Performance Comparison of CNN Based Hybrid Systems Using UC Merced Land-Use Dataset

Abstract

Remote sensing is the technology of collecting and examining data about the earth with special sensors. The data obtained are used in many application areas. The classification success of remote sensing images is closely related to the accuracy and reliability of the information to be used. For this reason, especially in recent studies, it is seen that Convolutional Neural Network (CNN), which has become popular in many fields, is used and high successes have been achieved. However, it is also an important need to obtain this information quickly. Therefore, in this study, it is aimed to get results as successful as CNN and in a shorter time than CNN. Hybrid systems in which features are extracted with CNN and then classification is performed with machine learning algorithms have been tested. The successes of binary combinations of two different CNN architectures (ResNet18, GoogLeNet) and four different classifiers (Support Vector Machine, K Nearest Neighbor, Decision Tree, Discriminant Analysis) have been compared with various metrics. GoogLeNet & Support Vector Machine (93.33%) is the method with the highest accuracy rate, while ResNet18 & Decision Tree (50.95%) is the method with the lowest accuracy rate.