Research Article
Deep Residual Network with Transfer Learning for High Spatial Resolution Remote Sensing Scenes Classification
@INPROCEEDINGS{10.1007/978-3-030-94551-0_26, author={Ziteng Wang and Wenmei Li and Guan Gui}, title={Deep Residual Network with Transfer Learning for High Spatial Resolution Remote Sensing Scenes Classification}, proceedings={Advanced Hybrid Information Processing. 5th EAI International Conference, ADHIP 2021, Virtual Event, October 22-24, 2021, Proceedings, Part I}, proceedings_a={ADHIP}, year={2022}, month={1}, keywords={Convolutional neural network Images classification High spatial resolution remote sensing Transfer learning}, doi={10.1007/978-3-030-94551-0_26} }- Ziteng Wang
Wenmei Li
Guan Gui
Year: 2022
Deep Residual Network with Transfer Learning for High Spatial Resolution Remote Sensing Scenes Classification
ADHIP
Springer
DOI: 10.1007/978-3-030-94551-0_26
Abstract
Deep residual network (DRN) is considered a promising image classification method for high spatial resolution remote sensing (HSRRS) images due to its great feature extraction capabilities. The classification performance of the DRN is greatly relies on the size of training samples. However, the sample size of HSRRS images is relatively small due to the high acquisition cost. Blindly increasing the sample size would requires huge computing resources and image annotation cost, but would not necessarily improve the classification performance of DRN. In this paper, a transfer learning-aided DRN method (TL-DRN) is proposed for a few shot learning to address the performance challenges associated with HSRRS with relatively small sample size and explore the impact of sample size on classification performance. In the experiment, the weights (shared knowledge) obtained by training the ImageNet datasets on the DRN model are transfered to the TL-DRN model. Experiments with ten different small-scale training sample sizes are conducted. Experimental results show that when the total training sample size is increased from 10 to 100, the classification performance of the TL-DRN model tends to be stable and the mean accuracy of its testing set has stabilized at around 94%. TL-DRN shows a superiority of up to 16% over DRN, in terms of classification accuracy.
