Joint Grid-Based Attention and Multilevel Feature Fusion for Landslide Recognition

Abstract

Landslide recognition (LR) is a fundamental task for disaster prevention and control. Convolutional neural networks (CNNs) and transformer architectures have been widely used for extracting landslide information. However, CNNs cannot accurately characterize long-distance dependencies and global information, while the transformer may not be as effective as CNNs in capturing local features and spatial information. To address these limitations, we construct a new LR network based on grid-based attention and multilevel feature fusion (GAMTNet). We complement CNNs by adding a transformer-based structure in a layer-by-layer fashion and improving methods for sequence generation and attention weight calculation. As a result, GAMTNet effectively learns global and local information about landslides across various spatial scales. We evaluated our model using landslide data collected from the southwest region of Jiuzhaigou County, Aba Tibetan, and Qiang Autonomous Prefecture, Sichuan Province, China. The results demonstrate that the proposed GAMTNet model achieves an F 1-score of 0.8951, a Kappa coefficient of 0.8807, and an MIoU of 0.8908, indicating its capability for the accurate landslide identification and its potential application in LR tasks.