Can Environmental Analysis Algorithms Be Improved by Data Fusion and Soil Removal for UAV-Based Buffel Grass Biomass Prediction?

Highlights

What are the main findings?

• Removing soil pixels consistently improved the performance of biomass prediction models.
• The model with the highest accuracy (CatBoost) was obtained using only RGB sensor data and the Boruta feature selection method.

What are the implications of the main findings?

• The model with the highest accuracy (CatBoost) was obtained using only RGB sensor data and the Boruta feature selection method.
• Vegetation cover area is a dominant predictor, suggesting that structural canopy metrics are more informative than complex spectral indices for forage biomass estimation.

Abstract

The growing demand for sustainable livestock systems requires efficient methods for monitoring forage biomass. This study evaluated spectral (RGB and multispectral), textural (GLCM), and area attributes derived from unmanned aerial vehicle (UAV) imagery to predict buffelgrass (Cenchrus ciliaris L.) biomass, also testing the effect of soil pixel removal. A comprehensive machine learning pipeline (12 algorithms and 6 feature selection methods) was applied to 14 data combinations. Our results demonstrated that soil removal consistently improved the performance of the applied models. Multispectral (MSI) sensors were the most robust individually, whereas textural (GLCM) attributes did not contribute significantly. Although the MSI and RGB data combination proved complementary, the model with the highest accuracy was obtained with CatBoost using only RGB information after Boruta feature selection, achieving a CCC of 0.83, RMSE of 0.214 kg, and R² of 0.81 in the test set. The most important variable was vegetation cover area (19.94%), surpassing spectral indices. We conclude that integrating RGB UAVs with robust processing can generate accessible and effective tools for forage monitoring. This approach can support pasture management by optimizing stocking rates, enhancing natural resource efficiency, and supporting data-driven decisions in precision silvopastoral systems.

1. Introduction

The growing demand for more productive, efficient, and sustainable agricultural systems has driven the adoption of advanced crop monitoring and management technologies [1,2]. In this context, the non-destructive estimation of forage biomass plays a strategic role, as it provides essential information on productivity, growth dynamics, and the carrying capacity of livestock systems [3,4]. Accurate biomass quantification is a crucial component of pasture management programs, enabling the optimization of stocking rates, enhancing natural resource use efficiency, and supporting data-driven precision agriculture practices [4,5].

In recent decades, advances in remote sensing and unmanned aerial vehicles (UAVs) have considerably advanced the monitoring of plant biomass [6,7]. The use of UAV-mounted sensors capable of capturing images with high spatial and temporal resolution offers an effective alternative to traditional methods, which are costly, destructive, and often unfeasible on a large scale [6,8,9]. In this context, vegetation indices derived from spectral imagery (particularly those based on visible and near-infrared bands) have been widely employed to infer the biophysical and biochemical parameters of crops, including biomass, leaf area, and chlorophyll content [9,10].

In the literature, it is well-established that spectral attributes obtained through remote sensing are directly related to vegetative vigor and biomass accumulation [5,9]. For example, previous studies have demonstrated a strong correlation between vegetation indices (VIs) such as the Normalized Difference Vegetation Index (NDVI) [9], Normalized Green-Red Difference Index (NGRDI), and Visible Atmospherically Resistant Index (VARI) [11], and the productivity of forage grasses. In addition to spectral attributes, structural and textural features extracted from high-resolution imagery have emerged as valuable complementary indicators, capturing spatial variations associated with canopy architecture and vegetation distribution [12,13,14]. Textural analysis based on the Gray-Level Co-occurrence Matrix (GLCM), for instance, has been successfully employed to enhance land cover class discrimination and estimate structural parameters in various agricultural systems [14,15].

Furthermore, methodological strategies have been proposed to mitigate background interference, particularly associated with exposed soil, which can compromise the vegetation’s spectral response in both homogeneous and heterogeneous agricultural scenes [16,17]. In such heterogeneous conditions, the coexistence multiple plant species, variable canopy densities, crop residues, shadows, and soil exposure increases spectral variability and makes vegetation–background separation more challenging [16]. Within this context, the use of thresholds applied to vegetation indices (VIs) stands out as a simple and operationally efficient technique that enables a more precise isolation of the canopy signal and has shown consistent results across different agricultural settings [17,18,19], especially when combined with the vegetation cover fraction [18,20]. Other approaches, such as object-based segmentation, supervised classification, and deep-learning methods, have also been applied to heterogeneous scenes; however, threshold-based procedures remain attractive due to their transparency, low computational cost, and ease of implementation, while still providing competitive performance in practical biomass-estimation workflows [10,16,21,22].

Despite recent advances in the integration of remote sensing, UAVs, and machine learning for estimating biophysical parameters in agricultural crops, their application to biomass quantification in forage species remains relatively underexplored, particularly regarding the number and diversity of species studied [10]. This gap is especially critical in the case of buffelgrass (Cenchrus ciliaris L.), an important species for livestock production in semiarid regions worldwide [23]. Owing to its drought tolerance, buffelgrass exhibits high competitive capacity, good productivity, and adaptability to diverse environmental conditions (e.g., soil types, altitudes, and rainfall ranging from 250 to 2700 mm), as well as a strategic role in the sustainability of production systems [11,23,24]. Therefore, studies that investigate the modeling of biomass in this and other forage grasses based on spectral and textural attributes represent a necessary advancement, with the potential to significantly expand the use of precision agriculture tools in forage crop management [7,10].

However, the optimized application of these technologies presents several challenges. First, although the superiority of multispectral sensors (MSI) is often assumed, RGB (Red, Green, Blue) sensors—when combined with advanced processing and variable selection techniques—can achieve comparable accuracy [25,26]. Second, the relative contribution of different sources of information, including spectral (RGB and MSI), textural (GLCM), and spatial (vegetation cover area) data, must be systematically compared [27]. Finally, a critical limitation of many approaches lies in the inclusion of a large number of predictors (e.g., dozens of VIs), which makes variable selection a fundamental step in the modeling process [28,29].

Accordingly, the objective of this study was to evaluate the accuracy of different combinations of spectral, textural, and area-based attributes derived from UAV imagery for predicting the biomass of buffelgrass, both with and without soil removal. Our hypothesis was that integrating these different sources of information, together with the use of variable selection techniques, would significantly improve the performance of predictive models compared to using each group of predictors in isolation. To achieve this, we implemented a comprehensive machine learning pipeline that tested multiple combinations of datasets and algorithms, as well as different feature selection methods. This approach not only enabled a direct comparison of the predictive potential of spectral, textural, and area-based information but also allowed for the assessment of the effect of soil removal. Consequently, this study provides an accessible and effective framework for forage monitoring, with significant potential to support data-driven decisions and optimize pasture management.

