Geoskill: Water Body Extraction

Automatically extract water bodies from Landsat 8/9 or Sentinel-2 satellite images using NDWI/MNDWI indices with optional Otsu thresholding and vector output.

ruiduobao@ruiduobao

Install

openclaw skills install @ruiduobao/geoskill-water-body-extraction

Water Body Extraction

Automatically extract water bodies from satellite imagery using spectral indices (NDWI / MNDWI). Designed for Landsat 8/9 and Sentinel-2 multi-band GeoTIFF inputs.

Features

NDWI (Normalized Difference Water Index): (Green - NIR) / (Green + NIR)
MNDWI (Modified NDWI): (Green - SWIR) / (Green + SWIR)
Otsu automatic thresholding — no manual threshold needed
Manual threshold mode for fine control
Batch processing of multiple images
Vector output — GeoJSON boundaries alongside raster mask
Sensor-aware band selection for Landsat 8/9 and Sentinel-2

Band Configuration

Sensor	Green	NIR	SWIR
Landsat 8/9	B3	B5	B6
Sentinel-2	B3	B8	B11

Usage

Single image extraction

bash

python scripts\water-body-extraction.py extract \
  --input image.tif \
  --sensor landsat8 \
  --index mndwi \
  --output water_mask.tif \
  --vector water_boundary.geojson

NEW (v0.2.0): one-shot by place name

bash

# Provide just --place; the skill auto-fetches a Sentinel-2 scene for you
python scripts\water-body-extraction.py extract \
  --place "成都市锦江区" \
  --preset water-urban \
  --start-date 2023-06-01 --end-date 2023-09-30 --max-cloud 20 \
  --output water_mask.tif \
  --vector water_boundary.geojson \
  --qa

Available presets:

water-urban (Sentinel-2 + MNDWI)
water-rural (Sentinel-2 + NDWI)
water-landsat (Landsat 8/9 + MNDWI)

Otsu threshold optimization

bash

python scripts\water-body-extraction.py threshold \
  --input image.tif \
  --sensor landsat8 \
  --index ndwi \
  --method otsu

Batch processing

bash

python scripts\water-body-extraction.py batch \
  --input-dir ./images/ \
  --sensor sentinel2 \
  --index mndwi \
  --output-dir ./masks/ \
  --vector-dir ./vectors/

Installation

bash

pip install rasterio>=1.3.0 numpy>=1.21.0 scipy>=1.7.0 shapely>=1.8.0 fiona>=1.8.0 tqdm>=4.64.0
# Or: pip install -r scripts/requirements.txt

Parameters

--input: Path to multi-band GeoTIFF
--sensor: Sensor type (landsat8, landsat9, sentinel2)
--index: Index to compute (ndwi, mndwi)
--threshold: Manual threshold value (default: auto via Otsu)
--output: Output raster mask path
--vector: Output vector GeoJSON path (optional)
--input-dir: Directory for batch processing
--output-dir: Output directory for batch masks
--vector-dir: Output directory for batch vectors
--method: Threshold method (otsu, manual)
--json: Output statistics as JSON

Output

Raster mask: Binary GeoTIFF (1=water, 0=non-water), same CRS/resolution as input
Vector: GeoJSON polygon boundaries of water bodies
Statistics: Water pixel count, percentage, threshold used

Dependencies

text

rasterio>=1.3.0
numpy>=1.21.0
scipy>=1.7.0
shapely>=1.8.0
fiona>=1.8.0
tqdm>=4.64.0

NDWI vs MNDWI Guidance

Index	Best For	Caution
NDWI	Open water, rural areas	Built-up areas may show false positives
MNDWI	Urban areas, mixed land cover	Suppresses built-up noise; may miss turbid water

Recommendation: Use MNDWI for urban/suburban scenes; NDWI for natural landscapes.

Cloud Masking Guidance

Clouds and cloud shadows cause false water detections. Best practices:

Pre-filter: Use only images with <10% cloud cover (check metadata)
QA band: Use the Landsat/Sentinel QA_PIXEL band to mask clouds before extraction
Post-processing: Apply morphological opening to remove small cloud-shadow false positives

Minimum Area Filtering

Remove small noise polygons with --min-area:

bash

python scripts\water-body-extraction.py extract \
  --input image.tif --sensor landsat8 --index mndwi \
  --output water_mask.tif --vector water_boundary.geojson \
  --min-area 1000

--min-area is in map units (m² for projected CRS). Typical values: 500–5000 m².

