Next Article in Journal
Effect of Pollination Methods on Fruit Development in Greenhouse Watermelon: Physiological and Molecular Perspectives
Previous Article in Journal
Comparing the Agronomic and Economic Aspects of Sewage Sludge Composting and Vermicomposting
Previous Article in Special Issue
YOLOv8n-RMB: UAV Imagery Rubber Milk Bowl Detection Model for Autonomous Robots’ Natural Latex Harvest
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Agriculture
Volume 15
Issue 21
10.3390/agriculture15212293
agriculture-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

Estimate the Pre-Flowering Specific Leaf Area of Rice Based on Vegetation Indices and Texture Indices Derived from UAV Multispectral Imagery

by
Jingjing Huang
1,2,
Sunan Wang
1,2,
Yuexia Pei
1,2,
Quan Yin
3,
Zhi Ding
1,2,
Jianjun Wang
1,2,*,
Weiling Wang
1,2,
Guisheng Zhou
4 and
Zhongyang Huo
1,2,*
1
Jiangsu Key Laboratory of Crop Genetics and Physiology/Jiangsu Key Laboratory of Crop Cultivation and Physiology, Agricultural College of Yangzhou University, Yangzhou 225009, China
2
Jiangsu Co-Innovation Center for Modern Production Technology of Grain Crops, Yangzhou University, Yangzhou 225009, China
3
College of Land Science and Technology, China Agricultural University, Beijing 100083, China
4
Joint International Research Laboratory of Agriculture and Agricultural Product Safety, Yangzhou University, Yangzhou 225009, China
*
Authors to whom correspondence should be addressed.
Agriculture 2025, 15(21), 2293; https://doi.org/10.3390/agriculture15212293
Submission received: 16 September 2025 / Revised: 23 October 2025 / Accepted: 30 October 2025 / Published: 3 November 2025
(This article belongs to the Special Issue Plant Diagnosis and Monitoring for Agricultural Production)
Browse Figures
Versions Notes

Abstract

Rice ranks among the most significant staple crops worldwide. Precise and dynamic monitoring of specific leaf area (SLA) provides essential information for evaluating rice growth and yield. While previous remote sensing studies on SLA estimation have primarily focused on crops such as wheat and soybeans, studies on rice SLA remain limited. This study aims to evaluate the predictive potential of several machine learning algorithms for estimating rice SLA across different growth stages, planting densities, and nitrogen treatments at the pre-flowering stage. By utilizing UAV-based multispectral remote sensing data, a high-precision rice SLA monitoring model was developed. The feasibility of using vegetation indices (VIs), texture indices (TIs), and their combinations to predict rice SLA was explored. VIs and TIs were derived from UAV imagery, and the recursive feature elimination was conducted on these indices individually as well as their combined fusion (VIs + TIs). Four machine learning algorithms were employed to predict SLA values. The results indicate that random forest-based models utilizing VIs, TIs, and their fusion can all predict rice SLA effectively with high accuracy. Among these models, the RF model utilizing the combined variables (VIs + TIs) exhibited the highest performance, with R2 = 0.9049, RMSE = 0.0694 m2/g, RRMSE = 0.1042, and RPD = 3.2419. This study demonstrates that individual VIs can provide effective spectral information for SLA estimation, especially during the crucial pre-flowering growth phase of rice. The fusion of VIs and TIs enhances the model’s adaptability to complex field conditions by integrating both canopy biochemical and structural characteristics, thus improving model stability. This technology offers a swift and efficient approach for monitoring crop growth in the field, offering a theoretical foundation for subsequent crop yield estimation.

1. Introduction

With the global population continuing to grow, the need for food is on the rise. In this framework, rice (Oryza sativa L.), a key staple crop, exerts a crucial function in ensuring grain security and boosting socio-economic growth, especially across Asia [1].
In rice growth monitoring and yield regulation research, specific leaf area (SLA) serves as a core indicator in the study of plant leaf traits and holds irreplaceable value. SLA is defined as the ratio between leaf area and dry weight, commonly represented in m2/g. Its value is affected by factors such as leaf thickness, shape, and weight. SLA not only reflects a plant’s resource utilization efficiency and environmental adaptability, but also quantifies crop growth and leaf physiological structure, varying by crop variety, growth stage, and leaf function [2,3]. From a physiological perspective, SLA determines the physiological cost of leaf construction and is a critical parameter in growth models. It directly correlates with the new leaf area generated per unit of biomass, influencing light capture and growth rates, thereby reflecting spatial variations in photosynthetic ability and leaf nitrogen levels (e.g., increased leaf nitrogen content often accompanies elevated SLA) [4,5,6,7,8].
In agricultural practice, the value of SLA is particularly prominent. It is a core parameter in crop models, such as CLM3.5 [9], WOFOST [10], and Noah-MP-Crop [11], which derive the aboveground biomass (AGB) and leaf area index (LAI) [12]. SLA also exhibits significant sensitivity in yield simulations. Models indicate that changes in SLA can serve as a potential yield indicator (A higher SLA under conditions of fertilization is linked to a 15% rise in yield), and it is a key factor influencing yield simulations under adversity conditions [3,13,14,15,16]. Notably, SLA values tend to decrease as the growing season advances, with values in the later stages being significantly lower than those in the early stages [17,18].
Traditional SLA measurements involve destructive sampling, which is both time-intensive and laborious, making it challenging to conduct large-scale monitoring. In recent years, remote sensing technologies have been increasingly employed to assess key indicators such as LAI, AGB, SLA, chlorophyll content, and moisture levels [19,20,21,22]. Among these technologies, satellite remote sensing enables large-scale, rapid, and non-destructive monitoring of crop SLA. However, it is constrained by limitations such as long revisit intervals and low spatial resolution [23]. Unmanned Aerial Vehicles (UAVs) have garnered significant attention owing to their capacity to capture various images through low-altitude flights. UAVs offer several advantages, including lower costs, high resolution, broader coverage, faster data collection, and enhanced dynamic monitoring capabilities. They are well-positioned to overcome the limitations of satellite and ground-based platforms, demonstrating substantial potential for applications in the agricultural sector [24,25]. For instance, Gong et al. [26] applied several common vegetation indices (VIs) derived from UAV imagery and determined that the multiplication of VIs and canopy height notably improved the prediction of LAI across the entirety of the rice cultivation period. Additionally, canopy texture features reflect crop structural information and have been extensively studied by many scholars. Li et al. [27] explored the potential of using VIs, texture indices (TIs), and their combinations to predict winter wheat LAI. The findings suggest that integrating spectral VIs and textural features from drone-acquired multispectral imagery can enhance the precision of LAI assessment during the vegetative stage of winter wheat.
Although SLA is closely related to canopy structure, the potential of texture features to provide supplementary information for SLA estimation in rice and the extent of their beneficial effects remain to be systematically validated. This study is the first to explore the contribution of texture features in the prediction of rice SLA. In the realm of SLA estimation research, the viability of estimating SLA using leaf and canopy reflectance through optimization and validation of vegetation indices has been established [28]. Mohsen et al. [29] employed Sentinel-2 data to compute VIs and utilized ML algorithms to estimate the LAI, SLA, leaf dry weight (LDW), and leaf specific weight (SLW) of wheat crops. Their findings indicated that the deep neural network (DNN) algorithm was remarkably successful in predicting foliar parameters across southern Iran, achieving SLA accuracy exceeding 72%. Additionally, de Paula et al. [30] have employed hyperspectral leaf technology to identify drought-tolerant soybean varieties. This research successfully predicted SLA at the leaf level using a partial least squares regression (PLSR) model, underscoring the significant promise of hyperspectral spectroscopy as a dependable, non-invasive, and precise approach for appraising physiological characteristics and identifying superior genotypes.
Although existing research has made some progress in assessing SLA in crops such as wheat and soybeans, studies on rice SLA remain quite limited. Precise monitoring of rice SLA faces two major technical challenges: first, the utilization of UAV multispectral technology in conjunction with machine learning for rice SLA estimation has not been fully explored, and the technical feasibility and accuracy boundaries remain unclear; second, existing crop SLA remote sensing models are often limited to single density or nitrogen levels, making them difficult to adapt to the complex field management conditions encountered in actual production (such as different planting densities and gradient fertilization strategies), resulting in insufficient generalizability and significant barriers to scaling up from experimental fields to large-scale production. More critically, a unified estimation model capable of simultaneously covering key growth stages such as tillering, jointing, and heading, while accommodating multiple density and nitrogen management scenarios remains an unexplored research area. To address these issues, this study proposes the hypothesis that in the Yangtze River Delta rice-growing region, by deriving VIs and TIs from drone multispectral imagery and integrating them with machine learning algorithms, accurate monitoring of rice SLA can be attained.
To validate this hypothesis, this study used rice as the research subject and designed a targeted multi-gradient experiment covering one rice variety, three planting densities, and 24 gradient fertilization levels, comprehensively simulating the complex management scenarios in actual production. By obtaining multispectral canopy imagery at various growth stages, extracting spectral indices along with multi-scale texture data, the study focused on addressing two core issues: (1) systematically comparing the performance differences in various machine learning models when combined with VIs, TIs, and performance differences in estimating SLA across multiple growth stages of rice; (2) developing a precise SLA estimation model for rice that can be applied to the tillering stage, jointing stage, and heading stage, and is compatible with different densities and nitrogen management strategies. This research seeks to provide technical support for non-destructive monitoring of rice SLA, thereby laying the foundation for precision cultivation and yield improvement in rice.

