Explainable ML for mapping minerals from CRISM hyperspectral data

This repository contains data and python code for training a Random Forest-based classifier for identifying minerals from their pre-processed CRISM spectra.

• It consists of the pre-processed dataset for training (a pickel dataframe): unified22MineralsinlierTrainingDatatset.pkl
• Exemplar spectra from the MICA dataset, which are used to compare detections made by the random forest with detections that have been validated through community consensus.: micaBandCentres.pkl
• Band Centres of the diagnostic absorption features in each MICA spectrum: micaBandCentres.pkl
• A numpy array containing the central wavelength of each band between 1 and 2.6 μm: wavelengthList.npy
  
   
   
• The python code for training the Random Forest, pre-processing the CRISM datacubes, classifying the pre-processed spectra, and post-processing the results to generate mineral maps: TRR3PlebaniPixelLevel-MedianFilter-SuperPixel.ipynb. The notebook also allows users to plot example spectra of the mineral detections made by the Random Forest and compare them with the MICA examplars, as well as generate explanations using SHAP. The output in the notebook were obtained using FRT1FD76 which can be downloaded from https://zenodo.org/records/15170534