UAV Monitoring of Crop Growth Across Multiple Fields by Integrating Aboveground Biomass Estimation and Phenological Stage Recognition

Accurate monitoring of crop aboveground biomass (AGB) using unmanned aerial vehicle (UAV) remote sensing is important for field-scale crop management. Surveys covering multiple fields often require higher flight altitudes, which reduce image resolution and may decrease AGB estimation accuracy. In addition, differences in crop phenological stages among fields can confound direct comparisons of crop growth. To address these issues, UAV RGB and multispectral imagery, together with ground-measured AGB data, were collected from wheat and rice fields in 2024 and 2025. Eight regression algorithms were evaluated using multispectral vegetation indices alone and in combination with RGB vegetation indices under repeated five-fold cross-validation. TabPFN achieved the best overall performance. After feature fusion, its R2 for wheat dry-weight AGB (AGBD) and fresh-weight AGB (AGBW) were 0.955 and 0.926, respectively, while those for rice were 0.764 and 0.768. After models trained on the 2024 data were updated with 20% of the 2025 samples, TabPFN achieved cross-year R2 of 0.667 for wheat AGBW and 0.831 for rice AGBW. Seven deep-learning models were further evaluated for phenological stage recognition using independent test data from different study areas. DenseNet achieved the highest accuracy on the rice test set (85.6%), whereas ViT-B achieved the highest accuracy on the wheat test set (85.2%). ViT-B was subsequently used as the teacher model to improve the wheat recognition performance of DenseNet through knowledge distillation. At the optimal distillation temperature of 4, the accuracy of DenseNet on the wheat test set increased from 56.6% to 73.6%. DenseNet contained only 6.959 M parameters, approximately 91.9% fewer than ViT-B. It was therefore adopted as the unified compact network architecture for phenological stage recognition in both crops. Finally, rice AGBW predictions were combined with phenological stage recognition results to produce stage-stratified biomass maps. Comparing fields within the same phenological stage reduced misinterpretation caused by normal stage-related differences in biomass and made anomalously low-growth areas easier to identify. The proposed workflow therefore integrates AGB prediction with phenological information to support more reliable field-scale crop-growth assessment and targeted management.