MRIFE: A Mask-Recovering and Interactive-Feature-Enhancing Semantic Segmentation Network For Relic Landslide Detection

Juefei He,Yuexing Peng,Wei Li,Junchuan Yu,Daqing Ge,Wei Xiang
2024-11-26
Abstract:Relic landslide, formed over a long period, possess the potential for reactivation, making them a hazardous geological phenomenon. While reliable relic landslide detection benefits the effective monitoring and prevention of landslide disaster, semantic segmentation using high-resolution remote sensing images for relic landslides faces many challenges, including the object visual blur problem, due to the changes of appearance caused by prolonged natural evolution and human activities, and the small-sized dataset problem, due to difficulty in recognizing and labelling the samples. To address these challenges, a semantic segmentation model, termed mask-recovering and interactive-feature-enhancing (MRIFE), is proposed for more efficient feature extraction and separation. Specifically, a contrastive learning and mask reconstruction method with locally significant feature enhancement is proposed to improve the ability to distinguish between the target and background and represent landslide semantic features. Meanwhile, a dual-branch interactive feature enhancement architecture is used to enrich the extracted features and address the issue of visual ambiguity. Self-distillation learning is introduced to leverage the feature diversity both within and between samples for contrastive learning, improving sample utilization, accelerating model convergence, and effectively addressing the problem of the small-sized dataset. The proposed MRIFE is evaluated on a real relic landslide dataset, and experimental results show that it greatly improves the performance of relic landslide detection. For the semantic segmentation task, compared to the baseline, the precision increases from 0.4226 to 0.5347, the mean intersection over union (IoU) increases from 0.6405 to 0.6680, the landslide IoU increases from 0.3381 to 0.3934, and the F1-score increases from 0.5054 to 0.5646.
Computer Vision and Pattern Recognition
What problem does this paper attempt to address?
### What problems does this paper attempt to solve? This paper aims to solve two main challenges in relic landslide detection: 1. **Visual Blur Problem**: - Due to long - term natural evolution and human activities, the surface features of relic landslides are very similar to those of the surrounding environment, which makes the optical features of the landslide area in High - Resolution Remote Sensing Images (HRSI) become blurred. This makes it difficult for models such as CNN and Transformer to distinguish the target from the background. 2. **Small - sized Dataset Problem**: - It is very difficult to construct a relic landslide dataset because accurately identifying landslides and marking boundaries requires a great deal of time and technical input. A small - sized dataset cannot fully support powerful model training and is prone to over - fitting, thus placing higher requirements on the model's learning and generalization abilities. To solve these problems, the authors propose a semantic segmentation network named **MRIFE (Mask - Recovering and Interactive - Feature - Enhancing)**. This model improves the ability of feature extraction and separation through the following methods: - **Contrastive learning and mask reconstruction methods**: By enhancing local salient features, the ability to distinguish the target from the background and represent the semantic features of the landslides is improved. - **Two - branch interactive feature - enhancing architecture**: Enrich the extracted features to solve the visual blur problem. - **Self - distillation learning**: Use the feature diversity within and between samples for contrastive learning, improve sample utilization, accelerate model convergence, and effectively solve the small - sized dataset problem. Through these improvements, the experimental results of the MRIFE model on the real relic landslide dataset show that it greatly improves the performance of relic landslide detection. Specifically, in the semantic segmentation task, compared with the baseline model, the accuracy is improved from 0.4226 to 0.5347, the mean IoU (mean Intersection over Union) is improved from 0.6405 to 0.6680, the landslide IoU is improved from 0.3381 to 0.3934, and the F1 score is improved from 0.5054 to 0.5646. ### Summary The core problem of the paper is to design a new semantic segmentation network MRIFE to deal with the visual blur and small - sized dataset challenges in relic landslide detection, thereby improving the accuracy and efficiency of detection.