2. Materials and Methods

2.1. Study Area

The study was conducted at the International Reference Center for Agrometeorological Studies in Cactus and Other Forage Crops, located at the Universidade Federal Rural de Pernambuco/Unidade Acadêmica de Serra Talhada (UFRPE/UAST), in Serra Talhada, Pernambuco, Brazil (−7.95458° S, −38.29509° W; Figure 1). The region is characterized by air temperatures ranging from 20.1 to 32.9 °C, an average annual rainfall of 642 mm, and a mean evapotranspiration demand of 1800 mm per year [30].

The experimental plots (6 × 4 m) consisted of areas previously established with irrigated buffalo grass (Cenchrus ciliaris L.). The experimental site is characterized by a Cambissolo Háplico Ta Eutrófico [31], classified as an Inceptisol according to the Soil Taxonomy classification [32]. The physical and chemical characterization of the 0.0–0.40 m layer indicated a pH in water of 6.90 and a soil texture composed of 77% sand, 10% silt, and 13% clay, the soil presented a sum of bases of 6.4 (cmol_c dm⁻³), with individual concentrations of K⁺ 0.61 (cmol_c dm⁻³), Na⁺ (0.04 cmol_c dm⁻³), Ca²⁺ (4.5 cmol_c dm⁻³), and Mg²⁺ (1.20 cmol_c dm⁻³), cation exchange capacity of 7.5 cmol_c dm⁻³, base saturation of 85%. Furthermore, the bulk and particle densities were recorded at 1.45 (g cm⁻³) and 2.57 (g cm⁻³), respectively.

Irrigation management was based on reference evapotranspiration (ET₀), which was calculated using the Penman–Monteith (Equation (1)) approach. Water was supplied through a drip irrigation system operating at 100 kPa, with emitters spaced 0.20 m apart and an average discharge of 1.57 L h⁻¹, presenting an application uniformity coefficient of 92%. Irrigation events were scheduled in the morning on Mondays, Wednesdays, and Fridays. Irrigation depths were applied at 50, 75, 100, and 125% of ET₀. When rainfall exceeded the irrigation depth required, water application was suspended, ensuring full replacement of ET₀. The irrigation water originated from a nearby artesian well and was classified as C3S1, with electrical conductivity of approximately 1.62 dS m⁻¹, sodium and potassium concentrations of 168.66 and 28.17 mg L⁻¹, respectively, and pH of 6.84, indicating high salinity and low sodium hazard [33]. The relationship between rainfall and the 100% ET₀ irrigation depth applied during the experimental period is presented in the Supplementary Material (Figure S1).

$${\mathrm{ET}}_0={\displaystyle \frac{0.408∆\left(\mathrm{Rn}-\mathrm{G}\right)+\mathrm{\gamma }\frac{900}{\mathrm{T}+273}{\mathrm{\mu }}_2({\mathrm{e}}_{\mathrm{s}}-{\mathrm{e}}_{\mathrm{a}})}{∆+\mathrm{\gamma }(1+0.34{\mathrm{\mu }}_2)}}$$

(1)

where ET₀ is the reference evapotranspiration (mm day⁻¹), Rn is the net radiation at the crop surface (MJ m⁻² day⁻¹), G is the soil heat flux (MJ m⁻² day⁻¹), T is the daily average air temperature at 2 m height above ground (°C), μ₂ is the wind speed at 2 m height (m s⁻¹), e_s is the vapor saturation pressure (kPa), e_a is the current vapor pressure (kPa), Δ is the slope of the vapor pressure versus temperature curve (kPa °C⁻¹), and γ is the psychrometric constant (kPa °C⁻¹).

Each plot was composed of four cultivation rows, but only the two central rows were considered for sampling to minimize border effects. Furthermore, each central row was subdivided into three equal segments (2 m each), resulting in the experimental plots (Figure 2). Thus, a total of six sampled units were obtained per plot, corresponding to the three segments of each of the two central rows. In cases where the row length was insufficient due to plant growth, subdivision was performed into two parts while maintaining representative sampling. Harvesting was carried out in full for each experimental unit, and the plant material was immediately weighed in the field to obtain fresh biomass (kg).

Two sampling events were conducted: the first on 5 January 2025 (90 experimental plots, 72-day cycle, starting on 25 October 2024), and the second on 3 June 2025 (94 experimental plots, 53-day cycle, starting on 11 April 2025). The average values for each treatment and date are described in

Table 1. Average values for each treatment and date.

Date	Irrigation Depths (% ET0)	Biomass (kg)
		per Treatment	per Date
5 January 2025	50	0.43	0.71
	75	0.78
	100	0.72
	125	0.90
3 June 2025	50	0.84	0.92
	75	0.99
	100	0.96
	125	0.88

Irrigation management was based on reference evapotranspiration (ET₀).

.

2.2. Aerial Imaging and Orthomosaic Generation

Aerial images were acquired on the day preceding biomass collection, specifically on 6 January 2025, and 3 June 2025 (Figure 1). Image acquisition was performed using a Mavic 3M UAV (DJI, Shenzhen, China) equipped with an RGB sensor (4/3 CMOS: 20 MP, 5280 × 3956 px) and four multispectral (MSI) sensors (1/2.8 CMOS: 5 MP, 2592 × 1944 px). The MSI bands included Green (G): 560 ± 16 nm, Red (R): 650 ± 16 nm, Red Edge (RE): 730 ± 16 nm, and Near-Infrared (NIR): 860 ± 26 nm. The UAV also featured a GNSS system (GPS + Galileo + BeiDou + GLONASS) and an RTK module (horizontal accuracy: 1 cm + 1 ppm; vertical accuracy: 1.5 cm + 1 ppm).

In this study, flight planning was conducted using the DJI Pilot 2 application (v.10.1.0.30), with flight parameters set at an altitude of 30 m, 80% front and side overlap, and a speed of 2 m s⁻¹. Flights were performed between 11:00 a.m. and 12:00 p.m. under clear skies and uniform lighting conditions. Ten ground control points (GCPs) were used for positional correction of the images. Radiometric calibration of the MSI bands was carried out using the integrated solar sensor. Orthomosaic generation, along with all geometric and radiometric correction steps, was performed in Agisoft Metashape software (v.2.1.2, Agisoft LLC, St. Petersburg, Russia) (Table S1). The final spatial resolution of the orthomosaics was approximately 0.8 cm for RGB and 1 cm for MSI.