2. Materials and Methods

2.1. Experimental Site and Design

The field trial was conducted from June to November 2024 at Jingxian Farm in Jiangyan District, Taizhou City, Jiangsu Province (Figure 1). Fertilization treatments included a basal fertilizer application immediately after transplanting, followed by two top-dressings during the tillering and heading stages. The experiment involved a single variety, Nanjeng 9108, with three planting densities (16 cm × 30 cm, 16 cm × 23 cm, 16 cm × 18.75 cm). There are 24 experimental plots with different fertilizer treatments for each of the three densities, repeated twice for a total of 144 plots. This experiment used 72 of these plots, each with an area of 15 square meters. The plots were separated by field ridges and waterproof boards, with independent drainage pipes.

2.2. Data Collection and Processing

2.2.1. UAV Image Acquisition and Processing

This study employed the DJI Phantom 4 Multispectral Edition (DJI Technology Co., Shenzhen, China), which offers high-precision positioning data, eliminating the need for ground control points. Data collection took place at three growth stages of rice plants: tillering (18 July 2024), jointing (1 August 2024), and heading (26 August 2024). Image capture was conducted using the DJI P4 multispectral UAV (SZ DJI Technology Co; Shenzhen, China), fitted with five multispectral detectors for the B (blue), G (green), R (red), Re (red edge), and NIR (near-infrared) wavelengths. The spectral resolution for each band is as follows: Blue band (450 ± 16 nm), Green band (560 ± 16 nm), Red band (650 ± 16 nm), Red-edge band (730 ± 16 nm), and Near-infrared band (840 ± 26 nm). The sensor’s radiometric resolution has a Digital Number (DN) value range of 0 to 65535. To minimize the impact of variations in solar elevation angle on experimental results, all drone flights were scheduled between 10:00 a.m. and 2:00 p.m. local time. Flights were conducted exclusively under clear, cloudless, and windless conditions to mitigate potential influences such as hotspots and shadow interference on image quality. Before each flight, two diffuse reflectance reference panels, with reflectance values of 50% and 75%, are placed on the ground. A linear relationship between the reflectance of these reference panels at different phases and the remote sensing imagery is utilized. Band-by-band operations are then performed to complete the radiometric calibration of the remote sensing images. The drone maintained a consistent elevation of 30 m and a speed of 3 m per second. Both along and across tracks adopt an 80% overlap setting. Following each flight, the collected images were processed using DJI Terra 2.3 software (https://enterprise.dji.com/cn/dji-terra, accessed on 1 September 2025) for two-dimensional multispectral integration and radiometric adjustment, producing corrected single-band reflectance images.

2.2.2. Specific Leaf Area Measurement

Field sampling was conducted during the tillering, jointing, and heading stages of rice growth. Two plants from each experimental plot were collected, placed in sampling bags, and labeled with the corresponding sampling point information. The samples were swiftly delivered to the laboratory. Rice plants under good growth conditions were selected, with 8 leaves randomly chosen from each plant. Each leaf was measured three times for total width using a leaf area measuring ruler. A sample segment was then cut from the center of each leaf. Both the sample segment and the remaining leaves were dried at 105 °C for 30 min, and subsequently dried at 80 °C for at least 48 h until constant weight was reached. The weight was recorded, and the leaf area at each sampling point was computed using the dry weight technique. The SLA value was established as the leaf area-to-dry weight ratio.
W a l l = W d r y 1 + W d r y 2
S a l l = S c u t / W d r y 1 × W d r y 1 + W d r y 2
S L A = S a l l   m 2 /   W a l l   g
where Wall represents the total dry weight, Wdry1 and Wdry2 represent the dry weight of the cut sample segment and the remaining sample segment, Sall represents the total leaf area, and Scut represents the area of the cut sample segment.

2.3. Feature Extraction from Remote Sensing Images

2.3.1. Vegetation Indices

VIs are simple yet effective indicators that reflect the characteristics of vegetation growth status and are widely utilized in monitoring biochemical and physiological parameters [31]. To calculate VIs and examine the correlations between different VIs, this study employed ENVI5.6 (ITExelis, Boulder, CO, USA) to retrieve the spectral reflectance data for individual experimental plots. To estimate SLA, 25 frequently used VIs were calculated, leading to a total of 30 spectral indices. Among these, a new vegetation index, the Red-Edge, Near-Infrared, Green-based Terrestrial Index (RE-NIR-GTI), is introduced in this study, as detailed in Table 1.

2.3.2. Texture Indices

The Gray-Level Co-occurrence Matrix (GLCM) [56] ranks among the most commonly utilized techniques for textural feature extraction, initially introduced by Haralick in 1973. Its multiscale characteristics and computational simplicity have garnered considerable attention [57]. In this research, ENVI 5.6 was applied to derive eight texture parameters—mean (Mean), variance (Var), homogeneity (homo), contrast (Con), dissimilarity (dis), entropy (Ent), second moment (SecMom), and correlation (Cor)—from multispectral UAV imagery based on GLCM. The feature extraction window size was configured as 3 × 3. Since the red-edge band and NIR band of rice SLA and UAV multispectral imagery are more sensitive [58], this study only selected the red-edge and NIR bands, selecting 16 TIs.

2.3.3. Background Removal

This study utilizes supervised classification techniques to eliminate background interference caused by varying growth stages and soil backgrounds, especially in the initial stages with sparse vegetation. The background removal was conducted through the use of ENVI software. Through analysis of UAV multispectral images, the images were categorized into three groups: vegetation, water bodies, and soil. By distinguishing different sample areas and merging similar objects, vector boundaries for non-vegetation and vegetation categories were generated. Masking operations were then performed to complete the background removal process.

2.4. Feature Selection

This research employs Recursive Feature Elimination (RFE) for selecting the feature subset most representative of the target variable. RFE repeatedly trains the model and removes the least significant features until reaching a predetermined feature count or another stopping criterion is met [59]. Cross-validation RFE is combined with the RF estimator to optimize model performance, reduce the dimensionality of features, and mitigate overfitting.
The number of parameters has a significant effect on the performance of machine learning models [60]. Identifying the most relevant attributes for a particular learning algorithm is often essential. To improve the effectiveness of the predictive regression model, the RFE learning curve is applied to determine the optimal number of remote sensing-derived variables. Subsequently, the hierarchical structure derived from the RFE feature importance ranking is used to screen the optimal remote sensing variables.

2.5. Model Construction and Accuracy Verification

The dataset was segmented into training and test subsets using stratified sampling, with a splitting ratio of 8:2. Optimize the model using K-fold cross-validation (K = 5, CV). This method divides the training data into K subsets, each acting as the validation set in turn, while the rest are used for training. By averaging the results across all folds, the model’s efficacy is improved, reducing training data error and preventing test data from influencing the training process. Model performance was assessed using four primary indicators: coefficient of determination (R2), RMSE (root mean square error), RRMSE (relative RMSE), and residual prediction deviation (RPD). Typically, higher R2 values (as shown in Formula (4)), lower RMSE values (as shown in Formula (5)), and lower RRMSE values (as shown in Formula (6)) indicate improved model performance [61]. Additionally, the model’s predictive capability is assessed using the RPD, which is calculated as the ratio between the standard deviation of measured values and cross-validation RMSE, Among these, SD denotes the standard deviation of SLA values in rice plants, as outlined in Formula (7) [62]. Following Rossel et al. [63], RPD values are interpreted as follows: 1.8 < RPD < 2.0 reflects a good estimate, 2.0 < RPD < 2.5 denotes a very good estimate, and RPD > 2.5 signifies an excellent estimate.
R 2 = 1 − ∑ i = 1 n y ^ i − y i 2 ∑ i = 1 n y i − y ¯ 2
RMSE = ∑ i = 1 n y ^ − y i 2 n
R R M S E = R M S E y ¯
RPD = SD RMSE

2.6. Machine Learning Regression Algorithms

Spectral reflectance data from simulations and field measurements have been widely utilized in machine learning. Machine learning methods are capable of estimating vegetation parameters and, by training on spectral reflectance data, exhibit resilience and enhanced prediction precision. This research utilized four commonly applied ML methods, namely RF, XGBoost, Gradient Boosting Decision Tree (GBDT), and PLSR, as base learners to assess the accuracy of yield prediction achieved through ensemble learning and data from multiple sensors integration. Among them, RF, XGBoost, and GBDT are used to capture the nonlinear relationship between remote sensing features and rice SLA, while PLSR adapts data characteristics by processing multicollinearity and dimensionality reduction capabilities. The four types of algorithms have the potential to capture nonlinear relationships and improve the accuracy of vegetation parameter estimation, and combine the linear and nonlinear information of coverage data to jointly ensure the accuracy and robustness of predictions. Figure 2 shows the workflow.

