多模态遥感AI融合 | Multimodal Remote Sensing AI Fusion

Normalizes multiple heterogeneous data sources (optical/SAR/thermal infrared, etc., supplied as multi-band rasters with each band = one modality) to a comparable scale, fuses them by user-specified or automatic weights, and performs joint classification on the fused image.

This skill is an offline numpy-equivalent implementation of a multimodal deep-learning fusion network: without relying on any deep-learning framework, it reproduces the multi-source fusion paradigm via "per-source min-max/z-score normalization → automatic inverse-noise-variance weighting (MAD Laplacian noise estimation) → weighted-average fusion → KMeans joint classification"; synthetic data guarantees that "fused noise below that of any single source" becomes a verifiable mathematical fact.

Dependencies / 依赖

pip install numpy rasterio scipy scikit-learn geopandas shapely

Usage / 使用方法

Example 1 (synthetic data, offline)

python geoskill-multimodal-fusion-ai.py --bbox 116.0 39.0 117.0 40.0 --synthetic --output-dir ./out

Example 2: synthetic dual-modal fusion (offline)

python geoskill-multimodal-fusion-ai.py --bbox 116.0 39.0 117.0 40.0 --synthetic --n-classes 3 --output-dir ./out

Example 3: fusion with specified weights

python geoskill-multimodal-fusion-ai.py --input multi.tif --weights 0.7,0.3 --output-dir ./out

Example 4: z-score normalization

python geoskill-multimodal-fusion-ai.py --input multi.tif --norm zscore --n-classes 5 --output-dir ./out

Output / 输出

Data Source / 数据源 / Source

Local multi-band GeoTIFF (each band = one modality), or --synthetic (optical-like + SAR-like dual observations of the same scene).

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-multimodal-fusion-ai description: '多源栅格标准化+加权融合+联合分类，输出融合分类栅格（离线numpy等价实现）'

多模态遥感AI融合 | Multimodal Remote Sensing AI Fusion

把多个异源数据（光学/SAR/热红外等，以多波段栅格输入，波段 = 模态）标准化到可比尺度，按用户权重或自动权重加权融合，并在融合图上做联合分类。

本 skill 是多模态深度学习融合网络的离线 numpy 等价实现：不依赖深度学习框架，用"逐源 min-max/z-score 标准化 -> 逆噪声方差自动加权（MAD 拉普拉斯噪声估计）-> 加权平均融合 -> KMeans 联合分类"复现多源融合范式；合成数据保证"融合噪声低于任一单源"成为可验证的数学事实。

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-multimodal-fusion-ai.py --bbox 116.0 39.0 117.0 40.0 --synthetic --output-dir ./out

示例 2：合成双模态融合（离线）

python geoskill-multimodal-fusion-ai.py --bbox 116.0 39.0 117.0 40.0 --synthetic --n-classes 3 --output-dir ./out

示例 3：指定权重融合

python geoskill-multimodal-fusion-ai.py --input multi.tif --weights 0.7,0.3 --output-dir ./out

示例 4：z-score 标准化

python geoskill-multimodal-fusion-ai.py --input multi.tif --norm zscore --n-classes 5 --output-dir ./out

输出

数据源 / Source

本地多波段 GeoTIFF（各波段 = 一个模态），或 --synthetic（同一场景的光学式 + SAR 式双观测）。

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