EVALUATION OF MACHINE LEARNING BASED ALGORITHMS FOR DETERMINING LAND COVER CHANGE WITH MULTI-TEMPORAL SENTINEL-2 SATELLITE IMAGES

Land cover maps created with remote sensing images are used for observation in many areas such as urban planning, land resources management, and biodiversity changes. The study of land cover change has a very important place in order to properly analyze the rapidly developing urbanization and to make the right city planning due to the increasing population and the migration wave that has surrounded the whole world. This study investigates the land cover change in the Istanbul province of Turkey, which receives the most irregular and continuous migration. The study was conducted over a five-year period between 2017 and 2021 with the use of Sentinel-2A satellite images. The classification was performed by use of two machine learning algorithms, which are Classification and Regression Trees and Random Forest on Google Earth Engine. The first level land cover class definitions of the CORINE legend were used through analysis. A comparative evaluation of two algorithms reported the higher performance of the random forest algorithm in terms of class-based accuracy, precision, recall, and F1-Score metrics. When the land cover changes are examined, it was determined that the urban land cover class faced a significant increase between 2017 and 2021 at the expense of a reduction in forest class. The biggest reason for this change can be related to the completion of the 3rd Airport project in this period and the construction of new settlements and transportation networks in Istanbul. In addition, water and agricultural land classes showed fluctuations over years, which are related to seasonal variations and cultivation patterns and are considered temporal changes. These results revealed that the use of machine learning algorithms on multi-temporal satellite images provides annual and periodical determination of land cover changes and can be used as a guide for further planning activities.