Integration of Unmanned Aerial Vehicle Spectral and Textural Features for Accurate Above-Ground Biomass Estimation in Cotton

Abstract

Timely and accurate estimation of Above-Ground-Biomass (AGB) in cotton is essential for precise production monitoring. The study was conducted in Shaya County, Aksu Region, Xinjiang, China. It employed an unmanned aerial vehicle (UAV) as a low-altitude monitoring platform to capture multispectral images of the cotton canopy. Subsequently, spectral features and textural features were extracted, and feature selection was conducted using Pearson’s correlation (P), Principal Component Analysis (PCA), Multivariate Stepwise Regression (MSR), and the ReliefF algorithm (RfF), combined with the machine learning algorithm to construct an estimation model of cotton AGB. The results indicate a high consistency between the mean (MEA) and the corresponding spectral bands in textural features with the AGB correlation. Moreover, spectral and textural feature fusion proved to be more stable than models utilizing single spectral features or textural features alone. Both the RfF algorithm and ANN model demonstrated optimization effects on features, and their combination effectively reduced the data redundancy while improving the model performance. The RfF-ANN-AGB model constructed based on the spectral and textural features fusion worked better, and using the features SIPI2, RESR, G_COR, and RE_DIS, exhibited the best performance, achieving a test sets R² of 0.86, RMSE of 0.23 kg·m⁻², MAE of 0.16 kg·m⁻², and nRMSE of 0.39. The findings offer a comprehensive modeling strategy for the precise and rapid estimation of cotton AGB.

1. Introduction

Cotton holds a pivotal role as an economic crop in China, serving as a strategic resource that significantly influences national welfare and the economy and occupying a prominent position within the national economic framework [1]. Above-Ground-Biomass (AGB) is indispensable in delineating crop growth and yield formation, serving as a critical indicator for assessing the quality and functionality of agricultural ecosystems [2]. Understanding the dynamics of cotton AGB is essential for predicting potential yield variations and evaluating crop phenotypic adaptation to environmental changes [3]. The conventional method of sample harvesting offers precise measurements of crop biomass and yield; however, it is time-consuming, labor-intensive, and detrimental to the agricultural ecosystem, rendering it unsuitable for continuous, large-scale observations [4]. As agricultural modernization and mechanization progress rapidly, large-scale, regionalized agriculture emerges as the future development trajectory. Consequently, the rapid and precise monitoring of crop AGB through scientific means becomes imperative for the intelligent and information-based management of agricultural ecosystems.

In recent years, the rapid advancement of low-altitude remote sensing technology, facilitated by unmanned aerial vehicles (UAVs), has provided a new avenue for swiftly and accurately assessing field crop conditions. Numerous studies have predominantly utilized sensors such as RGB cameras [5,6], multispectral [7,8], LiDAR [9], and hyperspectral [10,11] to capture spectral information of farmland. LiDAR sensors primarily serve to measure crop dimensions [12], whereas RGB sensors provide limited spectral bands [13]. On the other hand, multispectral and hyperspectral imaging offers a more comprehensive spectral band coverage, crucial for effectively capturing the growth status of crops [10]. However, hyperspectral sensors tend to have slower image acquisition rates compared to multispectral sensors, and they encounter additional computational challenges during data processing and analysis. In contrast, multispectral sensors swiftly and accurately provide sufficient spectral information, making multispectral imaging technology a promising tool for agricultural remote sensing [14].

Moreover, the extraction of crop traits from spectral data relies on the established relationship between the desired traits and the derived vegetation indices (VIs). Some scholars have observed that relying solely on vegetation indexes for estimating crop agronomic traits is prone to saturation [15]. Given that image textural features carry spatial information, they serve as valuable complements to spectral features and can effectively extend the model’s capability to estimate the saturation point’s location [16]. With the advancement of remote sensing technology, numerous modeling features have been developed. While multi-source features offer rich information, issues such as covariance and data redundancy often arise among these features. Many studies have employed Pearson’s correlation coefficient method to manually downscale features and filter out highly correlated ones [17,18]. However, traits with high covariance remain a challenge as Pearson’s correlation coefficient method (P) fails to address this issue [19]. Principal component analysis (PCA) excels in selecting the most important variables for known data but struggles with unknown and nonlinear data [20]. On the other hand, techniques like Multivariate Stepwise Regression (MSR) [21] and ReliefF (RfF) [22] can address various issues such as determining the importance of the information contained in features, the weights of each feature concerning the target trait, and the feature’s divergence [23]. However, given the abundance of remote sensing features available for selection [24], the choice of feature selection algorithm for dimensionality reduction is particularly important.

Furthermore, the modeling algorithm implements a degree of dimensionality reduction on the features to mitigate model complexity, thereby enhancing model performance. Deep learning algorithms are unsupervised and construct estimation models with high accuracy. However, more parameters are required, the model is larger, and computer performance is required [25]. Kasper Johansen et al. [26] employed the random forest (RF) algorithm, utilizing morphological, image, and spectral features as input variables to develop models for tomato biomass and yield. The artificial neural network model, based on fully connected layer connections, can capture numerous hidden features from the image and output a specified number of features to mitigate covariance among them, thus enhancing model accuracy. Yuan Huanhuan et al. compared the random forest model (RF), support vector machine (SVM), artificial neural network (ANN), and partial least squares (PLS) for the estimation of leaf area index in soybean and found that the RF and ANN models performed better in terms of trait estimation [27]. Artificial neural networks and random forest regression models excel in mitigating covariance issues across different data perspectives and possess robust data fitting and prediction capabilities. Despite their advantages, there is limited research on integrating feature selection algorithms with machine learning modeling algorithms to downscale features and devise optimal modeling strategies [28].

