Abstract: Soil erosion presents a substantial threat to ecological stability and agricultural sustainability across China's Loess Plateau, where check dams constitute vital infrastructure for comprehensive soil and water conservation initiatives. The precise delineation of check dam spatial extents coupled with accurate point positioning proves fundamentally indispensable for optimizing watershed management strategies and conservation planning frameworks. The Malian River Basin on the Loess Plateau is a typical representative of severe soil erosion and large-scale construction of check dams. The heterogeneous distribution pattern of dams and diverse land cover characteristics provide a large number of samples for the extraction of check dams. However, traditional extraction methods face substantial challenges including fuzzy boundaries in imagery and spectral confusion with terraced fields, while the limitations of manual interpretation hinder precise identification of check dam areas and points. To address these limitations, we propose a semantic segmentation model for check dam extraction (UNet-OBIA) based on object-based image analysis (OBIA) and the U-Net model. By constructing a multi-source remote sensing image dataset of check dams (RGB+DEM) that includes complex background information from four types of features: terraced fields , vegetation, bare land, and buildings, the optimal segmentation scale parameter is selected based on a scale parameter estimation method. The U-Net network performs an initial extraction of potential check dam areas. Subsequently, the object-based image analysis method is applied for multi-scale segmentation of the check dam boundaries, and the majority voting method is employed for feature fusion. Then the check dam areas are optimized by combining terrain parameters such as elevation. Finally, check dam body points are precisely located via hydrological analysis. We rigorously validated our method (UNet-OBIA) in three representative areas within the Malian River Basin on the Loess Plateau and comparative analyses against FCN, DeepLabv3+, PSPNet, and UNet++ models (all integrated with OBIA) were conducted to further validate the accuracy of our approach. The results showed that UNet-OBIA achieved significant performance metrics: Precision (92.47%), Recall (87.16%), OA (95.82%), F₁ (89.73%), and mIoU (81.41%). Compared with the baseline U-Net model, these metrics were improved by 2.83, 1.27, 1.25, 2.01, and 3.82 percentage points, respectively, demonstrating the optimal overall performance of UNet-OBIA. Experimental results demonstrate that integrating DEM topographic data with remote sensing imagery to construct a four-band dataset, combined with OBIA post-processing methods, leads to significant improvements in both the completeness of check dam boundary extraction and the accuracy of dam shape recognition in complex backgrounds. Furthermore, the hydrological analysis achieved a check dam point recognition accuracy of 80.3% and a completeness rate of 86.9%, all of which indicate the significant robustness and generalization capability of our proposed approach for check dam extraction in complex Loess Plateau terrain. This accurate and efficient extraction method provides vital data and methodological framework for optimizing the spatial layout of check dam systems, enhancing soil erosion control efforts, and improving hydrological simulations on the Loess Plateau watersheds.