geoskill-few-shot-classification

原型网络类中心距离+最近邻分类，输出分类栅格（离线numpy等价实现）

小样本遥感分类 | Few-Shot Remote Sensing Classification

Classifies an entire image using very few labeled samples (1–5 per class): the feature mean of each class's support samples serves as the "prototype", each pixel is assigned to the nearest prototype, and a softmax(−distance) confidence is provided; few-shot episode accuracy evaluation is supported.

This skill is an offline numpy-equivalent implementation of Prototypical Networks few-shot learning: without depending on torch, it reproduces the metric-learning paradigm through "feature extraction → support-set standardization → prototype computation → nearest-prototype classification by Euclidean distance"; unit tests cover 1-shot and multi-shot episode accuracy, probability normalization, and errors for insufficient samples.

Dependencies / 依赖

bash

pip install numpy rasterio scipy scikit-learn geopandas shapely

Usage / 使用方法

Example 1 (synthetic data, offline)

bash

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

Example 2: synthetic 3-way 3-shot (offline)

bash

python geoskill-few-shot-classification.py --bbox 116.0 39.0 117.0 40.0 --synthetic --n-shot 3 --output-dir ./out

Example 3: extreme 1-shot few-sample case

bash

python geoskill-few-shot-classification.py --bbox 116.0 39.0 117.0 40.0 --synthetic --n-shot 1 --output-dir ./out

Example 4: pseudo few-shot on real imagery (support selected by KMeans)

bash

python geoskill-few-shot-classification.py --input scene.tif --n-classes 4 --n-shot 5 --output-dir ./out

Output / 输出

File	Format	Description
classification.tif	GeoTIFF	Full-image nearest-prototype classification label map
few_shot_report.json	JSON	Episode accuracy/confidence/number of support samples
output-manifest.json	JSON	Run manifest (inputs/outputs/QA/exit code)

Data Source / 数据源 / Source

Local multi-band GeoTIFF, or --synthetic for a three-class spectrally separable scenario (ground truth used for episode evaluation).

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 ever uploaded.

License / License

MIT

---

---

name: geoskill-few-shot-classification description: '原型网络类中心距离+最近邻分类，输出分类栅格（离线numpy等价实现）'

小样本遥感分类 | Few-Shot Remote Sensing Classification

用极少量标注样本（每类 1~5 个）完成整幅影像分类：每类支持样本的特征均值作为"原型"，像元按最近原型归类，并给出 softmax(-距离) 置信度；支持少样本回合 (episode) 精度评估。

本 skill 是 Prototypical Networks 少样本学习的离线 numpy 等价实现：不依赖 torch，用"特征提取 -> 支持集标准化 -> 原型计算 -> 欧氏距离最近原型分类"复现度量学习范式；1-shot 与多 shot 回合精度、概率归一性、样本不足报错等均有单元测试覆盖。

依赖

bash

pip install numpy rasterio scipy scikit-learn geopandas shapely

使用方法

示例 1（合成数据，离线）

bash

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

示例 2：合成 3-way 3-shot（离线）

bash

python geoskill-few-shot-classification.py --bbox 116.0 39.0 117.0 40.0 --synthetic --n-shot 3 --output-dir ./out

示例 3：1-shot 极限小样本

bash

python geoskill-few-shot-classification.py --bbox 116.0 39.0 117.0 40.0 --synthetic --n-shot 1 --output-dir ./out

示例 4：真实影像伪少样本（KMeans 选支持）

bash

python geoskill-few-shot-classification.py --input scene.tif --n-classes 4 --n-shot 5 --output-dir ./out

输出

文件	格式	说明
classification.tif	GeoTIFF	整幅最近原型分类标签图
few_shot_report.json	JSON	episode 精度/置信度/支持样本数
output-manifest.json	JSON	运行清单（输入/输出/QA/退出码）

数据源 / Source

本地多波段 GeoTIFF，或 --synthetic 三类光谱可分场景（真值用于 episode 评估）。

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