Large File Guidance

For rasters >1 GB:

Process in tiles using --tile-size 4096 --overlap 256
Use --compress lzw to reduce output size
Ensure sufficient RAM: ~2× the uncompressed raster size

Vector Output Formats

In addition to GeoJSON, Shapefile and GeoPackage output are supported:

bash

python scripts\water-body-extraction.py extract \
  --input image.tif --sensor landsat8 --index mndwi \
  --output water_mask.tif --vector water_boundary.gpkg --vector-format gpkg

Accuracy Assessment

Validate results against reference data:

Collect ground-truth points or polygons
Compute confusion matrix: overall accuracy, kappa coefficient, F1 score
Compare with existing water datasets (e.g., JRC Global Surface Water, GlobeLand30)

Visualization

Visualize results in QGIS or Python:

Raster mask: Apply blue color map (water=blue, non-water=transparent)
Vector overlay: Semi-transparent blue fill with dark blue boundary
Before/after: Side-by-side comparison of original imagery and water mask

python

import rasterio
import matplotlib.pyplot as plt

with rasterio.open('water_mask.tif') as src:
    mask = src.read(1)
plt.imshow(mask, cmap='Blues')
plt.title('Water Body Extraction Result')
plt.show()

Citation

If this tool supports your research, please cite:

bibtex

@software{water_body_extraction,
  author  = {ruiduobao},
  title   = {Water Body Extraction Tool},
  url     = {https://github.com/ruiduobao/water-body-extraction},
  version = {0.1.0},
  year    = {2024},
}

Troubleshooting

Error	Cause	Solution
`ConnectionError`	Network issue	Check internet, retry
`HTTP 429`	Rate limit	Wait 60s, retry
`ValueError`	Invalid input	Check parameter format
Empty output	No data / all cloud	Try different image or adjust threshold
`ModuleNotFoundError`	Missing dep	Run pip install
`MemoryError`	Raster too large	Use tiling (--tile-size) or increase RAM
All water / no water	Threshold issue	Try manual threshold or different index

Data Source

Local raster processing — uses pre-downloaded satellite imagery. No external API calls.

Advanced Usage

Batch Processing with Shell Loop

bash

for img in scenes/*.tif; do
  python scripts\water-body-extraction.py extract     --input "$img" --index MNDWI --sensor sentinel2     --output "water_$(basename $img)"
  sleep 1
done

CI/CD Integration (GitHub Actions)

yaml

# .github/workflows/monthly-water.yml
name: Monthly Water Extraction
on:
  schedule:
    - cron: '0 0 1 * *'  # Monthly
jobs:
  extract:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: '3.11'
      - run: pip install numpy rasterio scipy
      - run: |
          python scripts\water-body-extraction.py extract \
            --input data/latest_scene.tif \
            --index MNDWI --sensor sentinel2 --otsu \
            --output data/water_latest.tif

Vector Export & PostGIS Import

bash

# Extract water and export as GeoJSON vector
python scripts\water-body-extraction.py extract   --input scene.tif --index MNDWI --vector --output water.geojson

# Import to PostGIS
ogr2ogr -f PostgreSQL PG:"dbname=gis_db" water.geojson -nln water_bodies

Performance Tips

Otsu thresholding is automatic; use --threshold 0.0 for manual control
Add --min-area 100 to filter small noise polygons
Use --cloud-mask with QA_PIXEL to improve accuracy in cloudy scenes

中文说明

使用光谱指数（NDWI / MNDWI）从卫星影像中自动提取水体。支持 Landsat 8/9 和 Sentinel-2 多波段 GeoTIFF 输入。

功能特性

NDWI（归一化水体指数）：(Green - NIR) / (Green + NIR)
MNDWI（改进型水体指数）：(Green - SWIR) / (Green + SWIR)
Otsu 自动阈值 — 无需手动设置阈值
手动阈值模式，精细控制
批量处理多张影像
矢量输出 — 同时输出 GeoJSON 边界和栅格掩膜
传感器感知 — 自动选择对应波段

波段配置

传感器	Green	NIR	SWIR
Landsat 8/9	B3	B5	B6
Sentinel-2	B3	B8	B11

使用示例

单幅影像提取

bash

python scripts\water-body-extraction.py extract \
  --input image.tif \
  --sensor landsat8 \
  --index mndwi \
  --output water_mask.tif \
  --vector water_boundary.geojson

