Optimized Martian Dust Displacement Detection Using Explainable Machine Learning

Ana Lomashvili, Kristin Rammelkamp, Olivier Gasnault, Protim Bhattacharjee, Elise Clavé, Christoph H. Egerland, Susanne Schröder, Begüm Demir, Nina L. LanzaWorkshops @ CVPR 2024 · DOI

Introduction

NASA’s Curiosity rover utilizes the ChemCam instrument to perform Laser-Induced Breakdown Spectroscopy (LIBS) on Martian rocks. The high-powered laser creates a plasma spark to determine elemental composition, a process that generates a shockwave often displacing the fine dust covering the target. This dust movement provides critical data on dust thickness, local atmospheric conditions, and geological cycles. However, identifying these events manually across thousands of targets is labor-intensive, and standard computer vision models frequently struggle with the complex, rugged textures of Martian geology.

Methodology

To address these challenges, researchers developed the Dust Displacement Detection (DDD) pipeline. The approach focuses on overcoming geological "noise" through a multi-stage process:

  1. Preprocessing (Texture Removal): Rather than using raw images, the pipeline crops and subtracts the "before" LIBS image from the "after" image. This "difference image" cancels out static rock textures, highlighting only the temporal changes caused by moved dust.
  2. Hybrid Architecture: The model employs the first 10 layers of a pretrained VGG16 network for feature extraction, followed by a Random Forest (RF) classifier with 1,000 trees. This hybrid design prevents overfitting on the relatively small dataset of 400 manually labeled images.
  3. Explainable AI (XAI) Optimization: Using SHAP (Shapley Additive exPlanations) and guided backpropagation, the team identified which feature maps contributed most to the classification. This allowed them to prune the model for potential onboard processing.
The DDD preprocessing pipeline: images are cropped, subtracted to remove static texture, and center-cropped for analysis.

Key Results and Discussion

The optimized DDD pipeline achieved a significant 92% accuracy, outperforming standard SVM and full VGG16 implementations. The XAI analysis revealed that the model was correctly focusing on dust contours rather than image artifacts.

Crucially for planetary exploration, where computational resources are limited, the researchers found that 67% of the feature extraction data could be discarded (reducing the 10th layer from 256 to 85 feature maps) with only a negligible 2% drop in accuracy. This pruning demonstrates the feasibility of running such models on rover hardware for real-time "edge" science.

Performance of the optimized DDD pipeline, showing a 92% overall accuracy and examples of classification errors.

When applied to a larger dataset of 1,925 targets, the model revealed that dust displacement is less frequent on the Gale Crater floor than on the slopes of Mt. Sharp, offering new insights into Martian climate patterns.

Conclusion and Future Work

The DDD pipeline successfully automates a tedious scientific task while maintaining high precision. Future work involves adapting this system for the SuperCam instrument on the Perseverance rover, which provides RGB data that could further enhance detection capabilities and enable more sophisticated autonomous science on Mars.

Summary updated

BibTeX
@inproceedings{lomashvili2024optimized,
  author = {Ana Lomashvili and Kristin Rammelkamp and Olivier Gasnault and Protim Bhattacharjee and Elise Clavé and Christoph H. Egerland and Susanne Schröder and Begüm Demir and Nina L. Lanza},
  title = {{Optimized Martian Dust Displacement Detection Using Explainable Machine Learning}},
  booktitle = {Workshops @ CVPR},
  year = {2024},
  doi = {10.1109/cvprw63382.2024.00671},
}
Figure