融合多传感器多时相无人机表型与机器学习预测地中海藜麦籽粒产量

Academic Editors: Hengbiao Zheng
and Xia Yao
Received: 17 September 2026
Revised: 3 November 2026
Accepted: 4 November 2026
Published: 7 November 2026
Copyright: © 2026 by the authors.
Licensee MDPI, Basel, Switzerland.
This article is an open access article
distributed under the terms and
conditions of the Creative Commons
Attribution (CC BY) license.
Article
Integrating Multi-Sensor, Multi-Temporal UA V Phenotyping and
Machine Learning for Grain Yield Prediction in Quinoa Across
Mediterranean Environments
Karen Gutter 1
, Iván Matus 2
, Dalma Castillo 2 , Josefa Pacheco 1 , Sara Silva 1 , Juan Felipe Alfaro-Quezada 1
,
Fernando Fuentes-Peñailillo 3
and Alejandro del Pozo 1 , *
1 Centro de Mejoramiento Genético y Fenómica Vegetal, Facultad de Ciencias Agrarias, Universidad de Talca,
Talca 3460000, Chile; kgutter@utalca.cl (K.G.); josefa.pacheco@utalca.cl (J.P .); s.silva08@ufromail.cl (S.S.);
juan.alfaro@utalca.cl (J.F.A.-Q.)
2 Instituto de Investigaciones Agropecuarias, Centro Regional de Investigaciones Quilamapu, Casilla 426,
Chillán 3810189, Chile; imatus@inia.cl (I.M.); dalma.castillo@inia.cl (D.C.)
3 Dirección de Investigación, Vicerrectoría Académica, Universidad de Talca, Talca 3460000, Chile;
ffuentesp@utalca.cl

• Correspondence: adelpozo@utalca.cl
  Abstract
  Quinoa (Chenopodium quinoa Willd.) is a promising crop for drought-prone Mediterranean
  agriculture, although environmental variability complicates grain-yield selection. This
  study evaluated the potential of multisensor, multitemporal unmanned aerial vehicle (UA V)
  phenotyping combined with machine learning to predict environment-specific best linear
  unbiased estimates (BLUEs) of grain yield across five site–year environments in central
  Chile. A panel of 196 quinoa accessions was evaluated in replicated α-lattice trials. RGB,
  thermal, and multispectral imagery acquired during four reference flights provided 60 pre-
  dictors. Because genotypes differed in phenological development, flight dates represented
  common acquisition times rather than synchronized developmental stages. Elastic Net (EN)
  and Random Forest (RF) were evaluated using repeated five-fold cross-validation in four
  environments with complete predictor data (n = 779 genotype × environment observations),
  followed by leave-one-environment-out (LOEO) to assess cross-environment transferability.
  A sensitivity analysis using 57 predictors included all five environments (n = 974). Grain
  yield varied substantially among environments with mean values ranging from 0.20 to
  3.21 t ha−1 and an approximate broad-sense heritability on an entry-mean basis of 0.49.
  Within-environment predictive performance was moderate, with mean R2 values of 0.49 for
  EN and 0.51 for RF. In contrast, RF performance under LOEO validation was substantially
  lower (R2 = −5.96 to 0.11), revealing limited transferability of absolute yield predictions
  to unseen environments. Environment-specific SHAP analyses identified differences in
  the relative importance of UA V-derived color, canopy temperature, and vegetation indices,
  highlighting the environmental dependence of yield-related predictors. Overall, integrating
  multisensor UA V phenotyping with machine learning provided useful within-environment
  grain yield predictions in quinoa.
  Keywords: Chenopodium quinoa; high-throughput phenotyping; NDVI; canopy temperature;
  random forest; elastic net; genotype-by-environment interaction; mediterranean environments
  Agriculture 2026, 16, 2436 https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 2 of 23

1. Introduction
   Quinoa (Chenopodium quinoa Willd.) is a traditionally Andean crop that has become
   a globally cultivated pseudocereal due to its exceptional nutritional composition and cli-
   matic resilience [1]. Quinoa grain is also valued for its high nutritional quality, providing
   protein, essential amino acids, fatty acids, minerals, vitamins, and dietary fibers, as well
   as hypoglycemic benefits [2–4]. These attributes have contributed to its growing popular-
   ity and consumption and have positioned quinoa as a strategic crop for enhancing food
   and nutritional security. Within the Chenopodium quinoa species, there are five ecotypes,
   each associated with distinct environments of origin and adaptation to contrasting climatic
   conditions, photoperiods, soils, salinity levels, pest pressures, and diseases [5,6]. In Chile,
   quinoa is primarily represented by two ecotypes: the salt-flat (also known as “salares”)
   and coastal-lowland ecotypes. These ecotypes are remarkably different, mainly in their
   adaptation to altitude, drought, and salinity, in their sensitivity to day length, and in their
   genetic backgrounds [7]. This genetic diversity, along with quinoa’s ability to grow success-
   fully across a wide range of environmental conditions, has contributed to its recent global
   expansion.
   Accordingly, accelerating breeding for drought-prone Mediterranean environments is
   of great importance [8–10], particularly in areas such as central Chile, where agricultural pro-
   duction is increasingly constrained by drought and other climate-related stresses [11]. How-
   ever, despite quinoa’s potential to contribute to future for food security, breeding progress
   in Mediterranean production zones remains limited by strong and complex genotype-by-
   environment (G × E) interactions [1]. Evidence from multi-environment testing indicates
   that G × E can dominate yield variability in quinoa, implying that selection decisions must
   explicitly account for environmental responsiveness rather than relying solely on average
   performance [12,13].
   In Chile, coastal/lowland quinoa germplasm has been characterized for breeding pur-
   poses and evaluated across Mediterranean environments to identify productive and stable
   varieties [14,15]. The “phenotyping bottleneck” remains a major constraint to accelerating
   genetic improvement because grain yield and stress tolerance are complex quantitative
   traits resulting from dynamic interactions among genotype, environment, and crop devel-
   opmental stage. Conventional field phenotyping is often labor-intensive, time-consuming,
   and difficult to scale across large breeding populations and repeated sampling dates [16,17].
   This limitation is particularly relevant under drought, where crop responses are highly
   dynamic and single-time-point measurements may fail to capture differences in the onset,
   intensity, and recovery from stress. Digital monitoring and automated control have been
   explored across agricultural production systems, including low-cost vertical-farming proto-
   types designed to improve resource-use efficiency and production continuity [18]. Within
   open-field breeding, unmanned aerial vehicle (UA V)-based high-throughput phenotyp-
   ing (HTP) offers an effective approach to overcome these limitations by rapidly acquiring
   high-resolution RGB, multispectral, and thermal imagery over large field experiments at
   plot-scale, allowing repeated monitoring at key phenological stages [17,19–21]. Very-low-
   altitude UA V RGB imagery has also been used to automate plant counting at emergence,
   illustrating the diversity of plot-level traits that can be extracted from aerial images [22].
   Repeated UA V measurements throughout the crop cycle enable the characterization of
   temporal trajectories in canopy development, greenness, senescence, and temperature,
   providing indirect indicators of biomass accumulation, photosynthetic activity, water sta-
   tus, and stress responses [19,23–25]. Among the spectral vegetation indices derived from
   multispectral imagery, the normalized difference vegetation index (NDVI) is widely used
   as indicators for canopy greenness and vegetation vigor [26,27]. Additionally, thermal
   https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 3 of 23
imaging provides estimates of canopy temperature, which can serve as indirect indicators
of stomatal regulation and transpiration cooling [28–30].
There is growing evidence that UA V-based phenotyping can capture agronomically
meaningful variation in quinoa and related crops [23,31,32]. For example, Ref. [33] used
UA V-based multispectral and thermal imagery to derive vegetation indices and canopy
temperature from eight quinoa varieties grown in W A, USA, highlighting that these remotely
sensed traits effectively discriminated among irrigation treatments and were significantly as-
sociated with seed yield and physiological responses related to crop performance. Likewise,
proximal and aerial sensing approaches have emerged as promising tools for linking plant
water use with yield in quinoa under reduced irrigation, highlighting canopy temperature
as a potentially valuable physiological trait for drought-focused selection [32,34,35]. The
transformation of UA V-derived phenotypes into actionable selection criteria is increasingly
relying on machine learning (ML), which is well suited to handle multicollinearity among
spectral features, nonlinear relationships, and interactions that are difficult to capture with
simple linear regressions [23,36]. Reviews of ML in agriculture and crop yield prediction
consistently highlight the rapid growth of supervised learning approaches using remote
sensing predictors, including regularized linear models (e.g., ridge, lasso, elastic net),
tree-based ensembles (e.g., random forest, gradient boosting), and deep learning [37,38].
In UA V-based yield estimation specifically, studies emphasize that optimal performance
often depends on (i) growth stage [39], (ii) sensor modality or multi-sensor fusion [40],
(iii) feature selection and model regularization [36], and (iv) evaluation strategies that test
generalization beyond a single environment or site–year [40].
In practical breeding-oriented studies, elastic net regression and random forest are
particularly attractive because they offer complementary strengths [41]. Elastic net can
stabilize prediction under highly correlated UA V traits by combining LASSO (L1) and
Ridge (L2) penalties (supporting both shrinkage and variable selection). Random forest
is a nonlinear machine-learning method based on an ensemble of decision trees and con-
structs many regression trees from different samples of the data and subsets of predictors.
