Advanced Techniques For The Classification Of Remote Sensing Data

With the development of remote sensing sensors,hyperspectral remote sensing images are now widely available.These images are usually collected by airborne earth observation instruments.The classification of hyperspectral images(HSI)is a crucial step in environmental monitoring,disaster management,agriculture management and military applications.Some of the key challenges in classification techniques are high dimensionality of data,limited number of training samples,and combining the spatial and spectral information.HSI classification techniques that use both spectral and spatial information are more suitable,effective,and robust than those using only spectral information.However,manual extraction of these features is difficult and time-consuming,especially when dealing with classification problems where the objects of interest are larger than the pixel size.Classification of HSI can benefit from deep learning models with deep architecture in remote sensing.It enables us to replace handcrafted features by automatically extracting informative features from the data itself and skipping the tedious manual feature extraction.In this study,a new method based on Convolutional Neural Networks(CNN)is proposed for the classification of HSI.Due to using more spatio-spectral features for the classification of HSI,the proposed method outperforms the existing state-of-the-art classification techniques.Our proposed method first reduces the dimension of hyperspectral images using Principal component analysis(PCA)to improve accuracy and diversity.The spatial and spectral features are then exploited by a fixed size convolutional filter to generate the combine spatio-spectral feature maps.Finally,these feature maps are fed into a Softmax classifier that predicts the class of the pixel vector.To validate the effectiveness of our proposed method,simulations are conducted using three datasets namely Indian Pines,Salinas,and Pavia University and comparisons with existing techniques are made.Our proposed method achieved an accuracy of 96.13%,97.32%and 97.11%for Indian Pines,University of Pavia and Salinas datasets respectively.Furthermore,the effect of parameters on the accuracy of our proposed method(different training sets,amount of features in each subset,compactness,scale parameters,shape)is examined as well in this thesis.