s2-water-unetplusplus-efficientnet-b4

Surface Water Detection — EfficientNet-B4 + UNet++

A semantic segmentation model for detecting surface water from Sentinel-2 multispectral satellite imagery, trained on the Earth Surface Water Dataset.

Model Description

Training Details

Performance

Tile-level Validation (during training)

Full-scene Validation (31 scenes, sliding window 512×512, overlap 256)

Usage

Installation

pip install geoai-py timm segmentation-models-pytorch

Inference from HuggingFace Hub

import geoai

geoai.timm_segmentation_from_hub(
    input_path="sentinel2_scene.tif",
    output_path="water_mask.tif",
    repo_id="giswqs/s2-water-unetplusplus-efficientnet-b4",
    window_size=512,
    overlap=256,
    batch_size=4,
)

Input Format

The model expects 6-band Sentinel-2 L2A GeoTIFF images:

Dataset

The Earth Surface Water Dataset contains 95 globally distributed Sentinel-2 scenes with binary water masks:

• Training: 64 scenes
• Validation: 31 scenes
• Source: Zenodo

Citation

@ARTICLE{Luo2021-te,
  title     = "{An applicable and automatic method for earth surface water
               mapping based on multispectral images}",
  author    = "Luo, Xin and Tong, Xiaohua and Hu, Zhongwen",
  journal   = "International Journal of Applied Earth Observation and
               Geoinformation",
  publisher = "Elsevier BV",
  volume    =  103,
  pages     =  102472,
  year      =  2021,
  url       = "http://dx.doi.org/10.1016/j.jag.2021.102472",
  doi       = "10.1016/j.jag.2021.102472",
  issn      = "1569-8432,1872-826X",
}

License

This model is released under the CC-BY-4.0 license, consistent with the training dataset license.