Data and Software for Automated Mineral Identification and Rock-Type Classification of Lunar Mare Basalts using SEM Images

This repository contains code for semantic segmentation of minerals in SEM images and subsequent rock-type classification for lunar mare basalt samples collected during Apollo missions. Overview This project implements a two-stage pipeline: 1. Semantic Segmentation: U-Net based models to segment minerals from SEM Back Scattering Electron (BSE) images 2. Rock Classification: Multiple machine learning classifiers to classify rock types based on modal mineral abundances Key Features - Three U-Net variants that handle pixel scale variability: - UNet_1: Baseline model without scale conditioning (loss: cross-entropy; CE+Dice and CE+GDL variants) - UNet_2: Dual UNet with routing based on pixel scale threshold (1.8 μm) - UNet_3: Continuous scale conditioning at bottleneck layer - Multiple classification approaches: - Rule-based baseline classifier - Gaussian Naive Bayes - Support Vector Machine (RBF) - Logistic Regression - Random Forest - Multilayer Perceptron - XGBoost Dataset The project uses two main datasets: 1. Segmentation Dataset (`/data-sem-label`) Contains SEM images and segmentation masks: - Input images: 256×256 grayscale SEM images - Output images: 256×256 segmentation masks with 10 mineral classes - Pixel scale files: Text files containing pixel size information (0.02-20 μm) - Structure: Organized in folders with `input-images/`, `output-images/`, and `input-features/` subdirectories - Mineral classes: - `c₀`: Void - `c₁`: Metallic Fe (Kamacite, Martensite) - `c₂`: FeS (Troilite) - `c₃`: Metal oxides (Ilmenite, Ulvospinel, Chromite) - `c₄`: Pyroxene group - `c₅`: Plagioclase - `c₆`: Silicate minerals (Quartz, Cristobalite, Glass) - `c₇`: Olivine - `c₈`: Late-stage melt compositions (Mesostasis, Groundmass) - `c₉`: Other (Phosphate, Zircon, etc.) 2. Classification Dataset (`data-lsc-modal.xlsx`) Excel file containing modal mineral abundance data for rock classification: Mineral abundance columns: - `Ol`: Olivine abundance - `Py`: Pyroxene abundance - `Pl`: Plagioclase abundance - `Ms`: Mesostasis abundance - `Si`: Silica minerals abundance - `Op`: Opaque minerals abundance (Fe-Ni, FeS, metal oxides) Rock type labels: - Ilmenite basalt (class 0) - Olivine basalt (class 1) - Pigeonite basalt (class 2) - Used for training and evaluating rock classification models based on modal mineral abundances extracted from segmentation results Training Details Segmentation Models - U-Net with encoder-decoder structure and skip connections - Input: 256×256 grayscale images (+ pixel scale (scalar)) - Output: 256×256 segmentation masks (10 classes) - Loss: Cross-entropy loss, Dice loss, Generalized dice loss - Metric: Pixel accuracy, intersection over union - Data augmentation: 90-degree rotations, brightness adjustment Classification Models - Features: Modal mineral abundances extracted from segmentation masks: - `Ol`: Olivine abundance - `Py`: Pyroxene abundance - `Pl`: Plagioclase abundance - `Ms`: Mesostasis abundance - `Si`: Silicate minerals abundance - `Op`: Opaque minerals abundance (Fe-Ni, FeS, metal oxides) - Classes: 3 rock types (Ilmenite basalt, Olivine basalt, Pigeonite basalt) - Workflow: 1. Segmentation model extracts mineral pixel counts from SEM images 2. Pixel counts are converted to modal abundances (percentages) 3. Classification models predict rock type from modal abundances