Removal of Soil and Vegetation Cover Fraction

The RGB and MSI orthomosaics were imported into QGIS software (v.3.28), and polygons corresponding to each experimental unit were manually delineated based on the boundaries of the planting rows (Figure 1e and Figure 2). For each polygon, the total area (m², area_total), vegetated area (m², area_crop), and bare soil area (m², area_soil) were calculated based on pixel classification. The separation between vegetation and soil was performed using predefined spectral index thresholds: NGRDI < −0.05 for RGB images and NDVI < 0.35 for MSI images. These threshold values were established through detailed visual inspection of the imagery. Including the total area allowed for representation of the effective size of each plot, acknowledging that buffelgrass grows without physical constraints imposed by the previously established planting rows.

2.3. Vegetation Indices

A total of 34 vegetation indices were calculated, including 18 derived from RGB bands and 16 from MSI bands (

Table 2. Vegetation indices and equations used.

Sensor	Index	Abbr.	Equations	References
MSI	Normalized Difference Vegetation Index	NDVI	(NIR − R)/(NIR − R)	[39]
	Normalized difference red edge	NDRE	(NIR − RE)/(NIR + RE)	[40]
	Normalized NIR index	NNIR	NIR/(NIR + RE + G)	[41]
	Ratio Vegetation Index	RVI	NIR/R	[42]
	Green Normalized Difference Vegetation Index	GNDVI	(NIR − G)/(NIR + G)	[43]
	Modified Chlorophyll Absorption in Reflectance Index	MCARI	((RE − R) − 0.2(RE − G)) × (RE/R)	[44]
	MERIS total chlorophyll index	MTCI	(NIR − RE)/(RE − R)	[12]
	Triangular Vegetation Index	TVI	0.5(120(NIR − G) − 200(R − G))	[45]
	Spectral Feature Depth Vegetation Index	SFDVI	((NIR + G)/2) − ((R + RE)/2)	[46]
	Soil adjusted vegetation index	SAVI	(NIR − R)(1 + 0.5)/(NIR + R + 0.5)	[47]
	Green optimal soil adjusted vegetation index	GOSAVI	(1 + 0.16) × (NIR − G)/(NIR + G + 0.16)	[48]
	Normalized green red difference index	NGRDIM	(G − R)/(G − R)	[49]
	Green-red ratio index	GRRIM	G/R	[50]
	Optimized Soil Adjusted Vegetation Index-Green	OSAVIgreenM	(1.5(G − R))/((G + R) + 0.16)	[41]
	Enhanced Vegetation Index 2-Green	EVI2greenM	(2.5(G − R))/(G + 2.4R + 1)	[41]
	Excess Red Vegetation Index	ExRM	1.4R − G	[51]
RGB	Normalized green red difference index	NGRDI	(G − R)/(G + R)	[49]
	Green-red ratio index	GRRI	G/R	[50]
	Optimized Soil Adjusted Vegetation Index-Green	OSAVIgreen	(1.5(G − R))/((G + R) + 0.16)	[41]
	Enhanced Vegetation Index 2-Green	EVI2green	(2.5(G − R))/(G + 2.4R + 1)	[41]
	Excess Red Vegetation Index	ExR	1.4R − G	[51]
	Excess Green vegetation index	ExG	2G − R − B	[51]
	Excess Blue Vegetation Index	ExB	1.4B − G	[51]
	Normalized Difference Vegetation Index RGB	NDVIrgb	((G + B) − R)/((G + B) + R)	[41]
	Green leaf index	GLI	(2G − R − B)/(2G + R + B)	[52]
	Normalized pigment chlorophyll ratio index	NPCI	(R − B)/(R + B)	[53]
	Visible atmospherically resistant index	VARI	(G − R)/(G + R − B)	[54]
	Woebbecke Index	WI	(G − B)/(R − B)	[51]
	Normalized Red	Rn	R/(R + G + B)	[50]
	Normalized Green	Gn	G/(R + G + B)	[50]
	Normalized Blue	Bn	B/(R + G + B)	[50]
	Color Intensity Index	INT	(R + G + B)/3	[55]
	Brightness Index	BI	((R2 + G2 + B2)/3)0.5	[56]
	Overall Hue Index	HUE	atan(2(B − G − R)/30.5(G − R))	[38]

Abbr.: Abbreviation; RGB index—R: Red, B: Blue, G: Green; MSI index—G: Green (560 ± 16 nm), R: Red (650 ± 16 nm), NIR: Near-infrared (860 ± 26 nm), and RE: Red edge (730 ± 16 nm).

). Some indices were adapted by band equivalence, allowing their calculation using either MSI or RGB bands. The mean value of each index was extracted for every delineated polygon, both with and without prior removal of soil pixels, to assess the effect of soil interference on the predictive models. All computations were performed in R software (v.4.3.2) [34], using the package raster (v.3.6-32) [35], terra (v.1.8-70) [36], sf (v.1.0-21) [37] and FieldImageR (v.0.6.0) [38].

2.4. Textural Analysis Using the Gray-Level Co-Occurrence Matrix (GLCM)

Textural metrics were derived from the RGB orthomosaics using the Gray-Level Co-occurrence Matrix (GLCM) approach [57]. The matrices were computed with a one-pixel displacement in four directions (i.e., 0°, 45°, 90°, and 135°) within an 11 × 11 moving window. The mean across directions was then calculated to ensure rotational invariance. Prior to analysis, original images with 256 gray levels were quantized into 32 gray levels to reduce computational cost and noise. The extracted metrics included contrast (CO), mean (ME), entropy (ENT), correlation (COR), variance (VAR), angular second moment (ASM), and maximum probability (MaxProb) (

Table 3. Description of the GLCM metrics.

Metrics	Description
Contrast (CO)	${\displaystyle \sum _{\mathrm{i},\mathrm{j}}^{\mathrm{N}-1}}{{\mathrm{P}}_{\mathrm{i},\mathrm{j}}\left(\mathrm{i}-\mathrm{j}\right)}^2$
Mean (ME)	${\displaystyle \sum _{\mathrm{i},\mathrm{j}}^{\mathrm{N}-1}}\mathrm{i}\times {\mathrm{P}}_{\mathrm{i},\mathrm{j}}$
Entropy (ENT)	${\displaystyle \sum _{\mathrm{i},\mathrm{j}}^{\mathrm{N}-1}}{\mathrm{P}}_{\mathrm{i},\mathrm{j}}(-\ln {\mathrm{P}}_{\mathrm{i},\mathrm{j}})$
Correlation (COR)	${\displaystyle \sum _{\mathrm{i},\mathrm{j}}^{\mathrm{N}-1}}{\mathrm{P}}_{\mathrm{i},\mathrm{j}}\left[{\displaystyle \frac{(\mathrm{i}-{\mathrm{\mu }}_{\mathrm{i}}\left)\right(\mathrm{j}-{\mathrm{\mu }}_{\mathrm{j}})}{\sqrt{{{\mathrm{\sigma }}_{\mathrm{i}}}^2{{\mathrm{\sigma }}_{\mathrm{j}}}^2}}}\right]$
Variance (VAR)	${\displaystyle \sum _{\mathrm{i},\mathrm{j}}^{\mathrm{N}-1}}{{\mathrm{P}}_{\mathrm{i},\mathrm{j}}\left(\mathrm{i}-{\mathrm{\mu }}_{\mathrm{i}}\right)}^2$
Angular Second Moment (ASM)	${\displaystyle \sum _{\mathrm{i},\mathrm{j}}^{\mathrm{N}-1}}{{\mathrm{P}}_{\mathrm{i},\mathrm{j}}}^2$
Maximum Probability (MaxProb)	$\underset{\mathrm{i},\mathrm{j}}{\max }{\mathrm{P}}_{\mathrm{i},\mathrm{j}}$

