Semi-supervised object-based classification of coastal dune vegetation land covers: A new hybrid labelling approach

Monitoring changes in coastal dunes is vital for effective coastal management, particularly given their role in climate change adaptation and mitigation. This study evaluates an object-based semi-supervised methodology with high-resolution satellite imagery from the Sentinel-2 mission and machine learning algorithms to map vegetation land cover of coastal dune systems. Unlike previous studies relying on very high-resolution pixel-based imagery, our approach uses object-based segmentation, which integrates spatial context and improves ecosystem analysis. The dataset includes annual Sentinel-2 composites, seasonal NDVI, and texture variables. The Multiresolution Segmentation (MRS) algorithm, optimized using ESP2 algorithm, was employed to group pixels into homogeneous objects. Training data were labelled through K-means clustering method, and the representativeness of subsets was evaluated with Random Forest algorithm. This workflow provides an improved and reproducible hybrid alternative by automating the initial labelling process and optimizing training subset representativeness. The final model with best-performing combination of Subsets of different Class Spectral Mixing was validated against an independent test set. Along the Andalusian coast (south Spain) the approach resulted in the identification of four classes in the Atlantic and Mediterranean regions, and five in the Mediterranean and southeast region. The optimal training subset achieved overall accuracies of 0.82 in the Atlantic and Mediterranean and southeast regions and 0.85 in the Mediterranean region, confirming its effectiveness in mapping vegetation land cover in southern Spain's coastal dune systems.