2.6.1. Random Forest (RF)

RF [64] represents an ensemble learning method frequently employed in classification and regression problems. It aggregates numerous decision trees to produce forecasts. In a random forest, each tree is built with stochastically selected data and attributes, which helps prevent overfitting and improves the model’s resilience and accuracy. RF is widely utilized in the processing of structured data, image recognition, and assessing feature importance. Owing to its ability to capture intricate relationships among features and its effectiveness in managing overfitting, this method is well-suited for estimating SLA values in rice cultivation and nitrogen fertilizer management strategies.

2.6.2. XGBoost

XGBoost incorporates several weak learners to form a robust predictor, leveraging a collection of decision trees to enhance the model’s ability to generalize. A key characteristic of XGBoost is the incorporation of a regularization term in the loss function to reduce the risk of overfitting. The algorithm uses a gradient boosting approach to iteratively refine the model, ensuring every successive step leads to enhanced performance [65]. In machine learning models, fine-tuning parameters is essential for optimizing performance. This research utilizes a combined approach of cross-validation and grid search to identify the optimal set of parameters and hyperparameters.

2.6.3. Gradient Boosting Decision Tree (GBDT)

GBDT [66] is a widely used ensemble learning technique in machine learning. GBDT functions as a supervised learning technique that integrates predictive results from a series of decision trees for developing a more robust predictive framework. It is capable of handling intricate data relationships, offering insights into feature importance, and performing well across diverse datasets, making it an essential tool in the machine learning domain.

2.6.4. Partial Least Squares Regression (PLSR)

PLSR [67] is a regression modeling technique utilized for forecasting a group of reliant variables relying on a group of autonomous variables. It combines the characteristics of multiple linear regression and principal component analysis. PLSR proficiently tackles multicollinearity among autonomous variables and is especially valuable for small sample datasets.

3. Results

3.1. Statistical Analysis of SLA

The differences in SLA values during the three growth stages before flowering in rice (tillering stage, jointing stage, and heading stage) were significant (Figure 3, Table 2). During the tillering stage, SLA values ranged from 0.57 to 1.33 m2/g, with an average of 0.85 m2/g. At the jointing stage, SLA values ranged from 0.42 to 0.91 m2/g, with an average of 0.64 m2/g. In the heading stage, SLA values ranged from 0.35 to 0.76 m2/g, with an average of 0.54 m2/g. From the tillering to the heading stage, SLA values gradually decreased as the growth stage progressed. This finding aligns with previous studies [17,18].

3.2. Correlation Analysis

This research investigated the correlation between various variables and rice SLA during the tillering, jointing, and heading stages (Figure 4). The results showed that 21 VIs were extremely significant (p < 0.01), and 2 VIs were significant (p < 0.05), but the correlation coefficients during the tillering, jointing, and heading stages were generally low (|r| < 0.35), r is the Pearson correlation coefficient. Furthermore, most VIs displayed an inverse relationship with rice SLA. In particular, the near-infrared (NIR) band from single-band data demonstrated the significant association with SLA, with a coefficient of −0.30. Among the VIs, CIRE showed the highest correlation, with a correlation coefficient of −0.33, which aligns with the sensitivity of near-infrared reflectance to leaf structure and chlorophyll content [68,69]. Next in order of strength are MTCI, GCI, and IRG. In contrast, VIs such as NPCI, NGRDI, R, G, and GI showed weak and non-significant correlations with SLA.
From the TIs extracted using GLMC, 12 indices reached the level of extreme significance (p < 0.01) (Figure 4), and three indices (Re-variance, Re-contrast, NIR-dissimilarity) showed significant correlations (p < 0.05). Among these, Re-correlation (r = 0.32) and NIR-correlation (r = 0.39) exhibited strong positive correlations, whereas Re-entropy (r = −0.31) and NIR-entropy (r = −0.29) showed strong negative correlations.
Overall, the correlation between VIs, TIs, and SLA is generally weak, mainly because the factors regulating SLA are complex and intertwined, including nitrogen availability, leaf age, and canopy structure [70]. This complexity thus highlights the need to incorporate diverse remote sensing variables to capture the multidimensional variation in SLA. This is demonstrated by the outstanding performance of the VIs + TIs fusion model in subsequent analyses, which validated the effectiveness of this strategy.

3.3. Selection of Remote Sensing Variables

In selecting the VIs, the learning curve was determined based on the RFE characteristics (Figure 5), which led to the identification of 29 as the ideal number of VIs. We employed feature importance ranking (Figure 6) to further refine the VIs, resulting in the selection of the following optimal VIs: G, B, RE-NIR-GTI, GB, R, ExB, Red Edge, NGI, NPCI, NDRE, SIPI, CIRE, NGRDI, EVI, RENDVI, MTCI, MCARI2, CI, GI, CCCI, NIR, TVI, OSAVI, NDVI, DVI, GNDVI, EVI2, IRG, and IPVI.
For the selection of texture indices (TIs), the learning curve indicated that 10 is the most favorable number of remote sensing variables. According to the feature importance ranking (Figure 6), the chosen optimal TIs include NIR-correlation, Re-Mean, NIR-Variance, NIR-Homogeneity, Re-dissimilarity, Re-homogeneity, NIR-contrast, Re-contrast, NIR-mean, and Re-correlation. After merging the 30 VIs and 16 TIs extracted and calculated during the initial processing (VIs and TIs fusion), and applying the RFE feature selection (Figure 6), 10 optimal remote sensing variables were selected for subsequent modeling. These variables include G, B, RE-NIR-GTI, Re-mean, NIR-correlation, R, NIR-homogeneity, NIR-variance, GB, and NIR-contrast. All remote sensing variables are derived from both VIs and TIs, rather than being confined to a single set of variables (VIs or TIs). This reflects the significant heterogeneity and complementarity of VIs and TIs in characterizing rice SLA.

3.4. Model Construction and Validation

Based on VIs, TIs, and their combination, optimal variables are selected for subsequent modeling, and model performance is assessed using the test dataset. This research utilized four different machine learning algorithms, RF, XGBoost, GBDT, and PLSR, to construct predictive regression models.
In the VIs-based modeling, the RF model demonstrated superior performance when evaluated on the combined training and test sets (Table 3). Specifically, the RF model demonstrated a high R2 on the test set, accompanied by lower values for RMSE and RRMSE. In the test dataset, RF attained an R2 of 0.9035, an RMSE of 0.0699 m2/g, an RRMSE of 0.1050, and RPD of 3.2184. Meanwhile, compared with PLSR with poor inversion performance, the RF model showed a 7.52% enhancement in R2 and a 2.34% reduction in RMSE, demonstrating significant advantages.
In TI-based modeling, the RF algorithm was employed to develop a rice SLA value estimation model with the highest estimation accuracy (Table 4). The R2 of this model for the training dataset was 0.8556, RMSE was 0.0711 m2/g, RRMSE was 0.0994, and RPD was 2.882. In the test set, the R2 was 0.9008, the RMSE was 0.0709 m2/g, the RRMSE was 0.1064, and the RPD was 3.1756.
In the modeling that integrates VIs and TIs, RF once again demonstrates the highest accuracy in estimating rice SLA (Table 5). Specifically, the training set R2 value of this model was 0.8525, RMSE was 0.0713 m2/g, RRMSE was 0.1067, and RPD was 2.6412. For the test dataset, the R2 value was 0.9049, the RMSE was 0.0694 m2/g, the RRMSE was 0.1042, and the RPD was 3.2419. The PLSR model is the second best inversion model, with an R2 value of 0.8606, an RMSE of 0.0841 m2/g, an RRMSE of 0.1262, and an RPD of 2.6780 in the test dataset.
To further confirm the accuracy of the mode, Figure 7 displays a scatter plot of the measured rice SLA values against the predicted values from the optimal models based on VIs, TIs, and the integration of VIs + TIs. The plot illustrates a clear clustering of data points along the 1:1 diagonal, indicating a high level of agreement between the measured and predicted values. This close match not only intuitively verifies the model’s reliability but also underscores its excellent predictive performance in rice SLA estimation.

