Precise classification of land cover is essential for effective environmental and urban planning, particularly in diverse landscapes with intricate spatial patterns. This study offers a comparative evaluation of pixel-based and object-oriented image classification techniques using Sentinel-2 satellite imagery of the Federal University of Technology, Akure (FUTA), Nigeria. The pixel-based classification applied the Maximum Likelihood Classification (MLC) method, which depended exclusively on spectral data, while the object-oriented approach integrated multi-resolution segmentation and contextual features such as shape and texture. Ground truth data were gathered from thirty (30) georeferenced locations using a mobile GPS for validation. Results show that while using the pixel-based method, the vegetation covers 2.515 km² (37%), compared to 2.266 km² (33%) from object-oriented classification; Farmland accounts for 1.917 km² (28%) versus 1.803 km² (27%); Bare Ground is recorded at 1.206 km² (18%) as opposed to 1.232 km² (18%); and Built-up is measured at 1.161 km² (17%) compared to 1.496 km² (22%) from the pixel-based classification. Accuracy assessments using confusion matrices revealed that the object-oriented method outperformed the pixel-based method, achieving an overall accuracy of 90% with a Kappa coefficient of 0.8663, compared to 80% accuracy for the pixel-based method. The object-oriented classification proved more effective in distinguishing built-up and bare ground areas, while both methods performed similarly in classifying vegetation. This study concludes that object-oriented classification is preferable for complex and urban environments where accuracy is critical. Expanding ground-truth data beyond thirty points and employing higher-resolution imagery would further enhance classification reliability and precision.
Nnamani et al. (Wed,) studied this question.