Automatic Identification Model for Landslide Disaster Using Remote Sensing Images Based on Improved Multiresunet

Abstract

The current imagery of landslides presents multiple challenges, including considerable scale variations among landslides, similarities in spectral characteristics to those of bare ground surfaces, and irregular edges. Despite significant progress in semantic segmentation achieved by convolutional neural networks (CNNs), the local sensory field of CNNs poses difficulties in differentiating between landslides and bare surfaces. As a result, efficient and accurate landslide extraction remains a challenging issue within the global research community. To address this problem, a novel automatic landslide hazard remote sensing image identification model, MultiResUNet-BFDC, is proposed. Initially, null convolutions with various null rates are introduced to replace some standard convolutions, thereby expanding the sensory field without increasing the parameter count and making the model more suitable for multi-scale landslide segmentation. Additionally, the Canny operator is employed to design a lightweight boundary-focused attention (BFA) mechanism, enhancing the model’s ability to emphasize landslide edge features. Furthermore, a new hybrid loss function, adaptive focal and Dice loss (AFD loss), is introduced through the adaptive AdaLoss algorithm by combining focal loss and Dice loss, improving the model’s ability to handle unbalanced samples. The experimental results indicate that the MultiResUNet-BFDC model demonstrates enhanced performance in landslide edge detail segmentation, with fewer misidentifications and omissions and superior functionality of the BFA attentional mechanism compared with other attentional mechanisms.

Keywords

• Landslide identification
• deep learning
• CNN
• attention mechanism
• remote sensing imagery