treatment consisted of twelve rows, with a
induced by location, year, and planting time. Additionally, the re­ length of 280 m and a row spacing of 0.76 m (Fig. 1). In 2023, the study
lationships between certain VIs and crop yield often decline in dense included two other field locations (Field 2 and Field 3), also using mid-
crop canopies due to saturation in VI values. This saturation renders maturing maize hybrids (Pioneer 0622, NK 1694, Innvictis A1548,
VI-based models unresponsive to small spatial or temporal variations Innvictis A1448, Innvictis A1857SS, and Innvictis A1687) planted on
within agricultural fields and thus leading to poor yield predictions seven different dates: February 28, March 3, March 14, March 23, March
(Brantley et al., 2011). 28, April 13, and April 18, and maintaining a similar experimental field
Alternatively, canopy spectral reflectance can directly provide in­ setup as in 2022 (Fig. 1). Both Field 1 and Field 2 were conventionally
formation about the physical conditions of the plants, such as leaf tilled, while a no-tillage practice was implemented in Field 3. In Field 1,
pigment content (Sims and Gamon, 2002) and leaf area index (Zhang maize had also been cultivated in previous year, whereas the previous
et al., 2021). However, the direct use of spectral reflectance data from crop was cotton in both Field 2 and 3. Additionally, the three fields
various wavebands to predict crop yield using linear models has been possess considerable spatial variation in elevation, as shown in Fig. 1. All
limited due to the highly variable nature of data and the influence of field locations were rainfed, and crop management practices followed
atmospheric effects (Gitelson et al., 2002b). In contrast to simple linear local recommendations in both years.
models, applying advanced machine learning (ML) techniques directly

2
B.R. Sapkota et al. Science of Remote Sensing 12 (2025) 100267

Fig. 1. Experimental field layout (left) and spatial elevation (right) for three fields (a, b and c). Numbers 1–7 depict different maize planting dates, with 1 as the
earliest and 7 as the latest. msl-mean sea level.

2.2. Multispectral data collection and processing crop, processing six maize rows at a time. The average speed during the
harvest was about 6.5 km h− 1. A Differential Global Positioning System
A UAS platform equipped with a Micasense RedEdge-MX Dual mul­ (DGPS) was used to record the geographic coordinates of each mea­
tispectral sensor (AgEagle Aerial Systems Inc.) was flown multiple times surement location. The raw yield data was pre-preprocessed in
throughout the maize growing season (Table 1) at an altitude of 30 m AgLeader’s SMS Farming software before being cleaned and corrected
above ground, close to solar noon (11:30 a.m. to 1:30 p.m.). Each flight for combine velocity, flow delays, field pass overlaps, and for potential
was conducted with an 80 % forward and 80 % sidewise overlap. The outliers using Yield Editor Software (Sudduth and Drummond, 2007;
sensor captures 10 multispectral bands — Coastal Blue (444 nm), Blue Adhikari et al., 2023). The yield was then converted to a standard
(475 nm), Green (531 nm), Green (560 nm), Red (650 nm), Red (668 moisture content of 10 % before further extraction and analysis.
nm), Red Edge (705 nm), Red Edge (717 nm), Red Edge (740 nm), and
Near Infrared (NIR) (842 mm) — at an image resolution of 1.2 MP per 2.4. Data extraction and segmentation
band. The sensor has a focal length of 5.424 mm and achieves a Ground
Sample Distance (GSD) of 2.08 cm per pixel when flown at 30 m. The The georeferenced point yield measurements from all three fields
camera is also equipped with a Downwelling Light Sensor (DLS 2) to were used to generate the point shape file in ArcGIS Pro software. A
improve the reflectance calibration during flights. Before and after every buffer with a radius of 1.52 m was created around each point, covering
flight, images of a calibrated reflectance panel (AgEagle Aerial Systems at least 4 center rows and ensuring that it remained within the six-row
Inc.) were captured to radiometrically calibrate the flight images during boundary. The buffered shape file was then converted into squares
pre-processing. Additionally, five permanent Ground Control Points (3.04 m by 3.04 m) and aligned with the inclination of the field using the
(GCPs) were installed in each field, and their coordinates were precisely ‘geopandas’ package in Python (Fig. 2). The median value for each
recorded using the Emlid Reach RS + Survey Kit (Emlid Tech Kft.) to reflectance band and VI was extracted from the corresponding square
serve as geo-reference points during pre-processing. After each flight, shape file using the python packages ‘rasterio’, ‘geopandas’, and
images were pre-processed using Pix4Dmapper (Pix4D S.A.) software to ‘pandas’, and the same process was followed for every UAS flight. The
obtain the geometrically and radiometrically calibrated reflectance average maize yield (Mg ha− 1) was also extracted for each square shape
mosaics for each of the 10 bands. file.
The extracted reflectance values and VIs were grouped according to
maize growth stages. UAS flights covered the crop growth from the fifth
2.3. Maize yield data measurements leaf stage to the maturity stage (Bell, 2017). Crop growth stages were
manually observed at each sampling. The fifth leaf stage (V5) to the
After the maize reached maturity, a John Deere combine harvester eighth leaf stage (V8) was classified as the ‘Mid- Vegetative’ phase, the
equipped with an AgLeader yield monitor system was used to harvest the ninth leaf stage (V9) to the thirteenth leaf stage (V13) as the ‘Late
Vegetative’ and the tasseling (VT) and silking (R1) as the ‘Flowering’
Table 1 phase. The blister (R2) to dough (R4) stage were classified as ‘Mid-
Flight dates of Unmanned Aircraft Systems (UAS) equipped with a multispectral Reproductive’, and the dent (R5) and maturity (R6) stage as ‘Late
sensor over three field locations in Temple, TX, during 2022–2023. Reproductive’ phase. Due to weather-related limitations, such as windy
Field 1 Field 2 Field 3 conditions, cloud cover, and precipitation, UAS data collection was
May 24, 2022 May 05, 2023 – feasible on average once every two weeks. Consequently, these growth
May 27, 2022 May 18, 2023 May 18, 2023 stage classifications were necessary to account for stages with similar
June 09, 2022 May 25, 2023 May 26, 2023 characteristics.
June 17, 2022 June 07, 2023 June 08, 2023
June 23, 2022 June 14, 2023 June 14, 2023
July 01, 2022 June 21, 2023 June 22, 2023
July 08, 2022 July 11, 2023 July 11, 2023

3
B.R. Sapkota et al. Science of Remote Sensing 12 (2025) 100267

Fig. 2. Illustration of sampling points chosen for spectral data extraction, ground-truth yield data, and yield estimation.

Fig. 3. Flow diagram illustrating the strategies used for data collection, pre-processing, calibration, data extraction, data segmentation, and machine learning
regression (ML) process for the yield prediction across different growth stages of maize. VIs: vegetation indices; R2: coefficient of determination; RMSE: root mean
squared error.

