机器学习土地覆被分类 | ML Land Cover Classification

Perform pixel-by-pixel supervised classification of multispectral imagery to produce a land use/land cover (LULC) class raster. Feature engineering is based on 6-band reflectance, from which NDVI and the local-variance texture of the near-infrared band are derived to form a per-pixel feature vector. The classifier supports Random Forest (rf) and Gradient Boosting (the xgboost option, implemented offline with scikit-learn's GradientBoostingClassifier).

Typical applications: regional land use/land cover mapping, rapid inventory of urban and cropland distribution, and pre-classification for change detection. After classification, an optional 3×3 majority filter removes salt-and-pepper noise, and overall accuracy (OA), Kappa and per-class producer/user accuracy are reported on a held-out validation set for quality control.

Synthetic mode automatically generates labeled multispectral scenes (water / vegetation / cropland / built-up / bare soil, each with a characteristic spectrum) fully offline, suitable for teaching, pipeline validation and unit testing.

Dependencies / 依赖

pip install 'numpy' 'rasterio' 'scipy' 'scikit-learn'

Usage / 使用方法

Basic Usage

python geoskill-lulc-classification-ml.py --bbox 116.0 39.0 117.0 40.0 --n-classes 5 --method rf

Example 1 (synthetic data, offline)

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

Example 2 (gradient boosting classifier)

python geoskill-lulc-classification-ml.py --bbox 121.0 31.0 122.0 32.0 --synthetic --method xgboost --output-dir ./out

Example 3 (read a local multispectral image)

python geoskill-lulc-classification-ml.py --input scene_sr.tif --n-classes 5 --output-dir ./out

Example 4 (disable majority filtering)

python geoskill-lulc-classification-ml.py --bbox 116 39 117 40 --synthetic --no-filter --output-dir ./out

Example 5 (3 classes + custom validation fraction)

python geoskill-lulc-classification-ml.py --bbox 116 39 117 40 --synthetic --n-classes 3 --test-fraction 0.3 --output-dir ./out

Output / 输出

Class indices: 0=water, 1=vegetation, 2=cropland, 3=built_up, 4=bare_soil.

Data Source / 数据源 / Source

• Local multispectral GeoTIFF (surface reflectance, ≥4 bands);
• --synthetic generates a physically consistent scene offline (no network, no account required).

Privacy / 隐私声明 / Privacy

• Runs offline by default; --synthetic mode requires no network at all.
• All processing is performed locally; no user data is uploaded.

License / License

MIT

name: geoskill-lulc-classification-ml description: '多波段特征逐像元 RandomForest/梯度提升土地覆被分类，含精度评估、面积统计与众数滤波去盐噪'

机器学习土地覆被分类 | ML Land Cover Classification

对多光谱影像执行逐像元监督分类，产出土地覆被（LULC）类别栅格。特征工程以 6 波段反射率为基础，派生 NDVI 与近红外波段局部方差纹理，构成逐像元特征向量； 分类器支持随机森林（rf）与梯度提升（xgboost 选项，离线等价实现为 scikit-learn 的 GradientBoostingClassifier）。

典型应用：区域土地利用/覆被制图、城市与农田分布快速摸底、变化检测的前置 分类。分类后可选 3×3 众数滤波去除盐噪，并在留出验证集上输出总体精度（OA）、 Kappa 与逐类生产/用户精度，便于质量把关。

合成模式自动生成带标签的多光谱场景（水体/植被/耕地/建成区/裸地各具特征光谱）， 完全离线，适合教学、流程验证与单元测试。

依赖

pip install 'numpy' 'rasterio' 'scipy' 'scikit-learn'

使用方法

基本用法

python geoskill-lulc-classification-ml.py --bbox 116.0 39.0 117.0 40.0 --n-classes 5 --method rf

示例 1（合成数据，离线）

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

示例 2（梯度提升分类器）

python geoskill-lulc-classification-ml.py --bbox 121.0 31.0 122.0 32.0 --synthetic --method xgboost --output-dir ./out

示例 3（读取本地多光谱影像）

python geoskill-lulc-classification-ml.py --input scene_sr.tif --n-classes 5 --output-dir ./out

示例 4（关闭众数滤波）

python geoskill-lulc-classification-ml.py --bbox 116 39 117 40 --synthetic --no-filter --output-dir ./out

示例 5（3 类 + 自定义验证比例）

python geoskill-lulc-classification-ml.py --bbox 116 39 117 40 --synthetic --n-classes 3 --test-fraction 0.3 --output-dir ./out

输出

类别索引：0=water, 1=vegetation, 2=cropland, 3=built_up, 4=bare_soil。

数据源 / Source

• 本地多光谱 GeoTIFF（地表反射率，≥4 波段）；
• --synthetic 离线合成场景（无需网络、无需账号）。

隐私声明 / Privacy

• 默认离线运行，--synthetic 模式完全无网络。
• 所有处理在本地完成，不上传用户数据。

License

MIT