语义分割 | Semantic Segmentation

Performs per-pixel semantic segmentation of multispectral remote sensing imagery and outputs a class raster with per-class area statistics. Two modes are supported: unsupervised (KMeans) and supervised (RandomForest, requiring a label raster via --labels).

This skill is an offline numpy equivalent implementation of FCN/U-Net semantic segmentation networks: without depending on any deep learning framework, it reproduces the semantic segmentation pipeline through "feature construction -> per-pixel sklearn classifier -> sliding-window tiled prediction and stitching -> majority-filter post-processing". Unit tests cover the consistency between tiled and whole-scene prediction, the denoising effect of post-processing, and the unsupervised class-remapping matching accuracy.

Dependencies / 依赖

pip install numpy rasterio scipy scikit-learn geopandas shapely

Usage / 使用方法

Example 1 (Synthetic Data, Offline)

python geoskill-semantic-segmentation.py --bbox 116.0 39.0 117.0 40.0 --synthetic --output-dir ./out

Example 2: Real-Image Tiled Segmentation

python geoskill-semantic-segmentation.py --input scene.tif --n-classes 5 --tile 64 --smooth 5 --output-dir ./out

Example 3: Supervised Random Forest (Synthetic Ground Truth)

python geoskill-semantic-segmentation.py --bbox 116.0 39.0 117.0 40.0 --synthetic --method rf --output-dir ./out

Example 4: Supervised Random Forest (Real Labels)

python geoskill-semantic-segmentation.py --input scene.tif --labels labels.tif --method rf --output-dir ./out

Output / 输出

Data Source / 数据源 / Source

A local multi-band GeoTIFF, or a --synthetic cube of three land-cover types (vegetation/soil/water) with built-in ground truth for accuracy QA.

Privacy / 隐私声明 / Privacy

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

License / License

MIT

name: geoskill-semantic-segmentation description: '逐像元sklearn随机森林分类+滑窗拼接+后处理，输出分类栅格（离线numpy等价实现）'

语义分割 | Semantic Segmentation

对多光谱遥感影像做逐像元语义分割，输出类别栅格与各类面积统计。支持无监督（KMeans）与有监督（RandomForest，需 --labels 标注栅格）两种模式。

本 skill 是 FCN/U-Net 语义分割网络的离线 numpy 等价实现：不依赖深度学习框架，用"特征构建 -> 逐像元 sklearn 分类器 -> 滑窗分块预测拼接 -> 众数滤波后处理"复现语义分割流程；分块预测与整幅预测的一致性、后处理去噪效果、无监督类别置换匹配精度均有单元测试覆盖。

依赖

pip install numpy rasterio scipy scikit-learn geopandas shapely

使用方法

示例 1（合成数据，离线）

python geoskill-semantic-segmentation.py --bbox 116.0 39.0 117.0 40.0 --synthetic --output-dir ./out

示例 2：真实影像分块分割

python geoskill-semantic-segmentation.py --input scene.tif --n-classes 5 --tile 64 --smooth 5 --output-dir ./out

示例 3：有监督随机森林（合成真值）

python geoskill-semantic-segmentation.py --bbox 116.0 39.0 117.0 40.0 --synthetic --method rf --output-dir ./out

示例 4：有监督随机森林（真实标注）

python geoskill-semantic-segmentation.py --input scene.tif --labels labels.tif --method rf --output-dir ./out

输出

数据源 / Source

本地多波段 GeoTIFF，或 --synthetic 合成三体物（植被/土壤/水体）立方体（自带真值用于精度 QA）。

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，不上传任何用户数据。

License

MIT