Identification of rural courtyards’ utilization status using deep learning and machine learning methods on unmanned aerial vehicle images in north China

Abstract

The issue of unoccupied or abandoned homesteads (courtyards) in China emerges given the increasing aging population, rapid urbanization and massive rural-urban migration. From the aspect of rural vitalization, land-use planning, and policy making, determining the number of unoccupied courtyards is important. Field and questionnaire-based surveys were currently the main approaches, but these traditional methods were often expensive and laborious. A new workflow is explored using deep learning and machine learning algorithms on unmanned aerial vehicle (UAV) images. Initially, features of the built environment were extracted using deep learning to evaluate the courtyard management, including extracting complete or collapsed farmhouses by Alexnet, detecting solar water heaters by YOLOv5s, calculating green looking ratio (GLR) by FCN. Their precisions exceeded 98%. Then, seven machine learning algorithms (Adaboost, binomial logistic regression, neural network, random forest, support vector machine, decision trees, and XGBoost algorithms) were applied to identify the rural courtyards’ utilization status. The Adaboost algorithm showed the best performance with the comprehensive consideration of most metrics (Accuracy: 0.933, Precision: 0.932, Recall: 0.984, F1-score: 0.957). Results showed that identifying the courtyards’ utilization statuses based on the courtyard built environment is feasible. It is transferable and cost-effective for large-scale village surveys, and may contribute to the intensive and sustainable approach to rural land use.