The effect of deep learning approach on classification performance in hyperspectral images

With the development of hyperspectral sensors, hyperspectral imaging has become a subject of interest in the field of remote sensing. Hyperspectral imaging, which allows measurement of reflected energy from the displayed surface materials at a narrow and adjacent plurality of wavelengths, provides extremely high dimensional and interrelated data. Storing, processing and interpreting and calculating this data is very difficult due to its complexity and processing load. Therefore, in the classification of hyperspectral data, size reduction methods are traditionally used as pre-processing step. However, conventional classifiers and dimension reduction methods are challenging in the spectral dimension and are inadequate in the extraction of distinctive features. There is also no definitive classifier and dimension reduction method selection method. In recent years, it has been a remarkable approach to classify the hyperspectral data with more robust, adaptable and extracted features from raw data by deep learning methods without reducing to subspace. Especially, the classification of hyperspectral images with convolutional neural networks, one of the deep learning methods, provides promising results. Within the scope of this thesis, samples of widely used hyperspectral data sets are classified by using one-dimensional, two-dimensional and three-dimensional convolutional neural networks by extracting spatial, spectral and spatial-spectral features. All the features provided by hyperspectral sensors are included in the classification effectively by using both separately and together spectral and spatial features. In addition, a comparative study and analysis is conducted between conventional classification and convolutional neural networks. Experimental studies have shown that convolutional neural networks have achieved very high classification rates. It has also shown that the proposed convolutional neural network architectures provide a better classification rate of 5% and 9% than the conventional methods.