Next Article in Journal
Combining UAS LiDAR, Sonar, and Radar Altimetry for River Hydraulic Characterization
Next Article in Special Issue
A Multi-Drone System Proof of Concept for Forestry Applications
Previous Article in Journal
Optimal Coverage Path Planning for UAV-Assisted Multiple USVs: Map Modeling and Solutions
Previous Article in Special Issue
Accurate Prediction of 327 Rice Variety Growth Period Based on Unmanned Aerial Vehicle Multispectral Remote Sensing
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Drones
Volume 9
Issue 1
10.3390/drones9010032
drones-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Improved Estimation of Aboveground Biomass in Rubber Plantations Using Deep Learning on UAV Multispectral Imagery

by
Hongjian Tan
1,
Weili Kou
1,2,
Weiheng Xu
1,2,
Leiguang Wang
3,
Huan Wang
1 and
Ning Lu
1,2,*
1
College of Big Data and Intelligent Engineering, Southwest Forestry University, Kunming 650223, China
2
Key Laboratory of National Forestry and Grassland Administration on Forestry and Ecological Big Data, Kunming 650223, China
3
College of Landscape Architecture and Horticulture, Southwest Forestry University, Kunming 650223, China
*
Author to whom correspondence should be addressed.
Drones 2025, 9(1), 32; https://doi.org/10.3390/drones9010032
Submission received: 19 November 2024 / Revised: 27 December 2024 / Accepted: 3 January 2025 / Published: 6 January 2025
(This article belongs to the Special Issue Feature Papers for Drones in Agriculture and Forestry Section: 2nd Edition)
Browse Figures
Versions Notes

Abstract

:
The accurate estimation of aboveground biomass (AGB) in rubber plantations is essential for predicting rubber production and assessing carbon storage. Multispectral sensors mounted on unmanned aerial vehicles (UAVs) can obtain high spatiotemporal resolution imagery of rubber plantations, offering significant advantages in capturing fine structural details and heterogeneity. However, most previous studies primarily focused on developing biomass estimation models for rubber using machine learning (ML) algorithms in conjunction with feature selection methods based on UAV-acquired multispectral imagery. The reliance on feature selection methods limits the model’s generalizability, robustness, and predictive accuracy. In contrast, deep learning (DL) exhibits considerable promise in extracting features from high-resolution UAV-based multispectral imagery without the need for manual selection. Nonetheless, it remains unclear whether DL can surpass traditional ML methods in improving the AGB estimation accuracy in rubber plantations. To address this, our study evaluated the performance of three ML algorithms (random forest regression, RFR; XGBoost regression, XGBR; categorical boosting regression, CatBoost) combined with feature selection techniques and a deep convolutional neural network (DCNN) using multispectral imagery obtained from UAV for the AGB estimation of rubber plantations. The results indicate that the RFR combined with a principal component analysis (PCA) for feature selection yielded the best performance (R2 = 0.81, RMSE = 11.63 t/ha, MAE = 9.27 t/ha) between the three ML algorithms. Meanwhile, the DCNN model derived from the G, R, and NIR spectral bands achieved the highest estimation accuracy (R2 = 0.89, RMSE = 6.44 t/ha, MAE = 5.72 t/ha), where it outperformed the other ML methods. Our study highlights the great potential of combining UAV-based multispectral imagery with DL techniques to improve AGB estimation in rubber plantations, offering a new perspective for estimating the physiological and biochemical growth parameters of forests.

1. Introduction

Rubber trees (Hevea brasiliensis), as an extensively cultivated economic crop in tropical regions, play a crucial role in natural rubber production. Natural rubber, an irreplaceable industrial raw material, is widely used in the manufacturing of tires, medical equipment, and various other industrial products [1]. In recent years, with the rapid expansion of rubber plantations, particularly in the Xishuangbanna region of China, over 20% of the land has been converted into rubber plantations. This expansion has not only boosted the local economic development but also provided vital support to the rubber industry [2].
The expansion of rubber plantations poses significant challenges to local ecosystems, particularly impacting biodiversity and ecosystem functioning. The conversion of old-growth forests and farmlands, along with the spread of rubber plantations, has substantially altered the distribution and accumulation of aboveground biomass (AGB), further disrupting the regional ecological balance and resource availability [3]. AGB, comprising approximately 70% to 90% of the total biomass in forests, plays a crucial role as a carbon pool within forest ecosystems. It also serves as a significant indicator of forest vegetation health and the associated successional stages. Therefore, an accurate estimation of the AGB is critical for predicting production and decreasing uncertainty in carbon stock assessments and is instrumental in shaping effective strategic plans for forest management [4].
Traditionally, the estimation of tree AGB has relied on destructive sampling methods, which involve felling trees in a specific area and directly measuring their biomass [5]. Although this method achieves the highest accuracy, it is time-consuming and labor-intensive and has a low operational efficiency in areas with complex terrain or poor accessibility. Additionally, destructive sampling causes direct damage to ecosystems, diminishing the ecological functions and biodiversity of rubber plantations [6]. In light of these limitations, employing a more efficient method is essential for accurately estimating the AGB in rubber plantations. In recent decades, remote sensing was proved to be a non-destructive technique and exhibited great potential in estimating the AGB [7]. For instance, Chen, et al. [8] developed a model for estimating the AGB in rubber plantations on Hainan Island using satellite imagery, which achieved an impressive performance (R2 = 0.83 and RMSE = 12.48 Mg/ha) and estimated a total AGB of approximately 5.40 × 107 Mg in 2017. However, this satellite-based AGB estimation method is susceptible to cloud cover and lower spatial resolution in small-scale rubber plantations in the Xishuangbanna region, which reduces the data accuracy [9]. In this context, the portability, ease of operation, and capability for low-altitude flights make unmanned aerial vehicles (UAVs) particularly advantageous for AGB estimation. The study of Liang, et al. [10] employed UAVs flying at low altitudes to avoid cloud interference and captured detailed spatial data to achieve fast, accurate, and non-destructive estimations of the AGBs in rubber plantations.
UAVs equipped with various sensors can provide highly accurate data for biomass estimation by capturing detailed spectral and structural characteristics of vegetation [11]. Recent studies demonstrated the successful estimation of AGB for urban vegetation, potato canopies, and Cinnamomum camphora using drones equipped with RGB, multispectral, hyperspectral, or LiDAR sensors [12,13,14]. However, due to the high cost of LiDAR equipment and data-processing technologies, especially when high-density point cloud data are required for large-scale surveys, both the budgetary and technical demands are significantly increased, limiting its widespread application in large-scale or long-term studies [15]. Similarly, RGB sensors, though cost-effective, capture only three broad spectral bands—red, green, and blue—offering limited spectral information and making it difficult to extract detailed biophysical parameters [16]. While hyperspectral sensors capture hundreds of narrow bands, providing richer spectral data, this advantage is offset by the challenge of large data volumes, which complicate the processing and analysis and significantly increase the computational costs [17]. In contrast, a multispectral camera presents an alternative among these sensors, where it is both cost-effective and user-friendly while also incorporating sensitive bands for assessing the vegetation growth vitality, which showed significant potential in estimating the AGB of oats [18] and LAI (leaf area index) of maize [19].
To improve the estimation accuracy of growth parameters in forests, numerous vegetation indices (VIs) derived from multispectral sensors were developed to reveal the difference in vegetation greenness, health, and canopy structure [14]. The normalized difference vegetation index (NDVI) is the most widely used index and has achieved promising accuracy in estimating the biomass of vegetation [20]. Although several studies employed the NDVI, difference vegetation index (DVI), and ratio vegetation index (RVI) to improve the AGB estimation accuracy of Carpinus betulus [21], VIs suffer from saturation issues in high-density or high-biomass forest surroundings [22]. To reduce the saturation effects of VIs, researchers attempted to extract texture features (TFs) from UAV-based multispectral imagery for AGB estimation in 2009 [23]. They found that compared with using only VIs, the use of TFs in addition to VIs significantly improved the estimation accuracy of temperate forest AGB. Similarly, the study of Liu, et al. [24] also demonstrated that the combination of VIs and TFs resulted in a reduction in RMSE values by 7.3% to 15.7% compared with using only VIs. However, these studies employed a large number of irrelevant or redundant variables, such as VIs and TFs, which not only complicated the model construction but also diminished the model’s predictive capability, increased the computational complexity, and reduced the generalization ability [2]. Thus, suitable feature selection algorithms showed great potential in determining sensitive feature variables and improving the accuracy of estimating crop growth parameters [25].
Principal component analysis (PCA), importance-based random forest regression (RFR), and variable selection using random forests (VSURF) were proved to effectively reduce redundant features from large sets of remote sensing variables [2,26]. PCA reduces the data redundancy by transforming a number of correlated variables into a set of uncorrelated variables while preserving the integrity of the information. Research by Wu, et al. [27] demonstrated that by using PCA to extract 8 principal components from the original 26 variables to estimate forest AGB, a cumulative explained variance of 95.69% was achieved. In contrast, importance analysis, as an effective method for evaluating the impact of features on the target variable, can identify key variables by quantifying the contribution of each feature to the model’s prediction results, which helps to reduce the feature dimensions and optimize the model performance. Genuer, et al. [28] proposed a strategy using a random forest to rank explanatory variables through a random variable introduction, assigning importance scores to each feature, thereby reducing the original set of 6,033 variables to 9 for interpretation and 6 for prediction. The latest study by Chen, et al. [29] developed the RFR-ANN-AGB model using these selected variables to achieve a high accuracy in estimating cotton AGB (R2 = 0.86, RMSE = 0.23 kg·m−2). In addition, VSURF is an efficient feature selection method designed to identify the most important features closely related to the target variable from high-dimensional data [26]. Previous studies showed that using variables selected through PCA, importance-based RFR, and VSURF in various machine learning (ML) models can greatly decrease the number of variables without compromising the model accuracy [30,31,32].
ML algorithms have been widely applied in vegetation AGB estimation based on remote sensing data. When addressing the nonlinear relationship between the AGB in rubber plantations and remote sensing parameters, ML effectively captures the underlying features and patterns within the data, significantly enhancing the accuracy of the AGB estimation [2]. For instance, Wang, et al. [14] employed the extreme gradient boosting (XGBoost) algorithm to estimate the AGB of C. camphora, which has a growth structure similar to rubber trees, where they achieved an excellent predictive performance (R2 = 0.929, RMSE = 587.746 kg·m−2). Although previous studies showed that ML techniques coupled with feature selection methods can achieve promising accuracy in estimating the physiological parameters of crops, the feature extraction process is not only time-consuming and complex but also heavily reliant on specialized domain knowledge and expertise [33,34]. Furthermore, the types and quantity of extracted feature variables are highly correlated with the feature selection algorithms, which will inevitably limit the capabilities and applicability of ML prediction models built on these feature variables [35,36]. In contrast, deep convolutional neural networks (DCNNs), based on deep learning (DL) techniques, eliminate the need for manual feature extraction from remote sensing images by automatically identifying significant features from large raw image datasets, and have been widely used in object classification, crop disease detection, and crop AGB estimation [37,38,39]. A recent study explored the use of DCNN to assess the impact of flight altitude and multi-scale remote sensing information on wheat AGB. The findings indicate that the DCNN model, developed using original high-spatial-resolution images, surpassed the ML models in estimating the wheat AGB (Zhu, et al. [40]). Although some studies attempted to estimate the AGB of crops, such as wheat, maize [41], and potatoes [42], using DL, it remains unclear whether a DCNN model constructed from UAV-based multispectral images will outperform traditional ML models in estimating the AGB of rubber trees, which, as tall trees, have canopy structures and other characteristics that are significantly different from those of field crops.
Therefore, the objectives of this study were as follows: (1) to systematically evaluate the advantages of variable selection methods combined with ML techniques for estimating the AGB in rubber plantations; (2) to determine the optimal model for estimating the AGB in rubber plantations using a DCNN in conjunction with various band combinations from UAV-based multispectral imagery; and (3) to compare the effectiveness of DCNN-based models to traditional ML models for AGB estimation in rubber plantations.

