Comparison of Deep Learning Classification Based Methods with Hyper Parameter Tuning on Hyperspectral Imagery

Abstract

Hyperspectral Imagery (HSI) with classification is the most significant and dominant area in remote sensing. The HSI's profuse information is used for various applications in mineralogy, agriculture, physics, astronomy, chemical imaging, surveillance, and environmental sciences. In this paper, a comparison of classification approaches on the hyperspectral image dataset of Indian Pines is carried out. The three methods considered for the study. First one is CNN-Convolutional Neural Network is used for encoding pixel's spectral and spatial information and conducted the classification. Second is the MLP-Multi-Layer Perceptron is a typical kind of neural network with several layers for classification and third is HybridSN is a spectral-spatial 3D-CNN along with spatial2D-CNN. A comparison of these deep learning classification-based methods is piloted. Hyperparameter tuning is performed to increase the model accuracy for the ideal model architecture.