Estimation of winter wheat aboveground biomass by combining unmanned aerial vehicle hyperspectral and textural features with deep learning method

Combining spectral and texture information from unmanned aerial vehicle (UAV) hyperspectral imagery with deep learning to construct winter wheat biomass inversion models is important for biomass monitoring and yield prediction. UAV hyperspectral imagery and biomass data were collected simultaneously across the jointing, flowering, and grain-filling stages. Optimal vegetation indices were integrated with texture metrics to construct a "graph-spectral" fusion index, optimized using competitive adaptive reweighted sampling (CARS). A 1D-CNN+BiLSTM model was developed and compared with support vector regression (SVR), back-propagation neural network (BPNN), and random forest (RF). Results showed that incorporating texture features significantly enhanced performance compared to spectral-only inputs, increasing validation R2 by 6.4% for 1D-CNN+BiLSTM and 5.2% for RF. The 1D-CNN+BiLSTM model based on CARS features achieved the highest accuracy, with modeling R2, RMSE, and RPD of 0.88, 3073.7 kg·hm-2 and 2.84, and validation values of 0.85, 3492.9 kg·hm-2 and 2.61, improving validation R2 by 2.4% and reducing RMSE by 3.6% compared to RF. Furthermore, the Diebold-Mariano test revealed significant differences in predictive accuracy between 1D-CNN+BiLSTM and BPNN, SVR, and RF. Additionally, CARS shortened training time from 7.836 to 6.397 s, improving model efficiency. These results provide a basis and reference for accurate UAV-based winter wheat biomass estimation.