For example, Ref. [42] integrated UA V multispectral indices and thermal features within
machine-learning frameworks (including elastic-net-type regularization), and multi-sensor
information improved grain yield prediction in winter wheat under different water stress
treatments, compared with single-sensor models, highlighting the value of combining spec-
tral vigor signals with canopy temperature under variable water status. Similarly, Ref. [43]
trained supervised ML models, including random forest and regularized regression variants,
on time-resolved UA V features in onions, and demonstrated that predictive accuracy and
model robustness were strongly influenced by predictor selection, phenological timing, and
rigor of the validation schemes. These findings support a conceptual shift from UA V phe-
notyping as a purely descriptive tool to an integrated “phenotype-to-prediction” pipeline
that can inform earlier, more efficient selection, especially when evaluation explicitly tests
generalization across environments. However, key gaps remain for quinoa breeding in
Mediterranean drought-prone regions.
Deep learning also plays an established role in high-throughput phenotyping and
breeding. Convolutional neural networks (CNNs) can learn image representations and
combine spatial image information with other predictors. For example, Zhou et al. [44] used
UA V imagery and a mixed CNN to estimate the yield of 972 soybean breeding lines under
drought stress [44]. Deep learning architectures that incorporate temporal information
extend this approach by representing changes during crop development. For instance,
Nevavuori et al. [45] evaluated CNN–LSTM, convolutional LSTM, and three-dimensional
CNN models using multitemporal UA V data to predict within-field crop yield [45]. These
approaches highlight the potential of combining automated image-based feature extraction
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 4 of 23
with temporal modeling to improve yield prediction beyond conventional regression models
based on vegetation indices.
Compared with major cereals, quinoa has fewer multi-environment, UA V-enabled
studies that (i) evaluate large diversity panels, (ii) combine RGB, multispectral (including
NDVI), and thermal phenotypes across post-anthesis stages, and (iii) quantify prediction
performance under validation strategies that reflect real breeding deployment. Address-
ing these gaps is essential because Mediterranean production zones are characterized by
strong spatial and interannual variability in rainfall timing, evaporative demand, and heat
events [46,47], which can modify genotype rankings and alter the physiological pathways
that determine yield [48].
This study evaluated the ability of multi-sensor and multitemporal UA V-derived traits
to predict environment-specific grain-yield BLUEs in quinoa. We compared regularized
linear and nonlinear tree-ensemble benchmarks, evaluated complementary sensor and
reference-flight information, and distinguished within-environment prediction from transfer
to an environment withheld during training. Additional analyses assessed BLUE precision,
genotype ranking across environments, and the consequences of using a reduced predictor
set across all five environments.
2. Materials and Methods
2.1. Plant Material and Multi-Environment Field Trials
The panel comprised 196 quinoa accessions maintained in the Germplasm Bank of the
Chilean Agricultural Research Institute (INIA). Most entries were accessions or collected
populations, whereas a small subset consisted of cultivars developed through mass selection
within existing populations. Because these materials were not generated through controlled
crosses, conventional crossing-derived pedigrees do not apply to the panel. All entries were
classified as coastal/lowland quinoa adapted to Mediterranean environments of central and
south-central Chile. Field experiments were conducted across five environments at three
experimental stations of the Chilean Agricultural Research Institute (INIA), located within
the Mediterranean climate region of central Chile: Cauquenes (35°56′44″ S, 72°16′58″ W)
and Hidango (34°06′ S, 71°47′ W), both managed under rainfed conditions, and Santa
Rosa (36°31′34″ S, 71°54′40″ W), managed under irrigated conditions (Figure 1). Trials
were conducted over two growing seasons (2023/2024 and 2024/2025). Environments were
defined as unique combinations of site × season, resulting in five environments: Cauquenes
2023–24 (CAU_2324), Hidango 2023–24 (HID_2324), Cauquenes 2024–25 (CAU_2425),
Hidango 2024–25 (HID_2425), and Santa Rosa 2024–25 (SR_2425).
Trials were established using an α-lattice incomplete block design and two replicates per
environment. Each replicate was subdivided into incomplete blocks to account for within-
field spatial heterogeneity. Genotypes were randomized within incomplete blocks in each
replicate. Sowing was performed at a rate of 15 kg ha−1 . Sowing dates are indicated in Table 1.
At sowing, 120 kg P2 O5 ha−1 (triple superphosphate, 261 kg ha−1 ) and 60 kg K2 O ha−1
(KCl/muriate of potash, 100 kg ha−1 ) were applied. Nitrogen was supplied as urea at
the 4–5 and 5–6 leaf stages, using 75 kg N ha−1 each (≈166 kg urea ha−1 ); in 2024–2025,
these applications were made on 1 October 2024 and 9 October 2024, respectively, while in
2023–2024, only the first application (20 November 2023) was applied due to drought.
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 5 of 23
Figure 1. Location of the experimental trials along with RGB, thermal and multispectral orthomosaics
acquired at Hidango, Cauquenes, and Santa Rosa.
Table 1. Environmental characterization and adjusted grain-yield performance across five quinoa
field environments in central Chile.
Environment Sowing Date Harvest Date Pp S–A (mm) Pp A–H (mm) Mean T S–A (°C) Mean T A–H (°C) GY_BLUE (t ha−1 )
CAU_2324 2 October 2023 10 March 2024 51.0 0.4 14.9 21.1 0.20
HID_2324 29 October 2023 6 April 2024 37.0 4.2 14.2 17.5 0.28
CAU_2425 3 September 2024 10 February 2025 56.2 0.0 13.5 20.1 0.50
HID_2425 27 August 2024 3 February 2025 48.0 0.4 12.8 16.9 0.59
SR_2425 10 September 2024 17 February 2025 127.7 0.6 12.9 18.7 3.21
Notes: S–A, sowing to anthesis; A–H, anthesis to harvest; Pp, cumulative precipitation. GY_BLUE
denotes environment-specific best linear unbiased estimates adjusted for the α-lattice design. The
anthesis date corresponds to the trial-level reference date used for climatic summarization and does
not represent the exact anthesis date of every genotype.
Analyses were restricted to the 196 genotypes represented across all five environments.
Field-design and yield records were linked using the composite identifier PlotID × Environ-
ment × Genotype. After excluding four genotypes evaluated only during the first season,
the analytical dataset contained 1960 plot records. Original replicate and incomplete-block
assignments were retained. Trait-specific missing values were not imputed and were treated
as missing. No statistical outlier exclusion was applied to numeric observations.
2.2. Ground-Based Agronomic Measurements and Grain Yield
For grain yield (GY) determination, the two central rows of each plot were harvested.
A 0.50 m section was excluded from each end of the harvested rows, giving a net harvested
area of 3.0 m2 per plot. Within each environment, all plots were harvested on a common
date, approximately one month after physiological maturity was considered to have been
reached at the trial level; harvest dates are reported in Table 1. Harvest was not scheduled
separately according to the physiological maturity of each genotype. Grain-yield samples
received no additional post-harvest drying. Grain moisture content was not measured, and
grain mass was not adjusted to a standard moisture content. GY was therefore expressed on
an as-harvested basis in t ha−1 with no moisture adjustment. Panicle density, harvest index,
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 6 of 23
thousand-kernel weight (TKW) and plant height (PH) were also recorded, as detailed
below.
PH (cm) was measured at the end of the growing period, one month after complete
plant maturation, from the soil surface to the upper end of the panicle. For harvest index
(HI), stems with their panicles from one linear meter of row were cut at ground level
and oven-dried at 65 °C for 48 h. HI was calculated as grain dry weight divided by shoot
dry weight and is dimensionless. This drying procedure applied to the HI subsample;
GY from the net harvested plot area was recorded on the as-harvested basis described
above. Seed counting for TKW (g) used a DATA Count S-JR seed counter (DATA Detection
Technologies Ltd., Jerusalem, Israel). Panicle density (panicles m−2 ) was calculated by
dividing the number of panicles harvested from a 3.0 m2 area by the sampled area (3.0 m2 ).
Grain yield was calculated from the grain mass harvested from the net plot area as:
GY (t ha−1) = m (kg)A (m2) × 10 (1)
where m is the harvested grain mass, A is the net harvested area (3.0 m2 ), and the factor
10 converts kg m−2 to t ha−1 . Grain yield therefore represents as-harvested grain mass and
was not adjusted to a standard moisture content.
2.3. UAV-Based High-Throughput Phenotyping
RGB and multispectral imagery was acquired using a DJI Mavic 3 Multispectral (M3M)
(SZ DJI Technology Co., Ltd., Shenzhen, China), while thermal imagery was acquired
using a DJI Mavic 3 Thermal (M3T) (SZ DJI Technology Co., Ltd., Shenzhen, China), to
capture plot-level spectral, color and temperature descriptors. Flights were conducted
on four dates per environment, operationally labelled according to trial-level reference
phenology as anthesis (ANT), watery-ripe seed (WRC), milk (MILK), and dough-stage
(DOUGH). Because the diverse genotypes included in the panel differed in phenological
development, these labels describe the approximate timing of each UA V campaign at the trial
level and were not intended to represent the exact developmental stage of every genotype.
Consequently, individual genotypes could be at different developmental stages on the same
acquisition date. Throughout the manuscript, ANT, WRC, MILK, and DOUGH are therefore
treated as reference flight labels rather than genotype-specific phenological observations.
Flights were conducted on clear days near solar noon to reduce confounding effects of
rapidly changing radiation and vapor-pressure deficit. Flight altitude was set to 30 m above
ground level, with 80% forward and lateral overlap.
