Abstract: High-precision crop identification models are fundamental for ensuring food security, optimizing planting structures, and advancing the development of smart agriculture. However, in plateau mountainous regions, remote sensing classification faces significant bottlenecks due to highly fragmented terrain, rugged topography, and extreme environmental complexity. These factors result in pronounced geospatial non-stationarity, where the relationship between spectral signatures and crop types varies across space, causing global models to suffer from a "homogenization constraint" that often suppresses localized signals. Furthermore, traditional spatial statistical methods, such as Geographically Weighted Regression (GWR), often struggle to characterize the high-dimensional nonlinear relationships inherent in multi-source remote sensing features and are frequently constrained by prohibitive computational overhead and a lack of model interpretability. To address these issues and systematically alleviate the impact of geospatial non-stationarity, this paper constructs a high-precision crop identification framework based on CCI-LightGBM-SHAP. First, using the Classification Confusion Index (CCI) based on sample spatial confusion characteristics, and integrating topographic factors with multi-source remote sensing features, the study area is recursively delineated into multiple spatial subdomains with internally homogeneous spectral and geomorphological characteristics. Second, ensemble learning models such as the Light Gradient Boosting Machine (LightGBM) are deployed as local "experts" within each subdomain to achieve fine-grained modeling of local feature structures by autonomously optimizing local feature subsets and decision rules, thereby bypassing global homogenization constraints. Finally, the SHapley Additive exPlanations (SHAP) method is employed to reveal the nuanced feature response mechanisms driving high-accuracy crop identification across different geographic partitions and complex environmental gradients The experimental results demonstrate that: (1) The overall classification accuracy was significantly enhanced, achieving an average Overall Accuracy (OA) of 0.928 and a Matthews Correlation Coefficient (MCC) of 0.915. Quantitative assessments indicate that partitioned local models consistently outperform the global baseline model in all subdomains. Among the evaluated ensemble learning algorithms, LightGBM exhibited the highest degree of stability and generalization capability, with F1-scores for key crop categories improving by as much as 6.28%. These metrics validate the effectiveness of the "partitioning + ensemble learning" strategy in mitigating classification errors induced by spatial heterogeneity. (2) The model possesses strong adaptability to spatial non-stationarity. SHAP analysis indicates that local models can autonomously adjust feature dependencies in response to varying geomorphological environments, capturing a dynamic decision-making transition from a "structure-dominated" to a "physiology-dominated" paradigm. (3) Multi-source remote sensing features exhibit differentiated synergistic driving mechanisms for various crop types. Specifically, rice identification is jointly influenced by topography and seasonal hydrological rhythms; corn relies on relatively homogeneous terrain and sufficient thermal conditions; rapeseed is primarily driven by spectral variations induced by phenological differences; whereas perennial crops such as orchards and tea gardens show stronger associations with stable environmental factors like altitude. In conclusion, the results suggest that the CCI-LightGBM-SHAP modeling approach mitigates the impact of spatial non-stationarity on identification accuracy by effectively adapting to localized environmental gradients. Through the integration of predictive precision and interpretability, this framework establishes a robust technical methodology for fine-grained remote sensing crop mapping. These findings provide highly valuable spatial decision-support data for agricultural management and food security assessments in complex, heterogeneous mountainous regions.