Seyreklik Temelli Sınıflandırmanın Hiperspektral Görüntülerde Uygulamaları
Seyreklik Temelli Sınıflandırmanın Hiperspektral Görüntülerde Uygulamaları
thumbnail.default.placeholder
Tarih
2015-07-08
Yazarlar
Çağlar, Halil
Süreli Yayın başlığı
Süreli Yayın ISSN
Cilt Başlığı
Yayınevi
Fen Bilimleri Enstitüsü
Institute of Science and Technology
Institute of Science and Technology
Özet
Teknolojinin gelişmesine paralel olarak sayısal algılamada kullanılan sensörler de gelişmektedir. Ölçülen ve üretilen verilerin kullanılabilmesi için aynı oranda iletim hızının ve saklama kapasitesinin artması beklenmektedir. Ancak sensör teknolojileri hızla ilerlerlerken aynı kapasitede bu iki teknoloji gelişememektedir. Her yıl bir önceki yıldan yüzde 58 daha fazla veri üretilmektedir. Sıkıştırılmış algılama bu sorunu çözmek için geliştirilmiştir.Sıkıştırılmış algılama sinyalin elde edilmesinde ve geri çatılmasında kullanılan etkili bir sinyal işleme tekniğidir. Sıkıştırılmış algılama günümüzde önemli ve çok popüler bir sinyal işleme modeli olarak kullanılmaktadır. Uzaktan algılamada yeryüzü yüzeyinden alınan görüntüler bant çözünürlüklerine göre sınıflandırılabilmektedirler. Hiperspektral görüntüler yüzlerce dar banttan oluşurlar ve çok yüksek miktarda spektral bilgi içerirler. Hiperspektral görüntülerde yer alan daha fazla spektral bilgi alındığı bölgenin daha iyi karakterizasyonu yansıtmaktadır. Arazi örtüsünün tanımlanmasında, ayırt edilmesinde ve sınıflandırmasında doğruluğu ve gürbüzlüğü diğer görüntülere göre daha yüksektir. Sınıflandırma uzaktan algılanarak elde edilen görüntülerin yorumlanabilir haritalara çevrilmesini sağlayan geliştirme sürecidir. Ayrıca, sınıflandırma coğrafi bilgi biliminin görüntü işlemede en önemli yönüdür. Bu tez kapsamında hiperspektral görüntüler üzerine seyreklik tabanlı sınıflandırıcılar uygulanması üzerine çalışılmıştır. Hiperspektral görüntüler üzerinde seyreklik yaklaşımını kullanan iki farklı algoritma, destek vektör makineleri ve en yakın komşuluk algoritmaları koşturularak yüzey örtüleri sınıflandırılmıştır. Açık kaynak olarak sunulan dört farklı veri seti kullanılmıştır. Hiperspektral veri setleri içerisindeki her ayrı sınıftan yüzde 10 eğitim için seçilirken geriye kalan yüzde 90 veri seti ise test verisi olarak kullanılmıştır. Başarımlar, doğruluk haritalarına göre karşılaştırılarak; her sınıf için ayrı ayrı ve genel ortalamalarda hesaplanmıştır. Seyreklik yaklaşımını kullanan iki algoritma gerçeklenmiştir. Gerçeklenen algoritmalardan biri sadece seyreklik yaklaşımını kullanırken diğer algoritma seyreklik yaklaşımının yanında destek vektör makinelerini de (Support Vector Machines – SVM) kullanan hibrit bir tekniktir. Ayrıca, bu hibrit algoritmada spektral bilgiler yanında uzamsal bilgiler de kullanılarak sınıflandırma başarımı arttırılmaya çalışılmıştır. Klasik sayılabilecek destek vektör makineleri ve en yakın komşuluk yöntemlerinin başarım çıktıları ile seyreklik tabanlı sınıflandırıcıların başarımları karşılaştırılmıştır. Seyreklik tabanlı sınıflandırıcıların klasik yöntemlerden daha iyi sonuçlar verdiği gözlemlenmiştir. Seyreklik yaklaşımında uzamsal bilgilerin de kullanılmasıyla sınıflandırma başarımının daha da arttığı görülmektedir. Seyreklik tabanlı algoritmaların işlem karmaşıklığının yüksek oluşundan dolayı çalışma süreleri klasik algoritmalardan daha fazla olduğu gözlemlenmiştir.
