Evaluating Different Machine Learning Methods on RapidEye and PlanetScope Satellite Imagery

Sa？etak Since the first satellite imagery of RapidEye and PlanetScope became available, numerous studies have been conducted. However, only a few authors have focused on evaluating the accuracy of more than two machine learning methods in land cover classification. This paper evaluates the accuracy of four different machine learning methods, namely: support vector machine, artificial neural network, naive Bayes, and random forest. All analysis was conducted on cities in Croatia, Vara？din and Osijek. On Vara？din area on RapidEye satellite imagery support vector machine achieved overall kappa value 0.80, artificial neural network 0.37, naive Bayes 0.84 and random forest 0.76. On Vara？din area on PlanetScope satellite imagery support vector machine achieved overall kappa value 0.77, artificial neural network 0.38, naive Bayes 0.76 and random forest 0.75. On Osijek area on RapidEye satellite imagery support vector machine achieved overall kappa value 0.75, artificial neural network 0.36, naive Bayes 0.85 and random forest 0.76. On Osijek area on PlanetScope satellite imagery support vector machine achieved overall kappa value 0.64, artificial neural network 0.23, naive Bayes 0.72 and random forest 0.63. Performance time of each method is also evaluated. Naive Bayes and random forest have best performance time in every scenario