RGB images were aligned and processed into orthomosaics using Agisoft PhotoScan
Professional Versión 1.4.5 build 7354 (64 bit); multispectral and thermal images were
processed using Pix4D Enterprise version 4.5.6. Individual plots were manually delineated
as polygonal regions of interest (ROIs) in ImageJ 1.54g, separately and independently
on each orthomosaic. For each sensor and acquisition date, descriptors were extracted
from the retained pixels of its corresponding plot ROI. Multisensor and multitemporal
integration therefore used plot-level descriptors; independent manual delineation does not
provide an estimate of pixel-level co-registration accuracy across sensors or dates. Polygon
boundaries were conservatively positioned inside plot margins, and border areas were
manually excluded to reduce edge effects, mixed pixels, exposed-soil contributions and
contamination from adjacent plots. No fixed numerical threshold or automated classification
rule was used for plot segmentation or soil–vegetation separation. Image-derived variables
were calculated from the pixels retained within each ROI; this procedure does not establish
that all retained pixels represented vegetation. The supplied RGB-derived indices were
hue, intensity, saturation, lightness, a*, b*, u*, v*, Green Area (GA), Green–Green Area
(GGA) and Crop Senescence Index (CSI). Canopy temperature (Tc) variables were the
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 7 of 23
mean, minimum and maximum temperatures from the retained pixels. NDVI is defined as
(NIR − Red)/(NIR + Red), using near-infrared and red bands; the supplied mean NDVI
values were summarized by genotype, environment and reference flight for analysis.
NDVI was retained as the representative multispectral vegetation index because the
study focused on multisensor and multitemporal yield prediction rather than on an exhaus-
tive comparison of spectral indices. Limiting the analysis to NDVI also reduced redundancy
among indices derived from overlapping spectral bands.
The available data contains the plot-level descriptors used in this reanalysis, but not
the raw-image archive, image-processing projects, masks, or a parameter log. Consequently,
the exact historical extraction settings cannot be reconstructed, and image processing was
not repeated. Published BreedPix definitions of GA, GGA and CSI are provided only as con-
textual reference conventions [49], not as verified historical procedures. The physiological
interpretation of these indices was therefore cautious, and a sensitivity analysis excluding
CSI was conducted.
2.4. Mixed-Model Analysis, BLUEs, and Heritability
Agronomic traits were analyzed at the plot level using linear mixed models that in-
cluded genotype, environment, replicate, and incomplete-block effects. Models used the
common panel of 196 genotypes, with the original block and replicate assignments retained
after removing the four first-season-only entries. Trait-specific missingness remained; no
measurements were imputed to force 196 estimable responses per environment. Across-
environment variance components were estimated with environment fixed and genotype,
genotype × environment, replicate within environment, and block within replicate and
environment random:
Yijkl = μ + Ei + Gj + (G × E)ij + Rk(i) + Bl(ki) + εijkl (2)
where Yijkl is the observation of genotype j in environment i, replicate k, and incomplete
block l; μ is the overall mean; Ei is the fixed effect of environment; Gj is the random genotype
effect; (G × E)ij is the random genotype-by-environment interaction; Rk(i) is the random
effect of replicate nested within environment; Bl(ki) is the random effect of incomplete block
nested within replicate and environment; and εijkl is the residual error. Because each site–
year environment was associated with a single water-management regime, water regime
was considered part of the environmental definition and was not fitted as an additional
independent factor. Variance components were estimated by restricted maximum likelihood.
The primary model assumed homogeneous residual variance. Residual standard deviations
were examined by environment and are supplied as diagnostics; a comparison with a
heterogeneous-residual model is not claimed.
Environment-specific genotype BLUEs were estimated using separate α-lattice models
within each environment:
Yjkl = μ + Gj + Rk + Bl(k) + εjkl (3)
Genotype and replicate were fixed effects, and incomplete block within replicate was
random. Models were fitted separately in each environment by REML. The BLUE of each
genotype was the equal-weight marginal mean over the two replicate levels, BLUE = L ̂β,
with covariance VBLUE = L Var( ̂β)L⊤ and SE = sqrt(diag(VBLUE)). Fixed-effect means and
their full covariance matrices were retained. Negative adjusted estimates were not truncated.
The primary ML analysis used these BLUEs without precision weights; a sensitivity analysis
used inverse-SE2 training weights, normalized to mean one within each training set. This
sensitivity does not propagate the complete covariance of the estimated responses.
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 8 of 23
Approximate broad-sense heritability on an entry-mean basis was calculated from the
primary fixed-environment REML model following [50]:
H2entry = σ2G
σ2G + σ2G×EnE + σ2enEnR (4)
where 𝜎2𝐺, 𝜎2𝐺×𝐸 and 𝜎2𝑒 are the genotypic, genotype × environment and residual variance
components; nE = 5 and nR = 2 nominal numbers of environments and replicates in the
experimental design, rather than estimated effective replication numbers. Because the
dataset contained missing observations and unequal estimation precision, 𝐻2entry should be
interpreted as an approximate entry-mean heritability under the nominal design rather
than as a generalized reliability measure.
For the relative contributions of variance components, only, a separate descriptive
REML model treated environment as random, with the remaining random terms as in
Equation (1). Its total variance included environment, genotype, genotype × environment,
replicate (environment), block (replicate, environment), and residual components, all
estimated jointly in that model. The environmental fraction is therefore not a variance
parameter from the fixed-environment model. This descriptive decomposition is supported
by five sampled environments.
Observed genotype ranking was compared between every pair of environments using
their shared identifiers. Spearman correlations and percentile 95% intervals were obtained
by resampling the shared identifiers 2000 times (seed 20261013). The intervals condition on
the estimated BLUEs and do not propagate their estimation errors. A descriptive identity,
Var(A − B) = (SDA − SDB)2 + 2 SDA SDB(1 − rPearson), was also used to separate scale
differences from imperfect association of the paired BLUEs; it is not a decomposition of true
genetic interaction.
2.5. Machine Learning Prediction
2.5.1. Prediction Target and Data Assembly
The prediction unit was genotype × environment and the response was GY_BLUE.
UA V variables were averaged over available field replicates for each genotype, environment
and reference flight before linkage to the corresponding BLUE. The full set contained
44 RGB predictors (11 variables × four flights), 12 thermal predictors (mean, minimum and
maximum temperature × four flights), and four mean-NDVI predictors. This produced one
response and predictor vector per observed genotype–environment combination. Genotype
labels were reconciled using the same confirmed name mapping as the field data. The design
contained 196 identifiers shared across all five environments, but only 195 had estimable
GY_BLUEs in all five because CQU 23 lacked observed GY in every 2024–2025 environment.
In total, 977 GY_BLUEs were available before predictor-completeness filtering.
The primary complete-case cohort comprised 779 genotype–environment records
across CAU_2324, HID_2324, HID_2425 and SR_2425. The same records and validation
partitions were used for every sensor and flight ablation. No missing-value imputation was
applied. CAU_2425 was excluded only from the complete 60-predictor analysis because
ANT thermal measurements were unavailable. An additional analysis included all five
environments and 974 records using the 57 predictors remaining after removing those three
ANT thermal variables.
2.5.2. Modeling Algorithms
Elastic Net (EN) used alpha = 0.01 and l1_ratio = 0.50 in the scikit-learn parameteriza-
tion, deterministic cyclic coordinate descent, tolerance 10−7 and at most 100,000 iterations.
Before fitting EN, each training partition was checked for numerically negligible predictor
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 9 of 23
variation: a column was excluded when its training standard deviation was ≤10−8 × max
(1, |training mean|). The retained columns were centered and scaled using training means
and standard deviations; the same transformation and column selection were then applied
to its test set. This safeguard prevents floating-point noise in nearly constant indices from
being amplified during scaling. Random Forest (RF) used 150 trees, max_features = √p,
bootstrap sampling and minimum leaf size one in the main analysis, LOEO and ablations.
These benchmark parameters were fixed across configurations; they were not optimized on
the held-out outcomes. The comparison therefore evaluates specified benchmarks, not the
best attainable performance of either model family.
2.5.3. Validation Strategies
Within-environment validation used five-fold CV repeated three times, with split seeds
7, 20 and 33. Both algorithms, sensor configurations and reference-flight configurations
used identical partitions within each cohort. Every genotype–environment record was
predicted once per repetition, and no fold was excluded based on its observed prediction
error. Leave-one-environment-out (LOEO) validation trained on the other environments
and tested on the withheld environment. Thus, the primary LOEO analysis trained on three
environments; the five-environment sensitivity trained on four. LOEO was not genotype-
disjoint and mostly evaluates transfer of observed materials to a new site–year.
For each cross-validation repetition and within each environment, the out-of-fold
predictions from the five test folds were pooled. Performance metrics were then calculated
once using the complete set of held-out predictions. For each metric, Table 2 will report the
mean and sample standard deviation (SD) across the three repetitions. The SD represents
variability among resampling repetitions and should not be interpreted as an inferential
confidence interval because the same observations were reused across repetitions. When
performance was averaged across environments, each environment was given equal weight,
regardless of its sample size. The coefficient of determination R2 was calculated as:
R2 = 1 − ∑𝑛𝑖=1(𝑦𝑖 − ̂y𝑖)2
∑𝑛𝑖=1(𝑦𝑖 − y)2 (5)
where yi and ŷi are the observed and predicted values for observation i, respectively; ȳ is
the mean of the observed values in the combined held-out dataset; and n is the number of
held-out observations.
The root mean square error (RMSE) was calculated as:
RMSE = √( 1𝑛) 𝑛∑𝑖=1(𝑦𝑖 − ̂y𝑖)2 (6)
Pearson’s correlation coefficient (r) was used to quantify the linear association between
observed and predicted values, whereas Spearman’s rank correlation coefficient was used
to quantify agreement in their ranking. For leave-one-environment-out (LOEO) validation,
prediction bias was calculated as:
BiasLOEO = ( 1𝑛) 𝑛∑𝑖=1( ̂y𝑖 − 𝑦𝑖) (7)
Positive bias values indicate systematic overprediction, whereas negative values indi-
cate underprediction. The accompanying record-level results tables retained the observed
and predicted values together with their corresponding genotype, environment, repeti-
tion, and fold identifiers. RF mean-decrease-in-impurity importance was retained. SHAP
provided complementary signed contributions to individual predictions [46], whereas
impurity importance supplies an unsigned split-based ranking and may favor predictors
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 10 of 23
with more candidate cut points [51]. To avoid interpreting only pooled environmental
contrasts, TreeSHAP was calculated separately within each environment for each held-out
fold of the main RF analysis. Mean absolute SHAP values were summarized across held-out
observations and repetitions.
