Deep Learning for Remote Sensing Image Superresolution: A comprehensive review and a new real-world benchmark

Abstract

Single-image superresolution (SISR) is a long-standing research topic in the low-level vision of remote sensing. Its goal is to reconstruct a corresponding high-resolution (HR) image from a low-resolution (LR) image. In recent years, SISR for remote sensing has attracted widespread attention due to its potential applications in small object detection, land cover classification, and precision agriculture, and it has coincided with the rapid development of deep learning (DL) methods. However, a comprehensive review of these methods and real-world datasets remains lacking. To address this gap, this article systematically and comprehensively reviews DL-based SISR papers from the past decade. Specifically, we first introduce the various DL network architectures used. Second, we provide a comprehensive summary of SISR algorithms for optical remote sensing images (RSIs), including methodology, systematic comparisons, the evolution timeline, commonly used loss functions, remote sensing datasets, and evaluation metrics. To further promote the field of remote sensing SR, we propose a global-scale, versatile, and abundant dataset for real-world SR, L8S2PS-SR, which consists of 195 precisely aligned Landsat 8–Sentinel-2–PlanetScope satellite image pairs captured at different seasons and geographical locations. Furthermore, we implement 15 representative methods on publicly available datasets and our proposed L8S2PS-SR dataset to evaluate their performance. Finally, some challenges and future research directions in the field of remote sensing SISR are discussed, aiming to provide research ideas for researchers. Our code and dataset are available at https://github.com/Meng-333/L8S2PS-SR.