In summary, this study utilized UAV multispectral technology to construct a cotton AGB estimation model. Mainly through the extraction of image spectral features and texture features. Screened the sensitive features of AGB using P, PCA, MSR, and RfF to explore its potential for accurate monitoring of AGB, aiming to provide reference materials for cotton growth monitoring.

2. Materials and Methods

2.1. Overview of the Study Area

The study area is located in Shaya County, Aksu Region, Xinjiang (Figure 1), characterized by a temperate continental arid climate, an average elevation of 988 m, an annual cumulative temperature of ≥10 °C of 4105 °C, an average annual sunshine of 3031.2 h, a frost-free period of 214 d. Precipitation is scarce in this area, with an average annual precipitation of 47.3 mm and an average annual evapotranspiration of about 2000.7 mm, and sub-film drip irrigation is the main high-yield cultivation technology in the area. The soil of experimental field was a sandy loam soil with an organic matter content of 9.8 g·kg⁻¹, total nitrogen of 0.6 g·kg⁻², a bulk density of 1.5 g·cm⁻³, and a soil pH of 8.3 [29].

The experiment spanned 2 years, commencing with a split-zone design in the first year (Figure 1). The main zone contained three horizontal irrigation levels such as W1 (3150 m³·ha⁻¹), W2 (4050 m³·ha⁻¹), and W3 (4950 m³·ha⁻¹). Meanwhile, the secondary zones encompassed seven mulching treatments, including ordinary PE mulch, high-burger mulch, no mulch, and four types of degradable mulches. The cotton variety (J206-5) was sown on 15 April 2021, utilizing the “one film, three tubes, and six rows planting” method, with a plant row configuration of [(10 + 66 + 10 + 66 + 10) + 66] × 10 cm. In the second year, one degradable film cover treatment was omitted, and same cotton variety was sown on 8 April 2022, while all other experimental conditions remained consistent with the first year.

2.2. Ground Data Acquisition

Following each image acquisition, three cotton plants were randomly selected from each plot (2.28 × 9 = 20.52 m²); the whole plant was harvested above the cotyledon node and packed in paper bags. Subsequently, the samples were subjected to a temperature of 105 °C for 30 min in an oven, followed by a reduction in temperature to 80 °C for drying until a constant weight was achieved, and then the data of dry weight was recorded. After removing outliers, the mean sample mass was utilized to compute the above-ground biomass (AGB) value per unit area (kg·m⁻²). In the first year of the experiment, 120 samples were collected at each stage, whereas in the second year, with one less treatment, 108 samples were collected at each stage, resulting in a total of 576 samples. These samples were randomly allocated into training and test sets at a ratio of 7:3, and the outcomes are detailed in

Table 1. Above ground biomass real measurement dataset in cotton.

Data Sets	Sample Size	Mean (kg·m−2)	Maximum (kg·m−2)	Minimum (kg·m−2)	Standard Deviation (kg·m−2)	Variance	Coefficient of Variation
Training Sets	403	1.12	2.87	0.19	0.61	0.37	54%
Test Sets	173	1.05	3.61	0.24	0.62	0.39	60%

.

2.3. UAV Image Acquisition and Feature Extraction

2.3.1. Image Acquisition

In the first year of the experiment, we employed the DJI M300 RTK UAV platform (Matrice 300 Real Time Kinematic, M300RTK) produced by DJI Technology Co., Ltd. in Shenzhen, China, equipped with the U.S. company MicaSense (Seattle, WA, USA), RedEdge-MX sensors to obtain low-altitude remote sensing imagery, and the sensors contain five bands: blue, green, red, red-edge, and near-infrared. In the second year of the experiment, the DJI Phantom 4 Multispectral version of the UAV platform (Phantom 4 Multispectral, P4M), produced by DJI Technology Co., Ltd. in Shenzhen, China, was used. The sensors contain one visible and five multispectral bands. Information about the UAV platforms and sensors is shown in

Table 2. UAV platform-specific information.

Date of Data Acquisition	Unmanned Aircraft Platforms	Sensor	Spectral Region (nm)	Flight Altitude (m)	Spacing Resolution (m)	Longitudinal and Lateral Overlap Rate
5 July 2021	M300RTK	RedEdge-MX	Blue: 475 ± 20	30	0.02117	75% + 75%
5 August 2021	M300RTK		Green: 560 ± 20	30	0.02117	75% + 75%
22 August 2021	M300RTK		Red: 668 ± 10	30	0.02117	75% + 75%
			RedEdge: 717 ± 10
			NIR: 840 ± 40
29 June 2022	P4M		Blue: 450 ± 16	30	0.02091	75% + 75%
			Green: 560 ± 16
7 August 2022	P4M		Red: 650 ± 16	30	0.02091	75% + 75%
			RedEdge: 730 ± 16
			NIR: 840 ± 26

.

Image was taken during clear, cloudless, and windless conditions at noon (13:00–15:00). Prior to commencing image capture, the UAV underwent DLS Configuration. Additionally, both the take-off and landing points were designated at the same location. An open and level terrain was carefully chosen for operation. Within the cotton field, a calibration plate (Ground Control Point, GCP) was strategically positioned to precisely determine the field’s location within the images. Flight routes were planned, preferably aligning with the direction of mulching. The UAV maintained a flight altitude of 30 m and a speed of 2.6 m per second. A 75% overlap rate was ensured in both side and heading directions. Manually, a set of photos of the reflectance correction plate was taken before each route mission. Images were captured at equidistant intervals along consistent flight paths.

