Geoskill: Change Detection

Detect and quantify land cover changes between two co-registered satellite images using NDVI difference, image differencing, or Change Vector Analysis (CVA).

ruiduobao@ruiduobao

Install

openclaw skills install @ruiduobao/geoskill-change-detection

Change Detection

Detect land cover changes between two satellite images using multiple methods: NDVI difference, image differencing, and Change Vector Analysis (CVA).

Features

NDVI Difference: ΔNDVI = NDVI_t2 − NDVI_t1 — vegetation change
Image Differencing: Per-band difference — general change
Change Vector Analysis (CVA): Multi-band magnitude + direction
Binary change mask with configurable threshold
Statistics report with change percentages
Supports Landsat 8/9 and Sentinel-2

Band Configuration

Sensor	Green	NIR	Red	SWIR
Landsat 8/9	B3	B5	B4	B6
Sentinel-2	B3	B8	B4	B11

Usage

Detect vegetation change (NDVI difference)

bash

python scripts\change-detection.py detect \
  --image-t1 2020.tif --image-t2 2023.tif \
  --sensor landsat8 --method ndvi-diff \
  --output change_magnitude.tif \
  --mask change_mask.tif

Change Vector Analysis

bash

python scripts\change-detection.py detect \
  --image-t1 2020.tif --image-t2 2023.tif \
  --sensor landsat8 --method cva \
  --output cva_magnitude.tif \
  --mask cva_mask.tif --threshold 0.15

Generate report

bash

python scripts\change-detection.py report \
  --mask change_mask.tif --json report.json

Installation

bash

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

Parameters

--image-t1: Path to time-1 image (earlier)
--image-t2: Path to time-2 image (later)
--sensor: Sensor type (landsat8, landsat9, sentinel2)
--method: Detection method (ndvi-diff, image-diff, cva)
--output: Output change magnitude GeoTIFF
--mask: Output binary change mask
--threshold: Change threshold (default: auto via Otsu)
--bands: Comma-separated band indices for CVA (default: all)
--json: Output statistics as JSON

Output

Change magnitude: Float GeoTIFF showing change intensity
Binary mask: 1=change, 0=no change
Statistics: Change pixel count, percentage, method used

Dependencies

text

rasterio>=1.3.0
numpy>=1.21.0
tqdm>=4.64.0

Method Selection Guidance

Method	Best For	Input Requirements
`ndvi-diff`	Vegetation change (growth, deforestation, crop shift)	Red + NIR bands
`image-diff`	General land cover change (any type)	Same bands in both images
`cva`	Multi-band change, direction matters	2+ matching bands

Recommendation: Start with ndvi-diff for vegetation studies; use cva when change direction (e.g., vegetation → urban) is important.

Co-registration Guidance

Misaligned images cause false changes. Before running detection:

Ensure both images share the same CRS and resolution
Use image registration tools (e.g., gdalwarp -tps, ENVI Auto-Registration, or QGIS Coregistration plugins)
Sub-pixel alignment (<0.5 pixel RMSE) recommended for pixel-based methods
Verify alignment by overlaying both images and checking stable features (roads, buildings)

CVA Direction Output

CVA produces two outputs:

Magnitude: Change intensity (float GeoTIFF) — higher = more change
Direction (angle): Indicates change type based on band contribution

Angle Range	Typical Interpretation
0°–90°	Increase in both bands (e.g., vegetation growth)
90°–180°	Band 1 decreases, Band 2 increases
180°–270°	Decrease in both bands (e.g., vegetation loss)
270°–360°	Band 1 increases, Band 2 decreases

Use --direction-output cva_direction.tif to save the angle raster.