4
B.R. Sapkota et al. Science of Remote Sensing 12 (2025) 100267

2.5. Data pre-processing Table 2

Details of multispectral bands and vegetative indices (VIs) used for maize yield
The classified data resulted in a varying number of observations predictions.
across growth phases. To enable meaningful performance comparison SN Type Name of factors Wavelength/Equation References
among growth-based models via ML regression methods, the number of 1 Band blue444 (B444) 444 ± 14 nm ​
data points used were standardized to a common number for each 2 Band blue475 (B475) 475 ± 16 nm ​
growth phase by randomly selecting observation using the ‘pandas’ 3 Band green531 (G531) 531 ± 7 nm ​
package in python. There were 18 to 20 independent harvest points in 4 Band green560 (G560) 560 ± 13.5 nm ​
5 Band red650 (R650) 650 ± 8 nm
each varietal line, and about 120 to 130 points per planting time block ​
6 Band red668 (R668) 668 ± 7 nm ​
(Fig. 2). The classified data resulted in a varying number of observations 7 Band red edge705 (RE705) 705 ± 5 nm ​
across growth phases, with ‘Mid-Vegetative’ having 2829 observations, 8 Band red edge717 (RE717) 717 ± 6 nm ​
‘Late Vegetative’ with 1915 observations, ‘Flowering’ phase with 2616 9 Band rededge740 (RE740) 740 ± 9 nm ​
observations, ‘Mid-Reproductive’ phase with 1700 observations, and 10 Band Near Infrared 842 842 ± 28.5 nm ​
(NIR842)
‘Late Reproductive’ phase with 2739 observations. The number of data 11 VI Normalized Difference (NIR842 - R668)/ Rouse et al.
points used were standardized to 1700 for each growth phase by Vegetation Index (NIR842 + R668) (1974)
randomly selecting observation. The selected data points for each crop (NDVI)
growth phase were further divided randomly in a 70:30 ratio as a cali­ 12 VI Normalized Difference (NIR842 - RE 717)/ Gitelson and
Red Edge Index (NIR842 + RE717) Merzlyak
bration dataset and external validation dataset (Fig. 3). The calibration
(NDRE) (1994)
dataset was used to train, calibrate, and cross-validate the ML models 13 VI Green Normalized (NIR842 - G560)/ Gitelson et al.
with 10 partitions of the data (k-folds). Difference Vegetation (NIR842 + G560) (1996)
Three types of datasets were used for ML regression modeling: (1) Index (GNDVI)
only reflectance bands, (2) only VIs, and (3) a combination of bands and 14 VI Normalized Green Red (G560 - R668)/(G560 Tucker (1979)
Difference Index + R668)
VIs as predictors. Each dataset was tested for multicollinearity among
(NGRDI)
the factors through pairwise comparisons using the ‘pandas’ package. A 15 VI Enhanced Vegetation (2.5 * (NIR842 - Huete et al.
subset of factors within a correlation threshold of -0.7 to 0.7 was created Index (EVI) R668))/((NIR842 + (2002)
to reduce multicollinearity, lowering computational complexity, and (6* R668) - (7.5 *
B475) + 1))
minimizing the risk of overfitting in the regression models (Gzar et al.,
16 VI Soil Adjusted (NIR842 - R668)/ Huete (1988)
2022). Removing high collinearity also helped reduce its adverse effect Vegetation Index ((NIR842 + R668 +
on coefficient estimates, variance in regression models, and the ranking (SAVI) 0.5) * (1 + L))
of the important features (Lindner et al., 2022). In this study, the 17 VI Modified Soil Adjusted (0.5 * (2 * (NIR842 + Qi et al. (1994)
response variable for all tested models was maize yield (Mg ha− 1), and Vegetation Index 1) - sqrt ((2 * NIR842
(MSAVI) + 1)2–8 * (NIR842 -
10 spectral bands and 11 commonly used VIs (Table 2) related to plant
R668))))
chlorophyll content, leaf area index (LAI), and yield were employed as 18 VI Triangular Vegetation (0.5* (120 * (NIR842 - Broge and
predictors (Barzin et al., 2020; Laliberte et al., 2011; Liu et al., 2012; Index (TVI) G650) - 200 * (R668 - Mortensen
Wang et al., 2019; Broge and Leblanc, 2001; Gitelson and Merzlyak, G 650) (2002)
19 VI Red Edge Triangulated (100 * (NIR842 - Haboudane
1994; Huete, 1988).
Vegetation Index RE717) – 10 * (NIR842 et al. (2004)
(RTVIcore) - G 650)
2.6. Machine learning algorithms 20 VI Visible (G560 - R668)/(G560 Gitelson et al.
Atmospherically + R668 – B475) (2002c)
Five types of ML regressors: Random Forest (RF), Extra Trees Re­ Resistant Index (VARI)
21 VI Plant Senescence (R668 - G650)/RE717 Merzlyak et al.
gressor (ETR), K-Nearest Neighbors (KNN), Support Vector Regressor
Reflectance Index (1999)
(SVR), and Linear Regression (LR) were evaluated using the python (PSRI)
package ‘sklearn’ (Pedregosa et al., 2011). RF is a non-parametric, su­
pervised ensemble learning method that combines prediction from
multiple decision trees. Each tree is built on a random subset of features 2.7. Model calibration, cross-validation, and external validation
and data points, which reduce variance and improve overall prediction
accuracy (Yuan et al., 2023; Breiman, 2001). Like RF, ETR also build Using the calibration dataset of each growth phase separately, sub­
ensemble decision trees but differ by randomly splitting features at each sets of factors within the collinearity threshold and containing the
node instead of using the best split, which reduces correlation between highest number of factors were identified. For example, if there were
trees and enhance performance in certain cases (Geurts et al., 2006). subsets with four factors each, while the other subsets contained fewer
KNN is a non-parametric, supervised learning classifier that predicts than four factors, the subsets with four factors were chosen. These
based on proximity of closest data points, using a distance metric to selected subsets were then used as the calibration factors for each ML
average their target values (Cover and Hart, 1967). SVR is a supervised regressors with the default settings in the ‘sklearn’ package (Table 3).
learning model that finds the hyperplane in high-dimensional space to For each ML regressor, the subset with the highest coefficient of
maximize the margin between the data points and the plane for good determination (R2) and the lowest root mean squared error (RMSE) for
generalization (Vapnik et al., 1997). SVR can handle linear as well as calibration and cross-validation was selected as the best-performing
non-linear relationships using kernel functions. Lastly, LR is a commonly subset. Hyperparameter tunning was then performed using the ‘Ran­
used parametric, supervised method that models the relationship be­ domizedSearchCV’ option with 20,000 fits to identify the best possible
tween features and a continuous target variable using a linear equation. combination of hyperparameters for each ML model. Randomized
It is interpretable, efficient to train, and a good baseline model. How­ search option was preferred over grid search to reduce the computation
ever, LR assumes a linear relationship between features and the target time, as it tests a random subset of parameter combinations rather than
variable, which often times limits its effectiveness in capturing the all possible combinations. The essential hyperparameters tuned for each
complex interactions that exist in real-world data. model are listed in Table 3.
After tuning the hyperparameters for each ML regressor model with
the selected subset of factors, the model was executed for calibration and
10-fold cross-validation using 70 % data. External validation dataset

5
B.R. Sapkota et al. Science of Remote Sensing 12 (2025) 100267

Table 3
Hyperparameters used for tuning the machine learning (ML)models with the
‘sklearn’ package in Python.
ML Hyperparameters tuning range Default
Regressor values

1. RF and ‘n_estimators’: range (100, 700) 100

ETR
​ ‘max_depth’: none + range (1,10) none
​ ’min_samples_split’: range (2, 10) 2
​ ’min_samples_leaf’: range (1, 5) 1
​ ’max_features’: [sqrt, log2, None] sqrt
2. KNN ‘n_neighbours’: range (1,30) 5
​ ‘weights’: [ uniform, distance] uniform
​ ‘leaf_size’: range (15, 45) 30
​ ‘algorithm’: auto auto
​ ‘p’: USDA-NASS, 2024; [1,2], where 1 = Manhattan 2
distance,
2 = Euclidean distance
3. SVR ’kernel’: [linear, poly, rbf, sigmoid] rbf
​ Regularization parameter 1
‘c’: [0.1, 1, 2, 3, 4, 5, 10, 20, 30, 40, 50]
​ ’epsilon’: [0.001,0.01,0.1] 0.1
​ ‘gamma’: [scale, auto] scale