2.3.2. Image Extraction and Parsing

Data preprocessing, including stitching and radiometric calibration of multispectral images, was executed using Pix4D Mapper software (version 4.5.6, Pix4D S.A., Prilly, Switzerland). Initially, the acquired remote sensing imagery was imported into the software, along with the POS geolocation information. Subsequently, the Ag multispectral template was selected, and either an aerial grid or a corridor-type route was chosen in the advanced settings match. Point cloud and 3D mesh texture generation followed, alongside the utilization of a reflectance grayscale correction plate for radiation processing and calibration of the reflection map. Finally, five single-band reflectance grayscale maps were generated and fused into one image.

Soil background mask was performed using ENVI (version 5.3, EXELIS., Boulder, CO, USA) software. The images were imported into the software, and the Band Math tool was used to calculate the NDVI vegetation index mapping, and then create a mask with a threshold of 0.3, and then the orthophotos were masked to exclude soil backgrounds.

Spectral information was extracted in bulk using ArcGIS (version 10.6, Esri Inc., Redlands, CA, USA) software. The reflectance grayscale map, with soil background removed, was imported into the software. Subsequently, a new region of interest, Shapefile, was created, delineating the test plot boundaries. The mean reflectance value within this region of interest was then extracted, serving as the spectral information for the plot.

Spectral features (SF) encompass both raw spectral bands and vegetation indices (VI). Vegetation indices are derived from linear or nonlinear band operations applied to the acquired spectral band data. This process aims to minimize the influence of soil background values, thereby enhancing the vegetation characteristics in the image and improving the accuracy of vegetation phenology estimation. In this study, spectral information from the blue, green, red, red-edge, and near-infrared bands (labeled B, G, R, RE, and NIR, respectively) was utilized to compute multispectral vegetation indices, resulting in a total of 20 indices. The formulas for calculating these vegetation indices are provided in

Table 3. Spectral features calculation formula.

Number	Spectral Features	Abbreviation	Formula	Reference
1	Difference Vegetation Index	DVI	$NIR-R$	[30]
2	Green-difference Vegetation Index (g)	GDVI	$NIR-G$	[31]
3	Normalized Difference Vegetation Index	NDVI	$\left(NIR-R\right)/\left(NIR+R\right)$	[32]
4	Green-normalized Difference Vegetation Index	GNDVI	$\left(NIR-G\right)/\left(NIR+G\right)$	[33]
5	Blue-normalized Difference Vegetation Index	BNDVI	$\left(NIR-B\right)/\left(NIR+B\right)$	[34]
6	Ration Vegetation Index	RVI	$NIR/R$	[35]
7	Green-Ration Vegetation Index	GRVI	$NIR/G$	[35]
8	Structurally Insensitive Pigment Index	SIPI2	$\left(NIR-G\right)/\left(NIR-R\right)$	[28]
9	Source Address Validation Improvement	SAVI	$\left(1+L\right)\left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\right)/\left(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}+L\right)$	[36]
10	Optimization of Soil Regulatory Vegetation Index	OSAVI	$1.16\times \left(\mathrm{N}\mathrm{I}\mathrm{R}-\mathrm{R}\right)/\left(\mathrm{N}\mathrm{I}\mathrm{R}+\mathrm{R}+0.16\right)$	[36]
11	Green-Optimization of Soil Regulatory Vegetation Index	GOSAVI	$1.16\times \left(NIR-G\right)/\left(NIR+G+0.16\right)$	[31]
12	Green Chlorophyll Index	CIgreen	$NIR/G-1$	[37]
13	RedEdge Simple Ratio Vegetation Index	RESR	$RE/R$	[38]
14	Enhanced Vegetation Index	EVI	$\left(2.5\times \left(NIR-\mathrm{R}\right)\right)/\left(NIR+6\times \mathrm{R}-7.5\times \mathrm{B}+1\right)$	[37]
15	Non-Linear Index	NLI	$\left({NIR}^2-R\right)/\left({NIR}^2+R\right)$	[31]
16	Atmospheric Resistance Vegetation Index	ARVI	$\left(NIR-2\times R+B\right)/\left(NIR+B\right)$	[39]
17	Transformed Difference Vegetation Index	TVI	$\sqrt[2]{\left(0.5+\left(NIR-R\right)/\left(NIR+R\right)\right)}$	[37]
18	Green-Renormalized Difference Vegetation Index	GRDVI	$\left(NIR-G\right)/\left(\sqrt[2]{\left(NIR+G\right)}\right)$	[31]
19	Modified Simple Ration Index	MSR	$\left(NIR/R-1\right)/\left(\sqrt[2]{\left(NIR/R\right)+1}\right)$	[40]
20	Green-Modified Simple Ration Index	GMSR	$\left(NIR/G-1\right)/\left(\sqrt[2]{\left(NIR/G\right)+1}\right)$	[31]

L: the soil adjustment coefficient, which is taken as 0.5 here.

.

Textural features (TF) pertain to the recurring local patterns and arrangements within an image, reflecting the periodic changes in the surface structure of objects. These features encompass spatial information and possess a strong anti-noise capability. Hence, in this study, we utilize the Gray-Level Co-occurrence Matrix (GLCM) in statistical methods. The Co-occurrence Measures tool in the ENVI 5.3 software is employed to extract textural features from the five channels of the multispectral image. A 3 × 3 window with a constant gray level is selected, with a computational step size of 1. The textural features are averaged over four angles, resulting in eight textural features calculated for each channel. These features include mean (MEA), variance (VAR), Homogeneity (HOM), Contrast (CON), Dissimilarity (DIS), Entropy (ENT), Angular Second-order Moments (ASM), and correlation (COR) [41].

