Deeply Fine-Tune a Convolutional Neural Network in Remote Sensing Image Classification: Easter Africa Countries (EAC)

Abstract

Remote sensing is resource data accessible and easy to get in different areas without time-consuming. The traditional image recognition task was unlimited to better classification. A convolutional neural network (CNN) was introduced to improve remote sensing image classification accuracy by eliminating the intra-class and class similarity. Training CNN from scratch requires a large annotated dataset that is occasional in the remote sensing area. Transfer learning of CNN weights from another large non-remote sensing dataset can occasionally help overcome typical RS image applications. Transfer learning consists of fine-tuning CNN layers to better the new dataset. In this paper, all of the experiments were done on nine categories for dataset collected in east Africa community countries (EAC) using three state-of-the-art architectures based on the effect of fine-tuning and pre-trained weights of CNN. Results indicate that fine-tuning the entire network is not always a significant way; we compared it with a process of using VGG16-DensNet pre-trained weights and RF as machine learning classified results can be improved up to 97.60. Alternatively, fine-tuning the top blocks can save computational power and produce a more robust classifier.