GDSNet: A gated dual-stream convolutional neural network for automatic recognition of coseismic landslides

Abstract

Automatic recognition of numerous coseismic landslides after a violent earthquake is crucial for emergency rescue and post-disaster reconstruction. Currently, deep learning techniques have achieved state-of-the-art performance in coseismic landslide recognition. However, Convolutional Neural Networks (CNNs) often lose detailed information during downsampling and cannot adequately learn changeable shapes, colors, and sizes of coseismic landslides. In addition, complicated backgrounds, e.g., bare slopes and dry riverbeds, are easily misidentified as coseismic landslides. Focusing on the above difficulties, this work proposes a Gated Dual-Stream Convolutional Neural Network (GDSNet) for landslide recognition, which contains two branches and a feature fusion module. One branch, called as CPSConv, can extract detailed landslide information of shapes, sizes, spectra, and textures and guarantee the accurate identification of small landslides. Another branch utilizes a gated convolution strategy to adjust feature weights and importance and to enhance landslide features and suppress background features. The feature aggregation and fusion module fuses the features from two branches to effectively improve the recognition accuracy of coseismic landslides. The GDSNet is applied to the landslide recognition of four earthquakes by model training and testing. In test dataset, compared with 9 state-of-the-art models, the mIoU, F1, Kappa coefficient values by the GDSNet are improved by at least 12.30%, 8.12%, and 16.25%, respectively. The recognition accuracy of small landslides is improved by 1.08%-37.68% than the other 9 deep learning models.

Keywords

• Coseismic landslide
• Automatic recognition
• Multi-spectral image
• Convolutional neural network