Table 2. Prediction of environment-specific GY_BLUE using elastic net (EN) and random forest
algorithms. GY_BLUE and RMSE refer to as-harvested yield in t ha−1 without moisture correction.
CV entries are mean ± sample SD across three repetition-level out-of-fold metrics (five folds per
repetition); LOEO entries are metrics from the single held-out environment. SD describes resampling
variability and is not a confidence interval.
Environment Model Strategy R2 r RMSE (t ha−1 )
CAU_2324 EN CV 0.609 ± 0.006 0.782 ± 0.005 0.085 ± 0.001
CAU_2324 RF CV 0.488 ± 0.013 0.707 ± 0.010 0.097 ± 0.001
CAU_2324 EN LOEO −55.208 −0.306 1.017
CAU_2324 RF LOEO −5.963 0.481 0.358
HID_2324 EN CV 0.438 ± 0.006 0.663 ± 0.004 0.104 ± 0.001
HID_2324 RF CV 0.483 ± 0.022 0.697 ± 0.017 0.100 ± 0.002
HID_2324 EN LOEO −3.688 0.410 0.302
HID_2324 RF LOEO −3.826 0.193 0.306
HID_2425 EN CV 0.292 ± 0.206 0.593 ± 0.092 0.240 ± 0.034
HID_2425 RF CV 0.432 ± 0.021 0.659 ± 0.017 0.216 ± 0.004
HID_2425 EN LOEO −0.737 0.582 0.378
HID_2425 RF LOEO 0.111 0.377 0.271
SR_2425 EN CV 0.601 ± 0.072 0.784 ± 0.039 0.706 ± 0.063
SR_2425 RF CV 0.644 ± 0.006 0.808 ± 0.005 0.669 ± 0.005
SR_2425 EN LOEO −1.207 0.484 1.666
SR_2425 RF LOEO −4.343 0.505 2.591
3. Results
3.1. Environmental Contrast and Quantitative-Genetic Context
Mean GY_BLUE ranged from 0.20 t ha−1 at CAU_2324 to 3.21 t ha−1 at SR_2425 (Table 1).
Figure 2 additionally displays the environment-specific BLUE distributions of HI and TKW,
allowing their dispersion to be compared with GY. The HI and TKW distributions are
generally narrower than the GY distribution in SR_2425; this visual comparison is descriptive
and should be interpreted together with their higher approximate entry-mean H2 values
(0.84 for HI and 0.83 for TKW; Figure 3A), rather than as a standalone test of stability. Based
on the recorded anthesis dates for each trial, cumulative precipitation from anthesis to
harvest did not exceed 4.2 mm in any environment.
The approximate entry-mean heritability of GY was 0.49, with σ2 G = 0.0441,
σ2 GE = 0.1578 and σ2 e = 0.1446 (Figure 3A). Corresponding H2 values for yield compo-
nents were 0.73 for panicle density, 0.84 for HI, 0.83 for TKW, and 0.63 for height (Figure 3A).
These values were calculated from the fixed-environment REML model using the nominal
five-environment, two-replicate approximation and homogeneous residual model. Variance
components differed among traits, with environmental variance accounting for the largest
proportion of total variance in all traits except panicle number, for which residual variance
predominated (Figure 3B).
Pairwise Spearman correlations of GY_BLUE ranged from 0.149 to 0.572, with
195–196 shared identifiers per pair. Thus, the observed rankings showed partial rather than
complete agreement. Correlations and their conditional bootstrap intervals, together with
the scale-versus-association decomposition, distinguish observed ranking differences from
changes in response magnitude. They do not establish error-free genetic crossover because
BLUE estimation uncertainty remains.
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 11 of 23
Figure 2. Environment-specific distributions of BLUEs for (A) grain yield (GY; t ha−1 ), (B) harvest
index (HI; dimensionless), and (C) thousand-kernel weight (TKW; g) across five field environments.
For each panel, violin width represents density; boxplots show the median and interquartile range;
black points show means; and labels give the number of estimable BLUEs. Negative adjusted estimates
were not truncated.
Figure 3. (A) Approximate broad-sense heritability on an entry-mean basis for grain yield (GY and
yield-related traits, estimated using restricted maximum likelihood (REML) model with environments
treated as fixed effects, assuming five environments (nE = 5) and two replicates (nR = 2). (B) Relative
contributions of variance components estimated using a separate REML model with environments
treated as random effects, including environment, genotype, genotype × environment (G × E) in-
teraction, replicate nested within environment, block nested within replicate and environment, and
residual variance. The two panels represent distinct statistical approaches with different purposes
and variance denominators. Panicles: number of panicles per square meter; HI: harvest index; TKW:
thousand-kernel weight; height: plant height.
3.2. Reference-Flight-Specific Pooled Associations Between UAV Traits and Grain Yield
Pearson correlation analysis across all environments revealed significant associations
between GY_BLUE and UA V-derived traits evaluated at four phenological stages: anthesis,
water-ripe grain, milky grain, and dough grain. In the pooled primary cohort, correlation
coefficients across these stages ranged from 0.86 to 0.91 for GA, 0.78 to 0.90 for GGA, and
0.82 to 0.91 for mean NDVI, whereas CSI consistently exhibited strong negative correlations
(r = −0.88 to −0.77) (Figure 4). Correlations depended on the reference flight: pooled
correlations ranged from −0.91 to −0.28 for mean temperature, −0.92 to −0.39 for minimum
temperature, and −0.80 to 0.23 for maximum temperature.
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 12 of 23
Figure 4. Heat map of Pearson correlation coefficients (r) between environment-specific grain-yield
BLUEs (GY_BLUE) and UA V-derived variables in the primary complete-case cohort (n = 779), pooled
across CAU_2324, HID_2324, HID_2425, and SR_2425. Columns are the four trial-level reference
flights: ANT, anthesis; WRC, water-ripe grain; MILK, milky grain; and DOUGH, dough grain. Rows
are RGB color descriptors (intensity, hue, saturation, lightness, a*, b*, u*, v*), RGB vegetation-fraction
indices (green area, GA; green–green area, GGA; crop senescence index, CSI), canopy-temperature
summaries (T Mean, T Min, T Max), and mean normalized difference vegetation index (NDVI). Cell
labels and colors give r: positive values (red) indicate higher GY_BLUE with higher values of the
UA V variable and negative values (blue) indicate the opposite. Correlations include both within- and
between-environment variation.
These results should be interpreted as descriptive associations across the pooled dataset.
Because the analysis combined environments with markedly different mean yield lev-
els, the correlation coefficients reflect both between-environment differences and within-
environment covariation. In addition, because the reference labels did not represent the
exact developmental stage of every genotype, differences among flight times may partly
reflect variation in the phenological status of genotypes at each acquisition date.
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 13 of 23
3.3. Within-Environment GY_BLUE Prediction and Cross-Environment Transferability
EN and RF algorithms produced positive within-environment CV R2 in all four envi-
ronments (Table 2; Figure 5). Table 2 gives the numerical CV and LOEO metrics for each
environment–model combination. Figure 5 visualizes only the corresponding CV results:
each bar is the mean metric across the three repeated CV runs and each error bar is their
sample SD. RF achieved an unweighted environment mean R2 of 0.512, Pearson r of 0.718
and RMSE of 0.271 t ha−1 ; the corresponding EN means were 0.485, 0.705 and 0.284 t ha−1 .
EN performed better in some environments and RF in others, so these results do not sup-
port universal superiority of either benchmark. All 15 validation folds per environment
contributed to each model.
Figure 5. Within-environment cross-validation performance of elastic net (EN) and random forest (RF)
models for prediction of GY_BLUE. Panels show (A) coefficient of determination (R2 ), (B) Pearson
correlation coefficient (r), and (C) root mean squared error (RMSE, t ha−1 ) for each of the four
environments. Bars are the mean metric calculated from out-of-fold predictions across three repeated
five-fold CV runs; error bars are the corresponding sample SD. These CV results are the numerical
CV entries reported in Table 2; LOEO results are reported only in Table 2.
Absolute transfer remained limited. RF LOEO R2 values were −5.963 in CAU_2324,
−3.826 in HID_2324, 0.111 in HID_2425, −4.343 in SR_2425, and EN values were −55.208 in
CAU_2324, −3.688 in HID_2324, −0.737 in HID_2425, and −1.207 in SR_2425. Negative R2
means squared error exceeded the deviation from the test-environment mean; this mean is
an evaluation baseline rather than a deployable predictor requiring no yield information.
The extreme EN LOEO R2 in CAU_2324 primarily reflected systematic underprediction.
The observed yield mean and SD were 0.204 and 0.136 t ha−1 , respectively, whereas the
predicted mean was −0.785 t ha−1 , resulting in a bias of −0.989 t ha−1 and an RMSE of
1.017 t ha−1 . Squared bias accounted for 94.5% of the mean squared error. Thus, the strongly
negative R2 mainly resulted from a large calibration shift relative to the limited within-
environment yield variation, with the unconstrained EN model extrapolating to physically
implausible negative yields.
