Buildings as one of the important man-made objects have various applications and need to be observed with aerial and satellite images. Deep learning models have often been used to automatically extract building footprints from aerial and satellite images. It is essential to evaluate and compare the features of different deep learning models in images with geometric and brightness variations. For this purpose, in this research the performance of three deep learning models called Mask-RCNN (Mask Region-based Convolutional Neural Network), U-Net and MA-(FCN) (Multi-scale Aggregation Fully Convolutional Network) is evaluated on two aerial and satellite datasets with F1-score and IOU metrics. The results of this research indicate that the model, quantity and quality of training samples and digital surface model affect the performance of these models. Also, using digital surface models alongside the 3 band RGB images is an effective way of improving the building footprint extraction with deep learning models. By using digital surface model, the IOU results of U-Net and MA-(FCN) models in building footprint extraction are increased 7. 46% and 5. 7% in satellite dataset and 3. 61% and 3. 34% in aerial dataset, respectively. U-Net and MA-(FCN) are more precise in building boundaries since they concatenate feature maps of encoder and decoder parts in producing final segmentation maps. Mask-RCNN is stable to overfitting because of using ResNet in its architecture.

In the last decade, several Convolutional Networks have been developed for the semantic segmentation, which have shown excellent performance in recognizing and labeling objects in images. Most of these Networks involve large-scale architectures that can detect tens or hundreds of predefined classes. With the exception of Fully Convolutional Networks, most applications use architectures that, after Convolutional layers, use a common classifier to classify the extracted features. In this paper, the method of converting a Network, which as classifier, has two flatten and dense layers (Fully connected), to a Fully Convolutional Network is described. The main advantage of this method is the ability to work on inputs of variable size and produce an output map instead of a number, which is the advantage of Fully Convolutional Networks. Newer models of the Deep Learning area generally use training images in which areas of interest are determined by masks, but in the proposed method only labeled images are given to the Network. The details of the proposed method are expressed in the form of a new problem of classification of boards with calligraphy of Shekasteh-Nastaliq and Suls, and classification of apple leaf diseases (as two-class problems) and the problem of identifying hand written Persian digits. For this purpose, first a Convolutional Network with the last Fully connected layer is designed and trained for square images. Then a new Fully Convolutional model is defined based on the previous model and the weights of the previous model are fed to the new model. The only difference between the two models is in the last layer, but the new model will be able to work on input images of any size. Experimental results show the efficiency of the proposed approach.

The development of automatic road and building detection systems in aerial imagery are always faced with challenges such as the appearance of buildings, illumination changes, imaging angles, and the density of roads and buildings in urban areas, to name a few. In recent years, employing multi-layered approach in artificial neural Networks, known as deep neural Networks, has attracted many researchers in this field (and the other fields alike), achieving stunning results. However, the use of Fully connected layers in this approach, significantly increases the average processing time and results in an overfitted model. In addition, in most of these methods, a single-class approach has been considered. That is, detecting the roads and the buildings from natural scenes is not possible at the same time, and therefore, it is necessary to build separate binary models for each of them. The main goal of this research is to design a new architecture by which the produced model can be able to simultaneously detect roads and buildings from natural scenes, and thus minimizing the complexity of the classification process. In addition, in the proposed architecture, excluding all Fully connected layers from the traditional multi-layered architectures is considered in order to reduce the average processing time. The results of the experiments performed on the Massachusetts dataset, show that the proposed architecture performs 38% faster than the other deep neural Network-based methods, and also increases the accuracy by an average of 2%.