Multi-sensor satellite-based vegetation mapping in the Antarctic Peninsula through machine learning

The ice-free areas of the Antarctic Peninsula are undergoing significant ecological changes due to regional warming, leading to the expansion of opportunistic vegetation, primarily mosses and lichens. Mapping these communities is essential for long-term ecological monitoring, yet it remains a challenge due to the high spatial fragmentation and spectral similarity of the land cover types.

This study evaluates the effectiveness of different satellite sensors and machine learning algorithms for automated vegetation classification in the ice-free areas of Barton Peninsula (King George Island). Using high-resolution WorldView-2 (2020) and medium-resolution Sentinel-2 and Landsat 8 (2023) imagery, we compared the performance of Support Vector Machines (SVM), Random Trees (RT), and k-Nearest Neighbours (kNN) classifiers through both pixel-based and object-based approaches.

Results indicate that the kNN classifier achieved the highest overall accuracy (OA = 0.91; Kappa = 0.87) when applied to WorldView-2 data, outperforming the traditionally favoured SVM in this specific environment. The study also highlights the limitations of coarser resolution sensors (Sentinel-2 and Landsat 8) in capturing small, fragmented patches of vegetation, where the "mixed pixel" effect remains a significant hurdle, assessing how the results can still be considered meaningful.

The developed methodology demonstrates that multi-sensor remote sensing is a robust tool for creating baseline vegetation maps. These maps are crucial for quantifying the "greening" of Antarctica and provide a scalable framework for environmental conservation efforts under the Antarctic Treaty System.