Self-Supervised Learning for Earth Observation: MAE & DINOv2

This repository contains the code for the project:

Self-Supervised Learning: MAE & DINOv2 Models and Practical Implementation
Nishant Gupta
MSc Artificial Intelligence, Technische Hochschule Ingolstadt
Supervised by Prof. Dr. Stefan Kugele and Dr. Mohamed Chouai
June 2025

Overview

This work investigates two state-of-the-art self-supervised learning (SSL) models—Masked Autoencoder (MAE) and DINOv2—applied to multispectral satellite imagery using the SSL4EO-S2C dataset. The project covers both a rapid prototype phase (100-patches) and a large-scale experiment (8-bit, 160GB+ dataset).

Contents

• /code/Prototype_Implementation.py — Implements both MAE and DINOv2 on the 100-patches subset (for rapid prototyping and comparison).
• /code/DINOv2_Implementation.py — Scaled-up implementation of DINOv2 on the full 8-bit SSL4EO-S2C dataset.

Datasets

The following publicly available datasets were used in this project:

• SSL4EO-S2C 100-patches subset:
  Google Drive Download Link
  A small subset (100 patches) used for rapid prototyping and comparison of MAE and DINOv2.

• SSL4EO-S2C 8-bit Full Dataset:
  Mediatum (TUM) Download Link
  The large-scale 8-bit version (40 GB compressed) used for full-scale DINOv2 training.

Please refer to the terms and conditions of the dataset providers for usage and citation requirements.