RESEARCH TEAM:
• Land cover classification and mapping using Sentinel-2 satellite imagery
Prof. Bogdan Zagajewski, Adam Waśniewski, MA
2D data and machine learning serving the visualised space? Is it possible to accurately represent the complex surface cover of our planet – from the tropical forests of Gabon, through the hills of Norway, to the mosaic of fields and meadows in Poland – using satellite imagery? Thanks to innovative image classification methods and the use of machine learning, the answer is: yes.
Our research focused on developing an effective approach to identifying and mapping what actually covers the Earth’s surface. We used satellite imagery and modern computer algorithms to test how well various methods perform in recognising complex landscapes. The analysis was carried out using data from three very different regions: dense equatorial forests in Gabon, the diverse landscape of Poland, and the cool, mountainous terrain of Norway. Two approaches were tested: a classical approach (all at once) and a hierarchical one (step-by-step, from easier to more difficult classes). The hierarchical approach proved more accurate – particularly in areas where land cover classes are difficult to distinguish, such as meadows, arable land, and wetlands. Our team also examined how the quality of reference data affects the final results. It turned out that well-selected samples – current, precise and representative – can improve classification accuracy by up to 40 percentage points. We also used additional data: a digital elevation model, vegetation indices, and satellite images that help differentiate various forest types. The results showed that artificial intelligence-based methods enable fast, accurate and scalable land cover mapping. They can be applied in environmental monitoring, natural resource management, spatial planning, and nature conservation – both locally and globally.
Upper graphics:
↑ Classification of forest types in Gabon based on Sentinel-2 imagery (Author: A. Waśniewski).
Lower graphics:
↑ FROM THE LEFT.
Visual comparison of results from hierarchical classification (a, d, g) and classical classification (b, e, h) with false-colour Sentinel-2 compositions (c, f, i) (Author: A. Waśniewski).
Land cover classification result for the Łódź Voivodeship and surrounding areas (Author: A. Waśniewski).
Land cover classification result for Viken County, Norway. Before (a, c, e) and after (b, d, f) verification of reference samples (Author: A. Waśniewski).
RESEARCH LOCATION: Africa / Europe (GABON / NORWAY / POLAND)