Sequence Image Datasets Construction via Deep Convolution Networks

Abstract

Remote-sensing time-series datasets are significant for global change research and a better understanding of the Earth. However, remote-sensing acquisitions often provide sparse time series due to sensor resolution limitations and environmental factors such as cloud noise for optical data. Image transformation is the method that is often used to deal with this issue. This paper considers the deep convolution networks to learn the complex mapping between sequence images, called adaptive filter generation network (AdaFG), convolution long short-term memory network (CLSTM), and cycle-consistent generative adversarial network (CyGAN) for construction of sequence image datasets. AdaFG network uses a separable 1D convolution kernel instead of 2D kernels to capture the spatial characteristics of input sequence images and then is trained end-to-end using sequence images. CLSTM network can map between different images using the state information of multiple time-series images. CyGAN network can map an image from a source domain to a target domain without additional information. Our experiments, which were performed with unmanned aerial vehicle (UAV) and Landsat-8 datasets, show that the deep convolution networks are effective to produce high-quality time-series image datasets, and the data-driven deep convolution networks can better simulate complex and diverse nonlinear data information.