Comparison of Methods of Artificial Neural Networks, Support Vector Machine and Decision Tree to Identify Clouds in Landsat 8 Satellite Images

This article compares the physical and data mining techniques in cloud detection. Presence of Clouds in optical satellite images, requires radiometric pre-processing in remote sensing applications. Usually it is possible to identify clouds in satellite images using supervised classification. In this article Landsat 8 satellite images of two areas in Alborz Mountains, using artificial neural network, support vector machine (SVM) and decision tree were classified and performance of these three methods in terms of overall accuracy, cloud and snow producer accuracy and the kappa coefficient were compared. In the first area, neural network and SVM showed higher overall accuracy and cloud producer accuracy than decision tree and in the second area, decision tree showed better performance for cloud identification. This indicates that in the first area the thresholds used in the decision tree, do not detect clouds well. While in the second area indices listed before, show higher numbers. The results show that decision tree can achieve comparable to or higher accuracy in comparison to data mining techniques (such as neural networks and SVM), however thresholds used in this method may not be suitable for all areas. In contrast, data mining methods, especially SVM can identify different cloud types in two classes’ case, increase in the number of classes can reduce cloud detection accuracy in these methods.