Deep Feature Learning And Classification Of Remote Sensing Images

Abstract

Hyperspectral imaging has been largely utilized in applications involving remote sensing to describe the composition of thousands of spectral bands in a single scene. Hyperspectral images (HSI) require an accurate training model for extracting the characteristics of scenes presented in an image. Image learning models involving spectral resolution present major challenges because of the complex nature of image frames. Several attempts have been made to address this complexity. Nevertheless, these models have failed to retain a deeper understanding of hyperspectral images. Since there are mixed pixels, limited training samples, and duplicate data, so the deep learning method solves the problem.In this method, spectral values (for every pixel) of the hyperspectral images are sequentially fed into spectral long-short-term memory (LSTM) through several routes to study the spectral features. Most of the existing state-of-the-art models are based on spectral-spatial frameworks. The added spatial features add more dimensions to hyperspectral images. However, these classification models do not take advantage of the sequential nature of these images. Due to the presence of mixed pixels, limited training samples, and redundant data, the utilization of deep learning techniques addresses the problems. This paper describes a method for the classification of hyperspectral images through spectral-spatial LSTM networks. For extracting the first principal constituent from such an image, principle component analysis (PCA) was used in spectral and spatial joint feature networks (SSJFN), as well as spectral and spatial individual extraction of the features via LSTM, to get the uniform end-to-end network. Furthermore, it was aimed to achieve the integration of all processes in a neural network by making a classifier to overcome the training error and backpropagation, which may lead to learning more features. During categorization, SoftMax classification considers the spatial and spectral characteristics of all the pixels independently to get two different outcomes. Afterwards, joint spectral-spatial results are gained by using the strategy of decision fusion. The classification accuracy improves by 2.69%, 1.53%, and 1.08% when compared to the rest of the state-of-art methods.