Remote Sensing Scene Classification Based on Improved GhostNet

Abstract

Nowadays, the design of convolutional neural network (CNN) models is getting deeper and wider. When traditional CNN is used to process limited data of remote sensing images, it will lead to overfitting. We will use lightweight and efficient models to classify remote sensing images. In order to improve the classification accuracy and reduce the intermediate parameters, we improved GhostNet and proposed a smaller CNN named Improved GhostNet. Meanwhile, we use image enhancement methods to enlarge the datasets and dropout, it will reduce the amount of parameters. We experimented on three datasets, such as AID, UC Merced, NWPU-RESISC45. Then, we used MobileNetV3-Small and GhostNet to compare with our CNN model. The classification accuracy of improved GhostNet achieves more than 91%, and the accuracy on the AID is improved by 2.05% compared to the original GhostNet. These results demonstrate the effectiveness and efficiency of improved GhostNet.