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 - 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.zip`) Contains SEM images and segmentation masks: - Input images: 256×256 grayscale SEM images - Output masks: 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 - 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本仓库包含针对阿波罗(Apollo)任务采集的月海玄武岩样品的扫描电子显微镜(Scanning Electron Microscope, SEM)图像中的矿物语义分割代码，以及后续的岩石类型分类代码。 ## 项目概述 本项目实现了两阶段处理流程： 1. 语义分割：基于U-Net的模型，从扫描电子显微镜背散射电子(Back Scattering Electron, BSE)图像中分割矿物相 2. 岩石分类：基于模态矿物丰度的多款机器学习分类器，用于岩石类型分类 ## 核心特性 - 三种适配像素尺度变异的U-Net变体： - UNet_1：无尺度调节的基线模型 - UNet_2：基于像素尺度阈值（1.8 μm）的双路U-Net - UNet_3：瓶颈层引入连续尺度调节的模型 - 多种分类方法： - 基于规则的基线分类器 - 高斯朴素贝叶斯(Gaussian Naive Bayes) - 径向基核支持向量机(Support Vector Machine, RBF) - 逻辑回归(Logistic Regression) - 随机森林(Random Forest) - 多层感知机(Multilayer Perceptron) - 极限梯度提升树(XGBoost) ## 数据集 本项目使用两个主要数据集： 1. 分割数据集（`data-sem-label.zip`） 包含SEM图像与分割掩码： - 输入图像：256×256灰度SEM图像 - 输出掩码：256×256的10类矿物分割掩码 - 像素尺度文件：包含像素尺寸信息（0.02~20 μm）的文本文件 - 目录结构：按`input-images/`、`output-images/`与`input-features/`子目录组织 - 矿物类别： - `c₀`：孔隙(Void) - `c₁`：金属铁（铁纹石(Kamacite)、马氏体(Martensite)） - `c₂`：陨硫铁(Troilite，化学成分为FeS) - `c₃`：金属氧化物（钛铁矿(Ilmenite)、钛铁尖晶石(Ulvospinel)、铬铁矿(Chromite)） - `c₄`：辉石族矿物 - `c₅`：斜长石 - `c₆`：硅酸盐矿物（石英(Quartz)、方英石(Cristobalite)、玻璃质(Glass)） - `c₇`：橄榄石 - `c₈`：晚期熔融相（充填物(Mesostasis)、基质(Groundmass)） - `c₉`：其他矿物（磷酸盐矿物(Phosphate)、锆石(Zircon)等） 2. 分类数据集（`data-lsc-modal.xlsx`） 用于岩石分类的模态矿物丰度数据Excel文件： 矿物丰度列： - `Ol`：橄榄石丰度 - `Py`：辉石丰度 - `Pl`：斜长石丰度 - `Ms`：充填物丰度 - `Si`：硅酸盐矿物丰度 - `Op`：不透明矿物丰度（Fe-Ni合金、陨硫铁、金属氧化物） 岩石类型标签： - 钛铁玄武岩（类别0） - 橄榄石玄武岩（类别1） - 易变辉石玄武岩（类别2） 该数据集用于基于分割结果提取的模态矿物丰度，训练与评估岩石分类模型。 ## 训练细节 ### 分割模型 - 采用编码器-解码器结构且带有跳跃连接的U-Net - 输入：256×256灰度图像 + 像素尺度标量参数 - 输出：256×256的10类分割掩码 - 损失函数：交叉熵损失 - 评估指标：像素准确率、交并比(Intersection over Union, IoU) - 数据增强方式：90°旋转、亮度调整 ### 分类模型 - 特征：从分割掩码中提取的模态矿物丰度，即`Ol`（橄榄石丰度）、`Py`（辉石丰度）、`Pl`（斜长石丰度）、`Ms`（充填物丰度）、`Si`（硅酸盐矿物丰度）、`Op`（不透明矿物丰度） - 类别：3种岩石类型（钛铁玄武岩、橄榄石玄武岩、易变辉石玄武岩） - 工作流程： 1. 分割模型从SEM图像中提取矿物像素计数 2. 将像素计数转换为模态丰度（百分比形式） 3. 分类模型基于模态丰度预测岩石类型