Comparison of Artificial neural network, Support vector machine and Random forest in mapping and detection of urban land use/ land cover change: The case of Dilla town

Abstract The accurate classifications of land use / cover (LULC) of satellite images are very important for monitoring the environment, planning of land use and detection of climate change. This paper compares the effectiveness of three machine learning algorithms for LULC classification using Google Earth explorer images from 2006 to 2022. The average kappa coefficient accuracy of ANN, SVM and RF was 0.81, 0.96 and 0.97 respectively. All classifiers performed more stably over different training sizes, but the RF classifiers achieved the highest precision in all of criteria. Change detection analysis showed that Agroforestry into urban was the largest change in the period, accounting for 2.28Km2. Between 2006 to 2022 study periods, 1.66km2 of land was changed in to urban area, a significant increase compared to the rest of LULC change. In this study field survey data, reference image, local community information and previous studies were used as reference to confirm the validation of the result. The study concluded integration of machine learning algorithm, especially RF classifier with object based approach is an effective means for land use and land cover mapping and change detection analysis. Therefor it recommended, the coming researchers to apply this effective technology in solving LULC related problem.