Automatic extraction of urban objects from high-resolution aerial images using convolutional neural networks (Study area: Ahvaz city)

Automatic extraction and reconstruction of these objects from aerial and satellite data can diminish the role of humans in the production of large-scale urban spatial information and reduce the cost and time, reduce their production drastically. Identifying target objects from aerial and satellite images using computer image processing techniques and neural networks is one of the most widely used algorithms for identifying urban features. Among these types of networks are Convolution Neural Networks, which have a high power to extract high-level features in all kinds of images. In this research, more than 850 educational examples of important urban structural features including buildings, roads and single trees have been prepared. In this research, among the different types of existing networks, due to the higher speed of processing, Convolution Neural Network based on one-shot multi-box detector with ResNet network base has been used in order to automatically discover and extract the range of these complications. The proposed model is trained by training samples based on 85% training data and 15% validation data, with 120 iterations and 93% accuracy. The trained model has been implemented on different images of the study area to discover the target complications. The evaluation results of the model are based on the accuracy criterion equal to 0.86, the recovery criterion equal to 0.82, and the F1Score criterion equal to 0.83. The results show that the proposed algorithm can be recruited in fields such as producing and updating large-scale urban maps and change detection in urban areas.