Assessing the Accuracy of Machine Learning Algorithms for Land Cover Classification in Kano State

The choice of image classification algorithm significantly affects the accuracy and reliability of land cover maps used for resource management decisions. This study evaluates the comparative accuracy of five machine learning classifiers including Random Forest, Support Vector Machine, Gradient Boosting, Artificial Neural Network, and K-Nearest Neighbour for land cover classification in Kano State's heterogeneous landscape. Sentinel-2 multispectral imagery with ten spectral bands is used as the primary data source, supplemented by terrain derivatives from SRTM elevation data. A stratified random sampling design generates 600 reference points across six land cover classes including cropland, urban, rangeland, water, bare soil, and forest. Each classifier is trained with 70 percent of reference samples and validated with the remaining 30 percent. Results demonstrate that Random Forest achieves the highest overall accuracy of 93.2 percent and a Kappa coefficient of 0.91, significantly outperforming K-Nearest Neighbour (82.4 percent). Feature importance analysis identifies the Red Edge bands and the Normalised Difference Vegetation Index as the most discriminating variables. Keywords: machine learning, land cover classification, Sentinel-2, accuracy assessment, Kano