(remaining 30 %) was used to assess the prediction performance of the

model. Along with the external validation, the most important factors
for each ML model were ranked using the ‘permutation importance’
method in ‘sklearn’ package, based on 200 permutations of the cross-
validation dataset. Permutation feature importance measures a
model’s dependence on a feature by shuffling its values and observing
the decrease in model performance (Altmann et al., 2010). All the fac­
tors in the selected subset were arranged from most important to least
important based on this approach.
Image processing, data extraction, and ML operations were all car­
ried out on a Windows 11 (Microsoft Corporation) computing system
equipped with an AMD Ryzen Threadripper Pro 5995WX 64-cores Fig. 4. Daily minimum temperature (T_min), maximum temperature (T_max),
processor and 512 GB memory. and precipitation recorded during the maize growing seasons in 2022 and 2023
at Temple, Texas.
2.8. Performance evaluation
3.2. Maize yield
The model performance matrices, including coefficient of determi­
nation (R2) and root mean squared error (RMSE), were noted for cali­ In Field 1, the mean maize yield was 2.11 Mg ha− 1, with the highest
bration, cross-validation and external validation to facilitate comparison yield of 5.85 Mg ha− 1 observed from the crop planted on March 10,
among the models. The prediction accuracy of the calibrated models was 2022, and the lowest yield of 0.33 Mg ha− 1 observed from the late
compared using cross-validation (R2CV) and external validation (R2V) planted crop on April 18, 2022 (Table 4). In Field 2, the mean maize
metrics. Models with a higher R2 were considered more effective than yield was 7.70 Mg ha− 1, with the maximum yield of 13.74 Mg ha− 1
others. In the conditions where R2 was equal among models, the model observed from the early-planted crop on February 28, 2023, and the
with the lower RMSE was deemed superior in performance. minimum yield of 1.33 Mg ha− 1 from the crop planted late on April 18,
2023 (Table 4). Similarly, in Field 3, the mean maize yield was 7.39 Mg
3. Results ha− 1, with the maximum yield of 14.04 Mg ha− 1 observed from maize
planted on March 7, 2023, and the minimum yield of 1.02 Mg ha− 1 from
3.1. Weather the late planted crop on April 18, 2023 (Table 4).

During the 2022 growing season, the spring months (March–May) 3.3. Yield prediction accuracy using reflectance band values
experienced relatively low precipitation, totaling 160 mm, with an
average maximum temperature of 27.6 ◦ C and an average minimum In the mid-vegetative phase, the ETR performed better (R2CV = 0.74
temperature of 15.0 ◦ C (Fig. 4). This was followed by a pronounced dry and R2V = 0.71) than other models in predicting yield, followed by RF
spell from June to August, during which there was only 19 mm of pre­ (R2CV = 0.71 and R2V = 0.69) (Table 5; Fig. 5). The worst performer among
cipitation, along with an average maximum temperature of 37.5 ◦ C and the tested models was found to be LR with a R2CV of 0.28 and R2V of 0.3.
an average minimum temperature of 24.0 ◦ C. In contrast, the 2023 NIR 842 was the most important factor for yield prediction in ETR
growing season began with a wet spring, receiving 326 mm of precipi­ model, with B475 and RE717 also being important factors across all
tation. The average maximum and minimum temperatures were 25.6 ◦ C regressors. In the late vegetative phase, ET again demonstrated superior
and 14.9 ◦ C, respectively. However, the summer months in 2023 (June performance (R2CV = 0.84 and R2V = 0.85), followed closely by RF (R2CV =
to August) remained dry, receiving only 32.45 mm of precipitation, with 0.83 and R2V = 0.83). In this growth phase, B475 emerged as the most
an average maximum temperature of 37.8 ◦ C and average minimum important factor in ETR model, with NIR842 and RE717 also playing
temperature of 24.5 ◦ C (Fig. 4). significant roles across all models.
In the flowering phase, the ETR outperformed (R2CV = 0.83 and R2V =
0.83) other models. RF was the second-best performing model (R2CV =

6
B.R. Sapkota et al. Science of Remote Sensing 12 (2025) 100267

Table 4
Maize yield statistics for three experimental fields during the 2022 and 2023 growing seasons.
Year (Field) Maximum yield (Mg ha− 1) Minimum Yield (Mg ha− 1) Average Yield (Mg ha− 1) Std Dev of Yield (Mg ha− 1)

2022 (Field 1) 5.85 0.33 2.11 1.06 ​

2023 (Field 2) 13.74 1.33 7.7 2.89 ​
2023 (Field 3) 14.04 1.02 7.39 2.31 ​

Combined 14.04 0.33 6.84 3.09 ​

0.83 and R2V = 0.82), followed by KNN, SVR and LR (Table 5). The most again the worst performer (R2CV = 0.24 and R2V = 0.21) among tested
important factor in the ETR model during the flowering phase was R560, models to predict maize yield. In this stage, TVI was the most important
with NIR842 and RE717 also serving as important factors in all the re­ factor in the ETR, while NGRDI and PSRI were consistently serving as
gressor models. Similarly, in the mid-reproductive phase, the best per­ important factors in other regressors.
forming model was ET (R2CV = 0.81 and R2V = 0.83), followed by RF and In the flowering phase, the ETR achieved the highest yield prediction
KNN. In mid-reproductive phase, G560 was the most critical factor accuracy (R2CV = 0.66 and R2V = 0.60), followed closely by RF regressor
across all regressors, and RE740 was second to repeat in most of the (R2CV = 0.66 and R2V = 0.59). LR remained the least accurate (R2CV = 0.37
models. During the late reproductive phase, the ETR and KNN regressor and R2V = 0.42) among the tested models (Table 6). Like late vegetative
both achieved superior results with R2CV and R2V of 0.82, outperforming phase, TVI was also identified as the most important factor in the ETR at
RF, SVR, and LR. The RE717 band was the most important feature in the the flowering phase, with PSRI also being another important factor
best performing ETR model, with NIR842 also consistently important across most models.
across most regressors. In the mid-reproductive phase, both ETR and RF regressors per­
formed best (R2CV = 0.73 and R2V = 0.74–0.75) in predicting maize yield.
Among VIs, PSRI was the most influential factor in the best performing
3.4. Yield prediction accuracy using commonly used vegetation indices RF regressor as well as in other high-performance models. ETR was the
best performing regressor model (R2CV = 0.82 and R2V = 0.83) to predict
During the mid-vegetative phase, yield prediction accuracy of KNN maize yield during the late reproductive growth phase, followed closely
regressor was the highest (R2CV = 0.49 and R2V = 0.53), followed closely by RF (R2CV = 0.82 and R2V = 0.82) and KNN (R2CV = 0.80 and R2V = 0.82).
by the RF (R2CV = 0.46 and R2V = 0.52), and ETR (R2CV = 0.48 and R2V = PSRI was the most important factor in the best performing ETR model.
0.50) (Table 6; Fig. 5). Whereas SVR and LR showed relatively lower
prediction accuracy, with LR found to be the poorest performer (R2CV =
0.26 and R2V = 0.29). PSRI was the most important VI in best performing 3.5. Yield prediction accuracy using combination of band values and
KNN regressor at mid-vegetative phase, with GNDVI and NGRDI also vegetation indices
being important factors in most regressors. In the late vegetative phase,
ETR performed best with an R2CV of 0.77 and R2V of 0.81, and LR was When using hybrid models combining reflectance bands and VIs,

Table 5
Calibration, cross-validation, and external validation statistics obtained for maize yield prediction using reflectance band values across different growth phases, via five
machine learning algorithms.
Growth Phase Method Calibration (n = 1190) Cross-Validation (n = 1190) External Validation (n = 510) Important Factors (Higher to lower)