3.5. Validate the Optimal Inversion Model

RF models incorporating VIs, TIs, and their combination exhibited the highest test set R2 and RPD, alongside the lowest RMSE and RRMSE, among all regression models, making them the optimal estimation models for rice SLA values. We evaluated the performance of the selected optimal RF (VIs, TIs, and their combinations) on the test dataset for each of all periods (Figure 8) to assess the accuracy of rice SLA value estimation.
As illustrated in the diagram presented in Figure 8, all regression methods can effectively predict the SLA values of rice in the RF models of VIs and TIs, as well as the fusion of VIs and TIs. Different colors correspond to different periods. From the figure, we observed that the data from different growth stages do not deviate significantly from the diagonal line, suggesting that the model has universal predictive capabilities for SLA values at each stage before the flowering stage of rice.
We also assessed the RMSE and RRMSE across different growth stages. The RMSE and RRMSE values were found to be lowest during the jointing stage, followed by the heading stage, with slightly higher values observed during the tillering stage. These findings are consistent with the visual patterns depicted in the scatter plot (Figure 8).

4. Discussion

4.1. The Impact of Various Remote Sensing Variables to SLA

The research results indicate that VIs, TIs and their combined features (VIs + TIs) all demonstrated high accuracy in estimating SLA, with test set R2 values exceeding 0.85 and RPD values greater than 2.5 (Table 5). These metrics suggest excellent predictive performance, as RPD > 2.0 generally indicates a model with strong practical application value [63]. Notably, the fusion model combining VIs and TIs performed the best: the test set R2 reached 0.9049, RPD reached 3.2419, and the RRMSE was as low as 0.1042, outperforming models using either VIs or TIs alone. This result indicates that integrating VIs with TIs improves the model’s capacity to capture intricate canopy features linked to SLA [26,27,28]. However, unlike previous studies where the fusion of features (VIs + TIs) significantly outperformed single features in estimating SPAD values or LAI, the performance of the integrated model in this investigation was only marginally higher than that of single-feature models. This suggests that in rice SLA estimation, using either VIs or TIs combined with machine learning alone can already achieve high accuracy.
Compared to prior SLA estimation studies, this model demonstrates a significant accuracy advantage. For instance, Ramon et al. [30] used PLSR models to predict soybean SLA for screening superior genes. The cross-validation R2 (CV) values at 20 and 28 days after drought were only 0.741 and 0.809, respectively. In the present study, the RF model utilizing the integration of VIs and TIs not only achieved a cross-validation R2 of 0.8525 in the training dataset, but also an R2 of 0.9049 in the test set, showing a significant improvement in accuracy. Additionally, another study estimated SLA using hyperspectral data, achieving an R2 of 0.88 and an RMSE of 13.30 cm2/g [71]. However, this research further demonstrates that high-precision estimation of rice SLA using multispectral data is equally feasible, providing a new approach for low-cost and large-scale monitoring. The effectiveness of vegetation indices arises from their sensitivity to leaf biochemical characteristics, such as leaf structure and chlorophyll content [72], which directly influence SLA. Analyses of the chosen vegetation indices indicate that many of them are associated with the near-infrared range, which has been proven to effectively capture dynamic changes in leaf characteristics [67]. For example, the CIRE calculated from the red edge and NIR wavelengths is indicative of the leaf area, while the MTCI index minimizes observational noise by simultaneously analyzing the NIR and red spectral bands. These factors likely explain why vegetation indices exhibit superior predictive performance compared to texture indices alone.
In this research, we introduce a novel vegetation index, the RE-NIR-GTI (Red-Edge, Near-Infrared, Green-based Terrestrial Index), which integrates reflectance information from the red-edge, NIR, and green-based bands, specifically designed to characterize key biophysical properties of vegetation in terrestrial ecosystems, such as chlorophyll content, leaf structure, growth status, and leaf area index. Its primary aim is to more accurately capture the physiological characteristics of rice at specific growth stages, particularly its correlation with SLA. The red-edge band was selected due to its high sensitivity to variations in chlorophyll content and leaf internal structure. Changes in its reflectance can be directly linked to SLA dynamics. The near-infrared band primarily reflects the cellular structure and moisture status of rice leaves, with healthy rice plants exhibiting higher reflectance in this band, particularly those with larger leaf areas. Thus, the near-infrared band is crucial for monitoring SLA changes in rice. The green light band provides valuable information on rice photosynthetic activity, as photosynthetic capacity is directly linked to leaf area size, further enhancing SLA estimation. Table 3 and Figure 6 validate the utility of this index: RE-NIR-GTI shows a highly significant correlation with SLA and is consistently selected as the optimal variable in the RFE screening, highlighting its importance in SLA prediction.
The spectral reflectance characteristics provide a fundamental basis for crop monitoring. Texture features are key spatial features that characterize crop canopy surface information, making them increasingly adopted in crop phenotyping research [73]. Hlatshwayo et al. [71] discovered that textural features in the red spectral band and NIR spectral bands showed significantly higher correlations with leaf characteristics than certain VIs. In the present study, we built a machine model for rice SLA using 16 texture features. The findings revealed that the texture model’s accuracy was lower than that of the VIs-based model. VIs are typically multi-band combinations, while TIs represent pixel statistics within a single band, which constrains their ability to predict vegetation parameters.

4.2. Performance of the Optimal Estimation Model

Among all algorithms, the RF model that integrates VIs and TIs following the RFE feature selection performed the best in predicting rice SLA values. Specifically, the training dataset obtained an R2 of 0.8525, with an RMSE of 0.0713 m2/g, while the test dataset registered an R2 of 0.9049, and an RMSE of 0.0694 m2/g. Additionally, the RPD of the test set was 3.2419, with a higher RPD indicating better model performance. This discovery is validated by Sun et al. [74], who employed VIs and TIs extracted from UAV data in combination with machine learning algorithms to predict LAI. Their study indicates such integrated models yield optimal performance. It also supports the results of Lourenço et al. [75], who demonstrated that after removing non-vegetation spectral interference through masking, the RF regression model, combined with VIs and TIs, exhibits significant effectiveness in biomass estimation.
Among all regression models, RF attained the highest R2 value and the lowest RMSE and RRMSE values, with a minimal difference between the predicted and actual rice SLA. Therefore, it is the most appropriate model for estimating rice SLA in this study. RF demonstrates robust predictive abilities for rice SLA, primarily due to its exceptional flexibility with small sample sizes. Even with limited data, the model can still provide accurate estimates. Moreover, its ability to identify and manage outliers and noise further emphasizes its robust performance. This stability is attributed to the fundamental principles of the algorithm. For example, the RF algorithm model effectively diminishes the vulnerability of a single decision tree to noise by utilizing ensemble learning methods [76], which helps the model preserve both accuracy and generalizability, even when faced with noisy data.
Furthermore, changes in SLA are jointly regulated by multiple factors, including nitrogen availability (which influences leaf thickness) and canopy structure (which affects light interception) [77]. Unlike linear models or single-tree algorithms (such as GBDT), RF excels at capturing the interactions between VIs (such as near-infrared indices) and TIs (such as entropy and contrast), and these interactions precisely reflect the regulatory mechanisms of SLA. The model’s stability on both the training and testing datasets (e.g., the RF model based on fused VIs and TIs: RMSE of 0.0713 m2/g for the training set and 0.0694 m2/g for the testing dataset) further confirms its applicability under field conditions, a core requirement for agricultural applications.
In summary, among all regression models, the RF model exhibits the greatest R2 and the smallest RMSE, thus most accurately reflecting the actual status of rice SLA. Therefore, it represents the optimal choice for estimating SLA values before rice flowering.

4.3. Limitations and Prospects for Future Research

Although our study has produced promising outcomes in the estimation of rice SLA, there are still limitations remaining to be addressed. Specifically, the study focused on a single rice variety and lacked validation for different varieties. this limitation could potentially affect the precision of our model when applied to different rice varieties and limit its suitability for multi-variety planting scenarios.
Future studies will include representative rice varieties from different genotypic groups, such as indica, japonica, and hybrid rice, to further validate the findings. Analyzing how varietal differences affect the accuracy of the SLA estimation model, while supplementing with model adaptability analysis under extreme cultivation conditions (ultra-high-density planting, low-nitrogen stress), exploring cost-optimization pathways for large-scale model application, and ultimately enhancing the model’s adaptability, universality, and practicality across diverse varieties and complex cultivation scenarios.

