Research on Recognition of Check Dams Considering Suitable Construction Areas and Microtopography Standard Deviation Based on Faster R-CNN

Accurate spatial identification of check dams is a key prerequisite for evaluating soil and water conservation benefits and optimizing dam system planning on the Loess Plateau. Current deep learning models face severe misclassification and omission issues under complex terrain due to the scarcity of check dam samples and the lack of prior geographic knowledge. This study proposes a recognition method based on Faster R-CNN, constrained by suitable areas and microtopography. The Xiliugou watershed in Inner Mongolia was selected as the study area. Based on Google Earth imagery and field survey data, a check dam sample dataset was constructed, integrating the morphological features of “linear dam body with a trapezoidal slope.” Using the construction suitable area constraints defined by the Technical Specifications for Check Dams and microtopography standard deviation (δ) derived from DEM as dual spatial filtering mechanisms, these were deeply embedded into the Faster R-CNN model to limit the search space and enhance geographic plausibility. Experimental results show that the constrained Faster R-CNN model achieved a precision and recall of 92.86% and 96.89%, compared with the accuracy rate of only deep learning model recognition (60.61%), which significantly increased by 32.25%, indicating that geographical constraints have an enhancing effect. Using this method, a total of 191 embankment dams were identified in the Xiliugou Basin. New 30 unrecorded embankment dams (21 small dams and 9 micro-dams) were discovered. The model’s good generalization ability was verified in the Han Tiechuan geographical isolation area, which contained 153 embankment dam samples, with an accuracy rate of 72.94%. Spatial analysis further revealed the “successive interception along tributaries” distribution pattern and strong spatial aggregation characteristics (box dimension D ≈ 0.36) of check dams in the Xiliugou watershed. This study confirms the critical role of suitable area and microtopography constraints in improving the accuracy and reliability of deep learning models and provides a transferable technical paradigm for automated, high-precision surveys of regional soil and water conservation projects.