Validated visual features of Multi-Perspective imagery with Explainable Machine learning for detecting rural vacant courtyards in North China

• Integrated horizontal and overhead imagery to detect intact vacant courtyards. • Select four key visual features from 390 models for optimized detection. • Apply SHAP to reveal feature asymmetry and spatial heterogeneity. As China’s rural population continues to decline and urbanisation accelerates, the number of vacant courtyards (VCs) has steadily increased. Existing methods for detecting VCs in China face significant challenges regarding data availability, spatial scale, resolution, and reliability, which hinder accurate assessment. This study establishes an innovative approach that evaluates courtyard utilisation status by integrating visual features from horizontal and overhead imagery, distinguishing it from traditional methods that rely solely on remote sensing textures or geometric information. We constructed a systematic set of visual features and employed an interpretable machine learning (XGBoost model) to detect courtyard utilisation status. We identified an optimal model comprising four primary features: from an overhead perspective, plants condition, enclosing wall edge cleanliness, and presence of solar water heaters; and from a horizontal perspective, door and window condition. The research further reveals the asymmetric relationship between these features and courtyard utilisation status, and the spatial heterogeneity underlying this relationship. The model significantly outperforms existing research, achieving an F1 score of 0.834 on the test set while maintaining high accuracy on the external validation dataset. This demonstrates the considerable advantages and potential of this relatively low-cost approach for rapidly detecting VC, providing theoretical support and reliable technical backing for implementing courtyard utilisation monitoring based on image data and sustainable development goals across a broad range of rural areas in the future.