Image super-resolution with dense-sampling residual channel-spatial attention networks for multi-temporal remote sensing image classification

Abstract

Image super-resolution (SR) techniques can benefit a wide range of applications in the remote sensing (RS) community, including image classification. This issue is particularly relevant for image classification on time series data, considering RS datasets that feature long temporal coverage generally have a limited spatial resolution. Recent advances in deep learning brought new opportunities for enhancing the spatial resolution of historic RS data. Numerous convolutional neural network (CNN)-based methods showed superior performance in terms of developing efficient end-to-end SR models for natural images. However, such models were rarely exploited for promoting image classification based on multispectral RS data. This paper proposes a novel CNN-based framework to enhance the spatial resolution of time series multispectral RS images. Thereby, the proposed SR model employs Residual Channel Attention Networks (RCAN) as a backbone structure, whereas based on this structure the proposed models uniquely integrate tailored channel-spatial attention and dense-sampling mechanisms for performance improvement. Subsequently, state-of-the-art CNN-based classifiers are incorporated to produce classification maps based on the enhanced time series data. The experiments proved that the proposed SR model can enable unambiguously better performance compared to RCAN and other (deep learning-based) SR techniques, especially in a domain adaptation context, i.e., leveraging Sentinel-2 images for generating SR Landsat images. Furthermore, the experimental results confirmed that the enhanced multi-temporal RS images can bring substantial improvement on fine-grained multi-temporal land use classification.