1.Changjiang River Scientific Research Institute of Changjiang Water Resources Commission,Wuhan 430010,China
2.Faculty of Resources and Environment Science,Hubei University,Wuhan 430062,China
3.Observation and Research Station for Small and Medium-Sized Rivers and Flash Flood Disasters in;Mountains of Southwest China,Ministry of Water Resources,Wuhan 430010,China
Objective As a distinct erosional landform, Benggang is often characterized by internal fragmented vegetation residues and irregular edges extending in digitated or claw-like shapes. Such strong spectral heterogeneity and similarity to the background make it challenging for traditional semantic segmentation models to balance global features and local details. Methods Based on unmanned aerial vehicle (UAV) orthophotos, this study proposed an improved UNet model (CA-UNet) integrated with the coordinate attention (CA) mechanism. By embedding CA modules during the decoding stage, the model enhanced its ability to capture cross-channel key information and long-range spatial dependencies, thereby significantly enhancing its feature perception of the complex morphology and fine-grained boundaries of Benggang. Results This study systematically compared the segmentation performance of CA-UNet with UNet, PSPNet, and DeepLabV3+. The results demonstrated that CA-UNet achieved a precision of 90.45%, recall of 95.17%, mean intersection over union (mIoU) of 79.48%, and an F1-Score of 92.56%, all of which outperformed the comparison models. Furthermore, CA-UNet exhibited superior performance in fine boundary delineation and error control. It effectively reduced patch adhesion caused by over-smoothing and internal hollows resulting from local information loss, thereby generating Benggang patches that better met monitoring requirements. Conclusion This study confirms that the proposed improved model can effectively identify complex geomorphological features of Benggang, providing a valuable reference for the intelligent, high-precision dynamic monitoring of Benggang erosion in the red soil regions of South China.
LIAOK T, LIUY, LIUQ, et al. Distribution characteristics and driving factors of Benggang erosion in Ganzhou City[J].Research of Soil and Water Conservation,2021,28(6):126-130.
LIUH H, QINF, CHENJ, et al. Accuracy of slope collapse data collected by artificial survey[J].Journal of Changjiang River Scientific Research Institute,2015,32(3):117-120.
LIUN, ZHANGH, DENGC L, WUJ, XIONGY. Distribution characteristics of Benggang in the middle and upper reaches of the Han River[J].Tropical Geography,2024,44(3):415-428.
LIUX, WANGS Y, YUANZ L, et al. Application progress of information technology in the study of dilapidated granite[J].Soil and Water Conservation in China,2021(6):55-58.
FANGL Y, YANGC H, SHUD, et al. Landslide detection in complex environments based on dual feature fusion[J].Remote Sensing for Natural Resources, 2025,37(5):91-100.
TIANP, RENY L, CHENY. Research progress on identification and extraction methods of soil and water conservation measures[J].Journal of Soil and Water Conservation,2024,38(5):1-9.
ZHOUX Q, YUH L, WEIY J, et al. Method for monitoring change in Benggang erosion based on oblique aerial images of UAV[J].Transactions of the Chinese Society of Agricultural Engineering,2019,35(21):51-59.
[21]
FUJ, LIUJ, JIANGJ, et al. Scene segmentation with dual relation-aware attention network[J].IEEE Transactions on Neural Networks and Learning Systems,2021,32(6):2547-2560.
[22]
LUOW J, LIY J, URTASUNR, et al. Understanding the effective receptive field in deep convolutional neural networks[CP].2017.
ZHANGJ X, CHENGD B, GUOF, et al. Risk assessment of Benggang in Shangyou County, Jiangxi Province based on matrix discriminant method[J].Bulletin of Soil and Water Conservation,2025,45(1):158-167.
[25]
KAMPFFMEYERM, SALBERGA B, JENSSENR. Semantic segmentation of small objects and modeling of uncertainty in urban remote sensing images using deep convolutional neural networks[C]∥2016 IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW) Nevada, USA: IEEE,2016.
ZHAOT, ZHANGS C, HEX N, et al. Improved DeepLabV3+ model for landslide identification in high-resolution remote sensing images after earthquakes[J].National Remote Sensing Bulletin,2024,28(9):2293-2305.
LIUX L, TANGC, ZHANGD L. Simulated runoff processes on colluvial deposits of Liantanggang Benggang and their water distributions[J].Transactions of the Chinese Society of Agricultural Engineering,2015,31(11):179-185.
WEIJ X, DENGY S, LIAOD L, et al. Spatial variation and influencing factors of soil limiting water content of granite collapsing gullies in southeast Guangxi[J].Acta Pedologica Sinica,2023,60(3):749-761.
ZHANGY T, HUANGD Q, WANGD W, et al. Review on research and application of deep learning-based target detection algorithms[J].Computer Engineering and Applications,2023,59(18):1-13.
[34]
HOUY W, LIUZ Y, ZHANGT, et al. C-UNet: Complement UNet for remote sensing road extraction[J].Sensors,2021,21(6):e2153.
[35]
LONGX D, ZHANGW W, ZHAOB. PSPNet-SLAM: A semantic SLAM detect dynamic object by pyramid scene parsing network[J].IEEE Access,2020,8:214685-214695.
FENGX, ZHANGX L, WANGJ X, et al. Multi crop classification extraction based on improved spatial-coordinate attention UNet[J].Transactions of the Chinese Society of Agricultural Engineering,2023,39(18):132-141.
[41]
HUOY, GANGS, DONGL, et al. An efficient semantic segmentation method for remote-sensing imagery using improved coordinate attention[J].Applied Sciences,2024,14(10):e4075.
LIH J, XIEH, XUW B, et al. Effects of spatial resolution of remote sensing image on the accuracy of extracted slope collapse data[J].Journal of Anhui Agricultural Sciences,2016,44(30):227-229.
[44]
SHENS Y, CHENJ S, ZHANGS Y, et al. Deep fusion of DOM and DSM features for benggang discovery[J].ISPRS International Journal of Geo-Information,2021,10(8):e556.
LIZH J, ZHONGL T, HUANGY H, et al. Monitoring: technology for collapse erosion based on the nap of the object photograph of UAV[J].Transactions of the Chinese Society of Agricultural Engineering,2021,37(8):151-159.