---
id: engines/ai4g-flood
about:
  label: AI4G Flood
  description: >-
    Open flood detection model from the Microsoft AI for Good Lab that compares
    Sentinel-1 radar images from before and during a flood, with a global record
    of detected floods from October 2014 to September 2024.
  tags:
    - flood
    - radar
    - sar
    - sentinel-1
    - flood extent
    - machine learning
tool:
  form: 'open-source model and Python scripts, with a global flood dataset'
  group: Water and flooding
  homepage: 'https://github.com/microsoft/ai4g-flood'
  licence: 'MIT (code, model and dataset)'
  cost: free
domains:
  - water
  - places
open:
  source: open
  data: open
  api: 'no'
  participation: open
  licence: MIT
  note: >-
    The trained model ships in the repository; the ten-year flood dataset is on
    Hugging Face under MIT.
relations:
  - predicate: answers
    target: questions/flooding
  - predicate: alternative_to
    target: engines/ml4floods
provenance:
  origin: agent
  sources:
    - title: microsoft/ai4g-flood (README)
      url: 'https://github.com/microsoft/ai4g-flood'
    - title: >-
        Mapping global floods with 10 years of satellite radar data (Nature
        Communications)
      url: 'https://www.nature.com/articles/s41467-025-60973-1'
    - title: >-
        Mapping Global Floods with 10 Years of Satellite Radar Data (arXiv
        2411.01411)
      url: 'https://arxiv.org/abs/2411.01411'
    - title: ai-for-good-lab/ai4g-flood-dataset (dataset card)
      url: 'https://huggingface.co/datasets/ai-for-good-lab/ai4g-flood-dataset'
    - title: ai4g_sar_model.ckpt (microsoft/ai4g-flood)
      url: >-
        https://github.com/microsoft/ai4g-flood/blob/main/models/ai4g_sar_model.ckpt
  created: '2026-10-21'
  reviewed: '2026-10-21'
  reviewed_by: 'Claude, from the sources listed'
  status: draft
---
Sentinel-1 carries a radar, which sees the ground through cloud and at night.[^4] This model takes a Sentinel-1 radar image from before a possible flood and one from during it, and marks the pixels that turned to water.[^1,2] It is a U-Net, a common network for labelling every pixel of an image, built on a small MobileNetV2 encoder.[^1]

The model comes from the Microsoft AI for Good Lab.[^2,4] Amit Misra, Kevin White, Simone Fobi Nsutezo and Juan Lavista of the lab, with William Straka III of the University of Wisconsin-Madison, published it in Nature Communications on 1 July 2025.[^2,3] They ran the model over every Sentinel-1 image from October 2014 to September 2024 to build a global record of floods.[^2,4] The paper uses it to find areas of Ethiopia that flood again and again, and to map the May 2024 floods in Kenya as they happened.[^2]

The scripts run in two ways. Give a country and a date range, and the script pulls processed Sentinel-1 images from the Microsoft Planetary Computer and writes one GeoTIFF map per scene, which a second script merges into one.[^1] Or download a before and after pair yourself, for example from the Alaska Satellite Facility.[^1] Those images need the same processing as the training data, or the results may differ.[^1] The trained model is a file of about 60 MB in the repository.[^5]

The ten-year record is a download of its own on Hugging Face.[^4] It holds detections at 20 metres in 3 by 3 degree tiles, with soil moisture, elevation, temperature and land cover alongside each one for weeding out false alarms.[^4] Recurrence maps count the months in which each place flooded.[^4] For a model of a place, that is a decade of observed floods to set against a flood model's predictions.

The model mistakes permanent lakes and rivers for flood unless they are masked out with a map of standing water such as ESA WorldCover or Global Surface Water.[^1] The paper's method widens each detection by 80 metres.[^1,2] The repository, trained model included, and the dataset are under the MIT licence.[^1,4] [ML4Floods](/engines/ml4floods/) does the same job from optical images.