Contextual modeling of remote sensing images with conditional random fields

---

Tezin Türü: Yüksek Lisans

Tezin Yürütüldüğü Kurum: Orta Doğu Teknik Üniversitesi, Mühendislik Fakültesi, Bilgisayar Mühendisliği Bölümü, Türkiye

Tezin Onay Tarihi: 2013

Öğrenci: GÜLCAN CAN

Danışman: FATOŞ TUNAY YARMAN VURAL

Özet:

Large within-class variance is a challenging problem for classification tasks in remote sensing. Contextual models are promising to address this problem. In this thesis, a contextual conditional random field model is proposed for target detection in satellite imagery. The proposed algorithm has three stages. First, contextual cues of the target that come from domain knowledge are identified by sparse auto-encoders and shown to be statistically consistent. The region represented by the most repetitive feature learned by sparse autoencoders is used as central node in the proposed model and called candidate region. Other nodes of the model are chosen as land-use land-cover classes in the surroundings of the candidate regions, since the spatial context of the target class is defined over expected and unexpected classes in its neighborhood. Secondly, regions that represent these classes are obtained by merging segments with the same label according to support vector machines. These regions are called meta-segments. In the last stage, the same features are extracted from the meta-segments and candidate region to be used as unary features in the conditional random fields model. Pairwise features in conditional random fields are essential for representing contextual relations and they are designed as class co-occurrence frequencies in three di erent neighborhoods of the candidate region. For each candidate region, a dynamic conditional random fields model is generated. The proposed method is robust in terms of being threshold-free and selecting contextual cues via sparse auto-encoders. Performance of the method is competitive to rule-based methods and segmentation-based classification methods.