Cross-Sensor Guidance

Comparing images from different sensors (e.g., Landsat 8 vs Sentinel-2):

Normalize radiometry: Apply relative radiometric normalization (e.g., histogram matching, IR-MAD)
Match resolutions: Resample to coarser resolution
Same season: Use images from similar phenological stage
Caution: Cross-sensor comparison introduces additional uncertainty

Cloud Masking Guidance

Clouds cause false change detections:

Pre-filter images with <10% cloud cover
Use QA_PIXEL band to mask clouds in both images
Consider temporal compositing (e.g., monthly median) to reduce cloud impact

Batch Processing

Process multiple image pairs with a CSV list:

bash

python scripts\change-detection.py batch \
  --pair-list pairs.csv --sensor landsat8 --method ndvi-diff \
  --output-dir ./results/

pairs.csv format: image_t1,image_t2,output_name

Vector Output

Convert change masks to vector polygons for GIS analysis:

bash

python scripts\change-detection.py vectorize \
  --mask change_mask.tif --output change_polygons.geojson --min-area 500

Outputs GeoJSON or Shapefile with change area attributes.

Visualization

Change magnitude: Apply diverging colormap (blue=no change, red=high change)
Binary mask: Overlay change areas in red on RGB composite
CVA direction: Use cyclic colormap (HSV) for angle visualization
Side-by-side: Show t1 image, t2 image, and change map together

Citation

bibtex

@software{change_detection,
  author  = {ruiduobao},
  title   = {Change Detection Tool},
  url     = {https://github.com/ruiduobao/change-detection},
  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 change detected	Lower threshold or check input
`ModuleNotFoundError`	Missing dep	Run pip install
Misregistration artifacts	Images not aligned	Co-register before detection
Striped output	Cloud shadow	Apply cloud masking

Data Source

Local raster processing — two co-registered GeoTIFF images required.

Advanced Usage

Batch Multi-Pair Detection

bash

for year in 2020 2021 2022 2023; do
  python scripts\change-detection.py detect     --image-t1 ${year}0101.tif --image-t2 $((year+1))0101.tif     --sensor landsat8 --method ndvi-diff     --output change_${year}_$((year+1)).tif
done

CI/CD Integration (GitHub Actions)

yaml

# .github/workflows/change-detection.yml
name: Change Detection Pipeline
on:
  schedule:
    - cron: '0 6 1 */3 *'  # Every 3 months
jobs:
  detect:
    runs-on: ubuntu-latest
    steps:
      - uses: actions/checkout@v4
      - uses: actions/setup-python@v5
        with:
          python-version: '3.11'
      - run: pip install numpy rasterio
      - run: |
          python scripts\change-detection.py detect \
            --image-t1 data/latest_2023.tif \
            --image-t2 data/latest_2024.tif \
            --sensor landsat8 --method ndvi-diff \
            --output data/change_latest.tif

Vectorize & Import to PostGIS

bash

python scripts\change-detection.py vectorize   --input change_mask.tif --output change.geojson

ogr2ogr -f PostgreSQL PG:"dbname=gis_db" change.geojson -nln change_areas

Performance Tips

Use --mask with QA band to exclude clouds before detection
CVA method requires more memory; use --bands to limit input bands
--threshold accepts 'otsu' (auto) or manual float value

中文说明

对两期卫星影像进行变化检测，支持 NDVI 差值、影像差值和变化向量分析（CVA）三种方法。

功能特性

NDVI 差值：ΔNDVI = NDVI_t2 − NDVI_t1 — 植被变化
影像差值：逐波段差值 — 通用变化
变化向量分析（CVA）：多波段变化强度 + 方向
二值变化掩膜，可配置阈值
统计报告，含变化百分比
支持 Landsat 8/9 和 Sentinel-2

波段配置

传感器	Green	NIR	Red	SWIR
Landsat 8/9	B3	B5	B4	B6
Sentinel-2	B3	B8	B4	B11

使用示例

植被变化检测（NDVI 差值）

bash

python scripts\change-detection.py detect \
  --image-t1 2020.tif --image-t2 2023.tif \
  --sensor landsat8 --method ndvi-diff \
  --output change_magnitude.tif \
  --mask change_mask.tif

变化向量分析

bash

python scripts\change-detection.py detect \
  --image-t1 2020.tif --image-t2 2023.tif \
  --sensor landsat8 --method cva \
  --output cva_magnitude.tif \
  --mask cva_mask.tif --threshold 0.15