2.3.3. Feature Selection Algorithm

In this study, four feature selection algorithms were employed to address the issue of covariance among the feature variables derived from the computed 5 original spectral bands, 20 vegetation indexes, and 40 textural features. These selected features are subsequently utilized as input variables in the model. The feature selection algorithms employed Pearson correlation, Principal Component Analysis, multiple stepwise regression, and ReliefF algorithm.

Pearson’s correlation coefficient method (P) [17] is a linear correlation coefficient, which is the most used type of correlation coefficient. Denoted as r, it is used to reflect the degree of linear correlation between two variables, with r values ranging from −1 to 1, and the larger the value, the stronger the correlation. The essence is to measure the degree of correlation between the variables through covariance while dividing by the product of their respective standard deviations for standardization.

Principal Component Analysis (PCA) [20] is a method that combines multiple correlated variables into a small number of new uncorrelated variables through orthogonal transformations, and the new variables can maintain the original information of the data as much as possible. In this study, the PCA was configured to retain principal component information with a variance contribution greater than 5%.

The fundamental principle of multiple stepwise regression (MSR) [21] involves systematically introducing independent variables one by one. Variables that demonstrate a significant impact on the target variable are retained, while those that do not maintain significance are systematically removed from the process. This process ensures that only independent variables with a significant effect on the target variable are included, thereby excluding variables that lack significant impact.

ReliefF (RfF) [22] is an enhancement of the RelieF algorithm designed to address multi-category problems. The RelieF algorithm itself is a widely utilized filtered Feature Weighting Algorithm that adjusts feature weights based on the correlation between each feature and the target variable. This correlation is determined by the feature’s capacity to distinguish between k close samples, thus updating the weights accordingly, with the default nearest neighbor distance (Near) set to 10. Notably, the algorithm’s runtime increases linearly with the number of samples drawn and the number of original features, ensuring efficient performance.

2.4. Model Construction and Evaluation

2.4.1. Model Building

In this study, the samples were randomly divided into training and test sets for the model in 7:3, and the cotton AGB was used as the target variable, while the feature set preferred by the feature selection algorithm was used as the input variable for constructing the model. Based on the Python (3.9.13) environment, three modeling algorithms implemented in the scikit-learn (1.1.1) package were Bayesian Ridge regression model, Random Forest regression model, and Artificial Neural Network model.

2.4.2. Bayesian Ridge Regression

Bayesian Ridge Regression (BRR) [42] is a linear model in machine learning. Linear model is an algorithmic model that constructs a hypothesis function (hypothesis) by combining the characteristic attributes of the target in a linear fashion and training it. This approach offers advantages such as simplicity of form, ease of modeling, and high interpretability. However, the data may face noise and covariance problems, and ridge regression, as a regularization method, mitigates the overfitting problem by introducing a regularization parameter, λ, that constrains the regression coefficients from a Bayesian perspective. In this study, the Bayesian ridge regression algorithm and parameters tol = 1.0 × 10⁻⁶, alpha_1 and alpha_1 as 1.0 × 10⁻⁶, and lambda_1 and lambda_2 as 1.0 × 10⁻⁶ were used.

2.4.3. Random Forest Regression

Random forest regression (RFR) [26] is a nonlinear model in machine learning, an algorithm that integrates multiple decision trees through the idea of integrated learning. Based on the Bootstrap method of resampling, multiple training sets are generated and input into each decision tree sub-model using a randomly selected set of split attributes, thus reducing the model variance to improve the model prediction performance. Random forests have good resistance to noise but are prone to overfitting and can handle high-dimensional data. In this study, decision tree depth, number of iterations, and other hyperparameters of stochastic Mori Ridge regression models were fine-tuned by grid search method using 5-fold cross-validation.

2.4.4. Artificial Neural Network

Artificial neural network (ANN) [27] represents a nonlinear model derived from a linear model foundation, offering distinct advantages for addressing nonlinear problems. Particularly useful when the relationship between feature variables and the target is ambiguous, ANN models excel in extracting high-dimensional features. Artificial neural network models can be internally categorized into 3 types: input layer, hidden layer, and output layer. High-dimensional feature information is extracted by increasing the number of layers in the hidden layer and the number of neurons in a single hidden layer, which in turn improves the accuracy of the model. The important parameters of this model are the number of hidden layers and neurons, activation function type, weight optimizer type, and maximum number of iterations. Determination of optimal hyperparameters is achieved through iterative loops utilizing the grid search method with 5-fold cross-validation.

2.4.5. Evaluation of Model Accuracy

In this study, coefficient of determination (R²) was chosen to evaluate the model fitting effect, Root Mean Square Error (RMSE) and Mean Absolute Error (MAE) were used to evaluate the model estimation error, and Normalized Root Mean Squared Error (nRMSE) was used to evaluate the efficiency of different models. A higher R² value indicates a better fit for the model, while smaller RMSE and MAE values signify lower model error. Additionally, a smaller nRMSE value suggests enhanced model generalization ability. The formulas [43] are as follows:

$$R^2=\frac{\sum _{i=1}^n{\left(x_i-\overline{x}\right)}^2{\left(y_i-\overline{y}\right)}^2}{n\sum _{i=1}^n{\left(x_i-\overline{x}\right)}^{2}n\sum _{i=1}^n{\left(y_i-\overline{y}\right)}^2}$$

(1)

$$RMSE=\sqrt{\frac{1}{n}{\sum }_{i=1}^n{\left(y_i-x_i\right)}^2}$$

(2)

$$MAE=\frac{1}{n}{\sum }_{i=1}^n\left|y_i-x_i\right|$$

(3)

$$nRMSE=\frac{RMSE}{\sqrt{{\sum }_i^n{\left(y_i-\overline{y}\right)}^2/n}}$$

(4)

where n is the sample size, i is the data of the ith sample, x_i is the measured value of the ith sample, $\overline{x}$ is the mean of the measured value, y_i is the predicted value of the ith sample, and $\overline{y}$ is the mean of the predicted value.