3.4. Effects of Sensor Ablation on GY_BLUE Prediction Performance of RF and EN Models
Across environments, the complete sensor-derived traits yielded a mean R2 = 0.512
for RF and 0.485 for EN (Figure 6). The highest mean R2 values were achieved using the
complete set of sensor-derived traits for both RF and EN models. The complete-sensor
configuration corresponded to the primary model summarized in Table 2, averaged equally
across environments. Reduced sensor sets therefore represent empirical performance–
measurement trade-offs within this cohort; small differences are descriptive and were not
tested as independent fold-level contrasts.
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 14 of 23
Figure 6. Sensor-ablation performance of RF and EN models for GY_BLUE prediction, evaluated using
identical primary-cohort records and cross-validation partitions: (A) coefficient of determination
(R2 ); and (B) Pearson correlation coefficient (r). Performance metrics were calculated from out-of-fold
predictions for each repetition and environment and subsequently averaged across environments
with equal weighting. Error bars represent the standard deviation of the environment-averaged
performance across three repetitions. Sensor configurations included RGB (11 derived indices),
thermal (3 indices), NDVI (1 index), and All (the complete set of 15 indices).
3.5. Predictive Value of Individual Reference UAV Flight Dates for GY_BLUE
The optimal RF reference-flight dates varied among environments, with the highest
predictive performance using all flights in CAU_2324, HID_2425, and SR_2425, and the
dough-grain stage (DOUGH) in HID_2324 (Figure 7). The all-flight configuration corre-
sponded to the primary RF model. These differences reflect the predictive information
captured at different UA V acquisition times rather than strictly defined phenological stages,
as genotypes were not necessarily at the same developmental stage on each reference-flight
date. Standard deviations across repeated cross-validation runs reflect variability associated
with data partitioning. Overall, these findings highlight the environment-dependent value
of UA V acquisition timing for grain yield prediction, without identifying a universally
optimal flight date.
Figure 7. RF cross-validation performance for GY_BLUE prediction using predictors from individual
reference flights or from all four reference flights (All) within each environment: (A) coefficient of
determination (R2 ); and (B) Pearson correlation coefficient (r).The All configuration is the primary
RF CV model reported in the RF/CV rows of Table 2; it is not an additional Table 2 entry. Metrics
were evaluated using the same primary-cohort records and repeated five-fold partitions. Error bars
represent the sample SD across three repetition-level out-of-fold metrics. ANT, WRC, MILK, and
DOUGH are trial-level reference-flight labels, not genotype-specific stages.
3.6. Environment-Specific Contributions of UAV-Derived Traits to GY_BLUE Prediction Across
Reference-Flight Dates
Held-out SHAP analysis was conducted separately for each environment to identify
the relative contributions of UA V-derived traits to RF predictions of GY_BLUE, at the dif-
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 15 of 23
ferent reference-flight dates. The predictors with the highest mean absolute SHAP values
differed among environments: b*_MILK in CAU_2324, Saturation_DOUGH in HID_2324, T
Mean_MILK in HID_2425, and NDVI (mean)_MILK in SR_2425 (Figure 8). These differ-
ences highlight the environment-specific importance of UA V-derived traits for grain yield
prediction and indicate that predictor rankings obtained from pooled analyses may not
accurately reflect their relative contributions within individual environments. SHAP values
quantify the contribution of each predictor to RF predictions relative to the model’s base-
line, with mean absolute SHAP values representing the magnitude of these contributions,
irrespective of their direction. The direction of individual effects can be further examined
using signed SHAP values in conjunction with predictor values and their correlations with
GY_BLUE. Importantly, the direction and magnitude of predictor contributions may vary
among environments, and SHAP attributions should not be interpreted as evidence of
causal physiological relationships with grain yield.
Figure 8. Environment-specific contributions of UA V-derived traits to GY_BLUE prediction across
reference-flight dates corresponding to different phenological stages: (A) CAU_2324; (B) HID_2324;
(C) HID_2425; and (D) SR_2425. Random Forest (RF) models were interpreted separately for
each environment using held-out TreeSHAP values obtained from repeated five-fold cross-validation.
Panels show the mean absolute SHAP values of the most influential UA V-derived predictors, identified
by their corresponding phenological stages and averaged across observations and repetitions. RF
models were fitted independently for each environment using 150 trees. ANT: anthesis; WRC: water-
ripe grain; MILK: milky grain; DOUGH: dough grain. The asterisk in a*, b*, u*, and v* is part of the
standard CIELAB (a* and b*) and CIELUV (u* and v*) color-space notation and does not indicate
statistical significance.
3.7. Sensitivity Analyses
The five-environment analysis included 974 genotype–environment records and 57 pre-
dictors. In the CAU_2425 environment, within-environment R2 was 0.280 for EN and 0.178
for RF. This sensitivity changes both predictor availability and the training environments,
so differences from the 60-predictor analysis cannot be attributed solely to sample size.
Inverse-SE2 training weights changed within-environment R2 by at most 0.0038 in absolute
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 16 of 23
terms across the eight environment–model comparisons. Excluding CSI changed R2 by at
most 0.0274. These checks assess robustness to marginal BLUE precision and one supplied
index.
4. Discussion
4.1. Environmental Conditions Constrained Grain Yield Predictions
The moderate approximate entry-mean heritability of GY (0.49) indicates that geno-
type differences were detectable in this multi-environment trial, but that their expression
was strongly dependent on environment and residual variation. The higher entry-mean
H2 values for HI (0.84), TKW (0.83), panicle density (0.73), and height (0.63) suggest
that these component traits may provide more repeatable auxiliary information than GY
under the present design; they do not make them direct substitutes for selection on GY.
Because the G × E component for GY exceeded the genotypic component, selection based on
one environment alone may not represent response elsewhere. This interpretation follows
the distinction between heritability and response to selection in multi-environment tri-
als [50]. It is consistent with reports of environmental sensitivity in Chilean coastal-lowland
quinoa [35,52] and with the need for multi-environment evaluation in Mediterranean
annual crops, including wheat [48].
RF explained an average 51.2% of the within-environment variation in the environment-
specific GY_BLUE values; EN explained 48.5%. These are averages of repeated out-of-fold
metrics, not coefficients of determination from a single pooled fitted model. R2 depends
on test-set variation, RMSE reports error on the yield scale, and Pearson r measures linear
association. Spearman correlation provides an additional ranking measure. A positive
Pearson correlation alone does not demonstrate preservation of genotype order or accurate
absolute calibration.
4.2. Sensor Complementarity and Phenological Timing
Higher values of UA V-derived canopy cover and greenness were strongly associated
with greater GY_BLUE across environments, whereas CSI exhibited negative associations
with grain yield (Figure 4). These relationships suggest that canopy development, veg-
etation cover, and spectral characteristics captured by UA V imagery may provide useful
indicators of quinoa productivity. Greater canopy cover and greenness may indicate a
larger photosynthetically active canopy area and increased light interception, potentially
supporting biomass accumulation and grain yield [32,53]. However, these associations do
not necessarily indicate prolonged photosynthetic activity or delayed canopy senescence,
as these physiological processes were not directly assessed. Similar relationship between
canopy traits and grain yield have been reported in Chilean quinoa germplasm [35,52],
while multispectral phenotyping has demonstrated its potential to characterize phenotypic
variability among quinoa accessions [54]. Nevertheless, the strong correlations observed in
the pooled analysis may partly reflect differences in environmental conditions and crop
development among trials rather than consistent genotypic relationships within individual
environments. Furthermore, variation in UA V acquisition dates and uncertainties in the
extraction procedures for some spectral indices may influence these associations. Therefore,
although UA V-derived traits show considerable potential as predictors of quinoa grain yield,
their physiological relevance and predictive stability should be further evaluated within
and across contrasting environments.
Canopy temperature can provide complementary information about energy balance
and water status. In annual crops with near-complete canopy cover, exposed-soil effects
are reduced; nevertheless, atmospheric conditions, canopy architecture, viewing geometry,
and acquisition timing still affect the temperature signal. UA V thermal/RGB measurements
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 17 of 23
have discriminated irrigation treatments in winter wheat [29] and combined multispectral
and thermal UA V traits have improved grain-yield prediction across wheat water-stress
treatments [42]. In quinoa, canopy-temperature-related phenotyping has likewise been
used to examine water-use and yield-related variation under reduced irrigation [32,35]. The
environment-specific importance estimates should therefore be interpreted in their local
context rather than as evidence that a pooled thermal variable always dominates genotype
discrimination. The available extracted values do not independently verify every image
calibration, masking, or registration setting.
The sensor-ablation analysis highlighted the complementary value of UA V-derived
spectral and thermal information for predicting quinoa grain yield across contrasting
environments (Figure 6). The combination of RGB-derived indices, NDVI, and thermal
traits provided the highest overall predictive performance for both Random Forest (RF; mean
R2 = 0.512) and Elastic Net (EN; mean R2 = 0.485), suggesting that integrating multiple
sensors captures complementary aspects of crop growth and physiological responses that
may not be fully represented by individual sensor types. The improved performance of the
combined sensor configuration suggests that these measurements provide complementary
information relevant to yield formation, particularly under contrasting water availability
and environmental conditions. However, the relative contribution of each sensor may vary
among environments, reflecting differences in the physiological processes limiting crop
productivity. These findings support the integration of multisensor UA V phenotyping to
improve grain yield prediction in quinoa, while highlighting the importance of evaluating
whether the additional predictive gains justify the greater costs and operational complexity
associated with multisensor data acquisition.
Similar considerations apply to annual crops. In wheat, UA V NDVI–yield relationships
changed across the growth cycle, with grain filling identified as an especially informative
period for yield prediction [55]. In quinoa, the current reference flights should likewise be
interpreted as acquisition times rather than synchronized genotype-level stages. These ex-
amples support incorporating phenological information into UA V-based prediction models,
while recognizing that the present study did not measure genotype-specific phenology at
each flight.
