On the analysis of deep convolutional neural networks applied to building detection in satellite images

Karagöz, Batuhan
Deep Learning has gained much interest recently, probably induced by the re- quirements to learn more complex and abstract concepts. As concepts to be learned become more abstract, their regions in the raw input space also become highly variational. In many cases, shallow architectures fail to learn highly varia- tional functions. One area of interest where concepts to be learned are complex is remote sensing. In this thesis, performance and suitability of deep architectures for recognition of building patches in satellite images are analyzed and discussed. Main architecture that is the subject of interest in this thesis is Deep Convo- lutional Neural Networks. Deep Convolutional Neural Networks has proven to be state of the art machine learning systems in several pattern recognition tasks such as bank check reading, handwriting recognition and face detection. We fo- cus on a particular CNN architecture and trained a Deep Convolutional Neural Network with fully supervised stochastic gradient descent. We obtained a clas- sification accuracy of 90 percent on average which is promising for deep learning implementations on the Remote Sensing Domain. Several measurements on the penultimate layer activations has been employed to reveal insights about what the models learn. Despite seemingly high accuracy results, these measurements put forward that the architecture we pick is unable to learn high level features.