In parallel to the development of technology, sensor technology used in digital detection is also advancing rapidly. In addition, similar to the development of sensor technology, there has been a constant increase in the detection frequency and data density. Huge amount of data, which has been acquired using these sensors must get transmitted and stored. Transmission and storage technology could not grow as fast as the development of sensors. These sensors produce data that is 58 percent more compared to the data from the previous year. That is a modern problem of technology. Compressed sensing can solve this problem. Sampling theorem says that if the sampling rate of a signal is more than twice of the signal's highest frequency, then the signal can be recover perfectly. In compressed sensing, the signal can be recovered far fewer samples than required by the sampling theorem. Compressed sensing is an effective signal-processing tool for getting the signals and reconstruction of that signal. It is an important and popular signal processing module. In remote sensing, images, which are collected from satellite, can be classified with respect to their spectral band resolution. These types are panchromatic, multispectral and hyperspectral images. Panchromatic images are composed only one narrow spectral band. Unlike Panchromatic images, Multispectral images are composed of tens of narrow spectral bands. Hyperspectral images are composed of hundreds of narrow spectral bands. Hyperspectral imaging systems look at objects using a huge portion of the electromagnetic spectrum. This system covers infrared, visible and ultraviolet area of the electromagnetic spectrum. Hyperspectral images contain huge amounts of information. Objects leave signature in the electromagnetic spectrum. This signature enables identification of the materials that make up a scanned object. Hyperspectral images reflect the characterization of the imaged area better than other imaging modalities. The rich spectral information in hyperspectral images allow them to be used in many field. These images can be used in identification and classification of land cover. The accuracy and robustness achieved using hyperspectral images is higher than other kind of image types. Classification is the process, which translates and interprets a map from an image obtained using remote sensing. Land cover could be forested, urban, agricultural and other types of features. Additionally, the classification is one the most important applications of the geographical science. There are two main image classification techniques. These techniques are supervised and unsupervised classification techniques. In supervised classification techniques, there is train data, which used to train the classifier and get the desired outputs whereas in unsupervised classification techniques no train datasets is provided. All data is clustered into different classes. Therefore, unsupervised classification techniques also known as clustering methods. In that thesis, implemented algorithms are supervised technique. Four different sets of hyperspectral data sets are used. Data sets are Indian Pines, Salinas, Pavia and Pavia University, which are open source. All data sets are taken with using AVIRIS (Airborne Visible/Infrared Imaging Spectrometer) sensors. AVIRIS is an optical sensor that takes spectral radiance in 224 contiguous spectral bands with wavelengths from 0.4 to 2.5 micrometers. The main objective of the AVIRIS project is to identify and monitor to the Earth's surface. These images are used to measure the performance of the implemented classification algorithms. Ten percent of hyperspectral image is randomly chosen for training data and remaining ninety percent of image is used for test data. In this thesis, sparsity based classification methods have been applied to hyperspectral images. There are four algorithms, which are applied to hyperspectral images. Two of these methods are sparsity based while other methods are support vector machines and k-nearest neighbors algorithms. There are two sparsity based algorithms implemented in this thesis. One of the implemented algorithms is based on only sparsity approach. First algorithm takes all the train data as an over-complete dictionary. Other algorithm, which named spatial-sparsity classification method, learns the over-complete dictionary from train data sets. In addition, this method uses a hybrid model that combine support vector machines with sparsity based approach. Additionally, this hybrid technique uses both spectral and spatial information. That approach tries to increase classification performance. The accuracy and robustness of sparsity based classification methods are compared with support vector machine and k-nearest neighbors algorithms. The performance of sparsity based classification algorithms in hyperspectral images has been observed to be better than these classical algorithms. The sparsity based classification algorithms have high complexity but promising performance. This complexity causes to increase run time of the algorithm.
