Using lightweight method to detect landslide from satellite imagery

Abstract

Accurate, rapid, and automated landslide detection is crucial for early warning, emergency management, and landslide mechanism analysis. Increasingly general-purpose detection models are being deployed for these complex and dynamic tasks involving features that are difficult to characterize. However, these models are computationally expensive and memory-hungry, while the accuracy and detection efficiency remain wanting. To address the above problems, this paper proposes an end-to-end model with high-precision and lightweight design for integrated landslide detection and segmentation. Here, we customized the backbone utilizing the advanced Efficient MOdel (EMO), and further used the linear cheap operation from GhostNet to reduce computational complexity. As a result, the total parameters of our models were reduced by up to 48.13%, compared to the baseline. Building on this, we employed a dynamic detection head with multiple attention mechanisms, and proposed a lightweight attention enhancement module for strengthened multi-scale feature extraction and fusion. The results demonstrate that our model outperforms the baseline on all metrics, achieving an outstanding F1 score of 96.75%.

Keywords

• YOLO
• Landslide detection
• Deep learning
• Efficient model
• Lightweight