Evaluating Landsat-8, Landsat-9 and Sentinel-2 imageries in land use and land cover (LULC) classification in a heterogeneous urban area

Land use and land cover (LULC) mapping is important for sustainable land management and has received great attention from researchers over the years. Classifying satellite imagery within urban environments poses challenges due to the spectral similarity among various LULC features. This study aims to evaluate Landsat 9, Landsat 8 and Sentinel-2 imageries in LULC classification in a heterogeneous urban area, using the city of Johannesburg as a case study. The objectives of the study were to examine the effectiveness of Random Forest (RF) and k-Nearest Neighbor (kNN) in the classification of Landsat 9, Landsat 8 and Sentinel-2 imageries in the study area. The benefits of integrating ancillary data and using post-classification correction (PCC) to generate precise LULC maps in the study area were also assessed. The performance of the multispectral bands for the satellite imageries was evaluated. The RF classifier performed better than kNN in LULC classification with high overall accuracies of 96%, 92% and 94% for Landsat 9, Landsat 8, and Sentinel-2 imageries, respectively. The kNN classifier produced overall accuracies of 95% (Landsat 9), 91% (Landsat 8) and 90% (Sentinel-2). The integration of additional data and the application of the PCC method led to enhanced accuracies in all three satellite imageries. For Landsat 9, both the RF and kNN classifiers exhibited a 1% improvement in accuracy. Notably, all overall accuracies demonstrated enhancements, with the maximum increase reaching 2%. The NIR, Red, and SWIR bands were the most influential with values of 100%, 94%, and 85%, respectively, in the LULC classification. The results of this study provide valuable information to land managers, municipalities, and stakeholders in understanding the spatial distribution of LULC classes, data, and classification methods to use in a heterogeneous urban environment.