C-PLES: Contextual Progressive Layer Expansion with Self-attention for Multi-class Landslide Segmentation on Mars using Multimodal Satellite Imagery

Document Type

Conference Proceeding

Publication Date

8-14-2023

Department

Department of Applied Computing

Abstract

Landslide segmentation on Earth has been a challenging computer vision task, in which the lack of annotated data or limitation on computational resources has been a major obstacle in the development of accurate and scalable artificial intelligence-based models. However, the accelerated progress in deep learning techniques and the availability of data-sharing initiatives have enabled significant achievements in landslide segmentation on Earth. With the current capabilities in technology and data availability, replicating a similar task on other planets, such as Mars, does not seem an impossible task anymore. In this research, we present C-PLES (Contextual Progressive Layer Expansion with Self-attention), a deep learning architecture for multi-class landslide segmentation in the Valles Marineris (VM) on Mars. Even though the challenges could be different from on-Earth landslide segmentation, due to the nature of the environment and data characteristics, the outcomes of this research lead to a better understanding of the geology and terrain of the planet, in addition, to providing valuable insights regarding the importance of image modality for this task. The proposed architecture combines the merits of the progressive neuron expansion with attention mechanisms in an encoder-decoder-based framework, delivering competitive performance in comparison with state-of-the-art deep learning architectures for landslide segmentation. In addition to the new multi-class segmentation architecture, we introduce a new multi-modal multi-class Martian landslide segmentation dataset for the first time. The dataset will be available at https://github.com/MAIN-Lab/C-PLES

Publication Title

IEEE Computer Society Conference on Computer Vision and Pattern Recognition Workshops

ISBN

9798350302493

Share

COinS