P_i,j: normalization of the matrix that represents an approximation of the probability of values i and j occurring in adjacent pixels within the defined window, i: value of a target pixel, j: value of the neighbor of pixel i, and N: number of rows or columns.

). Mean values of these metrics were extracted per polygon, both with and without the exclusion of soil pixels, to ensure comparability between the two scenarios. All analyses were performed in Python (v.3.10.18) [58], using the rasterio (v.1.4.3) [59], numpy (v.1.26.4) [60], and numba (v.0.61.0) [61] libraries.

2.5. Construction of the Datasets

Based on the extracted attributes, datasets were constructed for biomass modeling. Each dataset consistently incorporated area information (area_total, area_crop and area_soil) associated with each polygon. Seven primary combinations of data sources were created: (1) RGB indices and bands, (2) MSI indices and bands, (3) GLCM metrics, (4) RGB + GLCM, (5) RGB + MSI, (6) MSI + GLCM, and (7) RGB + MSI + GLCM. Each combination was produced in two versions: (i) with soil pixel removal and (ii) without soil pixel removal. This resulted in a total of 14 distinct datasets, enabling the evaluation of the impact of soil presence, different information sources, and their combinations on biomass modeling (

Table 4. Description of the datasets constructed for biomass modeling.

Nº	Data Sources	RGB Indices and Bands	MSI Indices and Bands	GLCM Metrics	Soil Pixel Removal
1	RGB	Yes	No	No	No
2	RGBrsoil	Yes	No	No	Yes
3	MSI	No	Yes	No	No
4	MSIrsoil	No	Yes	No	Yes
5	GLCM	No	No	Yes	No
6	GLCMrsoil	No	No	Yes	Yes
7	RGB + GLCM	Yes	No	Yes	No
8	RGBrsoil + GLCMrsoil	Yes	No	Yes	Yes
9	RGB + MSI	Yes	Yes	No	No
10	RGBrsoil + MSIrsoil	Yes	Yes	No	Yes
11	MSI + GLCM	No	Yes	Yes	No
12	MSIrsoil + GLCMrsoil	No	Yes	Yes	Yes
13	RGB + MSI + GLCM	Yes	Yes	Yes	No
14	RGBrsoil + MSIrsoil + GLCMrsoil	Yes	Yes	Yes	Yes

GLCM: Gray-Level Co-occurrence Matrix, MSI: Multispectral, RGB: Red, Green, and Blue, rsoil: soil removed.

).

2.6. Predictive Modeling of Biomass Using Machine Learning

The analytical pipeline was structured to ensure reproducibility, standardization, and comparability across models. Initially, the 14 datasets derived from combinations of predictors (Table 4) were partitioned into training (70%) and testing (30%) subsets. Stratified sampling ensured that the distribution of the response variable was preserved in both subsets. To prevent data leakage, feature scaling was performed using Min-Max normalization. The normalization parameters (minimum and maximum values) were estimated exclusively from the training set and subsequently applied to scale both the training and testing subsets. The test subset remained isolated throughout the entire training and optimization process, being used exclusively for the final generalization assessment.

Twelve regression algorithms were evaluated, representing different modeling paradigms: regularized linear regressions (Lasso, Ridge, Elastic Net); instance-based methods such as K-Nearest Neighbors (KNN); Support Vector Machines with Radial Basis Function (RBF-SVM) and Linear (L-SVM) kernels; and ensemble methods, including Decision Tree (DT), Random Forest (RF), Extra Trees (ET), Extreme Gradient Boosting (XGB), Light Gradient Boosting Machine (LGBM), and CatBoost. This diversity allowed for the exploration of both linear and non-linear relationships between the predictors and biomass.

To mitigate the performance bias associated with default parameters and ensure an equitable comparison among the algorithms, we performed a hyperparameter optimization process using k-fold cross-validation (k = 5) on the training set. As the hyperparameter optimization process can be computationally costly, particularly due to the large number of models, a small space of 20 parameter combinations (

Table 5. Tuned hyperparameters.

Models	Engines (Packages)	Hyperparameters
Ridge	glmnet	penalty
Lasso	glmnet	penalty
Elastic Net	glmnet	penalty, mixture
SVM Linear	kernlab	cost, margin
SVM Kernel RBF	kernlab	cost, rbf_sigma
K-Nearest Neighbors	kknn	neighbors
Random Forest	ranger	trees
Extra Trees	ranger	trees
Decision Trees	rpart	cost_complexity
XGBoost	xgboost	trees, learn_rate
LightGBM	lightgbm	trees, learn_rate
CatBoost	catboost	trees, learn_rate

) was selected, obtained through a maximum entropy space-filling design, to balance computational efficiency and search space coverage. The combination that minimized the mean Root Mean Square Error (RMSE) in the cross-validation was selected. The final model for each algorithm was then fitted using the entire training set.

2.7. Feature Selection

Given the high number of features extracted from the images, a feature selection step was incorporated to mitigate multicollinearity and reduce dimensionality, seeking to identify an optimized subset of predictor variables. Six feature selection methods were compared, divided into three groups: filter, wrapper, and embedded:

• Filter methods:

(1)

Correlation-based Elimination (Corr): Features with an absolute Pearson correlation > 0.8 were grouped using hierarchical clustering, retaining only the one most correlated with the response variable. In case of a tie, the secondary criterion was the feature with the highest variance, as it contains more information.

(2)

RReliefFF Algorithm: A method that estimates attribute quality based on its ability to discriminate neighboring instances [62,63]. As RReliefFF does not have a pre-defined threshold, features with a normalized importance > 40% of the maximum observed were selected after preliminary analyses in this study. Although this threshold was arbitrarily defined, we believe that setting it iteratively would deviate from the method’s intent.

• Wrapper methods:

(3)

Recursive Feature Elimination (RFE): A method that iteratively removes features until the best subset of variables is identified [64]. RF was used as the base model for evaluation with 5-fold cross-validation, choosing the subset that minimized the RMSE (root mean square error).

