关键词: GF-2, Google Earth影像, 农村宅基地, 绿视率, 机器学习

Abstract: To furnish data-driven decision support for rural settlement spatial restructuring,revitalization of underutilized homesteads,and precision land governance through accurate identification and classification of rural homestead utilization states.A rural homestead utilization identification framework driven by multi-source high-resolution remote sensing data is proposed,integrating deep learning and machine learning techniques.The findings indicate that:①The overall accuracy of homestead recognition based on GF imagery and Google Earth imagery exceeds 84%;②The XGBoost model demonstrates superior performance in identifying inhabited homesteads,achieving a precision of 94.6%,while the random forest (RF)model exhibits the best performance in recognizing idle homesteads,with a precision of 77.8%;③According to comprehensive evaluations using ROC and PR curves,the RF algorithm outperforms the others,with the green looking ratio derived from Google Earth imagery contributing 12.7%to feature importance.This study substantiates that fusing multi-source remote sensing and machine learning technologies constitutes an effective approach for homestead utilization mapping,thereby providing a robust technical foundation for advancing land resource intensification and sustainable rural land management.

Key words: GF-2, Google Earth imagery, rural homestead, green looking ratio, machine learning