Enhancing Land Cover Mapping with ThinkingEarth’s Self-Learning Techniques

News Item
30 January 2025

Mapping land cover is important for tracking environmental changes, managing natural resources, and making decisions about cities, farms, and climate. Traditional methods need a lot of manual work and labelled data, which can be slow and costly. ThinkingEarth's  self-supervised learning (SSL) technique can offer a better way to map land cover accurately and efficiently, using large amounts of unlabelled satellite data.
Mapping land cover is important for tracking environmental changes, managing natural resources, and making decisions about cities, farms, and climate. Traditional methods need a lot of manual work and labelled data, which can be slow and costly. ThinkingEarth's self-supervised learning (SSL) technique can offer a better way to map land cover accurately and efficiently, using large amounts of unlabelled satellite data.

How Self-Supervised Learning Helps with Remote Sensing

Self-supervised learning is a method where the system teaches itself by creating its own labels from unlabelled data. This helps models learn useful information without needing a lot of human input. In remote sensing, SSL helps in these ways:

  1. Finding Important Features: ThinkingEarth’s SSL technique pull out key details from satellite images, such as colours, shapes, and patterns, that might be missed by traditional methods.
  2. Less Need for Manual Labels: SSL reduces the need for large, labelled datasets, saving time and money. This makes it easier to map land cover in places where labelled data is hard to get.
  3. Stronger Models: SSL helps models work well even with data from different types of satellites and locations, making them reliable in various conditions.

Developing the First Copernicus Foundation Models

ThinkingEarth is working on creating the first Copernicus Foundation Models. This involves using self-supervised learning techniques specifically designed for Copernicus Sentinel datasets, including Sentinel-1, Sentinel-2, and Sentinel-3. The focus will be on improving contrastive learning and masked image modelling, which will help compress data effectively without losing important details.

Our approach aims to combine self-supervised data from different Sentinel sources to retain valuable information while balancing the strengths of each data type. We also plan to enhance the ability of deep learning models to perform well across different regions and time periods. This will be achieved by combining self-supervised learning machine learning, among other techniques, ensuring models stay accurate as new data becomes available.

Where This Technology is Useful

ThinkingEarth’s SSL-enhanced mapping can be used in many areas:

  • Environmental Tracking: Better detection of deforestation, urban growth, and wetland changes.
  • Farming: Improved crop monitoring and management.
  • Disaster Response: Quick and accurate mapping of areas affected by floods, fires, and storms.

Conclusion

ThinkingEarth’s self-supervised learning methods are changing how we map land cover. By using unlabelled data properly, mapping can be made faster, cheaper, and more accurate. This technology helps us better understand and protect our environment, now and in the future.

Share

Read next


How ThinkingEarth Leverages Large-Scale Datasets
News Item
27 March 2025

How ThinkingEarth Leverages Large-Scale Datasets

At ThinkingEarth, we harness the power of Self-Supervised Learning (SSL) and Graph Neural Networks (GNNs) to construct task-agnostic Copernicus Foundation Models and a graph-based representation of the Earth. By integrating diverse Earth Observation (EO) datasets, we aim to improve environmental monitoring, climate action, and sustainable development.
ThinkingEarth and Copernicus Foundation Models: Advancing Earth Observation
News Item
28 February 2025

ThinkingEarth and Copernicus Foundation Models: Advancing Earth Observation

At ThinkingEarth, we are developing advanced Artificial Intelligence (AI) techniques to better understand our planet as a complex system. Using Deep Learning (DL), eXplainable AI (xAI), and physics-aware Machine Learning (ML), we enhance Earth observation (EO).
Graph Neural Networks and ThinkingEarth
News Item
12 December 2024

Graph Neural Networks and ThinkingEarth

Graphs are a powerful way to represent complex relationships between different entities, making them essential for understanding intricate systems.
Newsletter of the project Thinking Earth

Stay tuned and subscribe to our quarterly newsletter

By submitting your e-mail address, you agree to our privacy policy for the site.