摘要： The paper presents an original neural network approach for region of interest detection and classification in multi-spectral satellite images. The proposed method uses a sequence of Pulse Coupled Neural Networks that identifies plausible regions of interest. These regions are passed to a dimension reduction algorithm, Principle Component Analysis, in order to generate the input data for a Support Vector Machine classifier, that validates the data. The algorithm's parameters are optimized using a Genetic Algorithm. The algorithm is designed to distinguish regions that are extremely similar, such as parks in a city that has entire districts made up of houses with yards. The algorithm has been tested on images provided by the Sentinel-2 satellite, and it proved that it can recall 76.85% of the pixels marked as park in the ground truth data, which was obtained from OpenStreetMap.