Manual plot delineation excluded borders but did not automatically isolate vegetation.
Soil or other background pixels retained within the ROIs may therefore influence color,
spectral and thermal descriptors, particularly where canopy cover is incomplete. The
extent of this contribution was not quantified. ROIs were independently delineated on
each orthomosaic, so the combined predictors represent plot-level summaries rather than
demonstrated pixel-aligned measurements across sensors and dates. Unverified calibration
settings and the documented index inconsistencies further limit physiological interpretation
and reproduction of the complete image-to-trait workflow.
4.3. Model Robustness and the Limits of Cross-Environment Transfer
EN and RF showed comparable mean within-environment performance, although their
relative performance varied among environments. Both algorithms were evaluated using
identical validation partitions, with all folds retained. Training-only removal of numerically
near-constant predictors prevented unstable scaling in EN while preserving the separation
between training and validation data.
LOEO performance remained limited because prediction errors were large relative to
the yield variability of some withheld environments and because environmental shifts pro-
duced substantial prediction bias. Negative R2 values therefore indicate performance below
the held-out-environment mean benchmark. Positive correlations in some environments
indicate partial association but not reliable absolute calibration. These results demonstrate
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 18 of 23
the need for independent site–year validation and explicit representation of environmental
covariates before model deployment.
Improving transfer will require explicit representation of the environment. Precipi-
tation and irrigation by phenological phase, growing-degree days, vapor-pressure deficit,
radiation, reference evapotranspiration, soil properties, and repeated soil-water measure-
ments could help separate the environmental yield level from genotype deviations around
that level. Integrating remote sensing with agrometeorological observations and ground-
based sensor networks has been proposed as a practical framework for representing crop
water demand and its spatial and temporal environmental drivers [56]. Multi-environment
and reaction-norm approaches that integrate environmental and genetic information pro-
vides a promising route for this purpose [57,58]. Reliable prediction in new environments
also requires data from more independent site–years. Evaluating many genotypes within a
small number of environments improves model fitting, but it provides limited evidence of
cross-environment generalization.
Deep learning approaches were not included in the present model comparison. The
dataset comprised 60 UA V-derived sensor indices collected across multiple flights rather
than raw image tensors, and the relatively small number of site–year environments limited
the independent data available for assessing cross-environment transferability. Therefore,
we focused on comparing EN, a regularized linear regression approach, and RF, a nonlinear
tree-based ensemble, as complementary benchmarks suited to the structure and size of
the dataset. Nevertheless, the exclusion of deep learning represents a limitation of the
model comparison and should not be interpreted as evidence of its inferior predictive per-
formance. Deep learning architectures, particularly convolutional neural networks (CNNs)
and temporal models such as long short-term memory (LSTM) networks, offer opportuni-
ties to capture spatial patterns and temporal dynamics that may not be fully represented by
extracted vegetation indices. Future studies should evaluate these approaches alongside
conventional machine learning models using larger datasets encompassing independent
genotypes and a broader range of environments, while applying consistent genotype- and
environment-aware cross-validation strategies to ensure robust comparisons of predictive
accuracy and generalization capacity [44,45].
4.4. Interpretation and Implications for Quinoa Breeding
The two-stage analysis predicts adjusted genotype-level means, rather than raw plot
yield. BLUE standard errors differ across environments and, for incompletely observed
entries, within environments. The primary models treat the estimated responses equally;
inverse-SE2 sensitivity weights account for marginal precision during training but do not
propagate full first-stage covariance or target-estimation uncertainty. Moreover, the across-
environment H2 summary assumes homogeneous residual variance despite differing resid-
ual scales. These limitations constrain the strength of genetic and physiological interpretation.
Grain yield was determined on an as-harvested basis without measuring or adjusting
grain-moisture content. Consequently, differences in absolute yield and prediction errors
may reflect variation in both grain dry matter and moisture content, although their relative
contributions could not be quantified. Furthermore, all genotypes were harvested on a
common date, approximately one month after physiological maturity at the trial level, de-
spite potential differences in maturity dates among genotypes. Thus, the interval between
physiological maturity and harvest may have varied among genotypes, potentially influenc-
ing grain moisture content at harvest. The available data do not allow these effects to be
quantified or distinguished from differences in dry matter accumulation. Accordingly, the
estimated environmental and G × E variance components reflect variation in the recorded
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 19 of 23
grain yield and cannot be attributed exclusively to differences in dry matter production or
specific environmental factors.
The validation strategy also determines the breeding scenarios in which the predictive
performance of these models can be reliably applied. Because LOEO separated the data
by environment rather than by genotype, it mainly evaluated the transfer of previously
observed genotypes to a new environment. Predicting newly introduced genotypes would
require genotype-disjoint validation, whereas the most stringent evaluation would exclude
both the test genotypes and the test environment from model training. For breeding ap-
plications, model usefulness should be assessed not only by R2 and RMSE but also by
genotype-ranking ability, overlap among the selected top-performing fractions, selection
differential, and expected response to selection. Because the panel consisted mainly of ac-
cessions and collected populations rather than pedigree-defined breeding lines, the present
results characterize predictive performance within a coastal/lowland diversity panel and
do not evaluate transfer among known families or crossing-derived lineages. Combining
genomic information, which connects untested material, with UA V observations, which
describe realized crop condition, and environmental covariates is a promising next step for
prediction-oriented breeding [57].
Within-environment cross-validation provides a retrospective assessment of predictive
performance, as grain yield observations from other genotypes within the same site–year
are available for model training. Consequently, this approach does not demonstrate the
ability to predict grain yield before harvest when no yield measurements are available
from the target environment. Implementing UA V-based yield prediction as a pre-harvest
selection tool would require models calibrated using historical or independent trials, fol-
lowed by prospective validation of predictive accuracy and the effectiveness of genotype
selection. Nevertheless, the present results demonstrate the potential of UA V-derived traits
to characterize genotypic variation in grain yield and support the identification and priori-
tization of promising genotypes for further evaluation, provided that appropriate model
calibration is available. Future research should prioritize standardized UA V acquisition and
image-processing protocols, expanded datasets encompassing independent site–year envi-
ronments, and modeling approaches that explicitly account for environmental variation and
prediction uncertainty. These advances will be essential to improve model transferability
and establish the practical value of UA V-based phenotyping for quinoa breeding programs.
5. Conclusions
Multisensor and multitemporal UA V phenotyping provided moderate predictive accu-
racy for quinoa grain yield within individual environments, with mean R2 values of 0.49 for
elastic net (EN) and 0.51 for random forest (RF). RGB-derived indices, NDVI, and canopy
temperature traits contributed complementary information to yield prediction, with the
highest predictive performance observed in the irrigated SR_2425 environment, which also
exhibited the greatest variation in grain yield. However, because phenological development
differed among genotypes, UA V acquisition dates represented reference flight times rather
than synchronized developmental stages. In contrast to within-environment predictions,
leave-one-environment-out validation revealed limited transferability of absolute yield
predictions to unseen site–year environments, although positive associations between ob-
served and predicted genotype rankings were maintained in some cases. These findings
indicate that UA V-derived traits can support within-environment genotype discrimination
when yield observations are available for model calibration, but predictive performance
under within-environment cross-validation does not guarantee reliable application to new
environments without local calibration. Improving cross-environment prediction will re-
quire more comprehensive environmental characterization, integration of environmental
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 20 of 23
covariates into predictive models, and larger datasets encompassing independent site–year
environments. Such advances, combined with rigorous environment-wise validation, will
be essential for developing robust and transferable UA V-based prediction tools to support
quinoa breeding and selection under contrasting Mediterranean growing conditions.
Author Contributions: Conceptualization, K.G. and A.d.P .; methodology, K.G. and A.d.P .; software,
F.F.-P .; validation, F.F.-P ., K.G. and A.d.P .; formal analysis, K.G. and F.F.-P .; investigation, K.G. and
A.d.P .; data curation, K.G., J.F.A.-Q., I.M., D.C., J.P . and S.S.; writing—original draft preparation,
K.G. and A.d.P .; writing—review and editing, K.G. and A.d.P .; supervision, A.d.P . and I.M.; fund-
ing acquisition, A.d.P . and I.M. All authors have read and agreed to the published version of the
manuscript.
Funding: This research was funded by the ANILLO project ATE220001 (Agencia Nacional de Investi-
gación y Desarrollo—ANID) and Beca de Doctorado Nacional (Agencia Nacional de Investigación y
Desarrollo—ANID), grant number 21212122.
Data Availability Statement: The raw data supporting the conclusions of this article will be made
available by the authors on request.
Conflicts of Interest: The authors declare no conflicts of interest.
References

1. Nazeer, S.; Akram, M.Z. Elucidating Genotypic Variation in Quinoa via Multidimensional Agronomic, Physiological, and
   Biochemical Assessments. Plants 2025, 14, 2332. [CrossRef] [PubMed]
2. Alandia, G.; Rodriguez, J.P .; Jacobsen, S.E.; Bazile, D.; Condori, B. Global Expansion of Quinoa and Challenges for the Andean
   Region. Glob. Food Secur. 2020, 26, 100429. [CrossRef]
3. Afzal, I.; Basra, S.M.A.; Rehman, H.U.; Iqbal, S.; Bazile, D. Trends and Limits for Quinoa Production and Promotion in Pakistan.
   Plants 2022, 11, 1603. [CrossRef]
4. Fathi, A.; Kardoni, F. The Importance of Quinoa (Quinoa Chenopodium Willd.) Cultivation in Developing Countries: A Review.
