Further Exploring Convolutional Neural Networks' Potential for Land-Use Scene Classification

Recently, with the success of deep convolutional neural networks (CNNs), many end-to-end learning algorithms have yielded excellent results. However, in the field of land use and land cover (LULC), very deep CNNs cannot be driven with even tens of thousands of images. In contrast to transferring methods that only employ a model pretrained with an irrelevant data set (e.g., ImageNet) and directly inherit parameters without refining, we explore an approach for effectively driving a deep CNN with a small data capacity. We propose a novel concept called the best activation model (BAM) in the end-to-end process for LULC image classification. BAM theoretically represents the best activation status for end-to-end networks with a small data set, taking both the data-capacity limitation and target-scene specificity into full consideration. The proposed method overcomes the problem of under-fitting and has optimal scene specificity for LULC scenes. Our approach greatly improves the time efficiency and yields excellent performance compared with state-of-the-art methods, obtaining averages of 99.0%, 98.8%, and 96.1% on the UC Merced Land-Use, WHU-RS19 data sets, and the Google data set of SIRI_WHU, respectively.