Otsu 阈值优化

bash

python scripts\water-body-extraction.py threshold \
  --input image.tif \
  --sensor landsat8 \
  --index ndwi \
  --method otsu

批量处理

bash

python scripts\water-body-extraction.py batch \
  --input-dir ./images/ \
  --sensor sentinel2 \
  --index mndwi \
  --output-dir ./masks/ \
  --vector-dir ./vectors/

安装

bash

pip install rasterio>=1.3.0 numpy>=1.21.0 scipy>=1.7.0 shapely>=1.8.0 fiona>=1.8.0 tqdm>=4.64.0
# 或: pip install -r scripts/requirements.txt

参数说明

--input: 多波段 GeoTIFF 路径
--sensor: 传感器类型（landsat8, landsat9, sentinel2）
--index: 计算指数（ndwi, mndwi）
--threshold: 手动阈值（默认 Otsu 自动）
--output: 输出栅格掩膜路径
--vector: 输出矢量 GeoJSON 路径（可选）
--input-dir: 批量处理输入目录
--output-dir: 批量输出掩膜目录
--vector-dir: 批量输出矢量目录
--method: 阈值方法（otsu, manual）
--json: 以 JSON 格式输出统计信息

输出结果

栅格掩膜: 二值 GeoTIFF（1=水体, 0=非水体），与输入相同的 CRS/分辨率
矢量: GeoJSON 多边形水体边界
统计信息: 水体像素数、占比、使用的阈值

依赖库

text

rasterio>=1.3.0
numpy>=1.21.0
scipy>=1.7.0
shapely>=1.8.0
fiona>=1.8.0
tqdm>=4.64.0

NDWI vs MNDWI 选择指南

指数	适用场景	注意事项
NDWI	开阔水体、农村地区	建筑区可能产生假阳性
MNDWI	城市区域、混合用地	抑制建筑噪声；可能漏检浑浊水体

建议：城市/郊区场景用 MNDWI；自然地貌用 NDWI。

云掩膜指导

云和云影会导致虚假水体检测。最佳实践：

预处理：仅使用云量 <10% 的影像（查看元数据）
QA 波段：使用 Landsat/Sentinel QA_PIXEL 波段在提取前掩膜云
后处理：形态学开运算去除小云影假阳性

最小面积过滤

使用 --min-area 去除小面积噪声多边形：

bash

python scripts\water-body-extraction.py extract \
  --input image.tif --sensor landsat8 --index mndwi \
  --output water_mask.tif --vector water_boundary.geojson \
  --min-area 1000

--min-area 单位为地图单位（投影 CRS 为 m²）。典型值：500–5000 m²。

大文件处理

对于 >1 GB 的栅格：

使用 --tile-size 4096 --overlap 256 分块处理
使用 --compress lzw 减小输出
确保足够内存：约为未压缩栅格大小的 2 倍

矢量输出格式

除 GeoJSON 外，还支持 Shapefile 和 GeoPackage 输出：

bash

python scripts\water-body-extraction.py extract \
  --input image.tif --sensor landsat8 --index mndwi \
  --output water_mask.tif --vector water_boundary.gpkg --vector-format gpkg

精度评估

使用参考数据验证结果：

采集地面真实点或多边形
计算混淆矩阵：总体精度、Kappa 系数、F1 分数
与现有水体数据集对比（如 JRC 全球地表水、GlobeLand30）

可视化

在 QGIS 或 Python 中可视化结果：

栅格掩膜：蓝色配色（水体=蓝色，非水体=透明）
矢量叠加：半透明蓝色填充 + 深蓝色边界
对比：原始影像与水体掩膜并排对比

引用格式

如果本工具支持您的研究，请引用：

bibtex

@software{water_body_extraction,
  author  = {ruiduobao},
  title   = {Water Body Extraction Tool},
  url     = {https://github.com/ruiduobao/water-body-extraction},
  version = {0.1.0},
  year    = {2024},
}

故障排除

错误	原因	解决方案
`ConnectionError`	网络问题	检查网络，重试
`HTTP 429`	速率限制	等待 60 秒后重试
`ValueError`	无效输入	检查参数格式
无输出	无数据 / 全是云	更换影像或调整阈值
`ModuleNotFoundError`	缺少依赖	运行 pip install
`MemoryError`	栅格过大	使用分块处理或增加内存
全水体 / 无水	阈值问题	尝试手动阈值或更换指数

数据来源

本地栅格处理 — 使用预下载的卫星影像，无外部 API 调用。