Convolutional Computation Performance Comparison for High Resolution Images

Extensive research is being conducted on high resolution satellite images in order to develop advanced artificial satellites and remote sensing equipment. When a deep neural network is trained using high resolution images, the training time is long due to the massive size of images and the convolutional computations performed by the filters. This study tested the convolutional performance over varying convolutional methods with filter sizes for global-scale satellite image data. The optimal filter size and convolutional method associated with the shortest training time were identified to apply the deep neural network in examining the concentration of chlorophyll.