Geospatial object recognition using deep networks for satellite images

---

Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Orta Doğu Teknik Üniversitesi, Mühendislik Fakültesi, Elektrik ve Elektronik Mühendisliği Bölümü, Türkiye

Tezin Onay Tarihi: 2018

Öğrenci: ONUR BARUT

Danışman: ABDULLAH AYDIN ALATAN

Özet:

Deep learning paradigm has been drawing significant interest during the last decade due to the recent developments in novel machine learning algorithms and improvements in computational hardware. Satellite image analysis is also an important scientific area with many objectives, such as disaster and crisis management, forest cover, road mapping, city planning, even military purposes. Spatial correlations of land cover or geospatial objects between different images lead to widely utilization of convolutional neural networks (CNN) in the literature for detection and classification of land cover and geospatial objects with manmade structures. Noting the increase in the availability of high resolution satellite images due to the existence of many new satellites up in the orbit, data to be used during train of such networks is also more reachable compared to the past. In this study, three main research directions for satellite image analysis is examined and tested through simulations. Land cover classification of image patches are first achieved by using conventional CNNs. Next, the segmentation of natural scenes into different land cover is obtained by deep networks that are capable of providing segment labels for every pixel on the image. Finally, detection of geospatial objects on satellite images is obtained by conventional object detection techniques based on deep networks. For all these purposes, a multispectral satellite image dataset is manually labeled for several geospatial and human-made objects. The effect of different architectures, training techniques and the training parameters are examined through simulations.