2. Materials and Methods

2.1. Study Area

This experiment was conducted in Jinghong City, Xishuangbanna Dai Autonomous Prefecture, located in southern Yunnan Province, China. The specific distribution of the sample plots and the under-canopy growth conditions of the rubber plantations are shown in Figure 1. This region experiences a tropical monsoon climate, with annual precipitation ranging from 1200 to 1700 mm and an average annual temperature of 23.5 °C. Rubber plantations are widely distributed in the low-altitude areas of Jinghong City, primarily concentrated at elevations between 500 and 1000 m. The cultivation of rubber trees is not only a major pillar of the local economy, bringing substantial income to Jinghong City, but it has also significantly improved the living standards of local residents. Additionally, the rubber industry provides stable employment opportunities and promotes the overall development of the rural economy. The extensive root systems of rubber trees play a crucial role in stabilizing soil and preventing erosion. During periods of frequent rainfall typical of the tropical monsoon climate, rubber plantations effectively reduce soil erosion, contributing to the sustainable use of the land.
In the study area, a total of 80 plots with varying tree ages and elevations were selected for field investigations, with a size of 20 × 25 m2 for each plot. These plots consisted of 40 from 2023 and another 40 from 2024.

2.2. Data Acquisition and Preprocessing

2.2.1. AGB Measurements

To measure the diameter at breast height (DBH) and coordinates of each rubber tree within the plots, a real-time kinematic instrument called ZHD V200 (RTK, Guangzhou Hi-Target Navigation Tech Co., Ltd., Guangzhou, China) was used to determine the boundaries of the sample plots and the coordinates of each rubber tree. Additionally, investigators measured the DBH of each rubber tree at 1.3 m above the ground using a diameter tape and manually counted the number of rubber trees within each plot. Typically, during the early growth stages of rubber trees, the tree height is positively correlated with the DBH, meaning that as the DBH increases, the tree height also increases, resulting in higher AGB values. However, as the rubber tree reaches the end of its rapid growth phase, the growth rate of the tree height gradually slows down, even though the DBH continues to increase [43].
To obtain the field-measured AGB of rubber trees, the traditional method involves destructive sampling. This process entails felling rubber trees in the sample area; separating and drying the tree trunk, branches, leaves, and other organs to a constant weight; and then measuring their dry weight to determine the AGB [44]. Although this approach can obtain an accurate AGB of rubber trees, it is impractical to cut down hundreds of rubber trees to attain the AGB of sampling plots due to the limitations in financial cost and its time-consuming nature, which also impacts the rubber yield. Thus, this study adopted the allometric growth equation model developed by Tang, et al. [45] for the Xishuangbanna rubber plantation to determine the total biomass of individual rubber trees (Equation (1)) and belowground biomass (Equation (2)). Although these AGB models only include the DBH parameter, they account for the effects of tree age and variety on the biomass, achieving a high accuracy (R2 > 0.97) in AGB calculations for rubber plantations in Xishuangbanna. Consequently, the AGB of individual trees was obtained by subtracting the belowground biomass from the total biomass (Equation (3)). The allometric growth equations for rubber forest biomass in Xishuangbanna are as follows:
W T = 0.136 D B H 2.437
W R = 0.108 D B H 1.948
W A G B = 0.136 D B H 2.437 − 0.108 D B H 1.948
where W T , W R , and W A G B represent the total biomass, belowground biomass, and AGB of the rubber tree, respectively, and DBH is the diameter (cm) measured at 1.3 m above the ground. In this study, four major rubber tree cultivars commonly planted in the region were selected for field surveys, with an elevation range of 600 to 900 m. The basic statistical information on the AGBs of the rubber plantations in the sample plots is presented in Table 1.

2.2.2. UAV Images Acquisition and Processing

In this study, a DJI Mavic 3 M (SZ DJI Technology Co., Shenzhen, China) quadcopter equipped with a multispectral camera and a visible light camera was used to capture images of the rubber plantations with a spatial resolution (~3 cm) at the 100 m flight height in May 2023 and May 2024, minimizing the impact of the differences in the rubber growth periods. The multispectral camera captured bands, including green (G): 560 nm ± 16 nm, red (R): 650 nm ± 16 nm, red edge (RE): 730 nm ± 16 nm, and near-infrared (NIR): 860 nm ± 26 nm, with an effective pixel count of ≥4.9 million. The visible light camera has an effective pixel count of ≥19 million, with a maximum photo resolution of 5280 × 3956 pixels. The UAV utilizes its built-in RTK system to accurately obtain GPS location information, and the captured images are saved in a JPG or TIF format. Additionally, to ensure high-quality imagery for both acquisition periods, the camera’s international standard organization (ISO) was set to 100, and the flight operations were conducted between 10:30 AM and 2:30 PM local time under clear, stable weather conditions with low wind speeds.
To generate mosaicked orthophotos for each sample plot, the UAV images were processed using DJI Terra software (version 3.2.0, DJI, Shenzhen, Guangdong, China). Reflectance correction was applied to the acquired multispectral images to mitigate the effects of seasonal variations, as well as illumination and environmental conditions. The workflow in DJI Terra (https://enterprise.dji.com/cn/dji-terra, accessed on 5 August 2024) included radiometric correction, point cloud generation, digital surface model (DSM) creation, and orthophoto generation [46]. The resulting orthophotos were saved in the TIF format for ease of subsequent processing.

2.3. Spectral and Textural Metrics Calculations

2.3.1. Spectral Indices Calculations

Given the AGB related to the LAI (leaf area index) and canopy coverage of crops, this study employed 20 vegetation indices (Table 2) that provide information on plant coverage and growth conditions for the AGB estimation of rubber plantations.

2.3.2. Textural Metrics Calculations

TFs can describe the spatial distribution and variation of pixel intensity values within an image, reflecting information about the vegetation. In this study, we employed 17 texture metrics derived from the gray-level co-occurrence matrix (GLCM) using parameter settings of a 135° orientation, 2 displacements, and a 7 × 7 window size based on previous studies [2,10] to estimate the AGB of the rubber plantations. The specific texture metrics were as follows: angular second moment (ASM), contrast (contrast), correlation (CORR), variance (VAR), inverse difference moment (IDM), sum average (SAVG), sum variance (SVAR), sum entropy (SENT), entropy (ENT), difference variance (DVAR), difference entropy (DENT), information measure of corr. 1 (IMCORR1), information measure of corr. 2 (IMCORR 2), dissimilarity (DISS), inertia (inertia), cluster shade (shade), and cluster prominence (PROM).