5. Conclusions

Accurate assessment of rice SLA values before the flowering stage is crucial for guiding nitrogen fertilizer application, thus promoting high rice yields and efficient nitrogen utilization. This study, based on rice field trials conducted in Jiangsu Province, China, involving one rice variety, three planting densities, and 24 fertilizer levels, with the objective of enhancing the precision of remote sensing monitoring of SLA during this critical growth period. The key findings are as follows:
(1)
VIs, TIs, and their fusion features (VIs + TIs) can all effectively estimate SLA, with fusion features performing the best. However, the accuracy of the fusion model in this research is only slightly higher than that of the single-feature models, which means that in rice SLA estimation, using only VIs or TIs combined with machine learning can already achieve high accuracy. This result highlights the synergistic mechanisms of multi-source remote sensing features under complex canopy conditions and provides new methods to improve the stability of SLA estimation. The findings can be used to construct SLA estimation models applicable to various critical growth phases prior to flowering (tillering stage, jointing stage, heading stage), various densities, and different nitrogen management strategies.
(2)
The RF model, selected through RFE, performed optimally, demonstrating high accuracy across VIs, TIs, and the integrated feature set (test set R2 > 0.90, RPD > 3.0). It also showed strong generalization ability across different densities and nitrogen management strategies. The model’s strength lies in its capacity to accurately capture nonlinear relationships and feature interactions, thereby enabling it to extract relevant information from diverse sources, effectively mitigate noise interference, and thus enhance the overall accuracy of SLA estimation.
This study provides an important reference for agricultural management practices: through dynamic monitoring of SLA before flowering, real-time guidance can be offered for nitrogen fertilizer application (e.g., timely fertilization when SLA decreases abnormally), which facilitates the achievement of both high yields and efficient nitrogen utilization. In future research, experiments should be conducted across more rice varieties and growth environments to further verify the stability of this method in crop breeding, thereby providing more comprehensive and accurate technical support for rice growth monitoring and evaluation.

Author Contributions

Conceptualization, J.W. and J.H.; methodology, J.H. and J.W.; formal analysis, J.H. and J.W.; investigation, J.H., S.W., Z.D., Y.P. and W.W.; data curation, J.H.; writing—original draft preparation, J.H., J.W., Q.Y. and G.Z.; writing—review and editing, J.H., J.W. and Q.Y.; visualization, J.H.; supervision, J.W. and Z.H.; funding acquisition, Z.H. All authors have read and agreed to the published version of the manuscript.

Funding

This work was funded by the Scientific and Technological Innovation Fund of Carbon Emissions Peak and Neutrality of Jiangsu Provincial Department of Science and Technology, China (BE2022424-2), the Jiangsu Agricultural Science and Technology Innovation Fund (CX(24)2008), the Taizhou Major Scientific and Technological Achievement Transformation Project (SCG202403), and the Priority Academic Program Development of Jiangsu Higher Education Institutions (PAPD), China.

Data Availability Statement

The data are available from the authors upon reasonable request as the data need further use.

Conflicts of Interest

The authors declare no conflicts of interest.

