Influence of Hyperparameter in Deep Convolution Neural Network Using High-Resolution Satellite Data

Abstract

In recent years, Deep Convolution Neural Network (DCNN) has evolved in the various fields of computer vision and performed state of the art in image classification task including feature extraction in high-resolution satellite imagery. During the training process, DCNN is required to achieve the highest accuracy but framework should be optimized with the help of hyperparameters (knobs) to reach that accuracy. In this study, we investigated the effect of various hyperparameters like batch size, learning rate, and other additional factors like test ratio w.r.t different variants of gradient optimizer algorithms, which can help to give the direction for the searching of better hyperparameter setting so that DCNN could achieve better classification accuracy with less computation time. Herein, we used SAT4 and SAT6 datasets to train the ALEXNET framework.