Classification of urban functional zones through deep learning

Nowadays, artificial neural networks (ANN) are models widely used in many areas; one of these is the classification of urban areas. This work aims to discuss a new framework for the delimitation of functional zones for the city of Naples through deep learning algorithms. More in detail, firstly, a segmentation approach is used to generate the urban zones from the satellite RGB image of interest; then, starting from an extrapolated OSM data, we develop a new labelled dataset used for the training of a convolutional neural network model. Finally, the urban zones are classified with a majority vote procedure. The innovative aspect of this methodology is the use of data provided for different purposes (that is, labelled OSM data) to compensate for the lack of data provided by experts in the field. For the experimentation, we compare two segmentation algorithms (FNEA and selective search) and three CNN models (AlexNet, ResNet-50 and a regularized version of AlexNet), providing good performances in the functional zone classification.