Comparative Evaluation of Supervised and Unsupervised Methods for Multitemporal Land-Use/Land-Cover Mapping in Lagos, Nigeria (2000–2024)

Abstract

The quantification of land cover change in rapidly expanding Nigeria cities remains a challenge despite the availability of satellite data. Land cover was classified in parts of Lagos for 2000 and 2024 using supervised and unsupervised techniques. We constrained training data using Global Mangrove Watch and JRC Surface Water Occurrence to improve the classification of coastal and aquatic zones. Random Forest produced the highest accuracy (84.8% in 2000, 91.5% in 2024), outperforming Support Vector Machines and the unsupervised classifiers in both years (McNemar p < 0.001). Built-up areas expanded by 119.6 km² (40.6% increase), while cropland and mangrove declined by 75% (20.9 km²) and 25% (20.5 km²) respectively. Water bodies and vegetation remained relatively stable in absolute terms but decreased proportionally as urban land expanded. Classification errors decreased across methods between 2000 and 2024 with the largest improvement being cropland (77% error reduction) and mangrove (30% error reduction). Results essentially reveal that mangrove loss has implications for reduced coastal protection capacity in a city already vulnerable to flooding and storm surges. Supervised classification with proper training data outweighs unsupervised methods when mapping ecologically critical classes.