A deep learning-driven change detection model for martian surface segmentation

This is the code for the paper: Detection of Detached Ice-fragments at Martian Polar Scarps Using a Convolutional Neural Network.

Introduction

This is a deep learning-driven change detection method for extracting the detached ice-fragments at martian polar scarps. The customized model is using:

• ResU-Net architecture, which combines ResNet and U-Net
• Siamese network architecture
• An augmented attention module
• A hybrid loss function: dice loss and focal loss

Datasets

A set of example is provided, but the entire dataset is currently not available to the public. However, users can easily apply our model to various training datasets available online for testing. The training dataset consists of three files:

• T1: the pre-detach image, can be grayscale or RGB
• T2: the post-detach image, can be grayscale or RGB
• Mask: class 1 represents the detached ice-fragments, while class 0 represents the background including unchanged areas and the changed shadows

References

If you find our work useful in your research, please consider citing:

@ARTICLE{10024321,
  author={Su, Shu and Fanara, Lida and Xiao, Haifeng and Hauber, Ernst and Oberst, Jürgen},
  journal={IEEE Journal of Selected Topics in Applied Earth Observations and Remote Sensing}, 
  title={Detection of Detached Ice-fragments at Martian Polar Scarps Using a Convolutional Neural Network}, 
  year={2023},
  volume={16},
  pages={1728-1739},
  doi={10.1109/JSTARS.2023.3238968}}