Explainable machine learning identifies key degradation factors in Robinia pseudoacacia shelterbelts

Abstract Introduction The degradation of Robinia pseudoacacia shelterbelts within the historic abandoned channels of the Yellow River compromises their ecological stability and critical ecosystem services, presenting an urgent need for effective restoration. Objectives This study aimed to develop a reproducible diagnostic system for this degradation to identify the key driving factors and critical thresholds for management intervention. Methods We constructed a Comprehensive Degradation Index (CDI) from indicators of stand structure, productivity, soil, and disturbance. Key variables were selected via correlation analysis, principal component analysis, Least Absolute Shrinkage and Selection Operator, and Random Forest (RF) feature importance. A spatially robust RF model was interpreted with SHapley Additive exPlanations analysis to identify degradation thresholds. Results The RF model demonstrated strong predictive performance and stability ( r 2 = 0.879). Analysis identified six key factors and their degradation thresholds: stand structural homogeneity ( H ′ < 0.58), uniform angle index ( W ≈ 0.50), soil total nitrogen (<0.30 g/kg), dead branch incidence (≈19.0%), canopy closure (≈69.0%), and disturbance level (≈3.29). Exceeding these thresholds leads to a marked increase in the CDI, revealing an accelerating degradation feedback loop. Conclusion Shelterbelt degradation is primarily driven by the interaction between stand structural simplification and elevated disturbance pressure. Surpassing the identified thresholds triggers an accelerated functional decline.