An evolutionary fuzzy classifier for satellite image classification

A boosted genetic fuzzy classifier (BGFC) is presented in this paper for land cover classification from VHR multispectral images. The proposed system is constructed through a two stage process. The first stage incrementally generates a fuzzy rule base, through repeated invocations of an evolutionary rule extracting algorithm. A boosting algorithm is employed, forcing new rules to focus on uncovered regions of the feature space. Each fuzzy rule is allowed to choose the features that best describe the patterns its antecedent cover, thus implementing a local feature selection mechanism. The resulting rule base is further fine-tuned in a subsequent stage, namely the genetic tuning stage, through an evolutionary algorithm (EA), increasing the classification performance of the final model. The BGFC is tested using an IKONOS satellite image, in a lake-wetland ecosystem of international importance. The results demonstrate the proposed system´s capabilities in handling complex multi-dimensional classification tasks.