Astragalus s, notably Astragalus microcephalus, is critical to semi-arid ecosystems, yet faces decline due to anthropogenic and environmental pressures. This paper introduces a novel reinforcement learning (RL) framework integrated with geographic information systems (GISs) to map Astragalus s susceptibility in the Tang-e Sayad Protected Area, a 5,584-ha semi-arid, mountainous region area in the Chaharmahal and Bakhtiari Province, Iran. Five RL models (Q-learning (QL), proximal policy optimization-based proximal updating (PU), deep Q-learning (DQL), RL-stack, and deep deterministic policy gradient (DDPG)) were implemented, with SHAP (SHapley Additive exPlanations) applied for interpretability. Using 196 field-validated points, 23 initial environmental factors were reduced to 15 after collinearity analysis. The PU model achieved exceptional performance (AUC: 0.99, Kappa: 0.92, RMSE: 0.21, R2: 0.81). SHAP analysis identified distance from farm lands (0.104), slope (0.071), and distance from dirt roads (0.048) as top predictors, highlighting anthropogenic and topographic influences on Astragalus microcephalus decline. Susceptibility maps, classified into five zones (very low to very high), showed PU allocating 2,228.1 ha (39.9%) to very low susceptibility and 856.1 ha (15.3%) to very high susceptibility, primarily near farmlands and dirt roads. Field observations confirmed Astragalus microcephalus vulnerability in high-risk zones due to grazing and agricultural pressures. This study's high-resolution maps and interpretable RL models provide actionable insights for conservation planning, demonstrating the potential of RL and explainable AI to advance ecological susceptibility modeling in semi-arid protected areas, with broader implications for sustainable ecosystem management.