Land use classification with sentinel-2 satellite data: Comparison of machine learning algorithms

In this thesis, land use/land cover (LULC) classification was carried out using Sentinel-2A satellite imagery with a spatial resolution of 10 m. Different machine learning algorithms were applied, and their classification accuracies were compared. The study area is located within the boundaries of Konya Province. By using satellite images from the years 2018, 2021, and 2024, both temporal analyses were performed and land cover changes were examined. The classification process was conducted on the Google Earth Engine platform using the JavaScript programming language. Random Forest, k-Nearest Neighbors, Naive Bayes, and Support Vector Machines with linear and polynomial kernels were employed, and the results were analyzed. The performance of each algorithm was separately evaluated using 70–30% and 80–20% training/testing data splits, enabling the analysis of whether classification performance was influenced by different data ratios. Classification results were assessed through confusion matrix-based accuracy metrics, including overall accuracy, Kappa coefficient, F-score, producer's accuracy, and user's accuracy. The results showed that the SVM with polynomial kernel achieved the highest accuracy, with 96% overall accuracy and a Kappa value of 0.94. Furthermore, comparisons of different years revealed notable changes in agricultural, residential, and industrial areas. The findings of this thesis highlight the effectiveness of machine learning algorithms in LULC classification and clarify the impact of different training/testing ratios on classification performance. The significance of this study lies in comparing four machine learning algorithms on the same study area with the same training data to examine performance differences. In related studies, either a single year or only one data split ratio (70–30% or 80–20%) was typically used. In contrast, this research employed images from three different years combined with two data ratios and evaluated them together. Moreover, land cover changes were examined through temporal analysis, and accuracy metrics obtained from the confusion matrix were interpreted and compared in detail. In this context, the findings of the thesis not only demonstrate the effectiveness of machine learning algorithms in LULC classification but also emphasize the impact of varying data ratios on accuracy levels..