Coconut fields classification using data mining on a large database of high-resolution Ikonos images

Abstract

Supervised classification of satellite images is a commonly used technique in Remote Sensing. It allows the production of thematic maps based on a training set chosen by domain experts. These training sets, called ROI (Regions Of Interest), statistically characterize each class (e.g. coconut, sand) of the satellite image. Thus, a set of ROI is manually created by domain expert for each image. When a large number of images with high resolution occurs, manual creation of ROI for each image can be very time and money consuming. In this paper, we propose a semi-automatic approach based on clustering to limit the number of ROI done by experts. Then, we use decision trees on a binary decomposition of RGB components to improve the classification. Experiments have been done on 306 high resolution images of Tuamotu archipelago.