(4)

Boruta: A method built upon RF, which compares the importance of real features with “shadow features” (randomly permuted copies of the original features) [65]. Only features with statistical significance (p-value < 0.01) were retained.

• Embedded:

(5)

Lasso Regression: Through L1 penalization (i.e., the sum of the absolute values of the coefficients), the coefficients of less relevant features are shrunk to zero [66].

(6)

Ridge Regression: Through the L2 penalty (i.e., the sum of the squared coefficients), it reduces the coefficients to be close to zero [66]. Similarly to RReliefF, we defined a threshold of 40% of the maximum value for the normalized coefficients. For both Lasso and Ridge, 5-fold cross-validation was employed.

2.8. Evaluation and Selection of Integration Methods, Models, and Features

The evaluation process was based on the seven previously defined dataset combinations. A structured stepwise analytical strategy was adopted (Figure 3):

• In Step 1, the individual data sources (RGB, MSI, and GLCM) were modeled both with and without soil pixel removal. This procedure allowed the assessment of whether soil removal improved predictive performance and, consequently, the selection of the most appropriate scenario regarding this source of interference.
• In Step 2, only the datasets retained from Step 1 were subjected to feature selection. Different feature selection techniques (filter, wrapper, and embedded) were applied, followed by new modeling runs. This step enabled the identification of the feature selection method that most effectively optimized the predictive performance for the selected scenario.
• In Step 3, the combinations of data sources were evaluated (RGB + GLCM, RGB + MSI, MSI + GLCM and RGB + GLCM + MSI). Two groups of combined datasets were constructed: one obtained directly from the original datasets without feature selection, and another obtained from the individual datasets already optimized in Step 2 through feature selection. Both groups of combinations were modeled and compared using machine learning algorithms. This approach enabled the simultaneous assessment of the effect of data fusion and the prior feature selection step, supporting the identification of the most appropriate dataset combination and the most accurate predictive models for biomass estimation.

Figure 3. Flowchart of the selection process for integration methods, modeling, and features. GLCM: Gray-Level Co-occurrence Matrix, MSI: Multispectral, RGB: Red, Green, and Blue, rsoil: soil removed.

Figure 3. Flowchart of the selection process for integration methods, modeling, and features. GLCM: Gray-Level Co-occurrence Matrix, MSI: Multispectral, RGB: Red, Green, and Blue, rsoil: soil removed.

All processing steps were performed strictly using the training set and internal cross-validation. The test subset remained completely isolated throughout the entire modeling and optimization workflow, being used exclusively for the final evaluation of the models’ generalization performance.

To quantify predictive accuracy, metrics widely recognized in the regression modeling literature were adopted. Lin’s concordance correlation coefficient [67] (CCC, Equation (2)) was used to simultaneously assess precision and accuracy, providing a robust measure of agreement between observed and estimated values. The mean absolute error (MAE, Equation (3)) and the root mean square error (RMSE, Equation (4)) were employed to measure the magnitude of deviations at different scales, allowing for the characterization of both mean discrepancies and the stricter penalization of large errors. Additionally, the mean absolute percentage error (MAPE, Equation (5)) was considered as a relative metric, enabling the interpretation of model performance in terms proportional to the variability of the observed data.

$$\mathrm{CCC}={\displaystyle \frac{2{\mathrm{\rho }\mathrm{\sigma }}_1{\mathrm{\sigma }}_2}{{\mathrm{\sigma }}_1^2+{\mathrm{\sigma }}_2^2+{\left({\mathrm{\mu }}_1-{\mathrm{\mu }}_2\right)}^2}}$$

(2)

$$\mathrm{MAE}={\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}\left|{\mathrm{y}}_{\mathrm{i}}-{\text{ŷ}}_{\mathrm{i}}\right|$$

(3)

$$\mathrm{RMSE}=\sqrt{{\displaystyle \frac{1}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}{\left({\mathrm{y}}_{\mathrm{i}}-{\mathrm{ŷ}}_{\mathrm{i}}\right)}^2}$$

(4)

$$\mathrm{MAPE}={\displaystyle \frac{100\%}{\mathrm{n}}}\sum _{\mathrm{i}=1}^{\mathrm{n}}\left|{\displaystyle \frac{{\mathrm{y}}_{\mathrm{i}}-{\text{ŷ}}_{\mathrm{i}}}{{\mathrm{y}}_{\mathrm{i}}}}\right|$$

(5)

All metrics were initially calculated to ensure a comprehensive evaluation. However, in certain situations, the decision was made to summarily present only the RMSE, as it expresses the magnitude of deviations in the same unit as biomass and more severely penalizes large errors. Since the other indicators (CCC, MAE, and MAPE) exhibited similar behavior, the selection of RMSE as the primary metric in these cases aimed to avoid redundancies, ensuring clarity in the presentation without compromising the consistency of the analysis. All models, selection methods, and evaluations were performed in the R environment (v.4.3.2), using the glmnet [68], caret [69], CORElearn [70] and Boruta [65] packages, and the tidymodels [71] (a collection of packages for modeling and machine learning). The ggplot2 [72] package was used for the analysis visualizations.

3. Results

The GLCM-based models exhibited the highest RMSEs with the KNN, LGBM, DT, and XGBoost algorithms, with values ranging between 0.25 and 0.32. For the other models, GLCM had results closer to the other scenarios (RGB, MSI), although its performance was still inferior (Figure 4). In contrast, the MSI and RGB predictors achieved the lowest RMSE values with tree-based models, such as RF and ET, with errors near 0.18 and 0.22, respectively. Overall, the highest RMSE values were observed for XGB and DT, whereas RF and ET showed the lowest values. The linear regression methods Ridge, Lasso, and Elastic Net also demonstrated promising results, with RMSE values below 0.22. A slight difference was observed in the GLCM data, where the SVM, Ridge, and Lasso algorithms performed better. The MAE, MAPE, and CCC analyses followed a similar pattern, with a greater difference among the MAPE values (Figure 4). The MSI predictor set yielded the highest CCC values overall and most consistently, with the majority of models reaching values between 0.75 and 0.88, reinforcing this dataset as the most promising. Analyzing the spatial distribution of the metrics, the removal of soil (variable that displays the subscript abbreviation rsoil) demonstrated a trend of performance improvement for most combinations of models and predictors. These tendencies were confirmed by the summary analysis presented in the Supplementary Table (Table S2).

