Mapping of Land Cover Over Highly Heterogeneous Areas in Yunnan Province With Active and Passive Remotely Sensed Data

Yunnan Province is one of the global biodiversity conservation hotspots. Due to the fragmented terrain and spatial heterogeneity of the plateau, land cover (LC) mapping in this region has always been challenging. In this study, we integrated 1593 scenes of Sentinel-1 and 8320 scenes of Sentinel-2 (S2) imagery, proposed a sample data generation strategy based on domain information rules, and developed a two-step feature optimization method. We also constructed a tile-based classification model to mitigate the impact of spatial heterogeneity in plateau regions. By comparing the performance of 45 different combinations of predictive variables, we analyzed the effectiveness of high-dimensional predictive variable optimization. We evaluated the impact of predictive variable sets, sensor types, classifier algorithms, and sample thresholds on LC mapping. The results indicate that: 1) the two-step feature optimization method can improve the classification accuracy of machine learning (ML). With optimal features, the model’s overall accuracy (OA) reached 92.21%, an increase of 0.64%–8.16% compared to using raw features alone; 2) the near-infrared band is a highly effective spectral predictor and slope significantly impacts the spatial distribution of LC in Yunnan Province; and 3) changes in sample thresholds affect the stability of prediction models and fivefold cross validation shows that under the optimal sample threshold scenario, prediction accuracy improves by 0.95%–2.76%; and 4) the tile-based classification model reduces the errors compared to traditional global models, mitigating heterogeneity effects. This study provides a new approach for classification in heterogeneous plateau regions and offers valuable data support for biodiversity conservation efforts.