A cross-spatial network based on efficient multi-scale attention for landslide recognition.

Landslide disasters are one of the frequently occurring geological hazards, posing a significant threat to human life and property safety. Swift and accurate identification of landslide areas is crucial for disaster prevention and mitigation. Current object detection algorithms have limitations in the localization and recognition of landslide areas. To address this issue, this paper proposes a cross-spatial network based on efficient multi-scale attention (EMA-Net) landslide recognition model. The proposed EMA-Net model incorporates the efficient multi-scale attention (EMA) for cross space learning, enhancing the model’s focus on landslide areas. Additionally, by employing convolution with absolute positioning (CoordConv), the positional information of features is retained to enhance the capability of multiscale feature extraction. The utilization of the SCYLLA-IoU (SIoU )loss function enhances regression learning ability for model prediction borders, thereby improving the efficiency and accuracy of the model. To assess its performance, EMA-Net is evaluated against other models, including Yolov5-\documentclass[12pt]{minimal}\usepackage{amsmath}\usepackage{wasysym}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{amsbsy}\usepackage{mathrsfs}\usepackage{upgreek}\setlength{\oddsidemargin}{-69pt}\begin{document}$$-$$\end{document}5.0, Yolov5-\documentclass[12pt]{minimal}\usepackage{amsmath}\usepackage{wasysym}\usepackage{amsfonts}\usepackage{amssymb}\usepackage{amsbsy}\usepackage{mathrsfs}\usepackage{upgreek}\setlength{\oddsidemargin}{-69pt}\begin{document}$$-$$\end{document}6.1, Yolov7, and Faster-R-CNN. The evaluation demonstrates that the proposed EMA-Net achieves a precision of 0.980, recall of 0.982, and mAP of 0.717, exhibiting clear improvement over the compared networks. Furthermore, through visualized analysis, the proposed network is capable of effectively identifying landslides within a smaller range. Comparative analysis of the aforementioned experiments validates the superiority of the proposed network. [ABSTRACT FROM AUTHOR]