Skip to content

Tile based classification

The tile based classification application takes a trained model in the ONNX format and does the inference on Sentinel-2 Level-1C data.

The model was trained using a Sequential Convolutional Neural Network (CNN) with Keras based on the benchmark dataset EuroSAT.

Training Data

The model is trained on the EuroSAT benchmark dataset which is based on Sentinel-2 satellite images and consists of a total of 27,000 labeled and geo-referenced images.

The dataset provides information on the following ten land cover / land use classes:

  • Annual Crop
  • Forest
  • Herbaceous Vegetation
  • Highway
  • Industrial
  • Pasture
  • Permanent Crop
  • Residential
  • River
  • Sea Lake

The benchmark dataset can be used to detect land cover/land use changes.

Inference

The inference is the process where the learned capabilities of a pre-trained model are put into practice and applied to a Sentinel-2 Level-1C acquisition.

This application foresees using Sentinel-2 Level-1C data converted to COG and structured as a STAC Catalog and Item.

There is a pre-processing step to stage and convert to STAC/COG a Sentinel-2 Level-1C acquisition then used for the inference process.

The inference step loads the pre-trained model and executes the inference process, which 'infers' land cover classes.