geoskill-change-detection-dl

深度学习变化检测 | Deep Learning Change Detection

Detects land surface changes from bi-temporal imagery (vegetation degradation, urban expansion, water body growth/shrinkage), outputting a change probability map, a binary change map, and change-region polygons as GeoJSON.

This skill uses a genuine deep learning model: a Siamese fully convolutional change detection network (FC-Siam-diff style, Daudt et al. 2018). The two epochs (red/nir two-channel) pass through a shared-weight encoder to extract multi-scale features, and the decoder fuses layer-wise feature differences |f1 − f2| to reconstruct pixel-level change probabilities (sigmoid, [0, 1]), which are then thresholded via --prob-thresh, aggregated into change regions through 8-neighborhood connected-component analysis, and geocoded. Both training and inference run on CUDA GPUs (torch ≥ 2.x, requiring cuDNN or automatically falling back to native CUDA convolutions).

The skill ships with pretrained weights cd_siamese_weights.pt (about 0.6 MB, trained on synthetic bi-temporal change pairs, holdout recall ≈ 0.999 / false alarm ≈ 0). If the weight file is missing, the first run automatically trains on synthetic data on the GPU (about 15 seconds) and caches the weights to disk.

Dependencies / 依赖

pip install numpy rasterio scipy torch

Usage / 使用方法

Example 1 (Synthetic Data, Offline)

python geoskill-change-detection-dl.py --bbox 116.0 39.0 117.0 40.0 --synthetic --output-dir ./out

Example 2: Two Files for Bi-temporal Input (Each with 2 Bands [red, nir] Reflectance)

python geoskill-change-detection-dl.py --input t1.tif --input2 t2.tif --prob-thresh 0.6 --output-dir ./out

Example 3: Single File with 4 Bands [red1, nir1, red2, nir2]

python geoskill-change-detection-dl.py --input pair.tif --min-area 16 --output-dir ./out

Output / 输出

In synthetic mode, QA writes synthetic_recall / synthetic_false_alarm (computed against the built-in ground truth).

Limitations / 局限（诚实声明）

• The bundled weights were trained on synthetic spectra (four reflectance classes: vegetation/bare soil/water/built-up); the probability outputs on real imagery have not been radiometrically calibrated or field-validated, so results on real data should be treated as screening-level rather than map-grade.
• --scale is the steepness parameter of the legacy classical baseline (1−exp(−scale·|dNDVI|)); it is kept only for CLI compatibility and is not used by the network probability path.

Data Source / 数据源 / Source

Local bi-temporal GeoTIFFs (each containing red/nir bands, or a single 4-band file), or a --synthetic pair (t1 fully vegetated, t2 degraded to bare soil at the center).

Privacy / 隐私声明 / Privacy

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

License / License

MIT

---

---

name: geoskill-change-detection-dl description: 'Siamese 全卷积网络(FC-Siam-diff 风格)双时相变化检测：GPU 训练/推理，输出变化概率、二值变化图与变化图斑'

深度学习变化检测 | Deep Learning Change Detection

从双时相影像检测地表变化（植被退化、城市扩张、水体消长），输出变化概率图、二值变化图与变化图斑 GeoJSON。

本 skill 使用真正的深度学习模型：Siamese 全卷积变化检测网络（FC-Siam-diff 风格， Daudt et al. 2018）。两个时相（red/nir 双通道）经共享权重编码器提取多尺度特征， 解码器融合逐层特征差 |f1 − f2| 重建像元级变化概率（sigmoid，[0,1]）， 再经 --prob-thresh 阈值二值化、8 邻域连通域聚合为变化图斑并地理编码。 训练与推理均在 CUDA GPU 上执行（torch ≥ 2.x，需 cuDNN 或自动退回 CUDA 原生卷积）。

随 skill 附带预训练权重 cd_siamese_weights.pt（约 0.6 MB，在合成双时相变化对上 训练，holdout recall ≈ 0.999 / false-alarm ≈ 0）。若权重文件缺失，首次运行时 自动在 GPU 上用合成数据训练（约 15 秒）并落盘缓存。

依赖

pip install numpy rasterio scipy torch

使用方法

示例 1（合成数据，离线）

python geoskill-change-detection-dl.py --bbox 116.0 39.0 117.0 40.0 --synthetic --output-dir ./out

示例 2：双文件双时相（各 2 波段 [red, nir] 反射率）

python geoskill-change-detection-dl.py --input t1.tif --input2 t2.tif --prob-thresh 0.6 --output-dir ./out

示例 3：单文件 4 波段 [red1,nir1,red2,nir2]

python geoskill-change-detection-dl.py --input pair.tif --min-area 16 --output-dir ./out

输出

合成模式 QA 会写入 synthetic_recall / synthetic_false_alarm（对内置真值计算）。

局限（诚实声明）

• 随附权重在合成光谱（植被/裸土/水体/建成区四类反射率）上训练，对真实影像的 概率输出未做辐射定标/外场验证；真实数据结果应视为筛查级而非制图级。
• --scale 为旧版经典基线（1−exp(−scale·|dNDVI|)）的陡峭度参数，保留仅为 CLI 兼容，网络概率路径不使用它。

数据源 / Source

本地双时相 GeoTIFF（各含 red/nir 波段，或单文件 4 波段），或 --synthetic 合成对 （t1 全植被，t2 中心退化为裸土）。

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