Impact of Gradient Descent Optimization Techniques for Classification of Hyperspectral Remotely Sensed Images Using 3D CNN

Rapid technological advances has seen massive developments in the application domains such as satellite, aerial and Unmanned Aerial Vehicles remote sensing. In the era of Fourth industrial revolution, both Artificial Intelligence and RS technologies play important role in society around the globe. The acquisition of RS image data from satellite RS can be in a form of visible, infrared, multispectral, and hyperspectral. Various resolution are derived from this different image acquisition sources. Hyperspectral RS however has emerged as a great topic in the RS research society at large. Deep Learning algorithms such as Convolutional Neural Networks (CNN) have gained a lot of attention for image classification tasks. CNN makes use of gradient descent for optimization such as Adam, SGD, Adamax, Adadeta, Nadam and Adagrad. The gradient-based algorithms affect the output performance of the classification algorithm. In this study, we evaluate the impact of six DL optimization techniques, and the model performance of a 3D CNN in terms of accuracy. The experimental simulation is performed using publicly available hyperspectral dataset, which is called University of Pavia (UP). To avoid overfitting, dropout was utilized on the model. The results show that implementation of a 3D CNN HSI classification, performed better with Adam and Nadam. In future, we plan to run the optimization models to check the computational efficiency and fine-tune the models network structure and improved accuracy.