References

  1. Li, C.; Teng, X.; Tan, Y.; Zhang, Y.; Zhang, H.; Xiao, D.; Luo, S. Spatio-temporal mapping of leaf area index in rice: Spectral indices and multi-scale texture comparison derived from different sensors. Front. Plant Sci. 2024, 15, 1445490. [Google Scholar]
  2. Garnier, E.; Shipley, B.; Roumet, C.; Laurent, G. A standardized protocol for the determination of specific leaf area and leaf dry matter content. Funct. Ecol. 2001, 15, 688–695. [Google Scholar] [CrossRef]
  3. Li, Z.; Zhan, C.; Hu, S.; Ning, L.; Wu, L.; Guo, H. Implementation of a dynamic specific leaf area (SLA) into a land surface model (LSM) incorporated crop-growth model. Comput. Electron. Agric. 2023, 213, 108238. [Google Scholar] [CrossRef]
  4. Abdi, A.; Kapourchal, S.A.; Vazifedoust, M.; Rezaei, M. A novel satellite-based methodology for retrieving specific leaf area of rice (Hashemi cultivar) at field scale. Environ. Eng. Manag. J. 2022, 21, 2093. [Google Scholar]
  5. Pierce, L.L.; Running, S.W.; Walker, J. Regional–scale relationships of leaf area index to specific leaf area and leaf nitrogen content. Ecol. Appl. 1994, 4, 313–321. [Google Scholar] [CrossRef]
  6. Witkowski, E.T.F.; Lamont, B.B. Leaf specific mass confounds leaf density and thickness. Oecologia 1991, 88, 486–493. [Google Scholar] [CrossRef]
  7. VAN Arendonk, J.J.C.M.; Niemann, G.J.; Boon, J.J.; Lambers, H. Effects of nitrogen supply on the anatomy and chemical composition of leaves of four grass species belonging to the genus Poa, as determined by image--processing analysis and pyrolysis–mass spectrometry. Plant Cell Environ. 1997, 20, 881–897. [Google Scholar] [CrossRef]
  8. Knops, J.M.; Reinhart, K. Specific leaf area along a nitrogen fertilization gradient. Am. Midl. Nat. 2000, 144, 265–272. [Google Scholar] [CrossRef]
  9. Göhler, M.; Mai, J.; Cuntz, M. Use of eigendecomposition in a parameter sensitivity analysis of the Community Land Model. J. Geophys. Res. Biogeosci. 2013, 118, 904–921. [Google Scholar] [CrossRef]
  10. Ma, H.; Huang, J.; Zhu, D.; Liu, J.; Su, W.; Zhang, C.; Fan, J. Estimating regional winter wheat yield by assimilation of time series of HJ-1 CCD NDVI into WOFOST–ACRM model with Ensemble Kalman Filter. Math. Comput. Model. 2013, 58, 759–770. [Google Scholar] [CrossRef]
  11. Zhang, Z.; Barlage, M.; Chen, F.; Li, Y.; Helgason, W.; Xu, X.; Liu, X.; Li, Z. Joint modeling of crop and irrigation in the central United States using the Noah--MP land surface model. J. Adv. Model. Earth Syst. 2020, 12, e2020MS002159. [Google Scholar] [CrossRef]
  12. Silva, L.d.Á.; Omena-Garcia, R.P.; Condori-Apfata, J.A.; Costa, P.M.d.A.; Silva, N.M.; DaMatta, F.M.; Zsögön, A.; Araújo, W.L.; de Toledo Picoli, E.A.; Sulpice, R.; et al. Specific leaf area is modulated by nitrogen via changes in primary metabolism and parenchymal thickness in pepper. Planta 2021, 253, 16. [Google Scholar] [CrossRef]
  13. Gilardelli, C.; Confalonieri, R.; Cappelli, G.A.; Bellocchi, G. Sensitivity of WOFOST-based modelling solutions to crop parameters under climate change. Ecol. Model. 2018, 368, 1–14. [Google Scholar] [CrossRef]
  14. Zhang, Y.; Homma, K.; Shi, L.; Wang, Y.; Qiao, H.; Zha, Y. Improving crop modeling by simultaneously incorporating dynamic specific leaf area and leaf area index: A two-year experiment. Agric. Syst. 2025, 227, 104357. [Google Scholar] [CrossRef]
  15. Terashima, I.; Miyazawa, S.-I.; Hanba, Y.T. Why are sun leaves thicker than shade leaves?—Consideration based on analyses of CO2 diffusion in the leaf. J. Plant Res. 2001, 114, 93–105. [Google Scholar] [CrossRef]
  16. Yao, H.; Zhang, Y.; Yi, X.; Zhang, X.; Zhang, W. Cotton responds to different plant population densities by adjusting specific leaf area to optimize canopy photosynthetic use efficiency of light and nitrogen. Field Crops Res. 2016, 188, 10–16. [Google Scholar] [CrossRef]
  17. Ji, W.; LaZerte, S.E.; Waterway, M.J.; Lechowicz, M.J. Functional ecology of congeneric variation in the leaf economics spectrum. New Phytol. 2019, 225, 196–208. [Google Scholar] [CrossRef]
  18. Zhou, H.; Zhou, G.; He, Q.; Zhou, L.; Ji, Y.; Zhou, M. Environmental explanation of maize specific leaf area under varying water stress regimes. Environ. Exp. Bot. 2020, 171, 103932. [Google Scholar] [CrossRef]
  19. Ač, A.; Malenovský, Z.; Olejníčková, J.; Gallé, A.; Rascher, U.; Mohammed, G. Meta-analysis assessing potential of steady-state chlorophyll fluorescence for remote sensing detection of plant water, temperature and nitrogen stress. Remote Sens. Environ. 2015, 168, 420–436. [Google Scholar] [CrossRef]
  20. Ali, A.M.; Darvishzadeh, R.; Skidmore, A.K. Retrieval of specific leaf area from landsat-8 surface reflectance data using statistical and physical models. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2017, 10, 3529–3536. [Google Scholar] [CrossRef]
  21. Chauhan, S.; Srivastava, H.S.; Patel, P. Wheat crop biophysical parameters retrieval using hybrid-polarized RISAT-1 SAR data. Remote Sens. Environ. 2018, 216, 28–43. [Google Scholar] [CrossRef]
  22. Wang, J.; Xiao, X.; Bajgain, R.; Starks, P.; Steiner, J.; Doughty, R.B.; Chang, Q. Estimating leaf area index and aboveground biomass of grazing pastures using Sentinel-1, Sentinel-2 and Landsat images. ISPRS J. Photogramm. Remote Sens. 2019, 154, 189–201. [Google Scholar] [CrossRef]
  23. Ma, Y.; Zhang, Q.; Yi, X.; Ma, L.; Zhang, L.; Huang, C.; Zhang, Z.; Lv, X. Estimation of Cotton Leaf Area Index (LAI) Based on Spectral Transformation and Vegetation Index. Remote Sens. 2022, 14, 136. [Google Scholar] [CrossRef]
  24. Osco, L.P.; Marcato, J., Jr.; Marques Ramos, A.P.; de Castro Jorge, L.A.; Fatholahi, S.N.; de Andrade Silva, J.; Matsubara, E.T.; Pistori, H.; Gonçalves, W.N.; Li, J. A review on deep learning in UAV remote sensing. Int. J. Appl. Earth Obs. Geoinf. 2021, 102, 102456. [Google Scholar] [CrossRef]
  25. Luo, S.; Jiang, X.; He, Y.; Li, J.; Jiao, W.; Zhang, S.; Xu, F.; Han, Z.; Sun, J.; Yang, J.; et al. Multi-dimensional variables and feature parameter selection for aboveground biomass estimation of potato based on UAV multispectral imagery. Front. Plant Sci. 2022, 13, 948249. [Google Scholar] [CrossRef]
  26. Gong, Y.; Yang, K.; Lin, Z.; Fang, S.; Wu, X.; Zhu, R.; Peng, Y. Remote estimation of leaf area index (LAI) with unmanned aerial vehicle (UAV) imaging for different rice cultivars throughout the entire growing season. Plant Methods 2021, 17, 88. [Google Scholar] [CrossRef]
  27. Li, W.; Wang, J.; Zhang, Y.; Yin, Q.; Wang, W.; Zhou, G.; Huo, Z. Combining texture, color, and vegetation index from unmanned aerial vehicle multispectral images to estimate winter wheat leaf area index during the vegetative growth stage. Remote Sens. 2023, 15, 5715. [Google Scholar] [CrossRef]
  28. Sibanda, M.; Mutanga, O.; Rouget, M.; Kumar, L. Estimating biomass of native grass grown under complex management treatments using worldview-3 spectral derivatives. Remote Sens. 2017, 9, 55. [Google Scholar] [CrossRef]
  29. Jamali, M.; Soufizadeh, S.; Yeganeh, B.; Emam, Y. Wheat leaf traits monitoring based on machine learning algorithms and high-resolution satellite imagery. Ecol. Inform. 2023, 74, 101967. [Google Scholar] [CrossRef]
  30. de Paula, R.G.; da Silva, M.F.; Amaral, C.; de Sousa Paula, G.; da Silva, L.J.; Pessoa, H.P.; da Silva, F.L. Leveraging leaf spectroscopy to identify drought-tolerant soybean cultivars. Smart Agric. Technol. 2024, 9, 100626. [Google Scholar] [CrossRef]
  31. Chen, L.; Xu, B.; Zhao, C.; Duan, D.; Cao, Q.; Wang, F. Application of multispectral camera in monitoring the quality parameters of fresh Tea Leaves. Remote Sens. 2021, 13, 3719. [Google Scholar] [CrossRef]
  32. Rouse, J.W., Jr.; Haas, R.H.; Schell, J.; Deering, D. Monitoring the Vernal Advancement and Retrogradation (Green Wave Effect) of Natural Vegetation; NASA/GSFC: Greenbelt, MD, USA, 1973.
  33. Chávez-Martínez, O.; Monjardin-Armenta, S.A.; Rangel-Peraza, J.G.; Sanhouse-García, A.J.; Mora-Felix, Z.D.; Plata-Rocha, W. Use of different vegetation indices for the evaluation of the kinetics of the cherry tomato (Solanum lycopersicum var. cerasiforme) growth based on multispectral images by UAV. Open Agric. 2024, 9, 20220357. [Google Scholar]
  34. Lemenkova, P.; Debeir, O. Computing Vegetation Indices from the Satellite Images Using GRASS GIS Scripts for Monitoring Mangrove Forests in the Coastal Landscapes of Niger Delta, Nigeria. J. Mar. Sci. Eng. 2023, 11, 871. [Google Scholar] [CrossRef]
  35. Haboudane, D.; Miller, J.R.; Pattey, E.; Zarco-Tejada, P.J.; Strachan, I.B. Hyperspectral vegetation indices and novel algorithms for predicting green LAI of crop canopies: Modeling and validation in the context of precision agriculture. Remote Sens. Environ. 2004, 90, 337–352. [Google Scholar] [CrossRef]
  36. Rondeaux, G.; Steven, M.; Baret, F. Optimization of soil-adjusted vegetation indices. Remote Sens. Environ. 1996, 55, 95–107. [Google Scholar] [CrossRef]
  37. Nguy-Robertson, A.L. The mathematical identity of two vegetation indices: MCARI2 and MTVI2. Int. J. Remote Sens. 2013, 34, 7504–7507. [Google Scholar] [CrossRef]
  38. Penuelas, J.; Baret, F.; Filella, I. Semi-empirical indices to assess carotenoids/chlorophyll a ratio from leaf spectral reflectance. Photosynthetica 1995, 31, 221–230. [Google Scholar]
  39. Fuentes, M.V.B.; Heenkenda, M.K.; Sahota, T.S.; Serrano, L.S. Analyzing Winter Wheat (Triticum aestivum) Growth Pattern Using High Spatial Resolution Images: A Case Study at Lakehead University Agriculture Research Station, Thunder Bay, Canada. Crops 2024, 4, 115–133. [Google Scholar] [CrossRef]
  40. Hancock, D.W. Spectral Reflectance of Canopies of Rainfed and Subsurface Irrigated Alfalfa. Ph.D. Thesis, University of Kentucky, Lexington, KY, USA, 2006. [Google Scholar]
  41. Kaufman, Y.J.; Tanre, D. Atmospherically resistant vegetation index (ARVI) for EOS-MODIS. IEEE Trans. Geosci. Remote Sens. 1992, 30, 261–270. [Google Scholar] [CrossRef]
  42. Gitelson, A.A.; Merzlyak, M.N. Remote estimation of chlorophyll content in higher plant leaves. Int. J. Remote Sens. 1997, 18, 2691–2697. [Google Scholar] [CrossRef]
  43. 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]
  44. Baret, F.; Guyot, G. Potentials and limits of vegetation indices for LAI and APAR assessment. Remote Sens. Environ. 1991, 35, 161–173. [Google Scholar] [CrossRef]
  45. Dash, J.; Curran, P. Evaluation of the MERIS terrestrial chlorophyll index (MTCI). Adv. Space Res. 2007, 39, 100–104. [Google Scholar] [CrossRef]
  46. Gitelson, A.A.; Viña, A.; Ciganda, V.; Rundquist, D.C.; Arkebauer, T.J. Remote estimation of canopy chlorophyll content in crops. Geophys. Res. Lett. 2005, 32, L08403. [Google Scholar] [CrossRef]
  47. Zarco-Tejada, P.; Morales, A.; Testi, L.; Villalobos, F. Spatio-temporal patterns of chlorophyll fluorescence and physiological and structural indices acquired from hyperspectral imagery as compared with carbon fluxes measured with eddy covariance. Remote Sens. Environ. 2013, 133, 102–115. [Google Scholar] [CrossRef]
  48. Khishigjargal, M.; Dulamsuren, C.; Lkhagvadorj, D.; Leuschner, C.; Hauck, M. Contrasting responses of seedling and sapling densities to livestock density in the Mongolian forest-steppe. Plant Ecol. 2013, 214, 1391–1403. [Google Scholar] [CrossRef]
  49. Guebsi, R.; Mami, S.; Chokmani, K. Drones in Precision Agriculture: A Comprehensive Review of Applications, Technologies, and Challenges. Drones 2024, 8, 686. [Google Scholar] [CrossRef]
  50. Silleos, N.G.; Alexandridis, T.K.; Gitas, I.Z.; Perakis, K. Vegetation indices: Advances made in biomass estimation and vegetation monitoring in the last 30 years. Geocarto Int. 2006, 21, 21–28. [Google Scholar] [CrossRef]
  51. El-Shikha, D.M.; Barnes, E.M.; Clarke, T.R.; Hunsaker, D.J.; Haberland, J.A.; Pinter, P., Jr.; Waller, P.M.; Thompson, T.L. Remote sensing of cotton nitrogen status using the canopy chlorophyll content index (CCCI). Trans. ASABE 2008, 51, 73–82. [Google Scholar] [CrossRef]
