Assessing the performance of machine learning algorithms in Google Earth Engine for land use and land cover analysis: A case study of Muğla province, Türkiye

Abstract

Regions with high tourism density are very sensitive to human activities. Ensuring sustainability by preserving the cultural characteristics and natural structure of these regions is of critical importance in order to transfer these assets to the future world heritage. Detecting and mapping changes in land use and land cover (LULC) using innovative methods within short time intervals are of great importance for both monitoring the regional change and making administrative planning by taking necessary measures in a timely manner. In this context, this study focuses on the creation of a 4-class LULC map of Muğla province over the Google Earth Engine (GEE) platform by utilizing three different machine learning algorithms, namely, Support Vector Machines (SVM), Random Forest (RF), and Classification and Regression Tree (CART), and on comparison of their accuracy assessments. For improved classification accuracy, as well with the Sentinel-2 and Landsat-8 satellite images, the Normalized Difference Vegetation Index (NDVI) and Normalized Difference Water Index (NDWI) are also derived and used in classification of the major land use classes, which are ‘built-up area & barren land’, ‘dense vegetation’, ‘water surface’, and ‘shrub, grassland & sparse vegetation’. Experimental results show that the most relevant algorithm is RF with 0.97 overall accuracy and 0.96 Kappa value, followed by SVM and CART algorithms, respectively. These results indicate that the RF classifier outperforms both SVM and CART classifiers in terms of accuracy. Moreover, based on the results of the RF classifier, 19% (2,429 km2) of the study region is classified as built-up area & barren land, 48% (6,135 km2) as dense vegetation, 2% (301 km2) as water surface and 30% (3,832 km2) as shrub, grassland & sparse vegetation class.