R2C RMSEC R2CV RMSECV R2V RMSEV

Mid-Vegetative (V5 - V8) RF 0.96 0.60 0.71 1.60 0.69 1.62 B475, NIR842, RE717
ETR 0.94 0.72 0.74 1.51 0.71 1.56 NIR842, B475, RE717
KNN 1.00 0.00 0.64 1.78 0.63 1.78 NIR842, B475, RE717
SVR 0.42 2.27 0.41 2.28 0.44 2.19 G560, RE740, RE717
LR 0.29 2.51 0.28 2.52 0.30 2.45 RE717, G560, RE717
Late Vegetative (V9 - V13) RF 0.92 0.91 0.83 1.28 0.83 1.31 B475, NIR842, RE717
ETR 0.98 0.4 0.84 1.25 0.85 1.26 B475, NIR842, RE717
KNN 1.00 0.00 0.76 1.54 0.79 1.48 NIR842, RE717, B475
SVR 0.61 1.97 0.61 1.97 0.62 1.98 NIR842, R650, RE717
LR 0.56 2.09 0.56 2.09 0.56 2.13 R560, NIR842, RE717
Flowering (VT - R1) RF 0.95 0.66 0.83 1.28 0.82 1.29 R668, NIR842, RE717
ETR 0.96 0.62 0.83 1.26 0.83 1.26 R560, NIR842, RE717
KNN 1.00 0.00 0.81 1.35 0.80 1.37 NIR842, RE717, R668
SVR 0.76 1.54 0.75 1.55 0.75 1.53 R668, NIR842, RE717
LR 0.55 2.08 0.54 2.10 0.59 1.94 R668, RE717, NIR842
Mid-Reproductive (R2 - R4) RF 0.88 1.11 0.80 1.42 0.83 1.29 G560, RE740
ETR 0.89 1.05 0.81 1.39 0.83 1.31 G560, RE740
KNN 1.00 0.00 0.79 1.45 0.81 1.37 G560, RE740
SVR 0.77 1.53 0.76 1.53 0.78 1.47 G560, RE740
LR 0.69 1.75 0.69 1.76 0.71 1.68 G560, NIR842
Late RF 0.90 1.05 0.82 1.38 0.81 1.46 RE717, NIR842
Reproductive (R5 - R6) ETR 0.90 1.00 0.82 1.36 0.82 1.42 RE717, NIR842
KNN 1.00 0.00 0.82 1.37 0.82 1.43 RE717, NIR842
SVR 0.77 1.57 0.76 1.57 0.78 1.58 G531, NIR842
LR 0.70 1.76 0.70 1.77 0.71 1.79 B444, RE740

V5 to V13 = Five leaf to thirteen leaf stage, VT = Tasseling stage, R1 = Silking stage, R2 = Blister stage, R4 = Dough stage, R5 = Hard Dough stage, R6 = Maturity
stage; RF = Random forest, ETR = Extra Trees Regressor, KNN = K-Nearest Neighbors, SVR = Support Vector Regressor, LR = Linear Regression; B = blue, G = Green, R
= Red, RE = Red Edge, NIR = Near infrared followed by the wavelength; R2C, R2CV, R2V = Coefficient of determination for calibration, cross validation and external
validation respectively, RMSEC, RMSECV, RMSEV = Root mean squared error for calibration, cross validation and external validation respectively.

7
B.R. Sapkota et al. Science of Remote Sensing 12 (2025) 100267

Fig. 5. Actual vs predicted maize yield results using reflectance values of bands (a, d, g, j, m), vegetative indices (b, e, h, k, n), and combination of bands and
vegetation indices (c, f, i, l, o) across different maize growth phases (Mid- Vegetative, Late Vegetative, Flowering, Mid- Reproductive, and Late Reproductive) using
the Extra Trees Regressor model. R2V = Coefficient of determination for external validation, RMSEV = Root Mean squared error for external validation.

8
B.R. Sapkota et al. Science of Remote Sensing 12 (2025) 100267

Table 6
Calibration, cross-validation, and external validation statistics obtained for maize yield using vegetation indices across different growth phases, via five machine
learning algorithms.
Growth Phase Method Calibration (n = 1190) Cross-Validation (n = 1190) External Validation (n = 510) Important Factors (Higher to lower)

R2C RMSEC R2CV RMSECV R2V RMSEV

Mid-Vegetative (V5 - V8) RF 0.82 1.28 0.46 2.18 0.52 2.04 PSRI, NGRDI
ETR 0.87 1.06 0.48 2.14 0.50 2.07 PSRI, GNDVI
KNN 1.00 0.00 0.49 2.11 0.53 2.00 PSRI, GNDVI
SVR 0.36 2.39 0.34 2.41 0.38 2.31 PSRI, NGRDI
LR 0.28 2.54 0.26 2.55 0.29 2.46 PSRI, NDVI
Late Vegetative (V9 - V13) RF 0.90 1.00 0.77 1.51 0.79 1.46 NGRDI, TVI, PSRI
ETR 1.00 0.00 0.77 1.49 0.81 1.38 TVI, NGRDI, PSRI
KNN 1.00 0.00 0.72 1.65 0.78 1.52 EVI, NGRDI, PSRI
SVR 0.59 2.03 0.57 2.05 0.61 2.01 NGRDI, PSRI, EVI
LR 0.25 2.74 0.24 2.74 0.21 2.84 TVI, NGRDI, PSRI
Flowering (VT - R1) RF 0.74 1.59 0.66 1.80 0.59 1.95 TVI, PSRI
ETR 0.72 1.63 0.66 1.80 0.60 1.93 TVI, PSRI
KNN 1.00 0.00 0.62 1.91 0.55 2.04 SAVI, PSRI
SVR 0.60 1.95 0.57 2.00 0.50 2.14 TVI, PSRI
LR 0.38 2.44 0.37 2.44 0.42 2.31 GNDVI, PSRI
Mid-Reproductive (R2 - R4) RF 0.80 1.40 0.73 1.64 0.75 1.58 PSRI, TVI
ETR 0.83 1.32 0.73 1.63 0.74 1.62 TVI, PSRI
KNN 1.00 0.00 0.71 1.68 0.74 1.59 VARI, PSRI
SVR 0.67 1.81 0.67 1.82 0.67 1.80 PSRI, VARI
LR 0.66 1.85 0.65 1.85 0.68 1.80 GNDVI, PSRI
Late RF 0.90 1.05 0.82 1.37 0.82 1.41 TVI, PSRI
Reproductive (R5 - R6) ETR 0.87 1.14 0.82 1.37 0.83 1.39 PSRI, TVI
KNN 1.00 0.00 0.80 1.42 0.82 1.43 VARI, PSRI
SVR 0.78 1.53 0.77 1.54 0.81 1.47 NDRE, PSRI
LR 0.73 1.68 0.73 1.68 0.76 1.63 GNDVI, PSRI

V5 to V13 = Five leaf to thirteen leaf stage, VT = Tasseling stage, R1 = Silking stage, R2 = Blister stage, R4 = Dough stage, R5 = Hard Dough stage, R6 = Maturity
stage; RF = Random forest, ET = Extra Trees Regressor, KNN = K-Nearest Neighbors, SVR = Support Vector Regressor, LR = Linear Regression; PSRI= Plant
Senescence Reflectance Index, NGRDI = Normalized Green Red Difference Index, GNDVI = Green Normalized Difference Vegetation Index, NDVI= Normalized
Difference Vegetation Index, TVI = Triangular Vegetation Index, EVI = Enhanced Vegetation Index, SAVI = Soil Adjusted Vegetation Index, VARI = Visible Atmo­
spherically Resistant Index; R2C, R2CV, R2V = Coefficient of determination for calibration, cross validation and external validation respectively, RMSEC, RMSECV, RMSEV =
Root Mean squared error for calibration, cross validation and external validation, respectively.

