Techniques based on support vector machines for cloud detection on QuickBird satellite imagery

Purpose of this work is the study of cloud detection techniques. This work identifies the cloud cover of optical images acquired by the QuickBird satellite, comparing these with others of the same area, acquired by Landsat 7 in which there are no clouds. The images are combined using an early fusion technique [1]. The tool exploits the neighborhood model [2] for increasing the amount of information for the training set and the Singular Value Decomposition for carrying out the feature extraction [3]. In order to introduce these structures into thematic classification tasks by SVMs it was necessary develop a tree kernel function based on tree kernel function defined in SVM-LightTK. The aim of the tree kernel function is evaluate the similarity level between a generic couples of tree structures.In this paper we report the results obtained comparing the performance of different approaches in cloud classification problem. The final purpose is the production of cloud cover maps. Throughout such different experimental setups we measured the capabilities of each algorithm under different points of view. First of all, we considered the classification accuracy by computing traditional parameter such as overall accuracy. A second analysis regarded the efforts that are required in the design of optimal algorithms. Indeed, these techniques are characterized by different parameters that have to be appropriately tuned in order to obtain the best performance. Finally the robustness of the techniques has been also considered. In particular the classification accuracy has been evaluated also for images not considered in the training phase