2.4. Regression Techniques

2.4.1. Random Forest Regression

Random forest regression (RFR) is a regression algorithm that combines multiple decision trees to enhance the model stability and accuracy. By aggregating predictions from these trees, RFR excels at handling high-dimensional, nonlinear data and effectively reduces overfitting, making it ideal for complex regression problems [64]. It is robust against noise and missing data, maintaining reliable performance even under imperfect conditions. Additionally, RFR is user-friendly, delivering strong results without extensive parameter tuning. Its ability to process diverse feature types and efficiently handle large-scale datasets further highlights its versatility, making RFR a preferred tool in fields like regression forecasting, environmental science, and medical research.

2.4.2. XGBoost Regression

XGBoost regression (XGBR) is a highly efficient and accurate regression algorithm built on the gradient boosting framework. By sequentially constructing decision trees, each tree corrects the errors of the previous ones, resulting in a well-optimized model [65]. XGBR is particularly effective for handling high-dimensional and sparse datasets, making it ideal for complex, nonlinear regression problems. It incorporates regularization to prevent overfitting and supports various loss functions, enhancing both its flexibility and robustness. XGBR is also highly scalable, efficiently processing large datasets with minimal computational resources. Additionally, it offers fine-grained control over the learning process through a wide range of hyperparameters, allowing users to tailor the model for specific tasks.

2.4.3. Categorical Boosting Regression

Categorical boosting (CatBoost) regression is a sophisticated regression technique within the gradient boosting framework, specifically designed to handle categorical features with enhanced efficiency. CatBoost builds a predictive model by sequentially adding decision trees, where each new tree focuses on correcting the residual errors left by the previous tree. This iterative process results in a robust ensemble model that excels at capturing complex, nonlinear relationships within the data, making it particularly adept at tackling intricate regression tasks [66]. One of CatBoost’s standout features is its ability to process categorical features directly, eliminating the need for extensive preprocessing. This capability not only helps to mitigate overfitting but also reduces the risk of data leakage. In addition to its handling of categorical features, CatBoost incorporates a variety of regularization techniques, including learning rate adjustments and tree pruning, to further combat overfitting and improve the model generalization. Its flexibility in choosing loss functions allows for optimization tailored to specific regression problems, enhancing performance across diverse applications. CatBoost’s precision and reliability underscore its effectiveness in predictive modeling and analytics, making it a powerful tool for addressing complex regression challenges.

2.4.4. Regression Techniques Based on DCNN

This study employed a DCNN based on the AlexNet architecture, which was proved to produce high accuracy and reliability when estimating crop AGB [40]. The AlexNet design comprises five convolutional layers, three max-pooling layers, and three fully connected layers (Figure 2). It utilizes the rectified linear unit (ReLU) activation function, a non-saturating function that, when compared with traditional tanh and sigmoid activation functions, effectively mitigates the vanishing gradient problem and accelerates the training process. The use of ReLU enables AlexNet to achieve faster convergence and improved performance when handling complex nonlinear data, thereby enhancing the overall model performance.
To investigate the performance of the DL models in estimating the AGB in the rubber plantations, we first performed image stitching on the multispectral UAV images, which had been corrected for reflectance, to generate orthophotos. Subsequently, regions of interest (ROIs) were determined based on the coordinates of the four corner points of the sample plots obtained during the field survey, and the reflectance imagery within these ROIs was extracted for the prediction of AGB in rubber plantations. Following the image fusion method proposed by Palsson, et al. [67], which involves stacking hyperspectral and multispectral images into a single input image for a DCNN, we extracted the matrix data from the multispectral bands at the same location, normalized the data, and stacked them together. Therefore, the dimensions of the input layer dynamically vary, defined as a 224 × 224 × 3 channel. Ultimately, the model outputs a regression value, specifically predicting the AGBs of rubber plantations based on the input reflectance imagery.

2.5. Features Selection and Models Assessment

2.5.1. Principal Component Analysis

PCA is a widely used dimensionality reduction technique that maps high-dimensional data to a lower-dimensional space through a linear transformation while retaining as much of the original data’s significant information as possible. PCA extracts the principal components that explain the maximum variance in the data by computing the eigenvalues and eigenvectors of the covariance matrix. Previous research often employed PCA to process remote sensing parameters, with the resulting principal components frequently showing high correlations with biomass [68]. This study employed PCA to extract the top 10 principal components from 20 VIs, 17 TFs, and their combinations in order to evaluate their performance when estimating the AGB of the rubber trees.

2.5.2. Importance Analysis

Importance analysis based on RFR and VSURF was used to identify key parameters related to the AGBs of the rubber plantations [63,69]. Unlike the variable selection method based on RFR importance analysis, the VSURF-based approach directly yields the best feature variable combination without the need for manual selection based on importance ranking [26]. This helps reduce model errors and improve the predictive performance of the model. This study evaluated the performance of RFR and VSURF in determining the optimal feature variables extracted from 20 VIs, 17 TFs, and their combination for enhancing the accuracy of the AGB estimation for the rubber plantations.

2.5.3. Accuracy Analysis

In this study, 80 plots were divided into a training set that consisted of 56 plots and a testing set that comprised 24 plots, following a 7:3 ratio through random stratified sampling methods. This approach aimed to ensure that a sufficient testing set was reserved for model performance evaluation, while also maintaining an adequate training set to support model learning [10]. To optimize the model’s performance, we employed a grid search method to determine the model’s hyperparameters [70]. Specifically, for both the RFR and XGBR models, the number of trees was set to 150, with a maximum depth of 16. For the CatBoost model, the number of iterations was set to 300, the learning rate to 0.01, and the maximum depth to 10. Moreover, to ensure the robustness of the predictive models, 5-fold cross-validation was used, with training samples serving as the foundation for the model development [2]. All these procedures and model establishments were implemented in the Python 3.8 environment. The model’s performance was assessed using the coefficient of determination (R2), root-mean-square error (RMSE), and mean absolute error (MAE). The equations for these three model evaluation metrics are as follows:
R 2 = 1 − ∑ i = 1 n y i − y i ^ 2 ∑ i = 1 n y i − y i ¯ 2
R M S E = ∑ i = 1 n y i − y i ^ 2 n − 1
M A E = 1 n ∑ i = 1 n y ^ i − y i
where y i is the measured value, y ^ i is the measured average value, y i ¯ is the sample mean value, and n is the number of samples.
The workflow for estimating the rubber plantation AGB is shown in Figure 3. First, we performed comprehensive processing on multiple UAV-captured images and field-sampled data to obtain the ground truth AGB measurements, as well as the VIs and TFs extracted from the orthophotos. Next, we divided these processed parameters into training and testing datasets to ensure the representativeness of the model training and validation. After the data partitioning, we conducted importance analysis and PCA to identify the most critical variables and principal components for estimating the rubber plantation AGB. Subsequently, we established three control groups—all variables, variables selected through importance analysis, and variables selected through PCA—in order to compare the performance of different variable selection methods. These three datasets were then input into three different ML models to construct AGB estimation models. To reduce the potential loss of key AGB-related information during manual feature extraction, we employed the original UAV images derived from five different band combinations to establish an optimal DCNN-based DL model and explored whether DL could improve the AGB estimation accuracy. Meanwhile, we employed a 5-fold cross-validation approach with repeated iterations to ensure the model robustness. Finally, three accuracy evaluation metrics were used to comprehensively assess the performance of each model, ultimately identifying the optimal method for estimating the biomass of rubber plantations.

3. Results

3.1. Machine Learning-Based Model for Estimating AGB in Rubber Plantations

To achieve an accurate AGB estimation in rubber plantations, this study evaluated the performance of three ML regressions (RFR, CatBoost, and XGBR) using all the VIs, TFs, and their combination without feature selection. As shown in Table 3, the CatBoost model outperformed the other two models, where it achieved the highest accuracy (R2 = 0.71, RMSE = 14.60 t/ha, MAE = 11.54 t/ha) when using the combination of VIs and TFs. Overall, when using a single feature set, the VIs more effectively reflected the relationship with the rubber plantations’ AGBs compared with the TFs. Figure 4 displays the assessment of the AGB estimation accuracy based on the combination of VIs and TFs using ML regression techniques. In contrast, combining both the VIs and TFs consistently yielded higher accuracies than using either feature alone, a trend observed across all three models.