ETR was the best performer in predicting maize yield at mid-vegetative on VI-based models, often exhibit inconsistencies due to variation in soil
phase (R2CV = 0.76 and R2V = 0.77), and LR was the worst performer (R2CV and environmental factors, particularly in rainfed environments.
= 0.29 and R2V = 0.34) (Table 7; Fig. 5). G560 was the most important This study evaluated the performance of various yield prediction
factor for the best performing ETR model, with NGRDI and NIR842 also approaches (reflectance-based, VI-based, and combination) across three
being other important factors across the models. Likewise, in the late different rainfed field locations over two years. Additionally, the
vegetative phase, ET performed best with an R2CV of 0.86 and R2V of 0.89, research assessed yield prediction models for maize planted at seven
while SVR was the poorest performer (R2CV = 0.60 and R2V = 0.62) at late different planting dates in these locations, accounting for the substantial
vegetative phase. B475 was the most important factor for the best per­ environmental variability experienced by crops. To our knowledge, no
forming ETR model, with NIR842, PSRI and NGRDI also serving as previous studies have evaluated the robustness of yield prediction
important factors in most models. models for maize across such a diverse set of growing conditions.
Similarly, ETR outperformed (R2CV = 0.85 and R2V = 0.84) all other Many prior studies on yield prediction have focused on using single
models in predicting maize yield in the flowering phase (Table 7). Most or multiple VIs to examine their relationships with crop yield (Shrestha
important factor in best performing ETR model was G560, with NIR842, et al., 2023; Ji et al., 2022; Li et al., 2019). During the early vegetative
GNDVI and PSRI also being significant factors for most models. In the stages, the relatively sparser canopy allows more light to penetrate and
mid- and late-reproductive phases as well, ETR performed best with an reflect, leading to more pronounced variations in spectral reflectance,
R2CV of 0.84–0.86 and R2V of 0.86. G560 remained the most important making soil-adjusted VIs like MSAVI sensitive during that phase
factor in the best performing ETR model, and NDRE and RE740 were (Buthelezi et al., 2023; Yan et al., 2023). However, as the canopy density
other important predictors in the flowering phase. Whereas PSRI was the increases, VIs like NDVI often reach a saturation point, reducing their
most important predictor in the best performing ETR at the late- effectiveness for yield estimation in later vegetative and reproductive
reproductive phase, with NDRE and RE740 being important factors in stages, even when advanced statistical methods such as ML are applied,
other high-performance models. as observed in this study.
Among the crop growth phases tested in this study, mid-vegetative
4. Discussion growth (V5-V8) is considered particularly important for early in-
season yield predictions and the development of timely crop manage­
Timely and accurate in-season crop yield predictions are critical for ment strategies. Reflectance-based models outperformed VI-based
enhancing the efficiency, profitability, and sustainability of agricultural models and generated more accurate yield predictions, thereby
productions systems. Early yield predictions enable farmers to establish demonstrating the potential of this alternative approach for achieving
management zones and make informed decisions regarding resource reliable early in-season maize yield predictions. The high importance of
allocation and management strategies during the crop growing season. the NIR842 band suggests a critical role of NIR region in assessing
Utilizing spectral information from UAS imagery can help achieve in- biomass and overall plant health at the mid-vegetative growth group.
season yield predictions and support crop management decisions. NIR bands are known to be sensitive to chlorophyll content and leaf area
However, traditional UAS-based yield models, particularly those relying index, which are indicative of vigorous vegetative growth (Xie et al.,

9
B.R. Sapkota et al. Science of Remote Sensing 12 (2025) 100267

Table 7
Calibration, cross-validation, and external validation statistics obtained for maize yield using combination of band values and vegetation indices across different
growth phases, via five machine learning algorithms.
Growth phase Method Calibration (n = 1190) Cross-Validation (n = 1190) External Important Factors (Higher to lower)
Validation (n = 510)

R2C RMSEC R2CV RMSECV R2V RMSEV

Mid-Vegetative (V5 - V8) RF 0.95 0.66 0.73 1.54 0.75 1.46 G560, NGRDI, NIR842, RE717
ETR 1.00 0.00 0.76 1.43 0.77 1.41 G560, NGRDI, NIR842, RE717
KNN 1.00 0.00 0.69 1.63 0.76 1.44 NGRDI, PSRI, NIR842, G560
SVR 0.49 2.15 0.47 2.17 0.51 2.05 NGRDI, PSRI, NIR842, G560
LR 0.31 2.49 0.29 2.49 0.34 2.38 PSRI, G560, VARI, NIR842
Late Vegetative (V9 - V13) RF 0.93 0.82 0.85 1.21 0.88 1.10 B475, NIR842, PSRI, RE717, NGRDI, NDVI
ETR 0.99 0.38 0.86 1.16 0.89 1.05 B475, NIR842, PSRI, RE717, NGRDI, NDVI
KNN 1.00 0.00 0.83 1.28 0.86 1.21 NIR842, NGRDI, PSRI, NDVI, RE717, B444
SVR 0.61 1.97 0.6 1.99 0.62 1.98 NIR842, NDVI, PSRI, NGRDI, RE717, R560
LR 0.62 1.94 0.62 1.95 0.63 1.95 R668, NGRDI, G560, RTVIcore, PSRI, RE740
Flowering (VT - R1) RF 0.96 0.60 0.84 1.23 0.83 1.23 G560, NIR842, GNDVI, PSRI
ETR 0.96 0.62 0.85 1.20 0.84 1.21 G560, NIR842, GNDVI, PSRI
KNN 1.00 0.00 0.84 1.22 0.83 1.25 NIR842, GNDVI, PSRI, RE717
SVR 0.74 1.58 0.73 1.61 0.75 1.51 R668, RE740, NGRDI, RE717
LR 0.63 1.88 0.62 1.89 0.67 1.73 NGRDI, G560, PSRI, RE740
Mid-Reproductive (R2 - R4) RF 0.96 0.66 0.83 1.30 0.85 1.23 G560, NDRE, RE740
ETR 0.96 0.65 0.84 1.25 0.86 1.18 G560, NDRE, RE740
KNN 1.00 0.00 0.79 1.43 0.82 1.35 GNDVI, RE740, RE717
SVR 0.76 1.55 0.76 1.56 0.77 1.50 G560, NDRE, RE740
LR 0.69 1.76 0.69 1.77 0.71 1.69 G560, NDRE, RE740
Late RF 0.94 0.79 0.85 1.23 0.84 1.33 NDRE, B444, RE740
Reproductive (R5 - R6) ETR 0.96 0.66 0.86 1.19 0.86 1.26 PSRI, MSAVI, RE740
KNN 1.00 0.00 0.86 1.21 0.85 1.28 NDRE, PSRI, RE740
SVR 0.82 1.38 0.81 1.40 0.83 1.37 GNDVI, R650, RE740
LR 0.74 1.64 0.74 1.64 0.77 1.59 GNDVI, PSRI

V5 to V13 = Five leaf to thirteen leaf stage, VT = Tasseling stage, R1 = Silking stage, R2 = Blister stage, R4 = Dough stage, R5 = Hard Dough stage, R6 = Maturity
stage; RF = Random forest, ET = Extra Trees Regressor, KNN = K-Nearest Neighbors, SVR = Support Vector Regressor, LR = Linear Regression; PSRI= Plant
Senescence Reflectance Index, NGRDI = Normalized Green Red Difference Index, GNDVI = Green Normalized Difference Vegetation Index, NDVI= Normalized
Difference Vegetation Index, MSAVI = Modified Soil Adjusted Vegetation Index, VARI = Visible Atmospherically Resistant Index, RTVIcore = Red Edge Triangulated
Vegetation Index, B = blue, G = Green, R = Red, RE = Red Edge, NIR = Near infrared followed by the wavelength; R2C, R2CV, R2V = Coefficient of determination for
calibration, cross validation and external validation respectively, RMSEC, RMSECV, RMSEV = Root Mean squared error for calibration, cross validation and external
validation respectively.

2018). activity as the crop progresses through its reproductive phase. Red edge
During the late vegetative growth group, the improvement in yield (RE740) also appeared frequently during both mid- and late-
prediction accuracy could be attributed to the increased differentiation reproductive phases, suggesting its utility in detecting decreased chlo­
in plant health and development, which becomes more pronounced at rophyll and protein content in leaves due to nutrient relocation to seed
this stage, making spectral differences more detectable. The dominance development during this stage (Hirel et al., 2007). As cited in previous
of the blue band (B475) highlights its effectiveness in detecting changes studies, PSRI was found to be a key VI across reproductive phases,
in canopy growth when the crop canopy rapidly expands. Blue light is effectively capturing the senescence of the maize crop (Hatfield and
also associated with trends of higher biomass production and photo­ Prueger, 2010).
synthetic capacity (Hogewoning et al., 2010). The consistent importance This study aligns with the previous findings that yield prediction
of NIR (NIR842) and red edge (RE717) bands further shows their role in accuracy can be weaker at mid-vegetative stage due to lower canopy
capturing vital information related to crop vigor during this critical cover compared to the late vegetative and flowering stages, where plants
growth phase. While VIs like NDVI often suffer from the saturation effect actively grow and achieve greater canopy coverage (Shrestha et al.,
when the canopy is vigorous, TVI, NGRDI and PSRI were found less 2023). Chlorophyll content is a vital pigment for photosynthesis in
affected, making them important for yield prediction at late growth plants and is well related to yield (Brewer et al., 2022). The NIR and red
stages of maize (Gao et al., 2023; Haboudane, 2004). However, it should edge bands are found to be more sensitive to the chlorophyll content at
be noted that TVI may be less reliable when the canopy cover is sparse, different phenological stages (Xie et al., 2018). In this study, these two
as its reflectance can be influenced by background soil (Prabhakara regions were also the most dominant in prediction models while using
et al., 2015). the reflectance band values only. Among VIs, PSRI, which incorporates
At the flowering stages, the red (R560), NIR (NIR842) and red edge the red edge band, and NGRDI, which includes green and red bands,
(RE717) bands were the most important spectral bands across all were the most influential in the prediction models while using VIs only.
models. Likewise, VIs such as TVI, GNDVI and PSRI employing red, NIR Additionally, when combining both spectral bands and VIs, these same
and red edge bands, were found consistently among important pre­ bands and indices remained dominant, leading to improved prediction
dictors. These findings highlight the persistent utility of electromagnetic accuracy across all models.
spectrum between 650 and 850 nm in monitoring crop growth for ac­ The findings of this study support that ML significantly facilitates the
curate yield prediction during the flowering stage which is characterized ability to handle complex, large-scale remotely sensed data for intricate
by maximum canopy coverage and peak chlorophyll content (Shrestha predictions such as crop yield. Traditional linear models, such as LR, are
et al., 2023). inadequate for addressing the high level of variability inherent in remote
For the mid-reproductive group, green band (G560) emerged as the sensing applications. ML algorithms, on the other hand, can efficiently
most important factor in the best-performing models, indicating its process vast datasets, uncovering patterns and insights that are chal­
relevance in capturing changes in plant health and photosynthetic lenging to detect using traditional methods. In this study, ETR was found

