Multi-decadal change detection of coastal dune landscapes using different supervised image classification algorithms

Abstract

Namibia’s coastal dune systems, vital to biodiversity, tourism, and coastal protection, are increasingly threatened by climate change and anthropogenic forces. This study investigates changes in these dynamic and fragile landscapes over a 30-year period (1994-2024) by comparing the performance of four supervised image classification algorithms: Maximum Likelihood, Minimum Distance, Spectral Angle Mapping, and Random Forest. Landsat 5 and 8 images for four epochs (1994, 2004, 2014, 2024) were pre-processed in QGIS, and classification was performed using the Semi-Automatic Classification Plugin (SCP). Virtual rasters were generated, and training samples were used to classify dune features and surrounding land cover types. Classification accuracy was assessed using 100 randomly generated reference points per year to ensure reliability and consistency. Results revealed Random Forest as the most accurate algorithm, outperforming the others in capturing spectral variability and complex land cover transitions, particularly near rapidly expanding urban centres like Swakopmund and Walvis Bay. The classification maps revealed notable spatial and temporal changes, including dune displacement, urban encroachment and fluctuations in bare rock coverage. This study addresses the current gap in long-term dune monitoring using advanced classification methods and contributes data-driven insights for conservation planning, environmental management and sustainable land use in Namibia’s coastal areas.