研究以河北省张家口市为研究区，针对较大范围农村居住用地闲置风险识别在复杂地形与城乡混杂背景下存在的漏检率高、背景干扰强及利用状态判别不足等问题，提出融合地形约束与多源时序活力特征的深度学习与机器学习协同识别方法，通过在U-Net模型中融合DEM地形通道与注意力机制，实现复杂地形条件下农村居住用地的像元级提取后，结合建成区边界与夜间灯光特征去除城市像元，构建农村居住用地掩膜，进而融合SDGSAT-1微光夜间灯光与Sentinel-2多时相NDVI时序特征，构建表征人类活动衰减与植被增强的多源协同特征，并利用机器学习模型实现农村居住用地闲置状态的后验概率估计与空间风险分异表达。结果表明：改进U-Net模型在复杂地形背景下具有更高的居住用地分割精度，IoU达到0.8994；在闲置风险识别阶段，轻量梯度提升机（LightGBM）模型综合性能最优（AUC=0.874，F1=0.826），空间自相关分析显示闲置概率像元呈显著正向空间集聚特征（Global Moran’s I=0.842）。研究基于“农村居住用地闲置风险在遥感时序上表现为夜间灯光活力衰减与庭院植被增强并存的复合信号”特征，所构建的闲置风险识别方法，实现了区域尺度农村居住用地闲置的像元级识别与概率化风险表达，可为农村低效建设用地监测与土地整治优先区划定提供参考依据。

Objective: Large-scale identification of idle rural residential land remains challenging in mountainous regions because of complex topographic conditions, fragmented spatial patterns, and heterogeneous land-use states. This study aims to develop a multi-source remote sensing framework for regional-scale identification of idle risk in rural residential land by integrating terrain-constrained spatial extraction with temporal indicators of human activity and vegetation dynamics. Method: Zhangjiakou, Hebei Province, China, was selected as the study area. An improved U-Net semantic segmentation model was developed by incorporating a Digital Elevation Model (DEM) channel and an attention mechanism into the conventional U-Net architecture to enhance the extraction of rural residential land under complex terrain. Sentinel-2 imagery and DEM data were jointly incorporated to strengthen the representation of spectral, spatial, and topographic information. Urban built-up area boundaries were subsequently introduced to exclude urban pixels and construct a rural residential land mask. Multi-source temporal remote sensing features were then derived from SDGSAT-1 low-light nighttime imagery and multi-temporal Sentinel-2 imagery. Nighttime light statistics were used to characterize human activity intensity and temporal variability, whereas Normalized Difference Vegetation Index (NDVI) statistics were employed to characterize vegetation conditions and spatial heterogeneity. Correlation analysis was performed to reduce feature redundancy, and ADASYN combined with SMOTE-Tomek was employed to alleviate class imbalance. Random Forest, XGBoost, and Light Gradient Boosting Machine (LightGBM) were comparatively evaluated, and the optimal model was used to estimate the posterior probability of rural residential land being idle. The resulting probabilities were spatially smoothed and classified into five risk levels, followed by independent stratified validation. Result: The improved U-Net achieved an Intersection over Union (IoU) of 0.8994 and an F1-score of 0.9470, while object-level evaluation yielded a patch integrity of 0.91, a relative area error of 0.11%, and an average boundary offset of 4.8 m. Among the machine learning models, LightGBM achieved the best overall performance, with an accuracy of 0.813, an area under the receiver operating characteristic curve (AUC) of 0.874, a recall of 0.867, a precision of 0.789, and an F1-score of 0.826. Feature importance analysis indicated that NDVI and nighttime light features provided complementary information for idle-risk identification. The integrated feature set achieved an AUC of 0.874, compared with 0.813 and 0.843 for nighttime light-only and vegetation-only features, respectively. Validation based on 500 independent stratified samples yielded an accuracy of 0.783, a precision of 0.745, a recall of 0.830, and an F1-score of 0.786. The proportion of samples exhibiting clear idle characteristics increased progressively from 18% in the low-risk class to 82% in the high-risk class. Global Moran’s I increased from 0.842 before spatial smoothing to 0.927 after smoothing, indicating significant positive spatial clustering of the estimated idle probabilities. Conclusion: The proposed framework integrates terrain-constrained deep learning, multi-source temporal remote sensing, and machine learning to achieve regional-scale identification and probabilistic characterization of idle rural residential land. The synergistic use of nighttime light and vegetation features effectively captures complementary signals associated with the attenuation of human activity and the enhancement of vegetation cover. The posterior probability provides a continuous spatial representation of idle status and reveals distinct spatial differentiation across Zhangjiakou. Relatively high idle-risk areas are mainly concentrated in the Bashang Plateau and some mountainous areas, whereas areas with stronger urban influence and better transportation accessibility generally exhibit lower risk. The proposed framework provides an effective remote sensing-based approach for monitoring inefficient rural residential land and supporting spatially differentiated land-use management and rural land consolidation.