   Cercet. Agron. Mold. 2020, 53, 337–356. [CrossRef]
5. Hafeez, M.B.; Iqbal, S.; Li, Y.; Saddiq, M.S.; Basra, S.M.A.; Zhang, H.; Zahra, N.; Akram, M.Z.; Bertero, D.; Curti, R.N. Assessment
   of Phenotypic Diversity in the USDA Collection of Quinoa Links Genotypic Adaptation to Germplasm Origin. Plants 2022, 11,
6. [CrossRef] [PubMed]
7. Böndel, K.; Schmid, K. Quinoa Diversity and Its Implications for Breeding. In The Quinoa Genome; Schmöckel, S.M., Ed.; Springer
   International Publishing: Cham, Switzerland, 2021; pp. 107–118.
8. Martínez, E.A.; Fuentes, F.F.; Bazile, D. History of Quinoa: Its Origin, Domestication, Diversification, and Cultivation with Particular
   Reference to the Chilean Context; John Wiley and Sons: Hoboken, NJ, USA, 2015.
9. Akram, A.; Abdullah, M.; Haider, D.; Tariq, T.; Ali, Q.; Hayat, M.K.; Haseeb, A.; Yasin, F.; Iqbal, A. Thriving in Adversity: Quinoa’s
   Response to Drought and the Promise of Optimized Watering. Agrobiol. Rec. 2025, 21, 78–98. [CrossRef] [PubMed]
10. Pradhan, A.; Rane, J.; Sammi Reddy, K. Inclusion of Quinoa in Cropping Systems for Ensuring Food and Nutrition Security in
    Drought Prone Semi-Arid Regions: An Energy-Water-Carbon-Food Nexus Approach. Energy 2025, 330, 136892. [CrossRef]
11. Maestro-Gaitán, I.; González-Bodí, S.; Hassinger-Lino, G.R.; Redondo-Nieto, M.; Rodríguez-Casillas, L.; Matías, J.; Zamarreño,
    Á.M.; García-Mina, J.M.; Granado-Rodríguez, S.; Bolaños, L.; et al. Transcriptional Reprogramming under Drought Reveals
    Divergent Adaptive Strategies in Quinoa. BMC Plant Biol. 2025, 25, 1718. [CrossRef] [PubMed]
12. Tramblay, Y.; Koutroulis, A.; Samaniego, L.; Vicente-Serrano, S.M.; Volaire, F.; Boone, A.; Le Page, M.; Llasat, M.C.; Albergel, C.;
    Burak, S.; et al. Challenges for Drought Assessment in the Mediterranean Region under Future Climate Scenarios. Earth-Sci. Rev.
    2020, 210, 103348. [CrossRef]
13. Nguyen, V .L.; Luu, H.N.; Phan, T.H.N.; Nguyen, V .L.; Chu, D.H.; Bertero, D.; Curti, N.; McKeown, P .C.; Spillane, C. Genotype by
    Environment Interaction across Water Regimes in Relation to Cropping Season Response of Quinoa (Chenopodium Quinoa). PLoS
    ONE 2024, 19, e0309777. [CrossRef] [PubMed]
14. Jokarfard, V .; Rabiei, B.; Laki, E.S.; Börner, A. Stability and Adaptability of Grain Yield in Quinoa Genotypes in Four Locations of
    Iran. Front. Plant Sci. 2024, 15, 1487106. [CrossRef] [PubMed]
15. Olguín, P .; Contreras, S.; Rojas, C.; Fuentes, F. Genotype-by-Environment Interaction Stability Analysis of New Quinoa
    (Chenopodium Quinoa Willd.) Varieties in the Mediterranean Zone of Chile. Plants 2025, 14, 3007. [CrossRef]
    https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 21 of 23
15. Madrid, D.; Salgado, E.; Verdugo, G.; Olguín, P .; Bilalis, D.; Fuentes, F. Morphological Traits Defining Breeding Criteria for Coastal
Quinoa in Chile. Not. Bot. Horti Agrobot. Cluj-Napoca 2018, 46, 190–196. [CrossRef]
16. Khuimphukhieo, I.; da Silva, J.A. Unmanned Aerial Systems (UAS)-Based Field High Throughput Phenotyping (HTP) as Plant
Breeders’ Toolbox: A Comprehensive Review. Smart Agric. Technol. 2025, 11, 100888. [CrossRef]
17. Gill, T.; Gill, S.K.; Saini, D.K.; Chopra, Y.; de Koff, J.P .; Sandhu, K.S. A Comprehensive Review of High Throughput Phenotyping
and Machine Learning for Plant Stress Phenotyping. Phenomics 2022, 2, 156–183. [CrossRef]
18. Carrasco, G.; Fuentes-Penailillo, F.; Perez, R.; Rebolledo, P .; Manriquez, P . An Approach to a Vertical Farming Low-Cost to Reach
Sustainable Vegetable Crops. In Proceedings of the 2022 IEEE International Conference on Automation/25th Congress of the Chilean
Association of Automatic Control: For the Development of Sustainable Agricultural Systems; ICA-ACCA 2022; Institute of Electrical and
Electronics Engineers Inc.: New York, NY, USA, 2022.
19. Volpato, L.; Pinto, F.; González-Pérez, L.; Thompson, I.G.; Borém, A.; Reynolds, M.; Gérard, B.; Molero, G.; Rodrigues, F.A.
High Throughput Field Phenotyping for Plant Height Using UA V-Based RGB Imagery in Wheat Breeding Lines: Feasibility and
Validation. Front. Plant Sci. 2021, 12, 591587. [CrossRef] [PubMed]
20. Feng, L.; Chen, S.; Zhang, C.; Zhang, Y.; He, Y. A Comprehensive Review on Recent Applications of Unmanned Aerial Vehicle
Remote Sensing with Various Sensors for High-Throughput Plant Phenotyping. Comput. Electron. Agric. 2021, 182, 106033.
[CrossRef]
21. Ali, N.; Bais, A.; Sangha, J.S.; Cuthbert, R.D.; Ruan, Y. High-Throughput UA V Phenotyping for Plot-Level Harvest Index Estimation
in Wheat Fields. Curr. Plant Biol. 2026, 45, 100582. [CrossRef]
22. Fuentes-Peñailillo, F.; Ortega-Farías, S.; de la Fuente-Saiz, D.; Rivera, M. Digital Count of Sunflower Plants at Emergence from
Very Low Altitude Using UA V Images. In Proceedings of the 2019 IEEE CHILEAN Conference on Electrical, Electronics Engineering,
Information and Communication Technologies (CHILECON); IEEE: New York, NY, USA, 2019.
23. Shu, M.; Fei, S.; Zhang, B.; Yang, X.; Guo, Y.; Li, B.; Ma, Y. Application of UA V Multisensor Data and Ensemble Approach for
High-Throughput Estimation of Maize Phenotyping Traits. Plant Phenomics 2022, 2022, 9802585. [CrossRef] [PubMed]
24. Yang, N.; Zhang, Z.; Yang, X.; Dong, N.; Xu, Q.; Chen, J.; Sun, S.; Cui, N.; Ning, J. Evaluation of Crop Water Status Using
UA V-Based Images Data with a Model Updating Strategy. Agric. Water Manag. 2025, 312, 109445. [CrossRef]
25. Liu, Y.; Fan, Y.; Yue, J.; Ma, Y.; Yang, F.; Fan, J.; Chen, R.; Bian, M.; Feng, H. UA V-Based Remote Sensing Estimation of above-Ground
Biomass in Different Crops: A Review. Int. J. Appl. Earth Obs. Geoinf. 2025, 144, 104938. [CrossRef]
26. Nițu, A.; Florea, C.; Ivanovici, M.; Racoviteanu, A. NDVI and Beyond: Vegetation Indices as Features for Crop Recognition and
Segmentation in Hyperspectral Data. Sensors 2025, 25, 3817. [CrossRef] [PubMed]
27. Deiss, L.; Batsaikhan, B.; Lavrenko, S.; Duarte Farias, G.; de Lima Menezes, J.; Alves, L.A.; Tiecher, T.; Martins, A.P .; Fioravan ço,
G.P .; da Silva, G.P .; et al. Normalized Difference Vegetation Index (NDVI) for Soybean Biomass and Nutrient Uptake Estimation
in Response to Production Systems and Fertilization Strategies. Front. Sustain. Food Syst. 2023, 6, 959681. [CrossRef]
28. Melandri, G.; Prashar, A.; McCouch, S.R.; Van Der Linden, G.; Jones, H.G.; Kadam, N.; Jagadish, K.; Bouwmeester, H.; Ruyter-Spira,
C. Association Mapping and Genetic Dissection of Drought-Induced Canopy Temperature Differences in Rice. J. Exp. Bot. 2020,
71, 1614–1627. [CrossRef]
29. Chandel, N.S.; Rajwade, Y.A.; Dubey, K.; Chandel, A.K.; Subeesh, A.; Tiwari, M.K. Water Stress Identification of Winter Wheat Crop
with State-of-the-Art AI Techniques and High-Resolution Thermal-RGB Imagery. Plants 2022, 11, 3344. [CrossRef] [PubMed]
30. Giménez-Gallego, J.; González-Teruel, J.D.; Blaya-Ros, P .J.; Toledo-Moreo, A.B.; Domingo-Miguel, R.; Torres-Sánchez, R. Automatic
Crop Canopy Temperature Measurement Using a Low-Cost Image-Based Thermal Sensor: Application in a Pomegranate Orchard
under a Permanent Shade Net House. Sensors 2023, 23, 2915. [CrossRef] [PubMed]
31. Yang, X.; Chen, J.; Lu, X.; Liu, H.; Liu, Y.; Bai, X.; Qian, L.; Zhang, Z. Advances in UA V Remote Sensing for Monitoring Crop