3. Results

3.1. Estimation of AGB in Cotton Based on Pearson’s Correlation Analysis

3.1.1. Correlation Analysis between Spectral Features of Cotton Canopy and AGB

The cotton canopy spectral features and AGB were correlated, and these results are shown in Figure 2. All vegetation indices were positively correlated with AGB, and the blue, green, red, and red-edge spectral bands were negatively correlated with AGB. Among the spectral features with high correlation with AGB were G and RESR, with correlation coefficients of −0.66 and 0.66, respectively.

The textural features of the five original band images of the cotton canopy were calculated and correlated with the AGB, and the results are shown in

Table 4. Correlation analysis between AGB and textural features of cotton.

Channel B	r	Channel G	r	Channel R	r	Channel NIR	r	Channel RE	r
B_MEA	−0.20 ***	G_MEA	−0.66 ***	R_MEA	−0.56 ***	NIR_MEA	0.26 ***	RE_MEA	−0.22 ***
B_VAR	0.06	G_VAR	−0.20 ***	R_VAR	−0.20 ***	NIR_VAR	0.10 *	RE_VAR	0.01
B_HOM	−0.08	G_HOM	0.16 ***	R_HOM	0.19 ***	NIR_HOM	−0.28 ***	RE_HOM	−0.11 **
B_CON	0.07	G_CON	−0.16 ***	R_CON	−0.20 ***	NIR_CON	0.25 ***	RE_CON	0.10 *
B_DIS	0.57 ***	G_DIS	0.03	R_DIS	0.11 **	NIR_DIS	0.43 ***	RE_DIS	0.33 ***
B_ENT	0.25 ***	G_ENT	0.25 ***	R_ENT	0.25 ***	NIR_ENT	0.25 ***	RE_ENT	0.25 ***
B_ASM	−0.25 ***	G_ASM	−0.25 ***	R_ASM	−0.25 ***	NIR_ASM	−0.25 ***	RE_ASM	−0.25 ***
B_COR	0.15 ***	G_COR	0.07	R_COR	0.20 ***	NIR_COR	0.01	RE_COR	0.04

***: p < 0.001, **: p < 0.01, and *: p < 0.05.

. The MEA feature and the ASM feature of each channel showed a highly significant negative correlation with AGB (p < 0.001), while the DIS feature, ENT feature, and COR feature showed positive correlation with AGB, with the DIS feature and ENT feature showing as highly significant (p < 0.001). The ENT and ASM of the five channels showed highly significant positive (r = 0.25, p < 0.001) and highly significant negative (r = −0.25, p < 0.001) correlation coefficients, as did the correlation coefficients of the AGB, respectively, showing the unique canopy textural features of cotton. The MEA features of the extracted textural features based on the five channels agree with the AGB correlation coefficients, and the five original spectral features agree with the cotton AGB correlation coefficients, revealing a degree of consistency between the cotton canopy’s textural and spectral features. The textural features with higher correlation are B_DIS, G_MEA, and R_MEA, with correlation coefficients of 0.57, 0.66, and 0.56, respectively, with G_MEA demonstrating the highest correlation coefficient.

3.1.2. Model Construction of Cotton AGB Estimation Based on Pearson’s Correlation Analysis

The model was developed using the features that exhibited high correlation coefficients with AGB [2]. The first two features were used in this study, and the results are shown in

Table 5. AGB estimation model for cotton based on Pearson’s correlation.

Features	Method	Feature Numbers	Feature Variables	Test Sets
				R2	RMSE (kg·m−2)	MAE (kg·m−2)	nRMSE
SF	BRR	2	G, RESR	0.56	0.41	0.32	0.93
	RFR			0.78	0.29	0.20	0.50
	ANN			0.82	0.26	0.18	0.48
TF	BRR	2	B_DIS, G_MEA	0.53	0.43	0.33	0.99
	RFR			0.67	0.36	0.24	0.63
	ANN			0.68	0.35	0.24	0.64
STF	BRR	4	G, RESR, B_DIS, G_MEA	0.69	0.35	0.25	0.71
	RFR			0.78	0.29	0.19	0.50
	ANN			0.80	0.28	0.19	0.50

Bolding indicates optimal results.

. When utilizing variables with high correlation, the nonlinear models (RFR and ANN) provided better fits to the AGB compared to the linear model (BRR). The use of spectral and texture feature fusion had a lower error (RMSE of 0.28–0.35 kg·m⁻², MAE of 0.19–0.25 kg·m⁻²) than using spectral feature or texture feature AGB models alone. Specifically, the ANN algorithm exhibited superior fitting in the test sets (R² of 0.80). On the other hand, the ANN algorithm showed less error between the estimated and measured values (RMSE of 0.26–0.35 kg·m⁻²). The ANN is more accurate among the three algorithmic models (nRMSE: 0.48–0.64). Consequently, the SF-ANN-AGB model proved to be more effective. In the test set, it showed R² of 0.82, RMSE of 0.26 kg·m⁻², MAE of 0.18 kg·m⁻², and nRMSE of 0.48.

