Transfer learning for high resolution aerial image classification

Abstract

With rapid developments in satellite and sensor technologies, increasing amount of high spatial resolution aerial images have become available. Classification of these images are important for many remote sensing image understanding tasks, such as image retrieval and object detection. Meanwhile, image classification in the computer vision field is revolutionized with recent popularity of the convolutional neural networks (CNN), based on which the state-of-the-art classification results are achieved. Therefore, the idea of applying the CNN for high resolution aerial image classification is straightforward. However, it is not trivial mainly because the amount of labeled images in remote sensing for training a deep neural network is limited. As a result, transfer learning techniques were adopted for this problem, where the CNN used for the classification problem is pre-trained on a larger dataset beforehand. In this paper, we propose a specific fine-tuning strategy that results in better CNN models for aerial image classification. Extensive experiments were carried out using the proposed approach with different CNN architectures. Our proposed method shows competitive results compared to the existing approaches, indicating the superiority of the proposed fine-tuning algorithm.