Capturing non-stationary driving mechanisms of Alpine meadow kNDVI dynamics via integrated Bi-LSTM-transformer and intelligent attribution

As a critical ecological barrier for the Yangtze River, the Upper Reaches of the Minjiang River (URMR) require a deep understanding of vegetation dynamics and their non-linear responses to climate. This study utilizes kNDVI data (2000–2025), a kernel-based index that effectively captures complex non-linear vegetation signals across diverse geographic gradients. We employ an integrated bidirectional long short-term memory (Bi-LSTM) and Transformer framework coupled with SHAP attribution and multi-scale lag analysis. Results reveal a significant greening trend across 90.54% of the area, though improvement is limited above 5000 m by thermal thresholds. SHAP attribution reveals that air temperature is the dominant driver (63.60%), exhibiting an “instantaneous triggering” mechanism (peak at Lag 0, r = 0.705). Conversely, hydrological forces show complex non-linearity; soil moisture displays a negative correlation in the current month (r = −0.377) but peaks at Lag 2 (r = 0.600), validating the “soil moisture memory effect.” Compared to traditional models, the Bi-LSTM-Transformer architecture provides a moderate yet consistent improvement in capturing “non-stationary” driving mechanisms. These findings establish a robust framework for high-precision vegetation monitoring and provide a scientific basis for adaptive management and ecological security in sensitive alpine ecosystems.