AI-Vision pipeline for satellite data analysis using Lunar Sinuous Rilles as an example

Overview

A MATLAB application that seamlessly integrates spectral and terrain remote sensing image data pre-processing steps like augmentation, enhancement, annotation & segmentation. Using semi-manual point-and-click selection of features by domain experts, one can create feature-specific training data of very high quality. The application is adaptable for any use case by altering the encoding for image bands and threshold limits. Using CNN-based architectures, the verification of fine-tuned models for the detection of lunar volcanic features like sinuous rilles was performed on the SELENE KAGUYA dataset.

💻 Software

  • Programming Language: MATLAB®
  • Development Environment: MATLAB R2023a
  • Tested On: MATLAB R2023b and R2024a releases
  • Required Platform: MATLAB® Desktop or MATLAB® Onlineâ„¢

📦 Required Tools & Libraries

| Curve Fitting Toolbox™ | Deep Learning Toolbox™ | Image Processing Toolbox™ | Optimization Toolbox™ | Signal Processing Toolbox™ | Statistics & Machine Learning Toolbox™ | Symbolic Math Toolbox™ | Text Analytics Toolbox™ | Parallel Computing Toolbox™ | MATLAB® Online™ | Simulink® Online™ |

  • Spectral data preprocessing
  • Grad-CAM visualisation

Impact

  • MATLAB application to reduce time and effort for preprocessing the spectral image data is fully adaptable for similar datasets across domains
  • High-resolution gold-standard data created is available for free use to the community, usable for future missions
  • Deep/ Transfer learning pipelines are customisable and reproducible for specific cases
  • Prediction accuracies on KAGUYA image data for sinuous rille classification were >95% accuracy
  • Research presented at Max Planck Institute & European Lunar Symposium
GitHub View GitHub Repository