Cross-regional estimation of leaf chlorophyll and soil moisture content in drip-irrigated citrus orchards using UAV data and transfer learning

Optimizing precision drip irrigation hinges on the synergistic management of water and fertilizer, which requires simultaneous monitoring of soil moisture content (SMC) and plant nitrogen status. However, the spatial generalization of estimation models for these key parameters remains a major challenge due to environmental heterogeneity. This study developed a transfer learning framework to achieve cross-regional estimation of both leaf chlorophyll content (LCC) and SMC in drip-irrigated citrus orchards using UAV multispectral data. Three attention mechanism-based deep learning models (CNN-LSTM-Attention (CLA), CNN-LSTM-Attention-Adaboost (CLA-A), and CNN-LSTM-Attention-XGBoost (CLA-X)) were constructed and compared. To enhance cross-regional generalization, we evaluated unsupervised (DANN, TCA) and semi-supervised (Fine-tuning, Multi-task Learning) transfer learning strategies. When directly transferred to the target domain, the models trained on the source domain exhibited performance degradation for both LCC (R 2 ranging from 0.158 to 0.308) and SMC (SMC10, SMC20, and SMC40), with R 2 generally below 0.20 across all depths, underscoring the limitation of direct model transfer. Notably, Fine-tuning strategies significantly enhanced the estimation accuracy in the target domain. The Fine-tuned CLA-X model achieved the best performance, yielding high accuracy for LCC (R 2 =0.751) and clear improvements for SMC (SMC10, SMC20, and SMC40). In contrast, unsupervised strategies (DANN, TCA) showed limited improvement (ΔR 2