10
B.R. Sapkota et al. Science of Remote Sensing 12 (2025) 100267

to be performing consistently well among tested algorithms for all types Declaration of competing interest
of datasets, regardless of whether reflectance-based or VI-based ap­
proaches were used across various growth stages (Fig. 5). ETR, an The authors declare that they have no known competing financial
ensemble regression method based on decision trees, is particularly interests or personal relationships that could have appeared to influence
unique in randomizing tree selection to achieve the best accuracies and the work reported in this paper.
reduce the risk of overfitting (Geurts et al., 2006), which proved to be
highly effective with the diverse remotely sensed data used in this study. Acknowledgements
Overall, findings indicate that using reflectance information for in­
dividual spectral bands directly, rather than relying on indices, could Authors thank Texas A&M AgriLife Research for the financial sup­
enhance maize yield prediction across different growth stages and var­ port through start-up funds. This work was partially supported by the
ied planting timing. ML could facilitate the generation of more accurate USDA National Institute of Food and Agriculture (NIFA), Multi-state
and consistent in-season yield predictions using multispectral bands Hatch project NC1210. We also acknowledge the USDA-ARS Grassland
information directly in models, thereby eliminating the need for VIs and Soil and Water Research Laboratory for conducting and maintaining
the associated complexity with crop density and varying environments. field experiments, as an equal opportunity provider and employer.
Notably, the differences in R2v for best model from each of two data Special thanks to Dr. Seth Murray for reviewing the manuscript and
groups were quite significant, with reflectance-based models out­ offering valuable feedback.
performing VI-based models during mid- and late-vegetative growth
stages, as well as at flowering and mid-reproductive stages. Although
Appendix A. Supplementary data
hybrids models, combining reflectance bands and VIs, showed slight
improvements over reflectance-based models in most cases, it is
Supplementary data to this article can be found online at [Link]
important to note that the approach generally requires more extensive
org/10.1016/[Link].2025.100267.
data pre-processing and greater computational resources.
The study also highlights the potential for increased integration of
Data availability
hyperspectral sensors in future applications. Furthermore, the spectral
bands information was scaled to align with satellite sources such as
Data will be made available on request.
PlanetScope, Sentinel-2A and Landsat8 (Román et al., 2022), incorpo­
rating 10 bands and a sampling area of approximately 3 m × 3 m, which
suggests that these satellite platforms could be directly adopted for yield References
predictions in future.
Adhikari, K., Smith, D.R., Hajda, C., Kharel, T.P., 2023. Within-field yield stability and
gross margin variations across corn fields and implications for precision
5. Conclusion conservation. Precis. Agric. 24, 1401–1416. [Link]
09995-7.
Altmann, A., Toloşi, L., Sander, O., Lengauer, T., 2010. Permutation importance: a
This study demonstrated the effectiveness of using reflectance data corrected feature importance measure. Bioinformatics 26, 1340–1347. [Link]
directly with ML models for in-season maize yield predictions, as org/10.1093/bioinformatics/btq134.
opposed to relying solely on VIs, which often saturate in dense canopies Attri, I., Awasthi, L.K., Sharma, T.P., 2023. Machine learning in agriculture: a review of
crop management applications. Multimed. Tool. Appl. 83, 12875–12915. https://
or become inconsistent across varied environments. When evaluating [Link]/10.1007/s11042-023-16105-2.
growth stage-based ML models on data from diverse maize-growing Barzin, R., Pathak, R., Lotfi, H., Varco, J., Bora, G.C., 2020. Use of UAS multispectral
environments, the reflectance approach outperformed the VI-based imagery at different physiological stages for yield prediction and input resource
optimization in corn. Remote Sens. 12, 2392. [Link]
approach in yield prediction accuracy, particularly during the mid-
Becker-Reshef, I., Barker, B., Whitcraft, A., Oliva, P., Mobley, K., Justice, C., Sahajpal, R.,
and late-vegetative growth stages, as well as during flowering and mid- 2023. Crop type maps for operational global agricultural monitoring. Sci. Data 10,
reproductive stages. Hybrids ML models that combined VIs and spectral 172. [Link]
Bell, J., 2017. Corn Growth Stages and Development. Amarillo.
bands achieved the highest predication accuracy but required extensive
Blumenthal, J.M., Thompson, W.H., 2009. Estimating corn grain yields (L-5515). http://
data pre-processing. NIR and red edge bands, along with their associated [Link]/CORN_SORGHUM/PUB_Estimating%20Corn%20Grain%
VIs like PSRI and TVI, were identified as important predictors for ML [Link]. (Accessed 7 May 2024).
models, emphasizing their critical role in capturing crop health and Brantley, S.T., Zinnert, J.C., Young, D.R., 2011. Application of hyperspectral vegetation
indices to detect variations in high leaf area index temperate shrub thicket canopies.
vigor across different maize growth phases. Remote Sens. Environ. 115, 514–523. [Link]
Furthermore, the study highlighted the utility of ML in processing Breiman, L., 2001. Random forests. Mach. Learn. 45, 5–32. [Link]
large datasets and identifying complex patterns, which is crucial for 1010933404324.
Brewer, K., Clulow, A., Sibanda, M., Gokool, S., Naiken, V., Mabhaudhi, T., 2022.
making precise yield predictions. Notably, the ETR consistently deliv­ Predicting the chlorophyll content of maize over phenotyping as a proxy for crop
ered the best performance across all growth stages, surpassing other health in smallholder farming systems. Remote Sens. 14, 518. [Link]
models in yield prediction accuracy, while the RF regressor also 10.3390/rs14030518.
Broge, N.H., Leblanc, E., 2001. Comparing prediction power and stability of broadband
demonstrated strong performance. The consistent performance of the and hyperspectral vegetation indices for estimation of green leaf area index and
ETR, with its ability to minimize overfitting and improve accuracy, canopy chlorophyll density. Remote Sens. Environ. 76, 156–172. [Link]
demonstrates the potential of advanced ML algorithms in enhancing 10.1016/S0034-4257(00)00197-8.
Broge, N.H., Mortensen, J.V., 2002. Deriving green crop area index and canopy
agricultural decision-making through better in-season yield predictions.
chlorophyll density of winter wheat from spectral reflectance data. Remote Sens.
Environ. 81, 45–57. [Link]
CRediT authorship contribution statement Buthelezi, S., Mutanga, O., Sibanda, M., Odindi, J., Clulow, A.D., Chimonyo, V.G.P.,
Mabhaudhi, T., 2023. Assessing the prospects of remote sensing maize leaf area
index using UAV-derived multi-spectral data in smallholder farms across the growing
Bala Ram Sapkota: Writing – original draft, Visualization, Meth­ season. Remote Sens. 15, 1597. [Link]
odology, Formal analysis, Data curation, Conceptualization. Gurjinder Camenzind, M.P., Yu, K., 2024. Multi temporal multispectral UAV remote sensing allows
S. Baath: Writing – review & editing, Supervision, Conceptualization. K. for yield assessment across European wheat varieties already before flowering.
Front. Plant Sci. 14. [Link]
Colton Flynn: Writing – review & editing. Kabindra Adhikari: Writing Cen, H., Wan, L., 2023. Crop yield estimation and prediction. In: Encyclopedia of Smart
– review & editing, Data curation. Chad Hajda: Writing – review & Agriculture Technologies. Springer International Publishing, Cham, pp. 1–13.
editing, Resources. Douglas R. Smith: Writing – review & editing, [Link]
Colomina, I., Molina, P., 2014. Unmanned aerial systems for photogrammetry and
Resources. remote sensing: a review. ISPRS J. Photogrammetry Remote Sens. 92, 79–97.
[Link]