3.2. Performance Assessment of Feature Selection Methods Using Machine Learning Regression Techniques

3.2.1. Performance of Feature Importance Analysis with ML Techniques

To investigate the feature importance, we conducted an analysis by dividing the VIs, TFs, and their combination into three groups. Considering the potential noise introduced by an excessive number of parameters, we selected the features that accounted for 98% of the importance from each group based on the importance analysis results derived from the RFR.
Figure 5 presents the selected features obtained through the importance analysis method based on the RFR, with a focus on the VIs, TFs, and their combination. For the VIs, the features with an importance score greater than 0.1 were the TVI (score = 0.33) and DVI (score = 0.14). In the case of the TFs, the features that exceeded an importance score of 0.1 included the contrast (score = 0.31), corr (score = 0.21), and var (score = 0.12). When the VIs and TFs were combined for the feature importance analysis, the features with scores above 0.1 were the TVI (score = 0.31) and DVI (score = 0.15). The features with the highest importance scores, TVI and contrast, along with their corresponding original RGB images, are displayed in Figure 6.
In this study, we performed feature selection on the VIs, TFs, and their combinations using the VSURF. The chosen VIs were the TVI and DVI, while the selected TFs included the SAVG and contrast. The final combination of the VIs and TFs comprised TCARI, DVI, GNDVI, RDVI, RRI, inertia, and SAVG. Additionally, we evaluated the performance of the feature selection based on the RFR- and VSURF-selected features in three machine learning models. As shown in Table 4, the results indicate that when using the combination of the VIs and TFs selected by the RFR as model parameters, the RFR model achieved the highest accuracy (R2 = 0.73, RMSE = 13.9 t/ha, MAE = 11.01 t/ha). The CatBoost model followed (R2 = 0.71, RMSE = 14.40 t/ha, MAE = 11.46 t/ha). In contrast, the models built with the VSURF-selected variables did not outperform those built with the RFR-selected variables in terms of accuracy (Figure 7).

3.2.2. Performance of PCA with ML Techniques

To reduce the dimensionality of numerous feature variables and build reliable predictive models, this study employed PCA to obtain principal components from the VIs, TFs, and their combinations, respectively. As shown in Figure 8, the findings indicate that five principal components derived from the VIs, three from the TFs, and six from their combination each contributed more than 1% to the variance. In particular, the PCA algorithm exhibited a stronger capability to extract information from the TFs compared with the VIs, which resulted in the reduction of all the TFs to just three principal components (Figure 8b).
This study evaluated the performance of the PCA based on the VIs, TFs, and their combination for estimating the AGBs of the rubber plantations by incorporating three ML techniques, with each principal component having a variance contribution greater than 1%. As shown in Table 5, the AGB estimation model that utilized the RFR achieved the highest accuracy (R2 = 0.81, RMSE = 11.63 t/ha, MAE = 9.27 t/ha) when both the VIs and TFs were used as model inputs, followed closely by the models that used CatBoost and XGBR. Notably, the VIs_PCA derived from the VIs using PCA outperformed the TFs_PCA extracted from the TFs in estimating the AGB of the rubber plantations when applied with the same ML regression algorithm. Figure 9 displays the assessment of the AGB estimation accuracy based on the combination of the VIs and TFs using PCA. Furthermore, it is clear that PCA demonstrated great potential in capturing the relationship between feature variables and the AGBs of the rubber plantations since it surpassed the importance analysis methods (Table 4).

3.3. Performance of Deep Learning for Estimating AGB in Rubber Plantations

To establish a high-accuracy and reliable AGB predictive model, this study evaluated the capabilities of RGB images and different spectral band combinations obtained from UAV-based multispectral sensors coupled with DL techniques to estimate the AGBs of the rubber plantations. Figure 10 illustrates the accuracy assessment of the AGB estimations of the rubber plantations derived from the RGB images and the DCNN. Obviously, the model developed using the original RGB imagery from UAVs combined with the DCNN demonstrated a lower estimation accuracy, where it achieved an R2 of 0.50, an RMSE of 12.78 t/ha, and an MAE of 10.78 t/ha.
Figure 11 illustrates the performance of four different combinations of the R, G, NIR, and RE bands derived from the multispectral imagery coupled with the DCNN in estimating the AGBs of the rubber plantations. The results show that the model based on the R, G, and NIR bands achieved the highest predictive accuracy (R2 = 0.89, RMSE = 6.44 t/ha, MAE = 5.72 t/ha), while the combination of G, NIR, and RE exhibited the lowest performance (R2 = 0.75, RMSE = 9.53 t/ha, MAE = 7.90 t/ha). Clearly, in this study, the DL-based model demonstrated a superior performance in the AGB estimation compared with the optimal model derived from ML techniques. In particular, the DL models not only offered higher accuracies in the predictions but also demonstrated good robustness across different band combinations. Even with optimized feature extraction and selection, the accuracies of the ML models remained lower than that of the DCNN model based on the R, G, and NIR band combinations. Our results indicate that the DL models exhibited greater potential for processing multispectral data, making them a superior choice for estimating the AGBs in the rubber plantations.

4. Discussion

4.1. Comparison of Importance Analysis and Principal Component Analysis

This study evaluated the performance of two importance-analysis-based methods combined with PCA and ML techniques for estimating the AGBs of rubber plantations. Although the VSURF-based feature selection method improved the model accuracy in the study by Virdi, et al. [71], it did not greatly enhance the predictive performance of the three ML models in estimating the the AGBs of the rubber plantations. However, the principal components derived from the PCA enhanced the accuracy of the AGB estimations for the rubber plantations. This suggests that dimensionality reduction methods could effectively reduce the redundancy between similar features, thereby enhancing the model performance. These findings align with the study of Huang, et al. [72], which demonstrated that selecting uncorrelated or redundant variables during model construction often leads to a decreased estimation accuracy. This highlights the importance of using a smaller, more sensitive set of variables when dealing with a large number of predictors. Additionally, the RFR algorithm exhibited great potential in estimating the AGBs of the rubber plantations, where it achieved the highest estimation accuracy between the three ML techniques evaluated, with an R2 of 0.75, an RMSE of 9.53, and an MAE of 7.90. This superior performance can be attributed to RFR’s ability to effectively establish nonlinear relationships between the features and the target variable by capturing the key features associated with the AGB. Recent research by Liu, et al. [73] further supports this finding, where they showed that the combination of RFR and PCA significantly improves the accuracy of AGB estimation for rubber plantations.

4.2. Optimal Spectral Band Combination of DCNN Model

This study indicated that the predictive accuracy of the DCNN model is related to spectral band combinations derived from UAV-based multispectral imagery. Although previous studies showed that RGB images combined with DL models can achieve promising predictive accuracy in estimating the AGB of Malania oleifera [74], the DCNN model derived from UAV-based RGB imagery obtained only a moderate precision when estimating the AGBs of the rubber plantations (R2 = 0.50, RMSE = 12.78 t/ha, MAE = 10.78 t/ha). The possible reason for this can be attributed to the fact that RGB imagery only includes visible light wavelengths and lacks sensitive bands that reflect crop vitality, which prevents them from accurately capturing variations in biomass [75,76]. Conversely, the combination of the green, red, and NIR bands acquired from UAV-based multispectral sensors coupled with the DCNN achieved the highest accuracy (R2 = 0.89, RMSE = 6.44 t/ha, MAE = 5.72 t/ha). A possible reason is that the near-infrared (NIR) band is typically unaffected by external light interference. The fusion of NIR and visible light bands (green and red) not only offers advantages that conventional RGB imagery cannot provide but also effectively reduces noise interference [77]. This is consistent with the findings of Liu, et al. [78], who demonstrated that combining features extracted from visible light imagery and the NIR band enabled DL models to achieve the highest accuracy in kiwifruit detection.

4.3. Advantages of a DCNN When Estimating the AGBs of Rubber Plantations

This study compared the performance of DL and traditional ML methods when estimating the AGBs in the rubber plantations. As shown in Figure 12, the AGB estimation model for the rubber plantations, based on PCA and RFR, achieved a relatively high accuracy (R2 = 0.81, RMSE = 11.63 t/ha, MAE = 9.27 t/ha), but it still did not reach the superior performance of the DCNN model. Although feature selection methods can improve model performance to some extent, they may fail to capture the complex relationships within the data, especially when dealing with multidimensional and heterogeneous remote sensing datasets [79]. In contrast, a DCNN can automatically extract multi-level features from raw data, which may result in superior adaptability when handling nonlinear and high-dimensional data. In addition, automated feature learning enables a DCNN to achieve a higher accuracy and robustness in complex and variable environments. This is consistent with the findings of Krizhevsky, et al. [80], who demonstrated that DCNN effectively manages complex feature spaces without relying on manually extracted features.

4.4. Limitations and Potential Applications

The DCNN-based model for estimating the AGB in rubber plantations demonstrated significant advantages in terms of accuracy and automated feature extraction. Compared with traditional ML models, the DCNN eliminates the need for manual feature extraction and can effectively improve the estimation accuracy. However, the generalization capability of DCNNs depends on large and representative datasets [81]. Although this study selected rubber tree samples from different varieties, altitudes, and stand ages, the 80 samples obtained were relatively few. This may lead to a weakened generalizability of the DCNN model developed in this research, potentially affecting its performance across different environments, conditions, and variables [82]. Furthermore, the bands derived from UAV-based multispectral sensors utilized in this study were broad, which may not capture sensitive information relevant to the AGB [83]. In the future, we will explore the potential of using narrow-band hyperspectral imagery to further improve the model’s performance in estimating the AGB in rubber plantations.
In this study, we compared the performance of ML models and DL models in estimating the AGBs of the rubber plantations. Although the prediction process of ML models is intuitive and easy to interpret, their estimation accuracy is influenced by the selection of the type and number of remote sensing feature variables. In contrast, while DL models require more computational resources and longer training times, they have a significant advantage in automatically learning features from input images, thereby reducing the need for manual feature extraction. Furthermore, consistent with the findings of Yu, et al. [83] on maize AGB estimation, the AlexNet-based DCNN model achieved the highest estimation accuracy in this study, outperforming the other DL models. Even though our research confirmed the great potential of DCNNs in estimating rubber AGB, further research is needed to explore whether other DL models, such as YOLO, VGG, and ResNet, can improve the estimation accuracy of crop growth parameters.
This study emphasized the advantages of combining a DCNN with various band combinations from UAV-based multispectral imagery for AGB estimation in rubber plantations. This suggests that the proposed approach may also be applicable to other physiological and biochemical parameter estimations of crops using hyperspectral imagery for further investigation.

