Remote Sensing Scene Classification with Lightweight Model

Abstract

Remote sensing has significantly progressed and are becoming more readily available than ever. There is a growing requirement for the automatic recognition and labelling of these images under various conditions. Many studies have applied deep learning networks like CNN to satellite image recognition. However, the cost of the models is high, which means they are inappropriate when computing resources are limited, for instance, satellites and mobile devices. Moreover, the quality of available satellite images varies, and the ability of the network to adapt to low-quality images is also critical. To solve these problems, this paper uses mobile-net V2 and traditional CNN networks to classify 21 kinds of satellite images from a public dataset and compare the results. Firstly, the dataset is increased from 100 to 500 per class through data augmentation. Secondly, mobile-net V2 is trained and then the performance is evaluated using the test set. Additionally, the quality of the images is reduced to figure out the influence of each model. To verify the effectiveness and accuracy of mobile-net V2, several traditional CNN networks are compared with validation data accuracy, test data accuracy, inference time, peak RAM, and flash usage. The experimental results show that the mobile-net V2 network is a low-cost and well adaptability model with high accuracy for this multiclassification work.