11
B.R. Sapkota et al. Science of Remote Sensing 12 (2025) 100267

Cover, T., Hart, P., 1967. Nearest neighbor pattern classification. IEEE Trans. Inf. Theor. rangeland environments. Remote Sens. 3, 2529–2551. [Link]
13, 21–27. [Link] rs3112529.
Darra, N., Anastasiou, E., Kriezi, O., Lazarou, E., Kalivas, D., Fountas, S., 2023. Can yield Lee, C., Herbek, J., 2005. Estimating corn yields (AGR-187). [Link]
prediction be fully digitilized? A systematic review. Agronomy 13, 2441. [Link] edu/agcomm/pubs/agr/agr187/[Link]. (Accessed 7 May 2024).
org/10.3390/agronomy13092441. Li, Y., Guan, K., Yu, A., Peng, B., Zhao, L., Li, B., Peng, J., 2019. Toward building a
Gao, S., Zhong, R., Yan, K., Ma, X., Chen, X., Pu, J., Gao, S., Qi, J., Yin, G., Myneni, R.B., transparent statistical model for improving crop yield prediction: modeling rainfed
2023. Evaluating the saturation effect of vegetation indices in forests using 3D corn in the U.S. Field Crops Res. 234, 55–65. [Link]
radiative transfer simulations and satellite observations. Remote Sens. Environ. 295, fcr.2019.02.005.
113665. [Link] Lindner, T., Puck, J., Verbeke, A., 2022. Beyond addressing multicollinearity: robust
Geurts, P., Ernst, D., Wehenkel, L., 2006. Extremely randomized trees. Mach. Learn. 63, quantitative analysis and machine learning in international business research. J. Int.
3–42. [Link] Bus. Stud. 53, 1307–1314. [Link]
Gitelson, A., Merzlyak, M.N., 1994. Quantitative estimation of chlorophyll-a using Liu, J., Pattey, E., Jégo, G., 2012. Assessment of vegetation indices for regional crop
reflectance spectra: experiments with autumn chestnut and maple leaves. green LAI estimation from landsat images over multiple growing seasons. Remote
J. Photochem. Photobiol., B 22, 247–252. [Link] Sens. Environ. 123, 347–358. [Link]
06963-4. Lobell, D.B., Asner, G.P., 2003. Climate and management contributions to recent trends
Gitelson, A., Kaufman, Y.J., Merzlyak, M.N., 1996. Use of a green channel in remote in U.S. agricultural yields. Science 299. [Link]
sensing of global vegetation from EOS-MODIS. Remote Sens. Environ. 58, 289–298. 1032–1032.
[Link] Ma, Y., Zhang, Z., Kang, Y., Özdoğan, M., 2021. Corn yield prediction and uncertainty
Gitelson, A.A., Kaufman, Y.J., Stark, R., Rundquist, D., 2002a. Novel algorithms for analysis based on remotely sensed variables using a Bayesian neural network
remote estimation of vegetation fraction. Remote Sens. Environ. 80, 76–87. https:// approach. Remote Sens. Environ. 259, 112408. [Link]
[Link]/10.1016/S0034-4257(01)00289-9. rse.2021.112408.
Gitelson, A.A., Stark, R., Grits, U., Rundquist, D., Kaufman, Y., Derry, D., 2002b. Maresma, A., Chamberlain, L., Tagarakis, A., Kharel, T., Godwin, G., Czymmek, K.J.,
Vegetation and soil lines in visible spectral space: a concept and technique for Shields, E., Ketterings, Q.M., 2020. Accuracy of NDVI-Derived corn yield predictions
remote estimation of vegetation fraction. Int. J. Rem. Sens. 23, 2537–2562. https:// is impacted by time of sensing. Comput. Electron. Agric. 169, 105236. [Link]
[Link]/10.1080/01431160110107806. org/10.1016/[Link].2020.105236.
Gitelson, A.A., Kaufman, Y.J., Stark, R., Rundquist, D., 2002c. Novel algorithms for Merzlyak, M.N., Gitelson, A.A., Chivkunova, O.B., Rakitin, V.Y.U., 1999. Non-destructive
remote estimation of vegetation fraction. Remote Sens. Environ. 80, 76–87. https:// optical detection of pigment changes during leaf senescence and fruit ripening.
[Link]/10.1016/S0034-4257(01)00289-9. Physiol. Plantarum 106, 135–141. [Link]
Grabowski, R., Self, S., 2022. Role of agricultural productivity growth in economic 3054.1999.106119.x.
development: the neglected impact on institutional quality in Africa. Evol. Inst. Nielsen, R.L., 2021. Estimating Corn Grain Yield Prior to Harvest. Corny News Network.
Econ. Rev. 19, 681–700. [Link] [Link]
Gzar, D.A., Mahmood, A.M., Abbas, M.K., 2022. A comparative study of regression (Accessed 7 May 2024).
machine learning algorithms: tradeoff between accuracy and computational Nielsen, R.L., Thomison, P.R., Brown, G.A., Halter, A.L., Wells, J., Wuethrich, K.L., 2002.
complexity. Mathematical Modelling of Engineering Problems 9, 1217–1224. Delayed planting effects on flowering and grain maturation of dent corn. Agron. J.
[Link] 94, 549–558. [Link]
Haboudane, D., 2004. Hyperspectral vegetation indices and novel algorithms for Pedregosa, F., Varoquaux, G., Gramfort, A., Michel, V., Thirion, B., Grisel, O.,
predicting green LAI of crop canopies: modeling and validation in the context of Blondel, M., Prettenhofer, P., Weiss, R., Dubourg, V., Vanderplas, J., Passos, A.,
precision agriculture. Remote Sens. Environ. 90, 337–352. [Link] Cournapeau, D., Brucher, M., Perrot, M., Duchesnay, É., 2011. Scikit-learn: machine
[Link].2003.12.013. learning in python. J. Mach. Learn. Res. 12, 2825–2830. [Link]
Haboudane, D., Miller, J.R., Pattey, E., Zarco-Tejada, P.J., Strachan, I.B., 2004. 1953048.2078195.
Hyperspectral vegetation indices and novel algorithms for predicting green LAI of Poncet, A.M., Knappenberger, T., Brodbeck, C., Fogle, M., Shaw, J.N., Ortiz, B.V., 2019.
crop canopies: modeling and validation in the context of precision agriculture. Multispectral UAS data accuracy for different radiometric calibration methods.
Remote Sens. Environ. 90, 337–352. [Link] Remote Sens. 11, 1917. [Link]
Hassan, M.A., Yang, M., Rasheed, A., Yang, G., Reynolds, M., Xia, X., Xiao, Y., He, Z., Prabhakara, K., Hively, W.D., McCarty, G.W., 2015. Evaluating the relationship between
2019. A rapid monitoring of NDVI across the wheat growth cycle for grain yield biomass, percent groundcover and remote sensing indices across six winter cover
prediction using a multi-spectral UAV platform. Plant Sci. 282, 95–103. [Link] crop fields in Maryland, United States. Int. J. Appl. Earth Obs. Geoinf. 39, 88–102.
org/10.1016/[Link].2018.10.022. [Link]
Hatfield, J.L., Prueger, J.H., 2010. Value of using different vegetative indices to quantify Qi, J., Chehbouni, A., Huete, A.R., Kerr, Y.H., Sorooshian, S., 1994. A modified soil
agricultural crop characteristics at different growth stages under varying adjusted vegetation index. Remote Sens. Environ. 48, 119–126. [Link]
management practices. Remote Sens. 2, 562–578. [Link] 10.1016/0034-4257(94)90134-1.
rs2020562. Román, A., Navarro, G., Caballero, I., Tovar-Sánchez, A., 2022. High-spatial resolution
Hirel, B., Le Gouis, J., Ney, B., Gallais, A., 2007. The challenge of improving nitrogen use UAV multispectral data complementing satellite imagery to characterize a chinstrap
efficiency in crop plants: towards a more central role for genetic variability and penguin colony ecosystem on deception island (antarctica). GIsci Remote Sens 59,
quantitative genetics within integrated approaches. J. Exp. Bot. 58, 2369–2387. 1159–1176. [Link]
[Link] Rouse, J.W., Haas, R.H., Deering, D.W., Schell, J.A., Harlan, J.C., 1974. Monitoring the
Hogewoning, S.W., Trouwborst, G., Maljaars, H., Poorter, H., van Ieperen, W., vernal advancement and retrogradation (green wave effect) of natural vegetation.
Harbinson, J., 2010. Blue light dose-responses of leaf photosynthesis, morphology, Earth Resources and Remote Sensing. NASA-CR-144661. [Link]
and chemical composition of Cucumis sativus grown under different combinations of ons/19750020419.
red and blue light. J. Exp. Bot. 61, 3107–3117. [Link] Sah, R.P., Chakraborty, M., Prasad, K., Pandit, M., Tudu, V.K., Chakravarty, M.K.,
Huete, A.R., 1988. A soil-adjusted vegetation index (SAVI). Remote Sens. Environ. 25, Narayan, S.C., Rana, M., Moharana, D., 2020. Impact of water deficit stress in maize:
295–309. [Link] phenology and yield components. Sci. Rep. 10, 2944. [Link]
Huete, A., Didan, K., Miura, T., Rodriguez, E.P., Gao, X., Ferreira, L.G., 2002. Overview s41598-020-59689-7.
of the radiometric and biophysical performance of the MODIS vegetation indices. Sarker, I.H., 2021. Machine learning: Algorithms, real-world applications and research
Remote Sens. Environ. 83, 195–213. [Link] directions. SN Comput Sci 2, 160. [Link]
00096-2. Shrestha, A., Bheemanahalli, R., Adeli, A., Samiappan, S., Czarnecki, J.M.P.,
Huete, A., Didan, K., van Leeuwen, W., Miura, T., Glenn, E., 2010. MODIS vegetation McCraine, C.D., Reddy, K.R., Moorhead, R., 2023. Phenological stage and vegetation
indices. In: Ramachandran, B., Justice, C., Abrams, M. (Eds.), Land Remote Sensing index for predicting corn yield under rainfed environments. Front. Plant Sci. 14.
and Global Environmental Change. Remote Sensing and Digital Image Processing. [Link]
Springer, New York, NY, pp. 579–602. [Link] Sims, D.A., Gamon, J.A., 2002. Relationships between leaf pigment content and spectral
26. reflectance across a wide range of species, leaf structures and developmental stages.
Ji, Z., Pan, Y., Zhu, X., Zhang, D., Dai, J., 2022. Prediction of corn yield in the USA corn Remote Sens. Environ. 81, 337–354. [Link]
belt using satellite data and machine learning: from an evapotranspiration 00010-X.
perspective. Agriculture 12, 1263. [Link] Soil Survey Staff, Keys to Soil Taxonomy, 2010. National Resources Conservation
Johnson, D.M., Rosales, A., Mueller, R., Reynolds, C., Frantz, R., Anyamba, A., Pak, E., Service, eleventh ed. Department of Agriculture, Washington, DC, U.S.
Tucker, C., 2021. USA crop yield estimation with MODIS NDVI: are remotely sensed Sudduth, K.A., Drummond, S.T., 2007. Yield editor: software for removing errors from
models better than simple trend analyses? Remote Sens. 13, 4227. [Link] crop yield maps. Agron. J. 99, 1471–1482. [Link]
10.3390/rs13214227. agronj2006.0326.
Kamilaris, A., Prenafeta-Boldú, F.X., 2018. Deep learning in agriculture: a survey. Teal, R.K., Tubana, B., Girma, K., Freeman, K.W., Arnall, D.B., Walsh, O., Raun, W.R.,
Comput. Electron. Agric. 147, 70–90. [Link] 2006. In-season prediction of corn grain yield potential using normalized difference
compag.2018.02.016. vegetation index. Agron. J. 98, 1488–1494. [Link]
Kross, A., McNairn, H., Lapen, D., Sunohara, M., Champagne, C., 2015. Assessment of agronj2006.0103.
RapidEye vegetation indices for estimation of leaf area index and biomass in corn Tucker, C.J., 1979. Red and photographic infrared linear combinations for monitoring
and soybean crops. Int. J. Appl. Earth Obs. Geoinf. 34, 235–248. [Link] vegetation. Remote Sens. Environ. 8, 127–150. [Link]
10.1016/[Link].2014.08.002. (79)90013-0.
Laliberte, A.S., Goforth, M.A., Steele, C.M., Rango, A., 2011. Multispectral remote TWDB, TexMesonet, 2024. Texas water development board, Texas. [Link]
sensing from unmanned aircraft: image processing workflows and applications for [Link]/DataProducts/CustomDownloads. (Accessed 10 April 2024).

