# A comparative study of loss functions and attention mechanisms in landslide semantic segmentation using U-Net

## TL;DR

This paper compares loss functions and attention mechanisms in U-Net models to improve landslide detection using satellite and elevation data.

## Contribution

The novel contribution is the integration of attention mechanisms and multi-modal data to enhance U-Net performance for landslide segmentation.

## Key findings

- The model with attention achieved an mIoU of 0.76, F1 score of 0.74, and accuracy of 0.94.
- Attention mechanisms improve segmentation by focusing on critical regions.
- Multi-modal data and diverse loss functions enhance landslide detection performance.

## Abstract

Advancement in landslide detection can be largely attributed to the introduction of deep learning, particularly semantic segmentation. Susceptible regions can be identified using satellite imagery and Digital Elevation Model (DEM) data. This study explores multi-modal data to improve the identification and detection of landslides. The U-Net model serves as a baseline that is further enhanced by the introduction of an attention mechanism that refines pixel-level predictions. Evaluation of various loss functions resulted in increased performance optimization. The Bijie landslide dataset, featuring high-resolution satellite images, DEM data, and ground truth masks, was used for training and evaluation. Precision, recall, F1 score, accuracy, mean intersection over union (mIoU), and Area Under the Curve (AUC) were metrics used to evaluate the performance. The model incorporating the attention mechanism achieved the highest mIoU of 0.76, F1 score of 0.74, and accuracy of 0.94, surpassing the base model. Attention mechanisms concentrate on critical regions and thus improve feature extraction by enhancing segmentation precision. The integration of multi-modal data and diverse loss functions contributes to better landslide detection.