The feature selection analysis revealed distinct patterns among the algorithms and the evaluated datasets. In the GLCM_rsoil dataset, the selection methods showed greater stability in dimensionality reduction, with selection rates varying between 20% (RFE, RReliefFF) and 40% (Lasso and Boruta), demonstrating the ability to identify a reduced subset of relevant features. The ME variable was the most consistently selected GLCM metric (identified by the majority of algorithms), followed by the area variables (area_crop, area_soil, and area_total). Conversely, the CO, ENT, MaxProb, and VAR features were not selected by any algorithm, indicating lower relevance for the predictive models (Figure 5).

For the MSI_rsoil dataset, the algorithms exhibited greater difficulty in selecting smaller sets of representative features, which may indicate lower redundancy of information gain among the indices used. The RFE method demonstrated the smallest reduction (4.3%), followed by Boruta (17.4%) and Lasso (34.8%) (Figure 5). Meanwhile, RReliefFF (78.3%), Corr (73.9%), and Ridge (69.6%) had the highest reduction rates. Furthermore, high agreement was observed among the methods in identifying relevant vegetation indices, highlighting RVI, MCARI, and GNDVI, followed by area_crop and area as the most frequently selected features by all six evaluated algorithms. Unlike GLCM and RGB, each feature was selected at least once by each algorithm. Likewise, this was the only case where a feature (RVI) was selected by all algorithms.

In the RGB_rsoil dataset, the algorithms also showed stability in dimensionality reduction. With the exception of Boruta, which reduced the set by only 37.5%—almost half of the second-lowest reduction by Lasso (66.7%)—the other methods performed substantial reductions (Figure 5). RReliefFF (87.5%) and Corr (83.3%) had the greatest feature reduction, followed by RFE (79.2%), Ridge (70.8%), and Lasso (66.7%) (Figure 5). The features ExR, WI, R, HUE, GRRI, and G, in addition to area_total and area_crop, stood out in the selection methods, with area_total being particularly notable as it was selected by five algorithms.

The application of feature selection methods resulted in variations in model performance, with both increases and decreases in RMSE observed, depending on the combination of the dataset, selection method, and machine learning algorithm (Figure 6). For the GLCM_rsoil dataset, we observed a consistent increase in error across all algorithms when Correlation and RReliefF selection were used, with values exceeding 0.05 in several cases and reaching up to 0.10 for RReliefF with GLCM. In contrast, the Boruta, RFE, Lasso, and Ridge methods predominantly resulted in decreases or small negative variations, notably below 0.03 in some models.

In the MSI_rsoil dataset, the Boruta method led to error reductions across different algorithms. However, selection by Correlation, RReliefF, and Ridge produced frequent increases in error, especially for the Correlation method, with positive relative differences in RMSE (ΔRMSE) in almost all models (Figure 6). RFE, on the other hand, yielded smaller fluctuations, with variations concentrated near zero and negligible positive increments, except for the reduction in XGB and the increase in LGBM.

For the RGB_rsoil dataset, selection by Boruta predominantly led to error decreases, with reductions observed in all evaluated algorithms, reaching values greater than 0.02 (Figure 6). The use of Correlation and RReliefF resulted in consistent increases, following the pattern of the other datasets. The Lasso and RFE methods predominantly showed reductions, with negative variations distributed among the different algorithms, generally remaining close to zero. Overall, the patterns indicated that the magnitude of the variations was more pronounced in the GLCM_rsoil and RGB_rsoil datasets, whereas MSI represented the greatest challenge for the selection methods, with infrequent performance gains.

Similarly to the simple datasets (GLCM, MSI, and RGB), applying feature selection methods resulted in varied impacts on model performance for the combined datasets when compared to the same combination without applying any selection, as measured by ΔRMSE (Figure 7a). The Boruta and RFE methods predominantly promoted a decrease in RMSE for most of the evaluated algorithms, regardless of the dataset combination. Clearly, the most expressive reductions in error were observed for the XGB, CatBoost, and RBF-SVM models, reaching a maximum reduction of 0.042 for RBF-SVM in the GLCM_rsoil + RGB_rsoil combination with RFE. In contrast, feature selection via Lasso showed mixed behavior, resulting in slight RMSE increases for some models and decreases for others. Overall, RFE presented better results when combinations were used, showing mostly reductions or less significant increases in RMSE. Consequently, the combined models based on RFE were retained.

The evaluation of different dataset combinations using RFE (Figure 7b) revealed that using combinations of subsets generally resulted in lower or very similar RMSE compared to using all available sets (All). Furthermore, the GLCM_rsoil + RGB_rsoil combination showed the best results in linear and hyperplane-based models, as well as DT. Meanwhile, MSI_rsoil + RGB_rsoil performed better for random tree-based and gradient boosting models. The GLCM_rsoil + MSI_rsoil combination was the least prominent among the two-source combinations.

Using the GLCM_rsoil predictors as a baseline (Figure 8a), the incorporation of only one of the variables promoted ΔRMSE reductions in most machine learning algorithms, with reductions exceeding 0.04. The GLCM_rsoil + RGB_rsoil combination demonstrated more consistent ΔRMSE reductions, particularly for non-linear models. Our results show prominence for XGB (0.065; GLCM_rsoil + MSI_rsoil) and CatBoost (0.055; All).

Employing MSI_rsoil as the reference (Figure 8b), there were less expressive variations in RMSE, with some improvements and deteriorations for the models. The most notable was the combination of MSI_rsoil with RGB_rsoil, which showed more consistent reductions for non-linear models. Meanwhile, the combination of all datasets (All) presented the smallest RMSE reduction, at 0.019 for LGBM. For the linear and SVM models, increases occurred in all combination cases involving MSI_rsoil when MSI_rsoil was used as the baseline.

Based on the RGB_rsoil predictors (Figure 8c), using the complete set of features (All) or only MSI_rsoil resulted in performance improvements for some algorithms (CatBoost, LGBM, RF, and ET), with ΔRMSE reductions of approximately 0.025. In contrast, the GLCM_rsoil + RGB_rsoil combination showed reductions in some specific models but increases in most. In general, combining GLCM or MSI with RGB_rsoil positively influenced the results of non-linear models that used only GLCM or MSI.

The comparative analysis among the different predictor sets allowed for the identification of three most promising scenarios for the final modeling stage (Figure 7 and Figure 8). Here, the isolated MSI_rsoil set stood out as the best overall performing dataset, presenting more robust and consistent results across the evaluated algorithms, with lower errors and higher concordance values. The MSI_rsoil + RGB_rsoil + RFE combination proved to be the most stable, promoting consistent gains, especially in tree-based models and boosting techniques, indicating strong complementarity between multispectral and visible-spectral information. Finally, RGB_rsoil + Boruta emerged as a strategic alternative for specific scenarios, particularly in linear and SVM models. Therefore, these three scenarios were selected for the model evaluation and selection, as well as for the assessment of feature influences.

