关键词: 冬油菜, 叶面积指数, mRMR算法, 交叉验证, 特征融合, 机器学习

Abstract:

[Objective] Leaf area index (LAI) is a vital agronomic parameter that reflects the structure of crop canopies, photosynthetic capacity, and population growth status. It holds significant importance for the precision cultivation and management of winter oilseed rape. Traditional methods for measuring LAI, such as destructive sampling or the use of costly instruments, are often constrained by low efficiency, high costs, and limited adaptability. In contrast, unmanned aerial vehicle (UAV) remote sensing technology offers a novel approach to rapid and non-destructive LAI monitoring due to its advantages in high resolution and flexibility. However, reliance solely on spectral vegetation indices (VIs) frequently results in saturation phenomena at elevated LAI levels, complicating accurate representation of complex canopy structures. Consequently, this study aims to investigate the integration of vegetation indices with texture features (TFs) using machine learning techniques to enhance the estimation accuracy of LAI throughout the entire growth cycle of winter oilseed rape. [Methods] The research was conducted at an experimental site in Gao'an city, Jiangxi province, where 81 plots were established with varying sowing dates, densities, and fertilization treatments. These plots encompassed four critical growth stages: seedling, bud elongation, flowering, and pod development. A DJI Phantom 4 RTK multispectral UAV was employed to acquire image data, while ground-truth LAI data were concurrently collected using an LAI-2200C plant canopy analyzer, resulting in a total of 324 valid samples. From the multispectral imagery obtained, eight vegetation indices, namely normalized difference vegetation index (NDVI), optimized soil adjusted vegetation index (OSAVI), normalized difference red edge (NDRE), enhanced vegetation index (EVI), green normalized difference vegetation index (GNDVI), simple ratio (SR), difference vegetation index (DVI), and canopy infrared reflectance estimate (CIRE), were derived alongside reflectance values from five spectral bands: blue, green, red, red-edge, and near-infrared. Additionally, 40 texture features were extracted based on the Gray-Level Co-occurrence Matrix. To select the ten most representative features with minimal redundancy among these variables, the minimum redundancy maximum relevance (mRMR) algorithm was utilized. Subsequently, three machine learning algorithms, multiple linear regression (MLR), extreme gradient boosting (XGBoost), and support vector machine regression (SVR), were applied to develop models for estimating LAI. To assess model generalizability and mitigate overfitting risks during evaluation processes, a 3-fold GroupKFold cross-validation approach was implemented to ensure that samples originating from the same plot remained intact between training and testing sets. The performance of each model was rigorously evaluated using several metrics, including the coefficient of determination (R²), root mean square error (RMSE), and mean absolute error (MAE). [Results and Discussions] The results indicated that the LAI of winter oilseed rape exhibited a dynamic trend throughout its growth cycle, characterized by being "low at the seedling stage, high at the bud elongation stage, decreasing at the flowering stage, and dropping again at the pod stage". The LAI values ranged from 0.90 to 6.39, with a uniform sample distribution observed within each developmental phase. Most vegetation indices and texture features demonstrated a highly significant correlation with LAI (P < 0.001), with the SR and near-infrared entropy (NIR-entropy) exhibiting the strongest correlation (r = 0.81). In terms of feature selection, vegetation indices such as SR, CIRE, and NDRE along with texture features like NIR-entropy and G-variance maintained high selection frequencies among the top ten mRMR-selected features. This indicated their stable contribution to model construction. Regarding model performance, the fused vegetation and texture features (VTFs) model outperformed all other models evaluated; specifically, the VTFs-SVR model achieved superior estimation accuracy across the entire growth cycle (R²=0.90, RMSE=0.38, MAE=0.27). When compared to models utilizing only vegetation indices or solely texture features, the fused model demonstrated particularly enhanced performance during high-coverage stages such as bud elongation, effectively addressing issues related to spectral saturation. Residual analysis further confirmed that the VTFs model exhibited a more concentrated residual distribution, indicating significantly greater prediction stability than single-feature models. [Conclusions] The fusion of vegetation indices and texture features extracted from UAV-based multispectral imagery, combined with machine learning modeling, particularly the SVR algorithm, enabled high-accuracy, non-destructive estimation of LAI throughout the entire growth cycle of winter oilseed rape. Texture features effectively supplemented canopy structural information, showing strong complementarity, especially during high-LAI stages where spectral data is prone to saturation. Through mRMR feature selection and group-based cross-validation, the model demonstrated good generalizability and practical application potential. This method could provide reliable technical support for monitoring winter oilseed rape growth status and precision agriculture management. Future research would further incorporate canopy structural parameters or multi-temporal features to enhance the model's estimation capability during stages dominated by non-leaf organs.

Key words: winter rapeseed, leaf area index, mRMR algorithm, nested cross-validation, feature fusion, machine learning