Land cover classification in an arid landscape of Iran using Landsat 8 OLI science products: Performance assessment of machine learning algorithms

Abstract

The spatial variation of several dynamic chemical soil characteristics is greatly influenced by land cover and land use. High-accuracy land use and land cover (LULC) classification have enormous promise for temporal scale evaluation of soil characteristics. The study aims to evaluate the performance of linear and non-linear classification methods in determining land cover classes by using remotely sensed time-series Landsat 8 OLI satellite data in an area where semi-arid agricultural activities are active. Four LULC classes were identified, and Landsat 8 images were classified using three supervised machine learning classifiers. When the producer’s accuracy, user’s accuracy, overall accuracy, and Cohen kappa coefficient were taken into account, it was observed that support vector machines (SVMs) and random forest (RF) algorithms produced more accurate results than multinomial logistic regression (MNLR). The SVMs had the highest overall classification accuracy of 96.00 % and a kappa coefficient of 0.93 on the test set. It is recommended to compare the efficiency of satellite data with different spectral and spatial resolutions.