Arbitrary-Scale Downscaling of Tidal Current Data Using Implicit Continuous Representation

TL;DR

This paper addresses the challenge of generating tidal current data at arbitrary resolutions without the heavy cost of numerical models. It introduces an implicit continuous representation-based downscaling framework comprising a feature extractor, an Arbitrary Scale Module, and an Auxiliary Train Module, augmented by Feature Map Splitting and learnable positional encoding to handle heterogeneity in tidal data. The method achieves substantial accuracy gains (MSE and MAE) and reduces compute (FLOPs) compared to a Baseline LIIF-based approach, and it can render outputs at scales not seen during training, including very high-scale out-of-distribution targets like 50x. The work has practical significance for coastal engineering and renewable energy planning by enabling fast, flexible generation of high-resolution tidal current fields.

Abstract

Numerical models have long been used to understand geoscientific phenomena, including tidal currents, crucial for renewable energy production and coastal engineering. However, their computational cost hinders generating data of varying resolutions. As an alternative, deep learning-based downscaling methods have gained traction due to their faster inference speeds. But most of them are limited to only inference fixed scale and overlook important characteristics of target geoscientific data. In this paper, we propose a novel downscaling framework for tidal current data, addressing its unique characteristics, which are dissimilar to images: heterogeneity and local dependency. Moreover, our framework can generate any arbitrary-scale output utilizing a continuous representation model. Our proposed framework demonstrates significantly improved flow velocity predictions by 93.21% (MSE) and 63.85% (MAE) compared to the Baseline model while achieving a remarkable 33.2% reduction in FLOPs.

Figures (5)

• Figure 1: Visualization of Tidal Current Data. Unlike images, tidal current data consists of two channels, U and V, representing the flow velocity, and one channel representing the water level. Both the features are highly correlated but show very different aspects.
• Figure 2: The overall framework of the proposed downscaling for tidal current data. Our framework consists of three parts: a feature extractor to extract features from low-resolution input, an arbitrary-scale module to predict arbitrary-scale downscaling prediction, and an auxiliary train module added for faster convergence and to prevent over-fitting of the feature extractor. Each output from the arbitrary-scale module and the auxiliary train module generates gradients and is used to train the feature extractor
• Figure 3: Feature extractor architecture. We use EDSR EDSR_2017 as a feature extractor and add learnable positional encoding before the first basic block.
• Figure 4: Train losses and validation MSE comparison between w/ and w/o the ATM. Both train losses and validation MSE are calculated between HR$^{arb}$ and GT.
• Figure 5: Visualization of the $\times$50 scale downscaling results. Green bounding boxes show overall texture and red bounding boxes show details of prediction. Since there is no huge difference between the INR-based methods(Baseline, Non-split, Ours) in texture, We only visualized Ours's result for the green bounding box between all results from the INR-based methods.