Improved Multi-Stage Rice Above-Ground Biomass Estimation Using Wavelet-Texture-Fused Vegetation Indices from UAV...

When estimating above-ground biomass (AGB) across multiple growth stages, vegetation indices (VIs) have limitations due to saturation under dense canopies and poor sensitivity to vertically growing organs (e.g., panicles). Discrete wavelet transform (DWT) can extract multi-directional, multi-frequency texture features reflecting canopy structure changes, but its application in crop biomass monitoring is underexplored. Therefore, to evaluate whether DWT-based textures can be used to estimate AGB across multiple growth stages and whether combining VIs can improve estimation accuracy, two-year field experiments involving four rice varieties and five nitrogen treatments were conducted. UAV multispectral images were acquired during the critical growth stages, from which Vis and wavelet textures (WTs) were extracted, and novel wavelet texture indices (WTIs) were constructed. Correlation analysis guided feature selection, and simple regression, multiple linear regression, and Optuna-optimized random forest were employed to develop rice AGB estimation models. The results indicated: (1) Compared to a single WT, the WTIs exhibited higher correlation with rice AGB across different growth stages. (2) Among the three models, the RF model performed best. Specifically, using only VIs to estimate AGB during pre-heading yielded relatively higher accuracy (R² = 0.713), while using WTIs to estimate AGB during post-heading and all-stage yielded higher accuracy (R² = 0.709 and 0.668). (3) Combining WTIs with VIs significantly improves the prediction accuracy of AGB at different growth stages (R² = 0.782, 0.769, and 0.732; RMSE = 114.655, 161.779, and 223.654 g/m²), with R² improving by 10-15% and RMSE decreasing by 13-17% compared to the VIs. The study demonstrates that DWT-based textures can effectively assist in the high-precision estimation of rice AGB. Moreover, integrating WTIs with VIs enables accurate and stable prediction of rice AGB under different management practices and varieties, providing an economical and efficient method for estimating rice AGB.