高光谱分类 | Hyperspectral Classification

Performs per-pixel supervised classification on hyperspectral image cubes (bands × H × W, typically 30+ bands). PCA first projects the high-dimensional spectra into principal-component space to reduce noise and redundancy; a classifier is then trained with random forest (RF) or support vector machine (SVM) to predict the land-cover class of each pixel.

Synthetic mode automatically generates N land-cover classes with characteristic spectral signatures (e.g., minerals/vegetation/soil/water), distributed as spatial patches with added noise; training samples are produced by stratified sampling from the ground-truth labels (70% training / 30% validation by default). The outputs include a classification map, accuracy assessment (overall accuracy, Kappa), and a confusion matrix. Suitable for hyperspectral mapping, mineral/vegetation mapping, and method-comparison teaching.

Dependencies / 依赖

pip install numpy rasterio scipy scikit-learn

Usage / 使用方法

Basic Usage (Synthetic Data, Offline)

python geoskill-hyperspectral-classification.py --bbox 116.0 39.0 117.0 40.0 --synthetic --output-dir ./out

Example 1: Random Forest, 4-Class Classification (30 Bands)

python geoskill-hyperspectral-classification.py \
    --bbox 116.0 39.0 117.0 40.0 \
    --synthetic --n-bands 30 --n-classes 4 --method rf \
    --output-dir ./rf_4class

Example 2: SVM Method Comparison

python geoskill-hyperspectral-classification.py \
    --bbox 121.0 31.0 122.0 32.0 \
    --synthetic --n-bands 30 --n-classes 4 --method svm \
    --output-dir ./svm_4class

Example 3: More Classes and Bands

python geoskill-hyperspectral-classification.py \
    --bbox 116.0 39.0 117.0 40.0 \
    --synthetic --n-bands 60 --n-classes 6 --method rf \
    --output-dir ./rf_6class

Example 4: Real Hyperspectral GeoTIFF Input

python geoskill-hyperspectral-classification.py \
    --input cuprite_subset.tif --method rf \
    --output-dir ./real_rf

Output / 输出

Data Source / 数据源 / Source

• Real mode: local hyperspectral GeoTIFF (multi-band rasters exported from ENVI/PRISMA/AVIRIS, etc.)
• Synthetic mode: locally generates characteristic spectral curves + spatial patches, with no external data source

Privacy / 隐私声明 / Privacy

• Runs fully offline by default and makes no network requests
• --synthetic mode reads no external data
• All computation is done locally; user data is never uploaded

License / License

MIT

name: geoskill-hyperspectral-classification description: 'PCA降维+RF/SVM逐像元高光谱监督分类'

高光谱分类 | Hyperspectral Classification

对高光谱影像立方体（bands × H × W，波段数通常 30+）执行逐像元监督 分类。先用 PCA 把高维光谱投影到主成分空间降噪去冗余，再用随机森林 （RF）或支持向量机（SVM）训练分类器，对每个像元预测地物类别。

合成模式自动生成 N 类具有特征光谱曲线的地物（如矿物/植被/土壤/水体）， 按空间斑块分布并叠加噪声；训练样本由真值标签分层抽样产生（默认 70% 训练 / 30% 验证），输出分类图、精度评估（总体精度、Kappa）与混淆 矩阵。适用于高光谱制图、矿物/植被填图、方法对比教学等场景。

依赖

pip install numpy rasterio scipy scikit-learn

使用方法

基本用法（合成数据，离线）

python geoskill-hyperspectral-classification.py --bbox 116.0 39.0 117.0 40.0 --synthetic --output-dir ./out

示例 1：随机森林 4 类分类（30 波段）

python geoskill-hyperspectral-classification.py \
    --bbox 116.0 39.0 117.0 40.0 \
    --synthetic --n-bands 30 --n-classes 4 --method rf \
    --output-dir ./rf_4class

示例 2：SVM 方法对比

python geoskill-hyperspectral-classification.py \
    --bbox 121.0 31.0 122.0 32.0 \
    --synthetic --n-bands 30 --n-classes 4 --method svm \
    --output-dir ./svm_4class

示例 3：更多类别与波段

python geoskill-hyperspectral-classification.py \
    --bbox 116.0 39.0 117.0 40.0 \
    --synthetic --n-bands 60 --n-classes 6 --method rf \
    --output-dir ./rf_6class

示例 4：真实高光谱 GeoTIFF 输入

python geoskill-hyperspectral-classification.py \
    --input cuprite_subset.tif --method rf \
    --output-dir ./real_rf

输出

数据源 / Source

• 真实模式：本地高光谱 GeoTIFF（ENVI/PRISMA/AVIRIS 等导出的多波段栅格）
• 合成模式：本地生成特征光谱曲线 + 空间斑块，无外部数据源

隐私声明 / Privacy

• 默认完全离线运行，不发起任何网络请求
• --synthetic 模式不读取任何外部数据
• 所有计算在本地完成，不上传用户数据

License

MIT