12
B.R. Sapkota et al. Science of Remote Sensing 12 (2025) 100267

USDA-ARS, 2024. Hydrologic data. [Link] Appl Earth Obs Remote Sens 11, 1482–1493. [Link]
-tx/grassland-soil-and-water-research-laboratory/docs/hydrologic-data. (Accessed JSTARS.2018.2813281.
10 April 2024). Yan, L., Liu, X., Jing, X., Geng, L., Che, T., Liu, L., 2023. Enhancing leaf area index
USDA-NASS, 2024. Quick Stats, Corn, Grain-production. [Link] estimation for maize with tower-based multi-angular spectral observations. Sensors
gov/. Accessed 10 April 2024. 23, 9121. [Link]
Vapnik, V.N., 1997. The support vector method. In: Gerstner, W., Germond, A., Yuan, Z., Sun, Y., Shasha, D., 2023. Forgetful forests: data structures for machine
Hasler, M., Nicoud, J.-D. (Eds.), Artificial Neural Networks — ICANN’97. Springer learning on streaming data under concept drift. Algorithms 16, 278. [Link]
Berlin Heidelberg, Berlin, Heidelberg, pp. 261–271. 10.3390/a16060278.
Wang, K., Huggins, D.R., Tao, H., 2019. Rapid mapping of winter wheat yield, protein, Zhang, J., Cheng, T., Guo, W., Xu, X., Qiao, H., Xie, Y., Ma, X., 2021. Leaf area index
and nitrogen uptake using remote and proximal sensing. Int. J. Appl. Earth Obs. estimation model for UAV image hyperspectral data based on wavelength variable
Geoinf. 82, 101921. [Link] selection and machine learning methods. Plant Methods 17, 49. [Link]
Xie, Q., Dash, J., Huang, W., Peng, D., Qin, Q., Mortimer, H., Casa, R., Pignatti, S., 10.1186/s13007-021-00750-5.
Laneve, G., Pascucci, S., Dong, Y., Ye, H., 2018. Vegetation indices combining the Zhu, W., Porth, L., Tan, K.S., 2019. A credibility-based yield forecasting model for crop
red and red-edge spectral information for leaf area index retrieval. IEEE J Sel Top reinsurance pricing and weather risk management. Agric. Finance Rev. 79, 2–26.
[Link]

13