Enhancing Satellite Imagery Resolution for Coastal and Ocean Engineering Applications Using Sub-pixel Convolutional Neural Networks and PixelShuffle Techniques

Deep neural networks (DNNs) for super-resolution (SR) address the limitations of low-resolution satellite imagery in coastal and ocean engineering analysis. Single Image Super-Resolution (SISR), an essential image restoration technique, reconstructs high-resolution (HR) images from low-resolution (LR) inputs. This is essential for applications such as environmental monitoring, where high-resolution data is critical, yet hardware/cost-limited. Image super-resolution (SR) evolved from traditional methods (e.g., prediction-based, edge-based, statistical, patch-based, sparse dictionary). The advent of deep neural networks (DNNs), particularly convolutional neural networks (CNNs), marked a breakthrough in SISR performance. Recent models (SRCNN, SRGAN, ESPCN, EDSR) leverage CNNs and Generative Adversarial Networks (GANs). These DNN-based models learn complex mappings from large datasets to reconstruct finer details, enhance edges, and reduce artifacts. In coastal and ocean engineering, SR has diverse and significant applications. Enhanced resolution enables more accurate mapping and monitoring of coastal areas, allowing the detection of subtle changes. SR provides engineers and scientists with higher-fidelity data to inform decisions and develop sustainable management strategies. Our research improves the accuracy of shoreline detection from freely available satellite imagery. The limited spatial resolution of existing methods hinders the precise extraction of shorelines. Using DNN-SR models, we significantly enhance spatial image detail, resulting in more refined and accurate mapping of the land-water interface, which improves coastal monitoring and management. We first down-sampled the 3-meter original PlanetScope imagery to 6-meter for DNN training. This is a standard technique to generate LR-HR pairs for training DNN-SR models. Specifically, this work focuses on upscaling LR PlanetScope imagery to HR (a 2x factor) using subpixel convolutional neural networks and pixel shuffle techniques for efficient upsampling. The ultimate goal is to apply this technique to enhance the resolution of freely available Sentinel-2 images for improving long-term historical analysis.