In parallel to the development of technology, sensor technology used in digital detection is also advancing rapidly. In addition, similar to the development of sensor technology, there has been a constant increase in the detection frequency and data density. Huge amount of data, which has been acquired using these sensors must get transmitted and stored. Transmission and storage technology could not grow as fast as the development of sensors. These sensors produce data that is 58 percent more compared to the data from the previous year. That is a modern problem of technology. Compressed sensing can solve this problem. Sampling theorem says that if the sampling rate of a signal is more than twice of the signal's highest frequency, then the signal can be recover perfectly. In compressed sensing, the signal can be recovered far fewer samples than required by the sampling theorem. Compressed sensing is an effective signal-processing tool for getting the signals and reconstruction of that signal. It is an important and popular signal processing module. In remote sensing, images, which are collected from satellite, can be classified with respect to their spectral band resolution. These types are panchromatic, multispectral and hyperspectral images. Panchromatic images are composed only one narrow spectral band. Unlike Panchromatic images, Multispectral images are composed of tens of narrow spectral bands. Hyperspectral images are composed of hundreds of narrow spectral bands. Hyperspectral imaging systems look at objects using a huge portion of the electromagnetic spectrum. This system covers infrared, visible and ultraviolet area of the electromagnetic spectrum. Hyperspectral images contain huge amounts of information. Objects leave signature in the electromagnetic spectrum. This signature enables identification of the materials that make up a scanned object. Hyperspectral images reflect the characterization of the imaged area better than other imaging modalities. The rich spectral information in hyperspectral images allow them to be used in many field. These images can be used in identification and classification of land cover. The accuracy and robustness achieved using hyperspectral images is higher than other kind of image types. Classification is the process, which translates and interprets a map from an image obtained using remote sensing. Land cover could be forested, urban, agricultural and other types of features. Additionally, the classification is one the most important applications of the geographical science. There are two main image classification techniques. These techniques are supervised and unsupervised classification techniques. In supervised classification techniques, there is train data, which used to train the classifier and get the desired outputs whereas in unsupervised classification techniques no train datasets is provided. All data is clustered into different classes. Therefore, unsupervised classification techniques also known as clustering methods. In that thesis, implemented algorithms are supervised technique. Four different sets of hyperspectral data sets are used. Data sets are Indian Pines, Salinas, Pavia and Pavia University, which are open source. All data sets are taken with using AVIRIS (Airborne Visible/Infrared Imaging Spectrometer) sensors. AVIRIS is an optical sensor that takes spectral radiance in 224 contiguous spectral bands with wavelengths from 0.4 to 2.5 micrometers. The main objective of the AVIRIS project is to identify and monitor to the Earth's surface. These images are used to measure the performance of the implemented classification algorithms. Ten percent of hyperspectral image is randomly chosen for training data and remaining ninety percent of image is used for test data. In this thesis, sparsity based classification methods have been applied to hyperspectral images. There are four algorithms, which are applied to hyperspectral images. Two of these methods are sparsity based while other methods are support vector machines and k-nearest neighbors algorithms. There are two sparsity based algorithms implemented in this thesis. One of the implemented algorithms is based on only sparsity approach. First algorithm takes all the train data as an over-complete dictionary. Other algorithm, which named spatial-sparsity classification method, learns the over-complete dictionary from train data sets. In addition, this method uses a hybrid model that combine support vector machines with sparsity based approach. Additionally, this hybrid technique uses both spectral and spatial information. That approach tries to increase classification performance. The accuracy and robustness of sparsity based classification methods are compared with support vector machine and k-nearest neighbors algorithms. The performance of sparsity based classification algorithms in hyperspectral images has been observed to be better than these classical algorithms. The sparsity based classification algorithms have high complexity but promising performance. This complexity causes to increase run time of the algorithm.
Açıklama
Tez (Yüksek Lisans) -- İstanbul Teknik Üniversitesi, Fen Bilimleri Enstitüsü, 2015
Thesis (M.Sc.) -- İstanbul Technical University, Instıtute of Science and Technology, 2015
Thesis (M.Sc.) -- İstanbul Technical University, Instıtute of Science and Technology, 2015
Anahtar kelimeler
Arazi Sınıflandırma,
Seyrek Yaklaşım,
Land Classification,
Sparse Approximations