Automatic building detection from remotely sensed images with u-net based convolutional neural network

With the increase in spatial resolution of satellite and aerial imagery, it is allowed to recognize relatively small objects such as buildings and trees. Thus, studies on object recognition and object segmentation have started to be performed in aerial images. It may be necessary to reduce the huge spectral dimension to increase processing speed and building segmentation accuracy in HS image processing. For this purpose, it is aimed to select a subset of the bands containing the most information by performing the band selection process in HS images. Various similarity criteria and search methods in the literature are investigated and two band selection methods using hierarchical clustering by taking structural information in images into consideration are proposed. Also a new metric is defined that measures the similarity of the two bands in the HS image. The results were compared on two different data sets: satellite and close-range images. After the spectral band selection process, classical and modern image processing based studies are carried out for building segmentation, it is investigated that which methods are used as popular and the stage of the studies in the literature. In the thesis, deep learning based methods which became very popular in solving different problems in image processing, are investigated. Two efficient Unet-based object segmentation architectures are proposed. Two different data sets consisting of MS and HS images are used for performance comparison, and it is shown that better performances have been achieved with the proposed architectures as a result of fewer training steps. Also, the results are analyzed using the proposed band selection and building segmentation methods on the labeled HS data set.