2 sonuçlar
Arama Sonuçları
Listeleniyor 1 - 2 / 2
Yayın Deep learning techniques for building density estimation from remotely sensed imagery(Işık Üniversitesi, 2019-04-05) Süberk, Nilay Tuğçe; Ateş, Hasan Fehmi; Işık Üniversitesi, Fen Bilimleri Enstitüsü, Elektronik Mühendisliği Yüksek Lisans ProgramıThis thesis is about point-wise estimation of building density on the remote sensing optical imageries by applying deep learning methods. The goal of the project is to reduce mean square error of the estimated density by applying architectural modi?cations on the deep learning network and using augmented training data. Recently, deep learning is one of popular ?eld of science and convolutional neural networks (CNNs) are well-known deep neural network. Recent studies indicate that some of the convolutional neural networks are highly e?ective in large scale image works such as recognition, semantic segmentation. There has been limited research in using deep networks to learn urbanization characteristics from remote sensing images. Remote sensing images could be used for regression problems and building density estimation is one of them. Building density information provides knowledge for real estate agents and urban planners, estimating disaster risk areas, environment protection and resource allocation. Our method provides a cheap and fast solution to these needs when there is no cadastral information. The main objective of this thesis is to achieve fast and accurate local building density estimation using high resolution remote sensing images. Deep learning methods based on CNN are applied in this project. Pre-trained visual geometry group (VGG-16) and fully convolutional network (FCN) are tested as convolutional neural network. We tested three di?erent modi?ed networks and then applied data augmentation in the train data to reduce mean square error value. The networks that we have performed simpli?ed original VGG-16 network for regression, VGG-16 network with sigmoid layer added and simpli?ed VGG-16 network with sigmoid layer. The best result (lowest mean square error) is obtained from sigmoid layer added VGG-16 network with data augmentation. Sigmoid layer added VGG-16 network gives us (?0,084) RMSE on building density estimation with the augmented train dataset. Original VGG-16 network gives (?0,105) RMSE, sigmoid layer added VGG-16 network gives (?0,095) RMSE and sigmoid layer added simpli?ed VGG-16 network gives (?0,090) RMSE on building density estimation with the small train dataset. FCN is one of the ideal network for classi?cation tasks so we have also applied fully convolutional network result to compare our results with its result. We have modi?ed the network to perform building density estimation in addition to semantic segmentation. The root mean square error of FCN is (?0,084) and our best result (lowest mean square error) is also (?0,084) RMSE at the same iteration number. Our results show that fast and accurate building density estimation is possible by using vanilla CNNs. Sigmoid layer addition, simpli?cation of the network for small dataset and data augmentation improves accuracy in the regression. Data augmentation is the most e?ective method to reduce RMSE in this thesis.Yayın Uzaktan algılanan görüntülerde bina yoğunluğu kestirimi için derin öğrenme(Institute of Electrical and Electronics Engineers Inc., 2019-09) Süberk, Nilay Tuğçe; Ateş, Hasan FehmiBu bildiri, derin öğrenme yöntemleri uygulayarak uzaktan algılamalı optik görüntülerde bina yoğunluğunun noktasal olarak kestirilmesi ile ilgilidir. Bu çalışma kapsamında, evrişimsel sinir ağına (ESA) dayalı derin öğrenme yöntemlerine başvurulmuştur. Önceden eğitilmiş, VGG-16 ve FCN-8s derin mimarileri bu probleme uyarlanmış ve ince ayar verilerek eğitilmiştir. Kestirilen değerler yerleşim bölgelerinde bina yoğunluk haritası oluşturmak için kullanılmıştır. Her iki mimarinin karşılaştırmalı benzetim sonuçları, güdümlü eğitim için binaları gösteren detaylı haritalara ihtiyaç duyulmadan hassas yoğunluk kestirimi yapılabileceğini göstermektedir.
