In the face of economic and environmental challenges, farmers in mountainous regions of China are increasingly willing to move to urban areas, leaving their rural residential land (RRL) permanently abandoned. Early and comprehensive identification of these vulnerable RRLs is crucial for rural land management and development planning. Given the advantages of machine learning techniques in predicting land use—owing to their efficiency and computational performance—this study aims to identify RRLs with a higher probability of abandonment in a more comprehensive and accurate manner by introducing multiple machine learning techniques and optimization strategies. Using various combinations of sampling strategies and balancing methods, the study derived the optimal technique–method–strategy combinations, compared their performance through model training and validation, and ultimately identified RRLs vulnerable to abandonment in Fang County, Hubei Province, China. The results show that the model using the random forest technique with a by-grid sampling strategy, along with both random oversampling and undersampling methods, achieved the highest accuracy (88.82 %). The model using the classification and regression tree (CART) technique with bagging, along with the same sampling strategy and methods, achieved a slightly lower accuracy (88.51 %). In contrast, models using CART with a by-grid sampling strategy and linear support vector machine with the same sampling strategy demonstrated even lower accuracy (76.84 %). Vulnerable RRLs identified through majority voting across multiple models were concentrated in the northeastern, midwestern, and southwestern regions of Fang County. Variable importance analysis revealed that parcel area, terrain, and accessibility exerted the greatest influence on sustainable RRL use. This study proposes a new protocol for predicting RRL use; its findings, along with future rural studies, can contribute to optimizing rural lands.
扫码关注我们
求助内容：
应助结果提醒方式：
京公网安备 11010802042870号