生成报告

bash

python scripts\change-detection.py report \
  --mask change_mask.tif --json report.json

安装

bash

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

参数说明

--image-t1: 时相 1 影像路径（早期）
--image-t2: 时相 2 影像路径（晚期）
--sensor: 传感器类型
--method: 检测方法（ndvi-diff, image-diff, cva）
--output: 输出变化强度 GeoTIFF
--mask: 输出二值变化掩膜
--threshold: 变化阈值（默认 Otsu 自动）
--bands: CVA 使用的波段索引（逗号分隔，默认全部）
--json: 以 JSON 输出统计信息

输出结果

变化强度: 浮点 GeoTIFF，显示变化强度
二值掩膜: 1=变化, 0=无变化
统计信息: 变化像素数、占比、使用方法

依赖库

text

rasterio>=1.3.0
numpy>=1.21.0
tqdm>=4.64.0

方法选择指南

方法	适用场景	输入要求
`ndvi-diff`	植被变化（生长、砍伐、作物变化）	Red + NIR 波段
`image-diff`	通用土地覆盖变化	两幅影像波段一致
`cva`	多波段变化，方向重要	2+ 匹配波段

建议：植被研究先用 ndvi-diff；需要变化方向信息时用 cva。

配准指导

未对齐的影像会导致虚假变化。检测前请确保：

两幅影像使用 相同的 CRS 和分辨率
使用影像配准工具（如 gdalwarp -tps、ENVI 自动配准、QGIS 配准插件）
像元方法建议亚像元对齐（<0.5 像元 RMSE）
叠加两幅影像检查稳定地物（道路、建筑）验证对齐

CVA 方向输出

CVA 产生两个输出：

强度：变化强度（浮点 GeoTIFF）— 值越大变化越强
方向（角度）：根据波段贡献指示变化类型

角度范围	典型解释
0°–90°	两个波段均增加（如植被生长）
90°–180°	波段 1 减少，波段 2 增加
180°–270°	两个波段均减少（如植被损失）
270°–360°	波段 1 增加，波段 2 减少

使用 --direction-output cva_direction.tif 保存角度栅格。

跨传感器指导

比较不同传感器影像（如 Landsat 8 vs Sentinel-2）：

辐射归一化：应用相对辐射归一化（如直方图匹配、IR-MAD）
分辨率匹配：重采样至较粗分辨率
同季节：使用相似物候期的影像
注意：跨传感器比较引入额外不确定性 |

云掩膜指导

云会导致虚假变化检测：

预过滤云量 <10% 的影像
使用 QA_PIXEL 波段掩膜两幅影像中的云
考虑时间合成（如月际中值）减少云影响

批量处理

使用 CSV 列表处理多对影像：

bash

python scripts\change-detection.py batch \
  --pair-list pairs.csv --sensor landsat8 --method ndvi-diff \
  --output-dir ./results/

pairs.csv 格式：image_t1,image_t2,output_name

矢量输出

将变化掩膜转换为矢量多边形用于 GIS 分析：

bash

python scripts\change-detection.py vectorize \
  --mask change_mask.tif --output change_polygons.geojson --min-area 500

输出 GeoJSON 或 Shapefile，含变化面积属性。

可视化

变化强度：发散色阶（蓝=无变化，红=高变化）
二值掩膜：在 RGB 合成图上叠加红色变化区域
CVA 方向：使用循环色阶（HSV）可视化角度
对比：同时展示时相 1、时相 2 和变化图

引用格式

bibtex

@software{change_detection,
  author  = {ruiduobao},
  title   = {Change Detection Tool},
  url     = {https://github.com/ruiduobao/change-detection},
  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	未检测到变化	降低阈值或检查输入
`ModuleNotFoundError`	缺少依赖	运行 pip install
配准伪影	影像未对齐	先配准再检测
条带状输出	云影	应用云掩膜

数据来源

本地栅格处理 — 需要两幅配准好的 GeoTIFF 影像。