  52. Peñuelas, J.; Gamon, J.; Fredeen, A.; Merino, J.; Field, C. Reflectance indices associated with physiological changes in nitrogen- and water-limited sunflower leaves. Remote Sens. Environ. 1994, 48, 135–146. [Google Scholar] [CrossRef]
  53. Gamon, J.A.; Surfus, J.S. Assessing leaf pigment content and activity with a reflectometer. New Phytol. 1999, 143, 105–117. [Google Scholar] [CrossRef]
  54. Xue, J.; Su, B. Significant remote sensing vegetation indices: A review of developments and applications. J. Sens. 2017, 2017, 1353691. [Google Scholar] [CrossRef]
  55. Courel, M.-F.; Chamard, P.C.; Guenegou, M.; Lerhun, J.; Levasseur, J.; Togola, M. Utilisation des bandes spectrales du vert et du rouge pour une meilleure évaluation des formations végétales actives. In Proceedings of the Congrès AUPELF-UREF, Sherbrooke, QC, Canada, 21–23 October 1991. [Google Scholar]
  56. Haralick, R.M.; Shanmugam, K.; Dinstein, I.H. Textural features for image classification. IEEE Trans. Syst. Man Cybern. 1973, SMC-3, 610–621. [Google Scholar]
  57. Hall-Beyer, M. Practical guidelines for choosing GLCM textures to use in landscape classification tasks over a range of moderate spatial scales. Int. J. Remote Sens. 2017, 38, 1312–1338. [Google Scholar] [CrossRef]
  58. Romero, A.; Aguado, I.; Yebra, M. Estimation of dry matter content in leaves using normalized indexes and PROSPECT model inversion. Int. J. Remote Sens. 2012, 33, 396–414. [Google Scholar] [CrossRef]
  59. Chen, L.; Xing, M.; He, B.; Wang, J.; Shang, J.; Huang, X.; Xu, M. Estimating soil moisture over winter wheat fields during growing season using machine-learning methods. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2021, 14, 3706–3718. [Google Scholar] [CrossRef]
  60. Denil, M.; Shakibi, B.; Dinh, L.; Ranzato, M.A.; De Freitas, N. Predicting parameters in deep learning. arXiv 2013. [Google Scholar] [CrossRef]
  61. Zha, H.; Miao, Y.; Wang, T.; Li, Y.; Zhang, J.; Sun, W.; Feng, Z.; Kusnierek, K. Improving unmanned aerial vehicle remote sensing-based rice nitrogen nutrition index prediction with machine learning. Remote Sens. 2020, 12, 215. [Google Scholar] [CrossRef]
  62. Williams, P.C. Variables affecting near-infrared reflectance spectroscopic analysis. In Near-Infrared Technology in the Agricultural and Food Industries; Williams, P.C., Norris, K., Eds.; American Association of Cereal Chemists, Inc.: St. Paul, MN, USA, 1987; pp. 143–167. [Google Scholar]
  63. Rossel, R.A.V.; Taylor, H.J.; McBratney, A.B. Multivariate calibration of hyperspectral γ-ray energy spectra for proximal soil sensing. Eur. J. Soil Sci. 2007, 58, 343–353. [Google Scholar] [CrossRef]
  64. Belgiu, M.; Drăguţ, L. Random forest in remote sensing: A review of applications and future directions. ISPRS J. Photogramm. Remote Sens. 2016, 114, 24–31. [Google Scholar] [CrossRef]
  65. Chen, T.; Guestrin, C. Xgboost: A scalable tree boosting system. In Proceedings of the 22nd Acm Sigkdd International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, 13–17 August 2016; pp. 785–794. [Google Scholar]
  66. Mansaray, L.R.; Zhang, K.; Kanu, A.S. Dry biomass estimation of paddy rice with Sentinel-1A satellite data using machine learning regression algorithms. Comput. Electron. Agric. 2020, 176, 105674. [Google Scholar] [CrossRef]
  67. Zhang, J.; Cheng, T.; Guo, W.; Xu, X.; Qiao, H.; Xie, Y.; Ma, X. Leaf area index estimation model for UAV image hyperspectral data based on wavelength variable selection and machine learning methods. Plant Methods 2021, 17, 49. [Google Scholar] [CrossRef] [PubMed]
  68. Prananto, J.A.; Minasny, B.; Weaver, T. Near infrared (NIR) spectroscopy as a rapid and cost-effective method for nutrient analysis of plant leaf tissues. Adv. Agron. 2020, 164, 1–49. [Google Scholar]
  69. Guyot, G.; Baret, F.; Major, D.J. High Specrral Resolution: Determination of Specrral Shifts Between the Red and Infrared. Int. Arch. Photogramm. Remote Sens. 1988, 11, 750–760. [Google Scholar]
  70. Wright, I.J.; Reich, P.B.; Westoby, M.; Ackerly, D.D.; Baruch, Z.; Bongers, F.; Cavender-Bares, J.; Chapin, T.; Cornelissen, J.H.C.; Diemer, M.; et al. The worldwide leaf economics spectrum. Nature 2004, 428, 821–827. [Google Scholar] [CrossRef]
  71. Ali, A.M.; Darvishzadeh, R.; Skidmore, A.K.; van Duren, I. Specific leaf area estimation from leaf and canopy reflectance through optimization and validation of vegetation indices. Agric. For. Meteorol. 2017, 236, 162–174. [Google Scholar] [CrossRef]
  72. 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. Geophys. Res. Lett. 2003, 30, 1248. [Google Scholar] [CrossRef]
  73. Li, M.; Tan, Y.; Pan, J.; Peng, S. Modeling forest aboveground biomass by combining spectrum, textures and topographic features. Front. For. China 2008, 3, 10–15. [Google Scholar] [CrossRef]
  74. Sun, X.; Yang, Z.; Su, P.; Wei, K.; Wang, Z.; Yang, C.; Wang, C.; Qin, M.; Xiao, L.; Yang, W.; et al. Non-destructive monitoring of maize LAI by fusing UAV spectral and textural features. Front. Plant Sci. 2023, 14, 1158837. [Google Scholar] [CrossRef] [PubMed]
  75. Lourenço, P.; Godinho, S.; Sousa, A.; Gonçalves, A.C. Estimating tree aboveground biomass using multispectral satellite-based data in Mediterranean agroforestry system using random forest algorithm. Remote Sens. Appl. Soc. Environ. 2021, 23, 100560. [Google Scholar] [CrossRef]
  76. Breiman, L. Random forests. Mach. Learn. 2001, 45, 5–32. [Google Scholar] [CrossRef]
  77. Reich, P.B.; Walters, M.B.; Ellsworth, D.S. Leaf life—Span in relation to leaf, plant, and stand characteristics among diverse ecosystems. Ecol. Monogr. 1992, 62, 365–392. [Google Scholar] [CrossRef]
Agriculture 15 02293 g001
Figure 1. Geographical location of the study area and distribution of 72 experimental plots.
Figure 1. Geographical location of the study area and distribution of 72 experimental plots.
Agriculture 15 02293 g001
Agriculture 15 02293 g002
Figure 2. Model Construction and Validation Process Flowchart. The figure illustrates the process of data collection, preprocessing, model training, model validation, and obtaining the optimal model.
Figure 2. Model Construction and Validation Process Flowchart. The figure illustrates the process of data collection, preprocessing, model training, model validation, and obtaining the optimal model.
Agriculture 15 02293 g002
Agriculture 15 02293 g003
Figure 3. SLA value distribution across the tillering, jointing, and heading stages of rice.
Figure 3. SLA value distribution across the tillering, jointing, and heading stages of rice.
Agriculture 15 02293 g003
Agriculture 15 02293 g004
Figure 4. (a) Correlation heatmap between VIs and SLA values; (b) Correlation heatmap between TIs and SLA values.
Figure 4. (a) Correlation heatmap between VIs and SLA values; (b) Correlation heatmap between TIs and SLA values.
Agriculture 15 02293 g004
Agriculture 15 02293 g005
Figure 5. RFE feature selection learning curve (based on MSE): (a) VIs; (b) TIs; (c) Vis + TIs. Note: Mean Square Error (MSE) is an indicator employed to assess regression performance. In scikit-learn, all evaluation metrics are converted to the format of ‘bigger is better’. For MSE, a lower value is preferable. To adhere to the convention of ‘bigger is better’, the negative MSE is used to conform to this convention.
Figure 5. RFE feature selection learning curve (based on MSE): (a) VIs; (b) TIs; (c) Vis + TIs. Note: Mean Square Error (MSE) is an indicator employed to assess regression performance. In scikit-learn, all evaluation metrics are converted to the format of ‘bigger is better’. For MSE, a lower value is preferable. To adhere to the convention of ‘bigger is better’, the negative MSE is used to conform to this convention.
Agriculture 15 02293 g005
Agriculture 15 02293 g006
Figure 6. RFE Feature Importance Ranking, presenting only the filtered optimal selected features: (a) VIs; (b) TIs; (c) Vis + TIs.
Figure 6. RFE Feature Importance Ranking, presenting only the filtered optimal selected features: (a) VIs; (b) TIs; (c) Vis + TIs.
Agriculture 15 02293 g006
Agriculture 15 02293 g007
Figure 7. Scatter plot between measured and predicted SLA values of the best-performing estimation model developed using various Remote sensing characteristic variables: (a) RF-VIs; (b) RF-TIs; (c) RF-(Vis + TIs). The diagonal denotes a 1:1 relationship. R2, RMSE, RRMSE, and RPD in the figure represent the evaluation accuracy of the test set.
Figure 7. Scatter plot between measured and predicted SLA values of the best-performing estimation model developed using various Remote sensing characteristic variables: (a) RF-VIs; (b) RF-TIs; (c) RF-(Vis + TIs). The diagonal denotes a 1:1 relationship. R2, RMSE, RRMSE, and RPD in the figure represent the evaluation accuracy of the test set.
Agriculture 15 02293 g007
Agriculture 15 02293 g008
Figure 8. SLA value RF prediction model built on various variables: (a) RF (VIs); (b) RF (TIs); (c) RF (VIs + TIs). All subplots depict scatter distributions for various growth stages in the validation dataset.
Figure 8. SLA value RF prediction model built on various variables: (a) RF (VIs); (b) RF (TIs); (c) RF (VIs + TIs). All subplots depict scatter distributions for various growth stages in the validation dataset.
Agriculture 15 02293 g008
Table 1. The 30 VIs and 8 TIs employed in this study to estimate SLA values.
Table 1. The 30 VIs and 8 TIs employed in this study to estimate SLA values.
VariablesSpectral IndexFormulaReferences
R\\
G\\
B\\
RedEdge\\
NIR\\
NDVI(NIR − R)/(NIR + R)[32]
IPVINIR/(NIR + R)[33]
DVINIR − R[34]
TVI0.5 × [120 × (NIR − G) − 200 × (R − G)][35]
OSAVI(NIR − R)/(NIR − R + L) (L = 0.16)[36]
MCARI21.2 × [1.2 × (NIR − R) − 2.5 × (R − G)][37]
SIPI(NIR − B)/(NIR + R)[38]
EVI22.5 × (NIR − R)/(NIR + 2.4 × R + 1)[39]
RENDVI(RE − R)/(RE + R)[40]
VIsSARVI(NIR − R)/(NIR + R + 0.5)[41]
RE-NIR-GTI1.5 (NIR − RE) − 2(RE − G)This study
NDRE(NIR − RE)/(NIR + RE)[42]
GCINIR/G − 1[43]
CI(NIR − RE)/(NIR + R)[44]
MTCI(NIR − RE)/(RE + R)[45]
CIRENIR/RE − 1[46]
NPCI(R − B)/(R + B)[47]
NGRDI(G − R)/(G + R)[48]
GNDVI(NIR − G)/(NIR + G)[49]
NGIG/(R + G + B)[50]
CCCI(NIR − RE)/(NIR + RE) × (NIR − R)/(NIR + R)[51]
ExB1.4 × B − G[52]
IRGNIR/G[53]
GBG/B[54]
GIG/R[55]
TIsGLCMMean, Var, Homo,
Cor, Dis, Ent,
SecMom, Con		[56]
Table 2. Descriptive statistics of rice SLA measurement.
Table 2. Descriptive statistics of rice SLA measurement.
PeriodNMax (m2/g)Min (m2/g)Mean (m2/g)SD (m2/g)CV (%)
Tillering stage721.330.570.850.1821.26
Jointing stage720.910.420.640.1422.17
Heading stage720.760.350.540.1324.24
All2161.330.350.680.2029.64
Table 3. Prediction performance of models utilizing VIs (Train/CV and Test sets).
Table 3. Prediction performance of models utilizing VIs (Train/CV and Test sets).
VIsModel Train(CV) Test
R2RMSE (m2/g)RRMSERPDR2RMSE (m2/g)RRMSERPD
RF0.85690.07020.10392.69600.90350.06990.10503.2184
XGBoost0.87950.06550.09283.00400.86440.08290.12442.7155
GBDT0.85830.07010.10292.68660.84800.08780.13172.5651
PLSR0.76270.09030.14201.94390.82830.09330.14002.4132
Table 4. Prediction performance of models utilizing TIs (Train/CV and Test sets).
Table 4. Prediction performance of models utilizing TIs (Train/CV and Test sets).
TIs-Model Train(CV) Test
R2RMSE_
(m2/g)		RRMSERPDR2RMSE_
(m2/g)		RRMSERPD
RF0.85560.07110.09942.88200.90080.07090.10643.1756
XGBoost0.88760.06300.09352.95830.85610.08540.12822.6359
GBDT0.86530.06830.10042.82410.86180.08370.12562.6898
PLSR0.71220.09960.14691.85630.80020.10060.15102.2369
Table 5. Model estimation accuracy derived from the integration of VIs and TIs (Train/CV and Test sets).
Table 5. Model estimation accuracy derived from the integration of VIs and TIs (Train/CV and Test sets).
VIs + TIs
-Model		 Train(CV) Test
R2RMSE_
(m2/g)		RRMSERPDR2RMSE_(m2/g)RRMSERPD
RF0.85250.07130.10672.64120.90490.06940.10423.2419
XGBoost0.88280.06410.09412.95840.85600.08540.12822.6355
GBDT0.86240.06910.10142.71120.85190.08660.13002.5986
PLSR0.76030.09150.13162.07310.86060.08410.12622.6780
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

