Evaluating the Performance of Supervised Classification Algorithms in Classifying SENTINEL-2A Images on Land Cover and Land Uses of a Part of Baghdad, Iraq

Nowadays, remote sensing (RS) technology is being recognized as the most up-to-date data source used in land cover and land use exploration. However, the sources of RS data, the type of algorithms applied, and the study area characteristics all influence how accurately relevant thematic maps are produced. This paper aims to assess the classification performance of supervised machine learning algorithms over regions affected by dust storm depositions. To this end, SENTINEL-2 imagery across the city of Baghdad, Iraq are considered. Data preprocessing is carried out using SNAP software. Maximum Likelihood (ML), Minimum Distance (MD), Random Forest (RF), K-Nearest Neighbor (KNN) and KD Tree KNN (KD-KNN) models are applied to produce land-use maps of five classes. Finally, a confusion matrix is implemented to examine the classification accuracy. The results revealed the superiority of the RF classifier with an overall accuracy of 83.33%. Followed by KNN and KD-KNN with an accuracy of 80% each. Arid lands scored the highest rating in most of the applied algorithms. The reported findings raise awareness regarding the selection of suitable algorithms for specific classification tasks.