Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

A Deep-Learning Framework for Land-Sliding Classification from Remote Sensing Image

Hieu Tang, Truong Vo, Dong Pham, Toan Nguyen, Lam Pham, Truong Nguyen

TL;DR

This work tackles landslide detection from multi-source remote sensing by addressing severe class imbalance with offline SMOTE using structural similarity, and by enriching training data online with CutMix/Mixup and other augmentations. It combines a pre-trained EfficientNetV2-Large backbone with a post-processing SVM classifier to improve generalization on imbalanced data, achieving a notable F1 of 0.8938 on the Zindi challenge dataset. Ablation studies and explainability analyses reveal that augmentations and the SVM head improve decision boundary quality, and that SAR bands contribute more to detection than optical bands in challenging conditions. The approach offers a practical, scalable pipeline for reliable landslide mapping across diverse environments using multi-band satellite imagery.

Abstract

The use of satellite imagery combined with deep learning to support automatic landslide detection is becoming increasingly widespread. However, selecting an appropriate deep learning architecture to optimize performance while avoiding overfitting remains a critical challenge. To address these issues, we propose a deep-learning based framework for landslide detection from remote sensing image in this paper. The proposed framework presents an effective combination of the online an offline data augmentation to tackle the imbalanced data, a backbone EfficientNet\_Large deep learning model for extracting robust embedding features, and a post-processing SVM classifier to balance and enhance the classification performance. The proposed model achieved an F1-score of 0.8938 on the public test set of the Zindi challenge.

A Deep-Learning Framework for Land-Sliding Classification from Remote Sensing Image

TL;DR

This work tackles landslide detection from multi-source remote sensing by addressing severe class imbalance with offline SMOTE using structural similarity, and by enriching training data online with CutMix/Mixup and other augmentations. It combines a pre-trained EfficientNetV2-Large backbone with a post-processing SVM classifier to improve generalization on imbalanced data, achieving a notable F1 of 0.8938 on the Zindi challenge dataset. Ablation studies and explainability analyses reveal that augmentations and the SVM head improve decision boundary quality, and that SAR bands contribute more to detection than optical bands in challenging conditions. The approach offers a practical, scalable pipeline for reliable landslide mapping across diverse environments using multi-band satellite imagery.

Abstract

The use of satellite imagery combined with deep learning to support automatic landslide detection is becoming increasingly widespread. However, selecting an appropriate deep learning architecture to optimize performance while avoiding overfitting remains a critical challenge. To address these issues, we propose a deep-learning based framework for landslide detection from remote sensing image in this paper. The proposed framework presents an effective combination of the online an offline data augmentation to tackle the imbalanced data, a backbone EfficientNet\_Large deep learning model for extracting robust embedding features, and a post-processing SVM classifier to balance and enhance the classification performance. The proposed model achieved an F1-score of 0.8938 on the public test set of the Zindi challenge.

Paper Structure

This paper contains 13 sections, 4 equations, 7 figures, 3 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Background
  3. Synthetic Minority Oversampling Technique (SMOTE):
  4. Cut Mix:
  5. Mixup:
  6. EfficientNetV2-Large:
  7. Support Vector Machine (SVM):
  8. Our proposed deep-learning framework
  9. Experiments and Discussion
  10. Zindi Challenge Dataset
  11. Experimental Settings
  12. Experimental Results and Discussions
  13. Conclusion

Figures (7)

  • Figure 1: Unit blocks in the EfficientNetV2-Large network architecture
  • Figure 2: The high-level architecture of proposed deep-learning framework
  • Figure 3: The training and inference processes of the proposed deep-learning framework
  • Figure 4: (a) Left: Anchor images; Middle: Nearest neighbors; Right: SMOTE-generated samples. (b): Class distribution on the training subset after applying the SMOTE algorithm.
  • Figure 5: Performance and parameter comparison among backbone deep neural network
  • ...and 2 more figures