Land Cover Classification of Satellite Imagery using Deep Learning

Abstract

In the assessment of remotely sensed imagery, hyper-spectral (HSI) image classifications are commonly employed.The hyper-spectral image includes various image bands.The Convolutional Neural Network (CNN) is an extensively useful deep learning algorithm for data visualization and processing.Both the Spatial along with Spectral information are important for HSI classes to be effective.Due to the higher computing complexity, only a few approaches have used 3D CNN.Hybrid Spectral Convolutional 2D-3D Network (HybridSN) is instituted for HSI classing in this paper.HybridSN involves a spatial and spectral 3D-CNN which is then trailed by a spatial 2D.A study of more abstract level spatial representation will continue with 2D-CNN over 3D-CNN.Furthermore, when compared to conventional CNNs, the employment of hybrid CNNs lessens the model's complexity.To see if this hybrid method works, a thorough HSI phase test was performed over Indian Pines, Salinas and Pavia University and results compared to ground truth.The presented HybridSN HSI classification model provides with best result.