The results showed that the lowest RMSE values were obtained by the RF (0.1857) and CatBoost (0.1858) models using RGB_rsoil, followed by CatBoost with MSI_rsoil (0.1879) and ET with RGB_rsoil (0.1895) (Figure 9a). Among the linear models, Lasso presented an RMSE of 0.1908, close to the values of Ridge (0.1909) and Elastic Net (0.1927), both with MSI_rsoil + RGB_rsoil. The highest errors were observed mainly for DT, with MSI_rsoil + RGB_rsoil (0.2365), MSI_rsoil (0.2471), and RGB_rsoil (0.2725). For the MAE metric, the lowest values were recorded for CatBoost with RGB_rsoil (0.1458), Lasso with MSI_rsoil + RGB_rsoil (0.1470), and Ridge with the same combination (0.1487) (Figure 9b). Ensemble models, such as RF and ET, also showed low absolute errors with RGB_rsoil (0.1489 and 0.1514, respectively). In contrast, XGB and DT presented the highest MAE, with 0.1945 for MSI_rsoil and 0.2077 for RGB_rsoil, respectively.

The MAPE values indicated the lowest percentage error for CatBoost (0.2047) and KNN (0.2098) with RGB_rsoil, and CatBoost with MSI_rsoil (0.2146). Random Forest with RGB_rsoil showed a MAPE of 0.2151, close to KNN with MSI_rsoil (0.2188) (Figure 9c). The highest values were obtained by L-SVM with 0.2751 for RGB_rsoil and LGBM with 0.2869 for MSI_rsoil + RGB_rsoil. In the CCC analysis, the highest values were observed for CatBoost with RGB_rsoil (0.8799) and with MSI_rsoil (0.8796), followed by L-SVM with MSI_rsoil + RGB_rsoil (0.8766). Furthermore, DT again stood out among the worst-performing models, with 0.8120 for MSI_rsoil + RGB_rsoil, 0.7983 for MSI_rsoil, and 0.7635 for RGB_rsoil (Figure 9d).

Overall, the CatBoost model stood out as the most consistent across all evaluated metrics, simultaneously exhibiting low RMSE, MAE, and MAPE values, as well as the highest concordance coefficients. Regarding the predictor sets, although the use of RGB_rsoil selected by Boruta stood out among the best performers, the results were found to vary according to the interaction between models and predictors, showing proximity in the obtained values and suggesting that different combinations can achieve competitive performance.

The detailed performance of the CatBoost model, which stood out in the comparative analysis, revealed high performance (Figure 10). In the training phase, the model achieved a CCC of 0.98, with an RMSE of 0.086, an MAE of 0.069, and a MAPE of 0.127. For the test dataset, the performance resulted in a CCC of 0.83, RMSE of 0.214, MAE of 0.161, and MAPE of 0.258. The scatter plot indicated a distribution of predicted and observed values close to the identity line for both datasets. The feature importance analysis (Figure 10b) identified the area_crop variable as the most significant predictor, contributing 19.94% to the model, followed by area_total with 13.83%. The remaining features showed lower contributions. Among them, ExR registered 7.76% importance, while WI, VARI, B, G, and R showed similar values, ranging from 6.41% to 6.04%. The features with the lowest predictive power were GRRI (2.96%), EVI2green (3.41%), NGRDI (3.88%), and BI (3.88%).

4. Discussion

In this study, we demonstrated the high capacity of machine learning models, powered by UAV data, to predict buffelgrass biomass. The results indicated that combining RGB sensors with robust preprocessing and feature selection strategies can generate models with accuracy comparable to, or even exceeding, that of more complex multispectral data [8,13,73,74] (Figure 9), demonstrating the viability of more accessible and scalable agricultural monitoring methodologies. The close performance and interchangeable rankings of the algorithms across the different metrics and evaluations suggest that even simpler approaches, when properly tuned, can rival more complex algorithms, highlighting that the processing steps are as important as the models themselves [75,76]. In particular, the success of CatBoost (Figure 9 and Figure 10) lies in its ability to optimize gradients to build robust models that capture feature interactions [77,78]. This superior performance, in alignment with other studies in the agricultural field, solidifies the standing of gradient boosting and decision tree-based algorithms [14,79,80,81,82].

However, although CatBoost achieved good absolute predictive performance, its apparent superiority should be interpreted with caution, as part of the gain resulted from overfitting (RMSE_train = 0.086; RMSE_test = 0.214, Figure 10). Stronger regularization strategies, such as restricting tree depth, increasing L2 penalization, reducing the learning rate, and implementing early stopping based on an independent validation set, represent appropriate avenues to mitigate this effect and will be considered in future developments [83]. More broadly, the results reinforce that predictive performance depends not only on the choice of the algorithm but also on adequate control of model complexity and rigorous validation procedures [83,84]. In addition, the small margin separating CatBoost from competing algorithms (Figure 9) indicates that these models are also likely to benefit from the aforementioned regularization and validation strategies.

The consistent improvement in model performance after soil exclusion (Figure 4) demonstrates that the soil background acts as spectral noise in sparse vegetation canopies [85,86,87]. Simple segmentation based on vegetation index thresholds (e.g., NGRDI and NDVI) proved to be a fundamental step, refining the relationship between the predictors and biomass [17]. Based on these findings, we reinforce the need for data extraction approaches that isolate the vegetation signal. Future research should focus on validating these models under more diverse conditions and investigating automated methods for spectral thresholding to further refine the pre-processing stage, such as Rana et al. [22] and Sarathamani et al. [16].

Although texture analysis is frequently employed to enhance prediction models [12,88,89], the comparison among data sources revealed that GLCM-derived attributes generally did not contribute to modeling biomass (Figure 4, Figure 7 and Figure 8). The main hypothesis for this divergence is that the mean of the textural parameters underestimated the canopy’s structural complexity, a characteristic that proved to be better represented by spectral data. Furthermore, the use of a single window and an average across directions may have obscured relevant information. Therefore, it is suggested that the predictive capacity of GLCM could be enhanced by extracting various alternative statistical aggregation metrics (e.g., maximum, minimum, standard deviation) [90] or by adopting deep learning architectures that process the raw image data, overcoming the limitations of a single summary metric [91]. The statistical summary of the GLCM metrics exhibited consistently high homogeneity and low variance (e.g., low mean contrast and a restricted entropy range) for buffelgrass cover, indicating a rapid saturation of textural information under dense cover conditions, considering the methods adopted in this study. This limited spatial variability at centimeter-scale resolution prevented the textural features employed from effectively capturing the structural differences that are critical for accurate biomass estimation (Table S3).

