Land classification using Sentinel-2 data

The aim is of the project is to give students an opportunity to experiment with the supervised multispectral classification of Earth observation data. The project should significantly expand the example given to students as a lab exercise.

The result of the project should be a pipeline of scripts from the transformation of the input data to the evaluation of the classification accuracy and demonstration of the classifiers on examples.

The implemented solution needs to satisfy at least the following criteria:

  • Use at least 3 types of classifiers. Experimentation is welcome – more examined classifiers and hyperparameter experimentation.
  • Demonstrate the following:
    • Multi-temporal pixels – an area captured at different times of the year; data entry consisting of pixels in multiple bands and multiple temporal steps
    • Data from multiple products in the training set
  • Students should try to visually represent the classification results, accuracy, and errors

The input of the project is Sentinel-2 data products and labeling. Students can decide on their own which dataset they use for labeling the data. The default option is to use the Urban Atlas 2018.

Some steps of the work

  1. Review state of the art in the classification of EO data. The review should inform the selection of classification methods used later in this project.
  2. Prepare a toy dataset; the size of the data is decided by students.
  3. Revise and refactor the example from the labs into code more suitable for batch processing and rapid modifications.
  4. Review, refactor and extend classification accuracy evaluation procedures.
  5. Experiment with various classifiers and their hyperparameters. The pipeline should start at the processing of the input data and end at the evaluation of the classification accuracy.
  6. Prepare a presentation of the results, including examples created using the toy dataset.

Computational resources

A remote computer for long-running code can be provided to students after discussion.

Semester 2020/2021

This project is assigned to Stanislav Kocan and Matúš Sabat