Land Use/Land Cover Classification and Its Variability Along the Jiangsu Coast Based on Pearson-SHAP-RFE Feature Selection from Landsat Imagery

The Jiangsu coastal zone has experienced substantial land use/land cover (LULC) changes under intensified anthropogenic activities, while spectral confusion caused by land–sea interactions increases the difficulty of accurate LULC classification. This study investigated LULC dynamics in the Jiangsu coastal zone in 2000, 2008, 2016, and 2024 using Landsat imagery and the Google Earth Engine platform. A total of 33 features, including spectral bands, spectral indices, texture, topographic variables, and nighttime light data, were extracted. To improve classification accuracy and model generalization, a Pearson-SHAP-RFE feature selection framework was developed by integrating Pearson correlation analysis, SHapley Additive exPlanations, and recursive feature elimination. Four classification algorithms, namely Random Forest, Extreme Gradient Boosting, Classification and Regression Trees, and Support Vector Machine, were compared. Feature selection improved classification accuracy across the four algorithms, with improvements ranging from 4.05% to 5.15% compared with the spectral feature baseline. Random Forest achieved the highest overall accuracy of 92.28%, with a Kappa coefficient of 0.9019. The long-term classification results showed that LULC changes mainly occurred among forest–grassland, construction land, and cultivated land. Water bodies showed the smallest centroid shift, whereas cultivated land exhibited the greatest spatial displacement.