In contrast, the combination of RGB and multispectral (MSI) information resulted in performance improvements in specific scenarios (Figure 7 and Figure 8), highlighting the contribution of the RE and NIR bands in capturing features complementary to visible spectrum data [92,93,94]. The stability observed in the combined MSI_rsoil + RGB_rsoil model (Figure 9) reinforces the value of sensor integration, even though the final model with the highest accuracy was based exclusively on RGB predictors. Studies conducted on oilseed rape [95], oat [29], and different grass species [96] indicate that fusing RGB data and multispectral characteristics can increase precision in estimating biomass and productivity. However, this gain is highly context-dependent, varying according to the cultivated species, the analyzed target, and environmental conditions [25,97]. Thus, the evaluation of this integration’s potential must be performed on a case-by-case basis, as demonstrated in this study by the proximity of the obtained metric values (Figure 9). For forage managers who prioritize low equipment cost, faster processing time, and high efficiency in large-scale monitoring, the RGB-only workflow is clearly the preferred choice. Conversely, for research scenarios or high-value precision agriculture applications that require the highest assurance of model stability under different conditions, the combination of MSI_rsoil + RGB_rsoil (although more computationally demanding) is justified.

Wrapper-type methods (RFE and Boruta) and Lasso (embedded) showed superior performance compared to filter methods (Correlation and RReliefFF) and Ridge (embedded), highlighting the relevance of strategies that automatically eliminate features (Figure 6 and Figure 7). The absence of predefined thresholds in Correlation, RReliefFF, and Ridge limits their evaluation capacity and reduces precision in feature selection. In contrast, Boruta and RFE, by estimating feature importance within the context of the modeling algorithm itself, identified more informative and non-redundant subsets. This approach resulted in reduced computational complexity and increased model robustness.

However, the effectiveness of feature selection proved to be strongly dependent on the nature of the input variables. Whereas in the GLCM_rsoil and RGB_rsoil datasets, high redundancy allowed methods like RFE and Boruta to identify subsets with information gain or without significant informational loss (Figure 6), in the MSI_rsoil dataset, dimensionality reduction was more limited (Figure 5), reflecting the complementarity of the chosen multispectral indices. In this scenario, feature selection proves to be an essential step in defining the best balance between reducing the number of attributes and preserving the model’s predictive capacity [10,28,29]. For example, Zhang et al. [29] also highlighted the importance of feature selection for ML models in estimating oat biomass, finding the best combination with RFE and a multilayer perceptron. Although filter methods, such as those based on correlation [28], are also employed in different contexts, model-integrated procedures exhibit greater robustness and constitute a more reliable starting point for defining the feature set [98].

The dominance of the area_crop (vegetation area) variable as the most important predictor in the final model (Figure 10b)—surpassing all vegetation indices and spectral bands—and its predominance in the feature selection methods (Figure 5) indicate that, for buffelgrass under the evaluated conditions, the vegetation cover fraction constitutes highly significant information, especially when associated with spectral data. This is particularly true in conjunction with the total available area_total, which was the second most important variable, characterizing a relationship between the available area and the area occupied by the crop. The estimate is complemented by spectral data, capable of capturing variations in canopy physiology and structure [10,94,99]. This finding has practical implications, suggesting that simpler models focused on the precise estimation of cover fraction can offer excellent cost-effectiveness in pasture monitoring, particularly for biomass [10,99,100,101]. This finding reinforces the fundamental importance of simple structural variables, indicating that spectral complexity is not the only path to greater accuracy in biomass estimation [101]. The high importance of the variables area_crop and area_total resulted, in part, from the experimental design since fresh biomass was measured at the whole-plot level. The use of absolute areas as predictors constitutes a limitation, as it introduces dependence between available area and total biomass, constraining extrapolation to larger operational scales. In practical applications, this constraint may be mitigated by previously estimating the fraction of vegetation cover through automatic image segmentation into regular grids or management zones, over which biomass is subsequently modeled.

Conducting a comprehensive comparative analysis, incorporating multiple algorithms, data sources, and feature selection methods, proved essential for critically evaluating the obtained results. This approach reduced the risk of biases associated with reliance on a single model or preprocessing technique, offering a robust basis for interpreting the observed patterns. The consistent performance across different algorithms confirmed that trends—such as the superiority of soil removal or the efficacy of certain feature selection methods—were not the result of model-specific behaviors but rather reflected generalizable responses from the dataset. These findings demonstrate that predictive performance emerges from the integration of data quality, preprocessing efficacy, and the appropriateness of the employed algorithms, reinforcing the interdependent nature of the modeling workflow components (Figure 4, Figure 6 and Figure 9).

Despite the promising results, we acknowledge the limitations of this study. The study was conducted at a single experimental site, and model transferability to other regions (with different soil types and management conditions) needs to be validated [102]. The thresholds for soil removal were defined by visual inspection and were effective for the present dataset; however, this procedure may introduce subjectivity, and the absence of a sensitivity analysis of these cut-offs remains a limitation of the workflow. Future work should evaluate data-driven procedures, such automated optimization methods, to define thresholds with greater robustness under different conditions. For application in heterogeneous environments, soil-normalization procedures based on soil-line concepts or local radiometric calibration can mitigate background effects on vegetation indices, whereas prioritizing highly informative structural variables, such as cultivated area and canopy area, may improve model stability relative to raw spectral bands under variable atmospheric and calibration conditions. Further research should also validate the models across broader spatial and temporal domains, integrate 3D structural metrics derived from Structure-from-Motion photogrammetry, and explore deep learning approaches, such as Convolutional Neural Networks, capable of jointly learning spectral and textural features from raw imagery to advance biomass estimation accuracy.

5. Conclusions

In conclusion, our study establishes a robust workflow for estimating buffelgrass biomass using UAV data and machine learning. The removal of soil pixels, for example, consistently improved model performance, reducing errors and increasing concordance for spectral data. In the evaluation of isolated data sources, multispectral (MSI) predictors proved to be the most robust, while textural (GLCM) attributes did not contribute significantly to the modeling. Although the combination of MSI and RGB data demonstrated strong complementarity in specific models, we have shown that RGB sensors, when associated with careful processing and feature selection, can produce high-precision results. The prominence of area calculations (area_total, area_crop and area_soil), surpassing the contribution of spectral and textural indices, demonstrates that the optimization of the analytical pipeline can be more decisive than sensor complexity. These findings pave the way for the development of low-cost tools for precision pasture management, with the potential to optimize animal production and promote sustainability in agricultural systems in semiarid regions.

Although the results are promising, the study was conducted at a single experimental site, indicating that model transferability needs to be validated under more diverse conditions. To translate these findings, future research should focus on integrating 3D structural information, such as canopy height obtained from photogrammetry, which can capture vertical biomass variability not fully represented by 2D data, as well as on exploring deep learning architectures to extract spectro-textural features directly from the images, further refining the estimation accuracy.