5. Conclusions

This study evaluated the performance of ML models and DL models in estimating the AGBs of the rubber plantations. For the ML models, we found that the PCA-based RFR model achieved the highest accuracy (R2 = 0.81, RMSE = 11.63 t/ha, MAE = 9.27 t/ha). In addition, the DL model based on the G, R, and NIR band combination achieved the highest estimation accuracy in this study (R2 = 0.89, RMSE = 6.44 t/ha, MAE = 5.72 t/ha). Our study found that integrating the DCNN with appropriate band combinations to extract features from UAV-based multispectral imagery exhibited great potential for improving the accuracy of AGB estimation in rubber plantations. The new DCNN-based method we proposed for estimating the AGB in rubber plantations is not only cost-effective but also reduces the need for manual feature extraction. This study provides a new perspective on the application of DL technology in estimating the physiological and biochemical parameters of forest growth.

Author Contributions

Investigation, Data curation, and Writing—original draft: H.T.; Conceptualization: W.K., L.W. and N.L.; Methodology: W.K.; Supervision: W.X.; Software and Validation: H.W.; Writing—review and editing: N.L. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (32160368, 32360435, 32360387, 32260391); the Key Laboratory of National Forestry and Grassland Administration on Forestry and Ecological Big Data, Southwest Forestry University (2022-BDK−02); the Joint Special Project for Agriculture of Yunnan Province (202101BD070001-059, 202301BD070001-160), Yunnan International Joint Laboratory of Natural Rubber Intelligent Monitor and Digital Applications (202403AP140001); and the Ten Thousand Talents Program Special Project for Young Top-notch Talents of Yunnan Province (YNWR-QNBJ-2019-270, YNWR-QNBJ-2020047).

Data Availability Statement

The datasets used and/or analyzed during the current study are available from the corresponding author upon reasonable request.

Acknowledgments

We would like to thank Limin Fuyang, Ziyi Yang, Yuying Liang, Yong Chen, Yanan Zhao, Yiqiang Yin, Maojia Gong, Xiaoqing Li, and Guiliang Chen for their help with the data collection. We also thank anonymous reviewers for their constructive comments.

Conflicts of Interest

The authors declare that they have no known competing financial interests or personal relationships that could have appeared to influence the work reported in this paper.