3.2. Estimation of AGB in Cotton Based on Principal Component Analysis

3.2.1. Contribution of Cotton Canopy Spectral and Textural Features Based on PCA

Principal Component Analysis was performed on the feature variables, and the results are shown in Figure 3. The spectral features displayed a variance contribution of more than 5%, primarily from the first two principal components (PC1 and PC2), which accounted for 83% and 11% of the variance, respectively (Figure 3a). Conversely, individual principal components of the textural feature variables contributed less, and the first four principal components (PC1 = 58%, PC2 = 21%, PC3 = 11%, and PC4 = 5%) contributed more than 5% of the variance, respectively (Figure 3b). As for the spectral and textural feature fusion, the first four principal components had variance contributions of 57%, 21%, 9%, and 6%, respectively (Figure 3c). The PCA algorithm is more effective in extracting the main information of spectral features than textural features.

3.2.2. Model Construction of Cotton AGB Estimation Based on PCA

AGB model construction feature variable selection of the principal component components with a variance contribution greater than 5%; the model results are shown in

Table 6. Accuracy of AGB estimation model based on PCA feature selection algorithm.

Features	Method	pc Numbers	Test Sets
			R2	RMSE (kg·m−2)	MAE (kg·m−2)	nRMSE
SF	BRR	2	0.51	0.44	0.34	1.00
	RFR		0.84	0.25	0.15	0.42
	ANN		0.86	0.23	0.15	0.38
TF	BRR	4	0.57	0.41	0.31	0.92
	RFR		0.69	0.35	0.23	0.66
	ANN		0.62	0.39	0.27	0.74
STF	BRR	4	0.64	0.37	0.27	0.79
	RFR		0.81	0.27	0.18	0.48
	ANN		0.84	0.25	0.25	0.42

Bolding indicates optimal results.

. Nonlinear models (RFR and ANN) outperformed the linear model (BRR). Notably, the PCA algorithm exhibited greater accuracy in constructing models based on spectral features compared to textural features. Moreover, in the PCA-based feature selection algorithm, the fusion of textural features with spectral features significantly enhanced the AGB model’s performance. Considering the potential saturation effect when using only a single feature, the STF-ANN-AGB model emerged as more effective among the AGB estimation models constructed based on the PCA feature selection algorithm. In the test set, it obtained an R² of 0.84, RMSE of 0.25 kg·m⁻², MAE of 0.25 kg·m⁻², and nRMSE of 0.42.

3.3. Estimation of AGB in Cotton Based on Multiple Stepwise Regression

The features were selected using a multiple stepwise regression (MSR) algorithm, wherein the feature variables underwent a series of analyses for significance and were ranked based on their association with cotton AGB. Subsequently, feature variables with the largest p-values (p > 0.05) were systematically phased out using an iterative approach. These selected feature variables were utilized as input variables for the model, and the model accuracy based on different numbers of features selected by the MSR algorithm is presented in

Table 7. Accuracy of AGB estimation model based on MSR algorithm.

Features	Method	Features Numbers	Features Variables	Test Sets
				R2	RMSE (kg·m−2)	MAE (kg·m−2)	nRMSE
SF	BRR	17	B, G, RE, R, DVI, NDVI, BNDVI, SIPI2, RVI, CIgreen, GOSAVI, RESR, SAVI, OSAVI, GRDVI, GMSR, MSR	0.87	0.22	0.15	0.37
	RFR			0.88	0.22	0.13	0.37
	ANN			0.90	0.20	0.12	0.33
TF	BRR	21	B_MEA, B_DIS, G_VAR, G_CON, G_COR, NIR_MEA, NIR_VAR, NIR_HOM, NIR_DIS, NIR_ENT, NIR_ASM, RE_MEA, RE_VAR, RE_CON, RE_DIS, RE_COR, R_MEA, R_VAR, R_HOM, R_ENT, R_ASM	0.82	0.26	0.17	0.46
	RFR			0.87	0.23	0.14	0.39
	ANN			0.89	0.20	0.13	0.33
STF	BRR	34	GDVI, NDVI, GNDVI, BNDVI, SIPI2, RVI, CIgreen, GOSAVI, EVI, OSAVI, NLI, GRDVI, GMSR, MSR, B_VAR, B_DIS, G_MEA, G_VAR, G_HOM, G_CON, G_DIS, G_COR, NIR_HOM, NIR_DIS, NIR_ENT, NIR_ASM, RE_VAR, RE_HOM, RE_CON, RE_DIS, RE_COR, R_CON, R_ENT, R_ASM	0.87	0.23	0.15	0.38
	RFR			0.86	0.23	0.12	0.40
	ANN			0.89	0.20	0.14	0.33

Italics and bolding indicate repeated features. Bolding indicates optimal results.

below. Among the selected features, there were 17 spectral features, 21 textural features, and 34 spectral and textural feature fusions, as determined by the MSR. Notably, some of the spectral features and textural features overlapped in the spectral and textural feature fusions, which are highlighted in italic and bold in Table 7, indicating strong consistency in this subset of features. The use of MSR selects more features, and the richness of information makes the model more accurate. The ANN algorithm constructs models with higher accuracy and lower error than the BRR and RFR algorithms. The SF-MSR-ANN-AGB model worked better for cotton AGB estimation with an R² of 0.90, RMSE of 0.20 kg·m⁻², MAE of 0.12 kg·m⁻², and nRMSE of 0.33.

