Comparison of CNNs for Remote Sensing Scene Classification

Nowadays, deep learning are used widely in many applications related to remote sensing i.e. earth observation, urban planning, earth's scene classification, and so on. The deep learning manner, especially CNNs, has proved its accuracy for these practical applications. Hence, in this article, CNNs models are reviewed and its five different architectures are applied for comparisons; namely, AlexNet, VGGNet, GoogleNet, Inception-V3, and ResNet-101. These models are carried out on seven different remote-sensing image datasets for image scene classification purpose; namely, WHU-RS19, UC-Merced Land Use, SIRI-WHU, RSSCN7, AID, PatternNet, and NWPU-RESISC45. These datasets have different spatial resolutions, ranging from 0.2 to 30, to differentiate the classification accuracy of the low and high resolution images. As well, the classification accuracy of each model is assessed by trying five different classifiers; namely, Naïve Bayes, Decision Tree, Random Forest, K-Nearest Neighbor (KNN), and Support Vector Machine (SVM). The best accuracy credits to ResNet-101 model with SVM classifier; it has reached about 98.6±0.02 % of the high resolution dataset, PatternNet.