FloodsML – Flood Detection with Deep Learning

FloodsML is a deep learning–based workflow for flood extent detection using Sentinel-1 SAR imagery and DEM data. The pipeline is packaged in Docker for reproducibility and portability, and it integrates seamlessly with STAC-compliant datasets.

---

🚀 Setup

1. Build Docker image

From the project root directory, build the Docker image:

docker build -t dev_floodsml -f dev_floodsml_Dockerfile .

2. Run container with mounted input data

Prepare an input folder on your local machine (e.g., C:\Users\<user>\Desktop\Dev FloodsML\input_src) and mount it into the container:

docker run -it --rm --name floods_demo_container ^
  -v "C:\Users\<user>\Desktop\Dev FloodsML\input_src:/app/input_src" ^
  dev_floodsml

---

📥 Input Data

The mounted folder must contain:

• Sentinel-1 before-flood acquisitions (VV and VH)
• Sentinel-1 after-flood acquisitions (VV and VH)
• DEM file(s) (one or more, named copdem.tif)

⚠️ All input data must follow the STAC item structure so that metadata, georeferencing, and provenance are preserved.

---

▶️ Running the model

Inside the container, run:

python /app/env/main.py \
  --aoi_wkt "POLYGON((...))" \
  [--hand] \
  [--buffer 5000] \
  [--treshold 5]

Parameters

• --aoi_wkt (required): AOI polygon in WGS84 (EPSG:4326). Must be of type POLYGON.
• --hand (optional): Apply HAND-based filtering in post-processing.
• --buffer (optional, default: 5000): Buffer distance for HAND calculation.
• --treshold (optional, default: 5): Threshold parameter for HAND filtering.

---

📤 Outputs

Results are published to:

/app/tmp/results

If you want to make the results visible on the host system, map the results directory when starting the container, e.g.:

docker run -it --rm --name floods_demo_container ^
  -v "C:\Users\<user>\Desktop\Dev FloodsML\input_src:/app/input_src" ^
  -v "C:\Users\<user>\Desktop\Dev FloodsML\results:/app/tmp/results" ^
  dev_floodsml

The outputs are:

• Cloud-Optimized GeoTIFF (COG) masks of flooded areas
• Wrapped as STAC items for interoperability

The masks contain three classes:

• 0 – non-water
• 1 – permanent hydrography
• 2 – flooded areas

This structure ensures compatibility with catalog-based workflows and easy integration with other disaster mapping products.

---

🔄 Workflow Overview

STAC input items → Preprocessing & Inference → STAC output items (COG masks)

The workflow ensures that both input and output remain fully standardized, enabling transparent use in broader EO platforms and services.

---

📌 Ready-to-use, portable, and interoperable — FloodsML can be deployed wherever flood mapping support is required.
"# floodsdl-mcube"