3.4. Estimation of AGB in Cotton Based on RfF Algorithm

3.4.1. Importance Analysis and Estimation of AGB in Cotton Based on RfF Algorithm

The ReliefF algorithm was employed for feature selection, assessing the importance of features regarding cotton AGB. Feature variables with low importance scores were excluded, and the feature variables were subsequently updated based on their weights relative to AGB. The results of the importance scores of the feature variables to AGB are depicted in Figure 4. Among spectral features, SIPI2 exhibited the highest importance score (0.33), followed by RESR (0.23). In contrast, textural features displayed high importance scores for G_COR (0.37) and RE_DIS (0.34). Additionally, Figure 4b reveals a notable consistency in importance scores across channel textural features, with ASM and ENT scoring lower while DIS and COR scored higher. Interestingly, the RfF algorithm assigned higher importance scores to textural features compared to spectral features.

The features, sorted according to their importance scores, are sequentially incorporated into the AGB model from highest to lowest, as depicted in Figure 5. The model accuracy results highlight the significant impact of the number of features on the AGB model’s performance: as the number of features increases, the model’s R² increases while the RMSE and nRMSE decrease. Notably, the contribution to model accuracy primarily stems from features with high importance scores, whereas features with lower importance scores make minimal contributions and tend to increase model complexity, thereby weakening the enhancement of model accuracy. Moreover, as the number of feature variables increases, the accuracy of the AGB models constructed by the three modeling algorithms tends to stabilize. In the AGB model constructed based on textural features, it is observed that the accuracy of the model using a small number of textural features is notably low, making it challenging to estimate AGB accurately. The RfF algorithm assigns higher importance scores to textural features compared to spectral features, and AGB model features constructed based on a small number of spectral and textural features fusion primarily rely on textural features, limiting the improvement in the accuracy of AGB estimation models. However, once the fourth feature, a spectral feature (RESR), is incorporated, the performance of the AGB model improves significantly. This further underscores the importance of multi-source features in compensating for the incomplete information of individual features.

3.4.2. Model Construction of Cotton RfR-AGB Estimation

Based on the results of the above analysis, in this study, the features with the top two importance scores of spectral features (SIPI2 (2nd), RESR (1st)) and the top two importance scores of textural features (G_COR (1st), RE_DIS (2nd)) were selected to construct the AGB model separately. Similarly, the features with the top two importance scores of spectral features and textural features in the spectral and textural features fusion (SIPI2 (17th), RESR (4th), G_COR (1st), RE_DIS (2nd)) were selected to construct the model AGB model, and the results are shown in

Table 8. Accuracy of AGB estimation model based on RfF feature selection algorithm.

Features	Method	Feature Numbers	Feature Variables	Test Sets
				R2	RMSE (kg·m−2)	MAE (kg·m−2)	nRMSE
SF	BRR	2	SIPI2, RESR	0.51	0.43	0.33	0.96
	RFR			0.82	0.26	0.19	0.47
	ANN			0.86	0.24	0.17	0.41
TF	BRR	2	G_COR, RE_DIS	0.08	0.60	0.51	2.77
	RFR			0.26	0.54	0.41	1.44
	ANN			0.13	0.58	0.48	2.40
STF	BRR	4	SIPI2, RESR, G_COR, RE_DIS	0.61	0.39	0.29	0.82
	RFR			0.83	0.25	0.17	0.47
	ANN			0.86	0.23	0.16	0.39

Bolding indicates optimal results.

. The fusion of spectral features with textural features demonstrated enhanced model performance. Specifically, the AGB model constructed using the ANN modeling algorithm exhibited more stable performance. Notably, the SF-RfF-ANN-AGB model and STF-RfR-ANN-AGB model displayed higher accuracy (R² of 0.86), but the STF-RfR-ANN-AGB model had lower error (RMSE of 0.23 kg·m⁻², MAE of 0.16 kg·m⁻² and nRMSE of 0.39).

3.5. Model Inversion for Cotton AGB Estimation Based on Optimal Modeling Strategy

Figure 6 shows the results of the cotton AGB model constructed based on spectral and textural feature fusions, a combined feature selection and modeling algorithm approach. Artificial neural network (ANN) models consistently outperformed other models in estimating cotton AGB, demonstrating high accuracy across the board. Among the models, the MSR-ANN-AGB model achieved the highest accuracy, with a test set R² of 0.89, RMSE of 0.20 kg·m⁻², MAE of 0.14 kg·m⁻², and nRMSE of 0.33. Notably, the cotton AGB model constructed based on the MSR feature selection algorithm selected numerous features (17–34), leading to higher accuracy compared to other feature selection algorithms. However, the optimal modeling strategy was observed in the RfF-ANN-AGB model, which employed a smaller number of features. This strategy reduced the utilization of redundant data and improved model performance, yielding an R² of 0.86, RMSE of 0.23 kg·m⁻², MAE of 0.16 kg·m⁻², and nRMSE of 0.39.

4. Discussion

4.1. Response of Cotton AGB Estimation Models to Spectral and Textural Features

The literature demonstrates that constructing models based on multimodal data fusion can enhance prediction accuracy and mitigate the saturation effect of the model [44,45]. For instance, Guo et al. achieved an RMSE of 5.77 days in estimating the date of summer maize tasseling by integrating spectral and image texture features using an improved adaptive feature weighting method [46]. Similarly, the findings of this study underscore the superiority of spectral and textural features fusion over models constructed solely with either spectral or image texture features, showcasing higher accuracy and stability while effectively delaying the saturation effect of spectral features. Furthermore, Wang et al. highlighted that combining image texture features can mitigate overestimation and underestimation phenomena observed in spectral index models alone [16]. Moreover, while the resolution of acquired images appears to have no impact on models constructed solely from spectral features [47], it does influence models constructed through spectral and textural feature fusion. This discrepancy stems from the fact that image texture features contain spatial information, making them highly sensitive to the spatial resolution of the image. Therefore, further research on spectral and textural feature fusions should consider the role of spatial resolution in image analysis.