Huang, J.; Wang, S.; Pei, Y.; Yin, Q.; Ding, Z.; Wang, J.; Wang, W.; Zhou, G.; Huo, Z. Estimate the Pre-Flowering Specific Leaf Area of Rice Based on Vegetation Indices and Texture Indices Derived from UAV Multispectral Imagery. Agriculture 2025, 15, 2293. https://doi.org/10.3390/agriculture15212293

AMA Style

Huang J, Wang S, Pei Y, Yin Q, Ding Z, Wang J, Wang W, Zhou G, Huo Z. Estimate the Pre-Flowering Specific Leaf Area of Rice Based on Vegetation Indices and Texture Indices Derived from UAV Multispectral Imagery. Agriculture. 2025; 15(21):2293. https://doi.org/10.3390/agriculture15212293

Chicago/Turabian Style

Huang, Jingjing, Sunan Wang, Yuexia Pei, Quan Yin, Zhi Ding, Jianjun Wang, Weiling Wang, Guisheng Zhou, and Zhongyang Huo. 2025. "Estimate the Pre-Flowering Specific Leaf Area of Rice Based on Vegetation Indices and Texture Indices Derived from UAV Multispectral Imagery" Agriculture 15, no. 21: 2293. https://doi.org/10.3390/agriculture15212293

APA Style

Huang, J., Wang, S., Pei, Y., Yin, Q., Ding, Z., Wang, J., Wang, W., Zhou, G., & Huo, Z. (2025). Estimate the Pre-Flowering Specific Leaf Area of Rice Based on Vegetation Indices and Texture Indices Derived from UAV Multispectral Imagery. Agriculture, 15(21), 2293. https://doi.org/10.3390/agriculture15212293

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop