【摘要】 准确获取淤地坝数目、位置及空间分布等信息是科学分析淤地坝减蚀和拦沙作用及其未来建设规划的重要基础。本研究以黄土高原延河流域淤地坝为研究对象，以高分二号影像以及Google Earth影像作为数据源分别制作坝体和坝地标签，选取深度学习目标检测(Faster R-CNN、YOLO v3、Cascade R-CNN和YOLOX)及语义分割(FCN、U-Net、PSPNet和DeepLab v3+)算法分别对淤地坝坝体及坝地进行识别提取，并结合地理分析方法实现流域尺度沟道缓冲区框定，在缓冲区内进行淤地坝提取。结果表明：在淤地坝坝体识别中，YOLOX模型表现最佳，当交并比(IoU)阈值为0.50～0.95时，平均精度(AP)达69.4%。基于该模型对延河流域影像进行识别，获得1 440个检测框，模型准确率为75.8%,召回率达85.7%。在淤地坝坝地提取中，4个模型表现均较好，平均交并比、精确率、召回率及F1分数均达0.85、90%、85%和88%以上。采用多数投票法结合4个模型预测结果分别提取了流域内92.0 km~2生产运行期坝地和10.6 km~2蓄水拦泥期坝地。通过引入地理分析方法，坝体识别准确率从75.8%提升至84.5%,识别提取效率提升1倍以上。该研究为黄土高原及其他地区的淤地坝快速准确识别提供了重要依据，淤地坝监测结果为流域综合治理和水土保持规划提供了科学借鉴和理论指导。更多还原

【Abstract】 Accurately obtaining information on the number, location, and spatial distribution of check dams is fundamental for scientifically analyzing their erosion reduction and sediment trapping effects, as well as for future construction planning. Focusing on the Yanhe River Basin in the Loess Plateau, utilizing high-resolution imagery from GF-2 and Google Earth as data sources, deep learning object detection algorithms(Faster R-CNN, YOLO v3, Cascade R-CNN, and YOLOX) and semantic segmentation algorithms(FCN, U-Net, PSPNet, and DeepLab v3+) were employed to identify and extract the check dam bodies and dam lands. Geographic analysis methods and comprehensive interpretation were used to optimize the extraction results of the check dams. The results were as follows: in the identification of check dam bodies, the YOLOX model performed the best, achieving an average precision(AP) of 69.4% at an intersection of union(IoU) threshold of 0.50～0.95. Using this model to identify images of the Yanhe River Basin, totally 1 440 detection boxes were obtained, with an accuracy of 75.8% and a recall rate of 85.7%. In the extraction of dam lands, all four models performed well, with an average IoU, precision, recall, and F1 score all exceeded 0.85, 90%, 85%, and 88%, respectively. The majority voting method was used to combine the predictions of the four models, resulting in the extraction of 92.0 km~2 of dam lands in the production and operation period and 10.6 km~2 in the water storage and sediment trapping period. By incorporating geographic analysis methods, the accuracy of check dam body identification was increased from 75.8% to 84.5%, with the efficiency of identification and extraction doubling. These results indicated that the proposed method can be rapidly and accurately applied to the identification of check dams in the Loess Plateau, which was crucial for analyzing their impact on the ecological environment and for comprehensive management of small watersheds.更多还原