4.2. Estimation of AGB in Cotton Based on Spectral and Textural Feature Fusions by Feature Selection Algorithm

In this study, it was found that the five original spectral features and textural features MEA corresponding to the channels exhibited agreement with the AGB correlation coefficients. This suggests that spectral features can, to some extent, characterize certain textural features, thereby highlighting the issue of covariance between features. Similarly, Jiang et al. found that models constructed after screening characteristic wavelength variables using CARS outperformed those utilizing full-wavelength, SPA, and Monte Carlo Uninformative Variable Elimination (MC-UVE) methods [48]. In this paper, four feature selection algorithms were compared and analyzed. The results indicated that feature selection based on Pearson’s correlation coefficient method failed to address the covariance problem between features, resulting in low model accuracy, with an R² of 0.80. Conversely, employing PCA to analyze spectral and textural feature fusions and selecting the greater 5% of the principal component components improved the model’s R² to 0.84. Moreover, the AGB model based on multiple stepwise regression retained more features, utilized the STF-MSR-ANN-AGB model strategy, and exhibited the highest accuracy with an R² of 0.89 and an RMSE of 0.20 kg·m⁻². The MSR method efficiently eliminated features insignificantly related to the AGB, thereby enhancing the model’s performance while also effectively downsizing the feature data. Additionally, it has been demonstrated that the ReliefF-RFE algorithm significantly improves plantation classification results for multi-period and single-period data compared to SVM, CART, and RF methods, achieving the highest accuracy [49]. This aligns with the results obtained using the RfF algorithm in this study. Utilizing the RfF algorithm in conjunction with the ANN model, the AGB estimation model constructed based on STF exhibited high accuracy, with an R² of 0.86, RMSE of 0.23 kg·m⁻², MAE of 0.16 kg·m⁻², and nRMSE of 0.39 for the test sets.

4.3. Impact of Machine Learning Algorithms on the Accuracy of Cotton AGB Estimation Models

Enhancing the accuracy of remote sensing estimation models has been studied extensively, primarily focusing on the integration of multi-source features and various modeling algorithms. Zhang et al. conducted a comparison between support vector machine and random forest regression for constructing a winter wheat LAI estimation model, revealing that the random forest regression modeling algorithm yielded superior accuracy [50]. Similarly, the findings of this study indicate that the RFR model provides a better fit, with significantly higher accuracy in both the training and test sets compared to the linear BRR model. This improvement is likely attributed to the inherent dimensionality reduction capabilities of the RFR and ANN algorithms. Furthermore, the ANN-MLP model predicted maize yields with optimum accuracy, achieving an R² of 0.98, RMSE of 0.23, and rRMSE of 20.44% [51]. In line with these findings, this paper suggests that utilizing the ReliefF algorithm for feature selection of spectral and textural features fusion, followed by integration with the ANN modeling algorithm, represents the optimal modeling strategy for constructing the cotton AGB estimation model.

Indeed, the feature selection algorithm employed in this study solely analyzes the spectral and textural features of multispectral images, overlooking the inclusion of high spatial resolution RGB images and high-dimensional hyperspectral data. This presents a path toward further research and exploration. We used full-fertility data to construct the cotton AGB model, and the specific availability at different fertility periods needs to be further explored. Spectral and textural feature fusion have shown promise in delaying the onset of spectral feature saturation points, but its effects on textural features of varying spatial resolutions remain unexplored. Furthermore, the impact of feature selection algorithms on the extent to which textural features delay spectral saturation points across different spatial resolutions warrants investigation. This avenue of inquiry holds the potential to advance our understanding of feature selection methods and their applicability across diverse remote sensing datasets.

5. Conclusions

The textural features MEA and the corresponding spectral features bands exhibited high consistency with AGB correlation. Models incorporating spectral and textural feature fusions demonstrated greater stability compared to models utilizing single spectral features or textural features alone. The cotton AGB estimation model (STF-RfF-ANN-AGB), constructed using the ReliefF algorithm for spectral and textural features fusion for feature selection combined with the ANN algorithm, proved to be the most effective. It showed a 7.44% increase in R² and a 16.68% decrease in RMSE, as well as a 13.50% decrease in MAE and a 22.68% decrease in nRMSE compared to the ANN-AGB model construction (STF-P-ANN-AGB) based solely on Pearson’s correlation coefficients. The STF-RfF-ANN-AGB model achieved an R² of 0.86, RMSE of 0.23 kg·m⁻², MAE of 0.16 kg·m⁻², and nRMSE of 0.39. The features used in this optimal modeling strategy were SIPI2, RESR, G_COR, and RE_DIS.

Conclusively, leveraging spectral and textural feature fusions alongside feature selection algorithms enhanced the generalization ability of cotton AGB estimation models. The RfF algorithm and ANN algorithm for constructing cotton AGB proved to be the most effective. The optimal modeling strategy was the STF-RfF-ANN-AGB model. This approach lays the strong foundation for improving remote sensing model accuracy through multi-feature fusion and feature selection algorithms, thus contributing to the advancement of digital and precision agriculture.