Intelligent Assessment of Suaeda salsa Growth Stress Factors Based on Multimodal Fusion and Transfer Learning

This paper proposes an improved machine learning algorithm based on multimodal fusion and transfer learning for intelligent assessment of Suaeda salsa growth stress factors. By integrating remote sensing images, near-surface images, and ground survey data, the assessment model is optimized through dynamic feature engineering. The results demonstrated that key stress factors affecting Suaeda salsa growth in Liaohe Estuary wetland, ranked by full-cycle comprehensive contribution, are: soil salinity (0.83) > crab burrow density (0.53) > mean NDVI (0.43) > precipitation (0.31) > soil moisture (0.26). Soil salinity is identified as the primary stress factor, consistent with existing studies; this work quantitatively evaluates the dual effects of crab burrow density for the first time, confirming its role as a critical biological stressor. The improved CNN-RF hybrid model achieves a mean square error (MSE) of 0.016 (R² = 0.941) under conventional scenarios, representing a 36% improvement in accuracy compared with traditional RF methods; under extreme compound stress conditions, the stability coefficient (SC) is 2.05, 4.7% lower than traditional RF approaches. Causal analysis based on the SHAP-LIME hybrid framework reveals a "salinity-moisture-crab burrow" ternary interaction mechanism. The GIS-integrated decision support system enables real-time visualization and multi-level early warnings of growth status, successfully identifying ecological degradation risks in the western high-salinity zone (local contribution 0.88) and northern crab burrow-intensive areas during empirical application in Liaohe Estuary. This research provides a data-driven intelligent assessment tool for coastal wetland ecological protection.