Water and Nutrient Status: Modeling Methods, Influencing Factors, and Challenges. Plants 2025, 14, 2544. [CrossRef] [PubMed]
32. Stanschewski, C.S.; Rey, E.; Fiene, G.; Craine, E.B.; Wellman, G.; Melino, V .J.; Patiranage, D.S.R.; Johansen, K.; Schmöckel, S.M.;
Bertero, D.; et al. Quinoa Phenotyping Methodologies: An International Consensus. Plants 2021, 10, 1759. [CrossRef] [PubMed]
33. Sankaran, S.; Espinoza, C.Z.; Hinojosa, L.; Ma, X.; Murphy, K. High-Throughput Field Phenotyping to Assess Irrigation Treatment
Effects in Quinoa. Agrosyst. Geosci. Environ. 2019, 2, 1–7. [CrossRef]
34. Tattaris, M.; Reynolds, M.P .; Chapman, S.C. A Direct Comparison of Remote Sensing Approaches for High-Throughput Pheno-
typing in Plant Breeding. Front. Plant Sci. 2016, 7, 1131. [CrossRef] [PubMed]
35. Dumschott, K.; Wuyts, N.; Alfaro, C.; Castillo, D.; Fiorani, F.; Zurita-Silva, A. Morphological and Physiological Traits Associated
with Yield under Reduced Irrigation in Chilean Coastal Lowland Quinoa. Plants 2022, 11, 323. [CrossRef]
36. Ndour, A.; Blasch, G.; Valente, J.; Gebrekidan, B.H.; Sida, T.S. Optimal Machine Learning Algorithms and UA V Multispectral
Imagery for Crop Phenotypic Trait Estimation: A Comprehensive Review and Meta-Analysis. Environ. Res. Commun. 2025, 7,
072002. [CrossRef]
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 22 of 23
37. Jabed, M.A.; Azmi Murad, M.A. Crop Yield Prediction in Agriculture: A Comprehensive Review of Machine Learning and Deep
Learning Approaches, with Insights for Future Research and Sustainability. Heliyon 2024, 10, e40836. [CrossRef] [PubMed]
38. Shawon, S.M.; Ema, F.B.; Mahi, A.K.; Niha, F.L.; Zubair, H.T. Crop Yield Prediction Using Machine Learning: An Extensive and
Systematic Literature Review. Smart Agric. Technol. 2025, 10, 100718. [CrossRef]
39. Zhou, H.; Yang, J.; Lou, W.; Sheng, L.; Li, D.; Hu, H. Improving Grain Yield Prediction through Fusion of Multi-Temporal Spectral
Features and Agronomic Trait Parameters Derived from UA V Imagery. Front. Plant Sci. 2023, 14, 1217448. [CrossRef] [PubMed]
40. Gade, S.A.; Madolli, M.J.; García-Caparrós, P .; Ullah, H.; Cha-um, S.; Datta, A.; Himanshu, S.K. Advancements in UA V Remote
Sensing for Agricultural Yield Estimation: A Systematic Comprehensive Review of Platforms, Sensors, and Data Analytics. Remote
Sens. Appl. Soc. Environ. 2025, 37, 101418. [CrossRef]
41. Liu, Z.; Di, L.; Yang, R.; Guo, L.; Zhang, C.; Li, H.; Shao, B. In-Season Crop Yield Prediction: State of the Art and Future Research
Direction. Int. J. Appl. Earth Obs. Geoinf. 2026, 146, 105129. [CrossRef]
42. Fei, S.; Hassan, M.A.; Ma, Y.; Shu, M.; Cheng, Q.; Li, Z.; Chen, Z.; Xiao, Y. Entropy Weight Ensemble Framework for Yield
Prediction of Winter Wheat Under Different Water Stress Treatments Using Unmanned Aerial Vehicle-Based Multispectral and
Thermal Data. Front. Plant Sci. 2021, 12, 730181. [CrossRef] [PubMed]
43. Wayal, S.M.; Parab, S.; Raj, A.; Khandagale, K.; Bhegde, S.; Dawale, M.; Bhangare, I.; Khaire, M.; Kadam, Y.; Shaikh, Z.; et al.
UA V Multispectral Sensing and Data-Driven Modeling for Precision Onion Yield Prediction. Front. Plant Sci. 2026, 16, 1696730.
[CrossRef] [PubMed]
44. Zhou, J.; Zhou, J.; Ye, H.; Ali, M.L.; Chen, P .; Nguyen, H.T. Yield Estimation of Soybean Breeding Lines under Drought Stress
Using Unmanned Aerial Vehicle-Based Imagery and Convolutional Neural Network. Biosyst. Eng. 2021, 204, 90–103. [CrossRef]
45. Nevavuori, P .; Narra, N.; Linna, P .; Lipping, T. Crop Yield Prediction Using Multitemporal UA V Data and Spatio-Temporal Deep
Learning Models. Remote Sens. 2020, 12, 4000. [CrossRef]
46. Devkota, M.; Patil, S.B.; Kumar, S.; Kehel, Z.; Wery, J. Performance of Elite Genotypes of Barley, Chickpea, Lentil, and Wheat
under Conservation Agriculture in Mediterranean Rainfed Conditions. Exp. Agric. 2021, 57, 126–143. [CrossRef]
47. De Ollas, C.; Morillón, R.; Fotopoulos, V .; Puértolas, J.; Ollitrault, P .; Gómez-Cadenas, A.; Arbona, V . Facing Climate Change:
Biotechnology of Iconic Mediterranean Woody Crops. Front. Plant Sci. 2019, 10, 427. [CrossRef]
48. Sellami, M.H.; Di Mola, I.; Ottaiano, L.; Cozzolino, E.; De Vita, P .; Mori, M. Assessing Temporal Variability in Durum Wheat
Performance and Stability through Multi-Trait Mean Performance Selection in Mediterranean Climate. Front. Agron. 2024, 6,
1466040. [CrossRef]
49. Kassim, Y.B.; Oteng-Frimpong, R.; Puozaa, D.K.; Sie, E.K.; Abdul Rasheed, M.; Abdul Rashid, I.; Danquah, A.; Akogo, D.A.;
Rhoads, J.; Hoisington, D.; et al. High-Throughput Plant Phenotyping (HTPP) in Resource-Constrained Research Programs: A
Working Example in Ghana. Agronomy 2022, 12, 2733. [CrossRef]
50. Holland, J.; Nyquist, W.; Cervantes-Martinez, C. Estimating and Interpreting Heritability for Plant Breeding: An Update. Plant
Breed. Rev. 2003, 22, 9–112. [CrossRef]
51. Strobl, C.; Boulesteix, A.L.; Zeileis, A.; Hothorn, T. Bias in Random Forest Variable Importance Measures: Illustrations, Sources
and a Solution. BMC Bioinform. 2007, 8, 25. [CrossRef] [PubMed]
52. del Pozo, A.; Ruf, K.; Alfaro, C.; Zurita, A.; Guerra, F.; Sagredo, B. Traits Associated with Higher Productivity and Resilience to
Drought-Prone Mediterranean Environments of Coastal-Lowland Quinoa (Chenopodium Quinoa Willd.). Field Crops Res. 2023,
299, 108985. [CrossRef]
53. Lupa-Condo, N.E.; Lope-Ccasa, F.C.; Salazar-Joyo, A.A.; Gutiérrez-Rosales, R.O.; Jellen, E.N.; Hansen, N.C.; Anculle-Arenas,
A.; Zeballos, O.; Llasaca-Calizaya, N.W.; Mayta-Anco, M.E. Phenotyping for Effects of Drought Levels in Quinoa Using Remote
Sensing Tools. Agronomy 2024, 14, 1938. [CrossRef]
54. Jiang, J.; Johansen, K.; Stanschewski, C.S.; Wellman, G.; Mousa, M.A.A.; Fiene, G.M.; Asiry, K.A.; Tester, M.; McCabe, M.F.
Phenotyping a Diversity Panel of Quinoa Using UA V-Retrieved Leaf Area Index, SP AD-Based Chlorophyll and a Random Forest
Approach. Precis. Agric. 2022, 23, 961–983. [CrossRef]
55. Hassan, M.A.; Yang, M.; Rasheed, A.; Yang, G.; Reynolds, M.; Xia, X.; Xiao, Y.; He, Z. A Rapid Monitoring of NDVI across the
Wheat Growth Cycle for Grain Yield Prediction Using a Multi-Spectral UA V Platform. Plant Sci. 2019, 282, 95–103. [CrossRef]
[PubMed]
56. Fuentes-Peñailillo, F.; del Campo-Hitschfeld, M.L.; Gutter, K.; Torres-Quezada, E. Data-Driven Integration of Remote Sensing,
Agro-Meteorology, and Wireless Sensor Networks for Crop Water Demand Estimation: Tools Towards Sustainable Irrigation in
High-Value Fruit Crops. Agronomy 2025, 15, 2122. [CrossRef]
https://doi.org/10.3390/agriculture16212436

Agriculture 2026, 16, 2436 23 of 23
57. Crossa, J.; Fritsche-Neto, R.; Montesinos-Lopez, O.A.; Costa-Neto, G.; Dreisigacker, S.; Montesinos-Lopez, A.; Bentley, A.R. The
Modern Plant Breeding Triangle: Optimizing the Use of Genomics, Phenomics, and Enviromics Data. Front. Plant Sci. 2021, 12,
651480. [CrossRef] [PubMed]
58. Costa-Neto, G.; Fritsche-Neto, R.; Crossa, J. Nonlinear Kernels, Dominance, and Envirotyping Data Increase the Accuracy of
Genome-Based Prediction in Multi-Environment Trials. Heredity 2021, 126, 92–106. [CrossRef] [PubMed]
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual
author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury
to people or property resulting from any ideas, methods, instructions or products referred to in the content.
https://doi.org/10.3390/agriculture16212436