References

  1. Cornish, K. Similarities and differences in rubber biochemistry among plant species. Phytochemistry 2001, 57, 1123–1134. [Google Scholar] [CrossRef]
  2. Fu, Y.; Tan, H.; Kou, W.; Xu, W.; Wang, H.; Lu, N. Estimation of rubber plantation biomass based on variable optimization from Sentinel-2 remote sensing imagery. Forests 2024, 15, 900. [Google Scholar] [CrossRef]
  3. Chen, H.; Yi, Z.-F.; Schmidt-Vogt, D.; Ahrends, A.; Beckschäfer, P.; Kleinn, C.; Ranjitkar, S.; Xu, J. Pushing the limits: The pattern and dynamics of rubber monoculture expansion in Xishuangbanna, SW China. PLoS ONE 2016, 11, e0150062. [Google Scholar] [CrossRef] [PubMed]
  4. Chen, L.; Ren, C.; Zhang, B.; Wang, Z.; Xi, Y. Estimation of forest above-ground biomass by geographically weighted regression and machine learning with sentinel imagery. Forests 2018, 9, 582. [Google Scholar] [CrossRef]
  5. Burt, A.; Calders, K.; Cuni-Sanchez, A.; Gómez-Dans, J.; Lewis, P.; Lewis, S.L.; Malhi, Y.; Phillips, O.L.; Disney, M. Assessment of bias in pan-tropical biomass predictions. Front. For. Glob. Chang. 2020, 3, 12. [Google Scholar] [CrossRef]
  6. Yang, X.; Blagodatsky, S.; Liu, F.; Beckschäfer, P.; Xu, J.; Cadisch, G. Rubber tree allometry, biomass partitioning and carbon stocks in mountainous landscapes of sub-tropical China. For. Ecol. Manag. 2017, 404, 84–99. [Google Scholar] [CrossRef]
  7. Ma, T.; Zhang, C.; Ji, L.; Zuo, Z.; Beckline, M.; Hu, Y.; Li, X.; Xiao, X. Development of forest aboveground biomass estimation, its problems and future solutions: A review. Ecol. Indic. 2024, 159, 111653. [Google Scholar] [CrossRef]
  8. Chen, B.; Yun, T.; Ma, J.; Kou, W.; Li, H.; Yang, C.; Xiao, X.; Zhang, X.; Sun, R.; Xie, G. High-precision stand age data facilitate the estimation of rubber plantation biomass: A case study of Hainan Island, China. Remote Sens. 2020, 12, 3853. [Google Scholar] [CrossRef]
  9. Doughty, C.L.; Cavanaugh, K.C. Mapping coastal wetland biomass from high resolution unmanned aerial vehicle (UAV) imagery. Remote Sens. 2019, 11, 540. [Google Scholar] [CrossRef]
  10. Liang, Y.; Kou, W.; Lai, H.; Wang, J.; Wang, Q.; Xu, W.; Wang, H.; Lu, N. Improved estimation of aboveground biomass in rubber plantations by fusing spectral and textural information from UAV-based RGB imagery. Ecol. Indic. 2022, 142, 109286. [Google Scholar] [CrossRef]
  11. González-Jaramillo, V.; Fries, A.; Bendix, J. AGB estimation in a tropical mountain forest (TMF) by means of RGB and multispectral images using an unmanned aerial vehicle (UAV). Remote Sens. 2019, 11, 1413. [Google Scholar] [CrossRef]
  12. Lin, J.; Chen, D.; Wu, W.; Liao, X. Estimating aboveground biomass of urban forest trees with dual-source UAV acquired point clouds. Urban For. Urban Green. 2022, 69, 127521. [Google Scholar] [CrossRef]
  13. Li, B.; Xu, X.; Zhang, L.; Han, J.; Bian, C.; Li, G.; Liu, J.; Jin, L. Above-ground biomass estimation and yield prediction in potato by using UAV-based RGB and hyperspectral imaging. ISPRS J. Photogramm. Remote Sens. 2020, 162, 161–172. [Google Scholar] [CrossRef]
  14. Wang, Q.; Lu, X.; Zhang, H.; Yang, B.; Gong, R.; Zhang, J.; Jin, Z.; Xie, R.; Xia, J.; Zhao, J. Comparison of machine learning methods for estimating leaf area index and aboveground biomass of Cinnamomum camphora based on UAV multispectral remote sensing data. Forests 2023, 14, 1688. [Google Scholar] [CrossRef]
  15. Bartmiński, P.; Siłuch, M.; Kociuba, W. The effectiveness of a UAV-based LIDAR survey to develop digital terrain models and topographic texture analyses. Sensors 2023, 23, 6415. [Google Scholar] [CrossRef]
  16. Adão, T.; Hruška, J.; Pádua, L.; Bessa, J.; Peres, E.; Morais, R.; Sousa, J.J. Hyperspectral imaging: A review on UAV-based sensors, data processing and applications for agriculture and forestry. Remote Sens. 2017, 9, 1110. [Google Scholar] [CrossRef]
  17. Zhou, H.; Liu, Z.; Huang, Z.; Wang, X.; Su, W.; Zhang, Y. ICTH: Local-to-global spectral reconstruction network for heterosource hyperspectral images. Remote Sens. 2024, 16, 3377. [Google Scholar] [CrossRef]
  18. Sharma, P.; Leigh, L.; Chang, J.; Maimaitijiang, M.; Caffé, M. Above-ground biomass estimation in oats using UAV remote sensing and machine learning. Sensors 2022, 22, 601. [Google Scholar] [CrossRef] [PubMed]
  19. Zhang, X.; Zhang, K.; Sun, Y.; Zhao, Y.; Zhuang, H.; Ban, W.; Chen, Y.; Fu, E.; Chen, S.; Liu, J. Combining spectral and texture features of UAS-based multispectral images for maize leaf area index estimation. Remote Sens. 2022, 14, 331. [Google Scholar] [CrossRef]
  20. Sánchez, N.; Plaza, J.; Criado, M.; Pérez-Sánchez, R.; Gómez-Sánchez, M.Á.; Morales-Corts, M.R.; Palacios, C. The second derivative of the NDVI time series as an estimator of fresh biomass: A case study of eight forage associations monitored via UAS. Drones 2023, 7, 347. [Google Scholar] [CrossRef]
  21. Moradi, F.; Darvishsefat, A.A.; Pourrahmati, M.R.; Deljouei, A.; Borz, S.A. Estimating aboveground biomass in dense Hyrcanian forests by the use of Sentinel-2 data. Forests 2022, 13, 104. [Google Scholar] [CrossRef]
  22. Matsushita, B.; Yang, W.; Chen, J.; Onda, Y.; Qiu, G. Sensitivity of the enhanced vegetation index (EVI) and normalized difference vegetation index (NDVI) to topographic effects: A case study in high-density cypress forest. Sensors 2007, 7, 2636–2651. [Google Scholar] [CrossRef] [PubMed]
  23. Meng, S.; Pang, Y.; Zhang, Z.; Jia, W.; Li, Z. Mapping aboveground biomass using texture indices from aerial photos in a temperate forest of Northeastern China. Remote Sens. 2016, 8, 230. [Google Scholar] [CrossRef]
  24. Liu, Y.; Liu, S.; Li, J.; Guo, X.; Wang, S.; Lu, J. Estimating biomass of winter oilseed rape using vegetation indices and texture metrics derived from UAV multispectral images. Comput. Electron. Agric. 2019, 166, 105026. [Google Scholar] [CrossRef]
  25. Zhang, Y.; Xia, C.; Zhang, X.; Cheng, X.; Feng, G.; Wang, Y.; Gao, Q. Estimating the maize biomass by crop height and narrowband vegetation indices derived from UAV-based hyperspectral images. Ecol. Indic. 2021, 129, 107985. [Google Scholar] [CrossRef]
  26. Genuer, R.; Poggi, J.-M.; Tuleau-Malot, C. VSURF: An R package for variable selection using random forests. R J. 2015, 7, 19–33. [Google Scholar] [CrossRef]
  27. Wu, Z.; Yao, F.; Zhang, J.; Liu, H. Estimating forest aboveground biomass using a combination of Geographical Random Forest and Empirical Bayesian Kriging models. Remote Sens. 2024, 16, 1859. [Google Scholar] [CrossRef]
  28. Genuer, R.; Poggi, J.-M.; Tuleau-Malot, C. Variable selection using random forests. Pattern Recognit. Lett. 2010, 31, 2225–2236. [Google Scholar] [CrossRef]
  29. Chen, M.; Yin, C.; Lin, T.; Liu, H.; Wang, Z.; Jiang, P.; Ali, S.; Tang, Q.; Jin, X. Integration of unmanned aerial vehicle spectral and textural features for accurate above-ground biomass estimation in cotton. Agronomy 2024, 14, 1313. [Google Scholar] [CrossRef]
  30. Hasan, M.A.M.; Nasser, M.; Ahmad, S.; Molla, K.I. Feature selection for intrusion detection using random forest. J. Inf. Secur. 2016, 7, 129–140. [Google Scholar] [CrossRef]
  31. Chandrashekar, G.; Sahin, F. A survey on feature selection methods. Comput. Electr. Eng. 2014, 40, 16–28. [Google Scholar] [CrossRef]
  32. Speiser, J.L.; Miller, M.E.; Tooze, J.; Ip, E. A comparison of random forest variable selection methods for classification prediction modeling. Expert Syst. Appl. 2019, 134, 93–101. [Google Scholar] [CrossRef]
  33. Le, Q.; Boydell, O.; Mac Namee, B.; Scanlon, M. Deep learning at the shallow end: Malware classification for non-domain experts. Digit. Investig. 2018, 26, S118–S126. [Google Scholar] [CrossRef]
  34. Sowmya, A.; Trinder, J. Modelling and representation issues in automated feature extraction from aerial and satellite images. ISPRS J. Photogramm. Remote Sens. 2000, 55, 34–47. [Google Scholar] [CrossRef]
  35. Yuan, Q.; Shen, H.; Li, T.; Li, Z.; Li, S.; Jiang, Y.; Xu, H.; Tan, W.; Yang, Q.; Wang, J. Deep learning in environmental remote sensing: Achievements and challenges. Remote Sens. Environ. 2020, 241, 111716. [Google Scholar] [CrossRef]
  36. Wang, L.; Zhang, J.; Liu, P.; Choo, K.-K.R.; Huang, F. Spectral–spatial multi-feature-based deep learning for hyperspectral remote sensing image classification. Soft Comput. 2017, 21, 213–221. [Google Scholar] [CrossRef]
  37. Simhadri, C.G.; Kondaveeti, H.K. Automatic recognition of rice leaf diseases using transfer learning. Agronomy 2023, 13, 961. [Google Scholar] [CrossRef]
  38. Du, X.; Sun, Y.; Song, Y.; Sun, H.; Yang, L. A comparative study of different CNN models and transfer learning effect for underwater object classification in side-scan sonar images. Remote Sens. 2023, 15, 593. [Google Scholar] [CrossRef]
  39. Zhang, L.; Zhang, L.; Du, B. Deep learning for remote sensing data: A technical tutorial on the state of the art. IEEE Geosci. Remote Sens. Mag. 2016, 4, 22–40. [Google Scholar] [CrossRef]
  40. Zhu, W.; Rezaei, E.E.; Nouri, H.; Sun, Z.; Li, J.; Yu, D.; Siebert, S. UAV flight height impacts on wheat biomass estimation via machine and deep learning. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2023, 16, 7471–7485. [Google Scholar] [CrossRef]
  41. An, J.; Li, W.; Li, M.; Cui, S.; Yue, H. Identification and classification of maize drought stress using deep convolutional neural network. Symmetry 2019, 11, 256. [Google Scholar] [CrossRef]
