Detection of airport in satellite images

In recent years, automatic detection of airport from satellite images has become a popular topic due to its importance in military and civil aviation fields. However, these images which have complex backgrounds such as buildings, mountains, rivers and roads, make it difficult to identify airports. In addition, since satellite images have a large size, it is crucially important to consider the complexity of the algorithm to the account in order to reduce the computational cost. In this thesis, two different approaches are proposed for automatic airport detection from satellite images. In the first method, initially the version of Line Segment Detector (LSD) we developed is used for the detection of potential airport candidate regions. Then efficient features are extracted from these candidate regions with Scale Invariant Feature Transform (SIFT). This features are fed into Fisher Vector (FV). In FV, Gauss Mixture Model (GMM) is used as visual dictionary. FV is obtained by combining first and second partial derivatives of all Gauss. FV is used as the feature vector and Support Vector Machine (DVM) is employed in the classifier stage. Accuracy, sensitivity and specificity criteria are used to evaluate the performance of the proposed method and yielded an accuracy of 94.6%. The superiority of the performance of the proposed first method has been proved by comparison with other studies previously suggested in the literature. We have also verified that the LSD we have developed for the detection of airport candidate regions has comparatively lower computational costs than other methods. In the second method, the usage of deep Convolutional Neural Networks (CNNs), which contain both feature extraction and classifier structure has been investigated for the automatic detection of airports from satellite images. Firstly, candidate airport regions have been detected by using the LSD method we developed. These regions are the same as the candidate regions in the first proposed method and used to train CNN architecture. The CNN model consists of five convolutions, three pools and three fully connected layers. Normalization and dropout layers have also been used to create an efficient architecture. Candidates in training data have been artificially increased to reduce overfitting. Accuracy, sensitivity and specificity criteria were also used to evaluate CNN performance. The second proposed method has accuracy of 95.21% and an increase of about 1% was achieved according to the first proposal.