Automatic Change Detection of Planetary Striped Landform on Mars Surface

Fengqi Zhang, Weifeng Hao, Mao Ye, Zhigang Tu, Fei LiIEEE JSTARS 2024 · DOI

TL;DR

Detecting surface changes on Mars is notoriously difficult due to poor image registration and shifting lighting conditions. Researchers have developed SLCD-Net, a lightweight Siamese deep learning model that uses multi-level feature fusion and spatial attention to accurately identify "striped" landform changes (like Dark Slope Streaks) while remaining robust to noise and misalignment.

What Problem Does This Paper Address?

In planetary science, "change detection" (CD) is the process of comparing images of the same location taken at different times to spot new features—like new impact craters or shifting sand dunes. On Earth, we have high-precision GPS and stable orbits, making it easy to align (register) images perfectly.

Mars is a different story. Satellite images often have significant spatial dislocation, and differences in sun angle or sensor noise create "pseudo changes"—visual differences that look like geological activity but are actually just lighting tricks. Existing automated methods often struggle with these artifacts or require heavy manual preprocessing. The authors specifically target striped landforms, such as Dark Slope Streaks (DSS) and Recurring Slope Lineae (RSL), which are critical for understanding Martian geomorphology and the potential presence of liquid water.

Methods

The authors introduce SLCD-Net (Striped Landform Change Detection Network), which moves away from traditional "early fusion" (stacking images before processing) to a more sophisticated Siamese architecture.

Key technical components include:

  • Dual-Stream SE-ResNet Backbone: The model processes the "before" and "after" images through identical parallel branches. It uses Squeeze-and-Excitation (SE) blocks within a ResNet framework to focus on important feature channels while keeping the model lightweight.
  • Multi-level Complementary Fusion: Instead of just comparing the final outputs, SLCD-Net fuses features at multiple depths. This allows the network to perform "mutual learning," where one branch uses information from the other to filter out noise and highlight actual temporal changes.
  • Spatial Attention Module: After fusion, this module helps the network focus on the specific spatial regions where changes are most likely to occur, further suppressing "pseudo changes" caused by misregistration.
  • Sliding Window Inference: While the model is trained on image patches, it can generate pixel-wise change maps by sliding a window across larger images, providing a more intuitive visualization of landform evolution.

Datasets

The study utilizes several key datasets to prove the model's mettle:

  1. Mars HiRISE RSL Dataset: A standard dataset for recurring slope lineae.
  2. New Mars HiRISE DSS Dataset: The authors constructed and released a new, more diverse dataset specifically for Dark Slope Streaks, featuring 396 test pairs and over 2,800 pairs for transfer learning.
  3. Cross-Planetary Generalization Sets: To test robustness, they evaluated the model on lunar impact data (LROC/CTX) and Earth-based desert imagery (PlanetScope).

Key Insights and Ideas

  • Siamese is Superior for Unregistered Data: The authors found that processing images separately before fusion prevents the "noise expansion" that happens when you stack misaligned images at the input layer.
  • The "Aha" of Complementary Fusion: By allowing the two branches to "talk" to each other at every level of the hierarchy, the model learns to ignore global lighting shifts (which affect the whole patch) and focus on local structural changes (the landform itself).
  • Lightweight Matters: Despite its complexity, SLCD-Net is designed to be efficient. It has a smaller memory footprint (~81MB) and faster training times than standard ResNet-50 based models, making it a candidate for future on-board processing.

Key Results

SLCD-Net significantly outperformed the previous state-of-the-art (SOTA) method, CATL:

  • Accuracy: Achieved 98.22% accuracy on the HiRISE RSL test set.
  • Robustness: On the more challenging and diverse DSS test set, it reached 73.23% accuracy, a major leap over baseline models which struggled to break 60%.
  • Generalization: The model successfully detected changes on the Moon and Earth without being specifically trained on those environments, proving its "feature-level" understanding of landform changes.
  • Visual Clarity: Qualitative results show that SLCD-Net produces much cleaner change maps with fewer false positives in areas of high shadow or slight misalignment compared to other methods.

Limitations and Future Work

While SLCD-Net is a major step forward, the authors note a few hurdles:

  • Patch-to-Pixel Bottleneck: The current pixel-wise detection relies on a sliding window approach, which is computationally expensive for massive planetary-scale images.
  • End-to-End Pixel Detection: Future work aims to develop a truly end-to-end pixel-wise network that doesn't require the sliding window middle-man.
  • Data Scarcity: There is still a dire need for more high-quality, pixel-labeled planetary datasets to move beyond "change/no-change" classification into more detailed semantic segmentation of Martian features.

Summary updated

BibTeX
@article{zhang2024automatic,
  author = {Fengqi Zhang and Weifeng Hao and Mao Ye and Zhigang Tu and Fei Li},
  title = {{Automatic Change Detection of Planetary Striped Landform on Mars Surface}},
  journal = {IEEE J. Sel. Top. Appl. Earth Obs. Remote. Sens.},
  year = {2024},
  doi = {10.1109/jstars.2024.3371015},
}
Figure