  42. Liu, Y.; Feng, H.; Yue, J.; Jin, X.; Fan, Y.; Chen, R.; Bian, M.; Ma, Y.; Li, J.; Xu, B. Improving potato AGB estimation to mitigate phenological stage impacts through depth features from hyperspectral data. Comput. Electron. Agric. 2024, 219, 108808. [Google Scholar] [CrossRef]
  43. Li, Y.Q.; Deng, X.W.; Huang, Z.H.; Xiang, W.H.; Yan, W.D.; Lei, P.F.; Zhou, X.L.; Peng, C.H. Development and evaluation of models for the relationship between tree height and diameter at breast height for Chinese-fir plantations in subtropical China. PLoS ONE 2015, 10, e0125118. [Google Scholar] [CrossRef]
  44. Revenga, J.C.; Trepekli, K.; Oehmcke, S.; Jensen, R.; Li, L.; Igel, C.; Gieseke, F.C.; Friborg, T. Above-ground biomass prediction for croplands at a sub-meter resolution using UAV–LiDAR and machine learning methods. Remote Sens. 2022, 14, 3912. [Google Scholar] [CrossRef]
  45. Tang, J.; Long, J.; Chen, M.; Guo, X.; Ceng, R. Biomass and its estimation model of rubber plantations in Xishuangbanna, Southwest China. Chin. J. Ecol. 2009, 28, 1942. [Google Scholar]
  46. Jarahizadeh, S.; Salehi, B. A comparative analysis of UAV photogrammetric software performance for forest 3D modeling: A case study using AgiSoft Photoscan, PIX4D Mapper, and DJI Terra. Sensors 2024, 24, 286. [Google Scholar] [CrossRef]
  47. Tucker, C.J. Red and photographic infrared linear combinations for monitoring vegetation. Remote Sens. Environ. 1979, 8, 127–150. [Google Scholar] [CrossRef]
  48. Rouse, J.W.; Haas, R.H.; Schell, J.A.; Deering, D.W. Monitoring vegetation systems in the Great Plains with ERTS. NASA Spec. Publ. 1974, 351, 309. [Google Scholar]
  49. Fitzgerald, G.; Rodriguez, D.; O’Leary, G. Measuring and predicting canopy nitrogen nutrition in wheat using a spectral index—The canopy chlorophyll content index (CCCI). Field Crops Res. 2010, 116, 318–324. [Google Scholar] [CrossRef]
  50. Qi, J.; Chehbouni, A.; Huerte, A.; Kerr, Y.; Sorooshian, S. A modified soil adjusted vegetation index. Remote Sens. Environ. 1994, 48, 119–126. [Google Scholar] [CrossRef]
  51. Ren, H.; Zhou, G.; Zhang, F. Using negative soil adjustment factor in soil-adjusted vegetation index (SAVI) for aboveground living biomass estimation in arid grasslands. Remote Sens. Environ. 2018, 209, 439–445. [Google Scholar] [CrossRef]
  52. Chen, J.M. Evaluation of vegetation indices and a modified simple ratio for boreal applications. Can. J. Remote Sens. 1996, 22, 229–242. [Google Scholar] [CrossRef]
  53. Pinty, B.; Verstraete, M. GEMI: A non-linear index to monitor global vegetation from satellites. Vegetatio 1992, 101, 15–20. [Google Scholar] [CrossRef]
  54. Rondeaux, G.; Steven, M.; Baret, F. Optimization of soil-adjusted vegetation indices. Remote Sens. Environ. 1996, 55, 95–107. [Google Scholar] [CrossRef]
  55. Daughtry, C.S.; Walthall, C.; Kim, M.; De Colstoun, E.B.; McMurtrey Iii, J. Estimating corn leaf chlorophyll concentration from leaf and canopy reflectance. Remote Sens. Environ. 2000, 74, 229–239. [Google Scholar] [CrossRef]
  56. Haboudane, D.; Miller, J.R.; Tremblay, N.; Zarco-Tejada, P.J.; Dextraze, L. Integrated narrow-band vegetation indices for prediction of crop chlorophyll content for application to precision agriculture. Remote Sens. Environ. 2002, 81, 416–426. [Google Scholar] [CrossRef]
  57. Gitelson, A.A.; Gritz, Y.; Merzlyak, M.N. Relationships between leaf chlorophyll content and spectral reflectance and algorithms for non-destructive chlorophyll assessment in higher plant leaves. J. Plant Physiol. 2003, 160, 271–282. [Google Scholar] [CrossRef] [PubMed]
  58. Gitelson, A.A.; Viña, A.; Arkebauer, T.J.; Rundquist, D.C.; Keydan, G.; Leavitt, B. Remote estimation of leaf area index and green leaf biomass in maize canopies-art. no. 1248. Geophys. Res. Lett. 2003, 30, 1248. [Google Scholar] [CrossRef]
  59. Clevers, J.G.; Kooistra, L.; Van den Brande, M.M. Using Sentinel-2 data for retrieving LAI and leaf and canopy chlorophyll content of a potato crop. Remote Sens. 2017, 9, 405. [Google Scholar] [CrossRef]
  60. Sims, D.A.; Gamon, J.A. Relationships between leaf pigment content and spectral reflectance across a wide range of species, leaf structures and developmental stages. Remote Sens. Environ. 2002, 81, 337–354. [Google Scholar] [CrossRef]
  61. Chen, S.; Chen, Y.; Chen, J.; Zhang, Z.; Fu, Q.; Bian, J.; Cui, T.; Ma, Y. Retrieval of cotton plant water content by UAV-based vegetation supply water index (VSWI). Int. J. Remote Sens. 2020, 41, 4389–4407. [Google Scholar] [CrossRef]
  62. Broge, N.H.; Leblanc, E. Comparing prediction power and stability of broadband and hyperspectral vegetation indices for estimation of green leaf area index and canopy chlorophyll density. Remote Sens. Environ. 2001, 76, 156–172. [Google Scholar] [CrossRef]
  63. Zhou, X.; Zhu, X.; Dong, Z.; Guo, W. Estimation of biomass in wheat using random forest regression algorithm and remote sensing data. Crop J. 2016, 4, 212–219. [Google Scholar]
  64. Li, Y.; Li, C.; Li, M.; Liu, Z. Influence of variable selection and forest type on forest aboveground biomass estimation using machine learning algorithms. Forests 2019, 10, 1073. [Google Scholar] [CrossRef]
  65. Zhang, Y.; Ma, J.; Liang, S.; Li, X.; Li, M. An evaluation of eight machine learning regression algorithms for forest aboveground biomass estimation from multiple satellite data products. Remote Sens. 2020, 12, 4015. [Google Scholar] [CrossRef]
  66. Palsson, F.; Sveinsson, J.R.; Ulfarsson, M.O. Multispectral and hyperspectral image fusion using a 3-D-convolutional neural network. IEEE Geosci. Remote Sens. Lett. 2017, 14, 639–643. [Google Scholar] [CrossRef]
  67. Lu, D.; Chen, Q.; Wang, G.; Liu, L.; Li, G.; Moran, E. A survey of remote sensing-based aboveground biomass estimation methods in forest ecosystems. Int. J. Digit. Earth 2016, 9, 63–105. [Google Scholar] [CrossRef]
  68. Shi, L.; Westerhuis, J.A.; Rosén, J.; Landberg, R.; Brunius, C. Variable selection and validation in multivariate modelling. Bioinformatics 2019, 35, 972–980. [Google Scholar] [CrossRef] [PubMed]
  69. Stuke, A.; Rinke, P.; Todorović, M. Efficient hyperparameter tuning for kernel ridge regression with Bayesian optimization. Mach. Learn. Sci. Technol. 2021, 2, 035022. [Google Scholar] [CrossRef]
  70. Virdi, J.S.; Peng, W.; Sata, A. Feature selection with LASSO and VSURF to model mechanical properties for investment casting. In Proceedings of the 2019 International Conference on Computational Intelligence in Data Science (ICCIDS), Gurugram, India, 6–7 September 2019; pp. 1–6. [Google Scholar]
  71. Huang, N.; Li, R.; Lin, L.; Yu, Z.; Cai, G. Low redundancy feature selection of short term solar irradiance prediction using conditional mutual information and Gauss process regression. Sustainability 2018, 10, 2889. [Google Scholar] [CrossRef]
  72. Liu, Z.; Michel, O.O.; Wu, G.; Mao, Y.; Hu, Y.; Fan, W. The potential of fully polarized ALOS-2 data for estimating forest above-ground biomass. Remote Sens. 2022, 14, 669. [Google Scholar] [CrossRef]
  73. Gong, M.; Kou, W.; Lu, N.; Chen, Y.; Sun, Y.; Lai, H.; Chen, B.; Wang, J.; Li, C. Individual tree AGB estimation of Malania oleifera based on UAV-RGB imagery and MASK R-CNN. Forests 2023, 14, 1493. [Google Scholar] [CrossRef]
  74. Hunt Jr, E.R.; Hively, W.D.; Fujikawa, S.J.; Linden, D.S.; Daughtry, C.S.; McCarty, G.W. Acquisition of NIR-green-blue digital photographs from unmanned aircraft for crop monitoring. Remote Sens. 2010, 2, 290–305. [Google Scholar] [CrossRef]
  75. Tilly, N.; Hoffmeister, D.; Cao, Q.; Huang, S.; Lenz-Wiedemann, V.; Miao, Y.; Bareth, G. Multitemporal crop surface models: Accurate plant height measurement and biomass estimation with terrestrial laser scanning in paddy rice. J. Appl. Remote Sens. 2014, 8, 083671. [Google Scholar] [CrossRef]
  76. Jung, C.; Zhou, K.; Feng, J. Fusionnet: Multispectral fusion of RGB and NIR images using two stage convolutional neural networks. IEEE Access 2020, 8, 23912–23919. [Google Scholar] [CrossRef]
  77. Liu, Z.; Wu, J.; Fu, L.; Majeed, Y.; Feng, Y.; Li, R.; Cui, Y. Improved kiwifruit detection using pre-trained VGG16 with RGB and NIR information fusion. IEEE Access 2019, 8, 2327–2336. [Google Scholar] [CrossRef]
  78. LeCun, Y.; Bengio, Y.; Hinton, G. Deep learning. Nature 2015, 521, 436–444. [Google Scholar] [CrossRef] [PubMed]
  79. Krizhevsky, A.; Sutskever, I.; Hinton, G.E. Imagenet classification with deep convolutional neural networks. In Advances in Neural Information Processing Systems; MIT Press: Cambridge, MA, USA, 2012; Volume 25. [Google Scholar]
  80. Alzubaidi, L.; Zhang, J.; Humaidi, A.J.; Al-Dujaili, A.; Duan, Y.; Al-Shamma, O.; Santamaría, J.; Fadhel, M.A.; Al-Amidie, M.; Farhan, L. Review of deep learning: Concepts, CNN architectures, challenges, applications, future directions. J. Big Data 2021, 8, 1–74. [Google Scholar] [CrossRef] [PubMed]
  81. Signoroni, A.; Savardi, M.; Baronio, A.; Benini, S. Deep learning meets hyperspectral image analysis: A multidisciplinary review. J. Imaging 2019, 5, 52. [Google Scholar] [CrossRef] [PubMed]
  82. Liu, T.; Wang, J.; Wang, J.; Zhao, Y.; Wang, H.; Zhang, W.; Yao, Z.; Liu, S.; Zhong, X.; Sun, C. Research on the estimation of wheat AGB at the entire growth stage based on improved convolutional features. J. Integr. Agric. 2024, in press. [Google Scholar] [CrossRef]
  83. Yu, D.; Zha, Y.; Sun, Z.; Li, J.; Jin, X.; Zhu, W.; Bian, J.; Ma, L.; Zeng, Y.; Su, Z. Deep convolutional neural networks for estimating maize above-ground biomass using multi-source UAV images: A comparison with traditional machine learning algorithms. Precis. Agric. 2023, 24, 92–113. [Google Scholar] [CrossRef]
Drones 09 00032 g001
Figure 1. Map of the study area.
Figure 1. Map of the study area.
Drones 09 00032 g001
Drones 09 00032 g002
Figure 2. The architecture of AlexNet used for this study.
Figure 2. The architecture of AlexNet used for this study.
Drones 09 00032 g002
Drones 09 00032 g003
Figure 3. The workflow for estimating the rubber plantation AGB based on multispectral images.
Figure 3. The workflow for estimating the rubber plantation AGB based on multispectral images.
Drones 09 00032 g003
Drones 09 00032 g004
Figure 4. The 1:1 relationship between the predicted and measured AGB values based on the combination of VIs and TFs using ML regression techniques. (a) RFR, (b) XGBR, and (c) CatBoost.
Figure 4. The 1:1 relationship between the predicted and measured AGB values based on the combination of VIs and TFs using ML regression techniques. (a) RFR, (b) XGBR, and (c) CatBoost.
Drones 09 00032 g004
Drones 09 00032 g005
Figure 5. Feature selected through the importance analysis based on VIs, TFs, and their combination. (a), (b), and (c) represent the importance scores of the VIs, TFs, and their combination, respectively.
Figure 5. Feature selected through the importance analysis based on VIs, TFs, and their combination. (a), (b), and (c) represent the importance scores of the VIs, TFs, and their combination, respectively.
Drones 09 00032 g005
Drones 09 00032 g006
Figure 6. Images of the RGB, TVI and contrast from a sample plot. (a) Original RGB image, (b) TVI, and (c) contrast derived from the red band.
Figure 6. Images of the RGB, TVI and contrast from a sample plot. (a) Original RGB image, (b) TVI, and (c) contrast derived from the red band.
Drones 09 00032 g006
Drones 09 00032 g007
Figure 7. The 1:1 relationship between the predicted and measured AGB values based on selected VIs and TFs using machine learning regression techniques. The first and second rows represent the RFR and VSURF methods, respectively. (a) RFR, (b) XGBR, and (c) CatBoost.
Figure 7. The 1:1 relationship between the predicted and measured AGB values based on selected VIs and TFs using machine learning regression techniques. The first and second rows represent the RFR and VSURF methods, respectively. (a) RFR, (b) XGBR, and (c) CatBoost.
Drones 09 00032 g007
Drones 09 00032 g008
Figure 8. PCA based on the VIs, TFs, and their integration. (a) VIs, (b) TFs, and (c) the combination of the VIs and TFs.
Figure 8. PCA based on the VIs, TFs, and their integration. (a) VIs, (b) TFs, and (c) the combination of the VIs and TFs.
Drones 09 00032 g008
Drones 09 00032 g009
Figure 9. The 1:1 relationship between the predicted and measured AGB values based on variables selected from the VIs and TFs using PCA. (a) RFR, (b) XGBR, and (c) CatBoost.
Figure 9. The 1:1 relationship between the predicted and measured AGB values based on variables selected from the VIs and TFs using PCA. (a) RFR, (b) XGBR, and (c) CatBoost.
Drones 09 00032 g009
Drones 09 00032 g010
Figure 10. The accuracy assessment of the AGB estimation derived from the DCNN with the RGB imagery.
Figure 10. The accuracy assessment of the AGB estimation derived from the DCNN with the RGB imagery.
Drones 09 00032 g010
Drones 09 00032 g011
Figure 11. The accuracy assessment of the AGB estimation derived from the DCNN with the multispectral imagery. (a) The combination of G, R, and NIR; (b) the combination of G, R, and RE; (c) the combination of R, NIR, and RE; and (d) the combination of G, NIR, and RE.
Figure 11. The accuracy assessment of the AGB estimation derived from the DCNN with the multispectral imagery. (a) The combination of G, R, and NIR; (b) the combination of G, R, and RE; (c) the combination of R, NIR, and RE; and (d) the combination of G, NIR, and RE.
Drones 09 00032 g011
Drones 09 00032 g012
Figure 12. Accuracy comparison between machine learning and deep learning models.
Figure 12. Accuracy comparison between machine learning and deep learning models.
Drones 09 00032 g012
Table 1. Basic statistical data on AGB (t/ha) of rubber plantations in sample plots.
Table 1. Basic statistical data on AGB (t/ha) of rubber plantations in sample plots.
CultivarsPlanting YearAltitude (m)AGB (t/ha)
MinMaxMeanSDCV (%)
GT11997–2000844–90277.13171.95101.8126.150.26
RRIM6001994, 2003631–696114.36137.74123.598.700.07
Yunyan 77-41995–2010640–87669.69184.02116.5227.080.23
Yunyan 74-721994, 2002636–701104.94138.71119.6814.190.12
Note: SD and CV represent standard deviation and coefficient of variation, respectively.
Table 2. Summary of vegetation indices derived from the UAV-based multispectral imagery for the AGB estimation of rubber plantations.
Table 2. Summary of vegetation indices derived from the UAV-based multispectral imagery for the AGB estimation of rubber plantations.
VINameFormulaReference
NDVINormalized difference vegetation index ( N I R − R ) / ( N I R + R ) [47]
RVIRatio vegetation index N I R / R [48]
NDRENormalized difference red-edge index ( N I R − R E ) / ( N I R + R E ) [49]
MSAVIModified soil-adjusted vegetation index 2 × N I R + 1 − ( 2 × N I R + 1 ) 2 − 8 × ( N I R − R ) 2 [50]
SAVISoil-adjusted vegetation index 1.5 ( N I R − R ) / ( N I R + R + 0.5 ) [51]
MSRModified simple ratio ( N I R / R − 1 ) / N I R / R + 1 [52]
NLINonlinear index ( N I R 2 − R ) / ( N I R 2 + R ) [53]
RDVIRenormalized difference vegetation index ( N I R − R ) / N I R + R [52]
DVIDifference vegetation index N I R − R [48]
OSAVIOptimized soil-adjusted vegetation index 1.16 ( N I R − R ) / ( N I R + R + 0.16 ) [54]
MCARIModified chlorophyll absorption ratio index R E − R − 0.2 ( R E − G ) × ( R E / R ) [55]
TCARITransformed chlorophyll absorption in reflectance index 3 R E − R − 0.2 × R E − G × R E / R [56]
GCVIGreen chlorophyll vegetation index N I R / G − 1 [57]
RNDVIRed-edge normalized difference vegetation index ( R E − R ) / ( R E + R ) [58,59]
GNDVIGreen normalized difference vegetation index ( N I R − G ) / ( N I R + G ) [57]
CIREChlorophyll index from red-edge N I R / R E − 1 [60]
RRIRed-edge ratio index N I R / R E [61]
NGRDINormalized green–red difference index ( G − R ) / ( G + R ) [62]
GIGreen index G / R [62]
TVITriangular vegetation index 60 N I R − G − 100 R − G [63]
Note: G, R, RE, and NIR represent the reflectance values in the green, red, red-edge, and near-infrared bands, respectively.
Table 3. AGB estimation model based on VIs, TFs, and their combination with three machine learning regression techniques.
Table 3. AGB estimation model based on VIs, TFs, and their combination with three machine learning regression techniques.
MethodFeaturesTest Sets
R2RMSE
(t/ha)		MAE
(t/ha)
RFRVIs0.6117.6413.16
XGBR0.5818.3214.23
CatBoost0.5817.9613.85
RFRTFs0.3423.0517.87
XGBR0.2524.6819.67
CatBoost0.2423.4917.14
RFRVIs and TFs0.7115.1912.10
XGBR0.6117.7113.70
CatBoost0.7114.6011.54
Table 4. The accuracy assessment of the AGB-estimation model derived from feature selection based on importance analysis using RFR combined with different machine learning algorithms.
Table 4. The accuracy assessment of the AGB-estimation model derived from feature selection based on importance analysis using RFR combined with different machine learning algorithms.
Regression MethodFeaturesTest Sets
Feature Selection Method
RFRVSURF
R2RMSE
(t/ha)		MAE
(t/ha)		R2RMSE
(t/ha)		MAE
(t/ha)
RFRVIs0.6715.4412.540.5617.8615.08
XGBR0.5218.6715.160.4020.7116.63
CatBoost0.6116.8213.050.5419.2214.02
RFRTFs 0.2922.6918.900.2123.9218.44
XGBR0.2922.6616.680.1924.0218.78
CatBoost0.2423.5017.150.1326.3920.34
RFRVIs and TFs 0.7313.9011.010.7014.6911.93
XGBR0.6416.2012.830.5218.1814.87
CatBoost0.7114.4011.460.6417.1212.37
Table 5. The accuracy assessment of the AGB estimation model derived from PCA combined with three machine learning algorithms.
Table 5. The accuracy assessment of the AGB estimation model derived from PCA combined with three machine learning algorithms.
MethodFeaturesTest Sets
R2RMSE
(t/ha)		MAE
(t/ha)
RFRVIs_PCA0.7413.7710.84
XGBR0.5817.4413.42
CatBoost0.6216.6612.88
RFRTFs_PCA0.3222.3217.59
XGBR0.3122.3217.60
CatBoost0.2922.7817.69
RFRVTs_PCA0.8111.639.27
XGBR0.7612.9610.11
CatBoost0.8012.1010.02
Note: the acronyms VIs_PCA, TFs_PCA, and VTs_PCA correspond to principal components extracted from the VIs, TIs, and their combination, respectively.
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Tan, H.; Kou, W.; Xu, W.; Wang, L.; Wang, H.; Lu, N. Improved Estimation of Aboveground Biomass in Rubber Plantations Using Deep Learning on UAV Multispectral Imagery. Drones 2025, 9, 32. https://doi.org/10.3390/drones9010032

AMA Style

Tan H, Kou W, Xu W, Wang L, Wang H, Lu N. Improved Estimation of Aboveground Biomass in Rubber Plantations Using Deep Learning on UAV Multispectral Imagery. Drones. 2025; 9(1):32. https://doi.org/10.3390/drones9010032

Chicago/Turabian Style

Tan, Hongjian, Weili Kou, Weiheng Xu, Leiguang Wang, Huan Wang, and Ning Lu. 2025. "Improved Estimation of Aboveground Biomass in Rubber Plantations Using Deep Learning on UAV Multispectral Imagery" Drones 9, no. 1: 32. https://doi.org/10.3390/drones9010032

APA Style

Tan, H., Kou, W., Xu, W., Wang, L., Wang, H., & Lu, N. (2025). Improved Estimation of Aboveground Biomass in Rubber Plantations Using Deep Learning on UAV Multispectral Imagery. Drones, 9(1), 32. https://doi.org/10.3390/drones9010032

Article Metrics

Back to TopTop