CASCADE: Cross-scale Advective Super-resolution with Climate Assimilation and Downscaling Evolution
Alexander Kovalenko · 2026
Read CASCADE: Cross-scale Advective Super-resolution with Climate Assimilation and Downscaling Evolution by Alexander Kovalenko (2026) in the Tower of Babel, a free online library of public-domain works preserved in clean, searchable text. No sign-up. The full text is served as a plain-text file at a predictable URL, where text is present: https://antarctic-labs.com/tower-of-babel/arxiv-2603-00109.txt.
Introduction
Kovalenko presents CASCADE, a system that sharpens coarse geophysical fields by moving detail along learned flow paths instead of inventing pixel detail. It splits motion into large and subgrid parts, mirroring the closure problem in weather prediction, and keeps coarse consistency through an assimilation correction step. Tested on SEVIR radar data at 4x sharpening of severe convective storms, it beats baselines on continuous scores and threshold skill scores. Its visible velocity and correction fields make its workings interpretable, and its outputs stay coherent in time and true to mass conservation.
Catalog record
| CREATOR | Alexander Kovalenko |
|---|---|
| YEAR | 2026 |
| COLLECTION | PAPERS |
| FORMAT | TXT |
| FILE SIZE | 48544 |
| LICENSE | http://creativecommons.org/licenses/by/4.0/ |
| RIGHTS | CC BY |
| SOURCE | arXiv |
Source: https://arxiv.org/abs/2603.00109
Browse the full library · About the Tower of Babel
Full text
The opening of the work, from the library's hosted text:
CASCADE: Cross-scale Advective Super-resolution with Climate Assimilation and Downscaling Evolution Alexander Kovalenko Abstract Super-resolution of geophysical fields presents unique challenges beyond natural image enhancement: fine-scale structures must respect physical dynamics, conserve mass and energy, and evolve coherently in time. These constraints are especially critical for extreme events, where rare, localized, high-intensity features drive impacts and where temporally inconsistent “hallucinated” detail can misrepresent hazards. We introduce CASCADE (Cross-scale Advective Super-resolution with Climate Assimilation and Downscaling Evolution), a framework that reframes spatiotemporal super-resolution as an explicit transport process across scales. Rather than hallucinating high-frequency content per pixel, CASCADE reconstructs fine structure by iteratively advecting coarse information along learned, flow-conditioned velocity fields through semi-Lagrangian warping. The architecture decomposes motion into resolved (large-scale) and subgrid (unresolved) components, mirroring the closure problem in numerical weather prediction, and enforces low-resolution consistency through an assimilation-style innovation step. Evaluated on SEVIR radar data for 4 × \times super-resolution of severe convective storms, CASCADE outperforms strong baselines across both continuous metrics (PSNR, SSIM, MAE) and threshold-based skill scores (CSI, HSS, POD) while providing interpretable diagnostics through visualizable velocity and correction fields. By encoding advection as the fundamental operator rather than learning it implicitly, CASCADE produces temporally coherent, physically consistent, and mass-conserving reconstructions well suited to advection-dominated extremes in atmospheric and oceanic applications. 1 Introduction Over the past decade, deep neural networks have achieved remarkable performance across perception and language tasks by learning multiscale representations from data. These advances have naturally extended to scientific computing, where machine learning (ML) models increasingly serve as fast, accurate surrogates for traditional numerical methods and as components in hybrid modeling pipelines. In atmospheric science, ML-based models are now used both as rapid approximations to numerical weather prediction (NWP) systems and as post-processing tools. Recent global deep-learning forecasters, including FourCastNet [ 15 ] , Pangu-Weather [ 2 ] , and GraphCast [ 7 ] , demonstrate that training on large reanalysis archives can produce medium-range forecasts with skill competitive to traditional NWP at a fraction of the computational cost. Complementary efforts explore foundation-model approaches that unify tasks across weather and climate domains. For instance, ClimaX pretrains a transformer on climate-model output and successfully transfers learned representations across multiple variables, lead times, and downstream tasks [ 14 ] . Despite these advances, extreme weather events remain a major challenge. Extremes are rare, strongly nonlinear, and often controlled by intermittent small-scale processes that are difficult to represent in coarse-resolution datasets. Standard training objectives such as mean-squared error tend to bias models toward smooth conditional means, systematically under-representing the tails of the distribution that matter most for hazard assessment and emergency response. This has motivated growing interest in models specifically designed to capture extremes, including high-resolution precipitation forecasting systems and nowcasting architectures such as MetNet-3 [ 13 ] . To improve both robustness and physical consistency, physics-informed neural networks (PINNs) and related physics-guided ML approaches incorporate known structure, differential constraints, conservation laws, symmetries, directly into the learning process [ 6 , 3 ] . While scaling classical PINN formulations to high-dimensional chaotic flows remains challenging, the underlying principle is sound: encode the right operators rather than forcing networks to rediscover fundamental physics from scratch. Super-resolution (SR) methods developed for natural images typically prioritize perceptual realism, accepting that multiple high-frequency completions may be visually plausible for a given low-resolution input. Geophysical SR, by contrast, faces stricter constraints. Fine-scale structures must be consistent with both the coarse-scale state and the subsequent spatiotemporal evolution. A visually convincing precipitation filament is scientifically meaningless if it advects in the wrong direction, violates mass conservation, or fails to match coarse-scale observations when averaged. These requirements motivate SR approaches that explicitly incorporate temporal context, physical transport mechanisms, and assimilation-style constraints rather than relying solely on per-pixel hallucination. For extreme events, these requirements are non-negotiable: impact-relevant quantities (e.g., peak precipitation rates, storm cores, sharp fronts) are precisely the high-frequency structures most likely to be smoothed by regression or spuriously created by purely generative detail. A useful downscaling method must therefore (i) move intense features in physically plausible directions, (ii) preserve their integrated magnitude as they advect, and (iii) remain consistent with the observed coarse-scale state. CASCADE is designed around these constraints by making transport the core operator and using an assimilation-style correction to prevent drift. For extreme events, these requirements are non-negotiable: impact-relevant quantities (e.g., peak precipitation rates, storm cores, sharp fronts) are precisely the high-frequency structures most likely to be smoothed by regression or spuriously created by purely generative detail. A useful downscaling method must therefore (i) move intense features in physically plausible directions, (ii) preserve their integrated magnitude as they advect, and (iii) remain consistent with the observed coarse-scale state. CASCADE is designed around these constraints by making transport the core operator and using an assimilation-style correction to prevent drift. This paper introduces CASCADE ( C ross-scale A dvective S uper-resolution with C limate A ssimilation and D ownscaling E volution), a framework for super-resolution and dynamical downscaling of spatiotemporal geophysical fields. CASCADE reconstructs fine-scale structure through iterative, flow-guided semi-Lagrangian advection across scales. In its dynamical downscaling variant, it incorporates an explicit innovation step analogous to data assimilation, ensuring temporal consistency while respecting coarse-scale observations. 2 Related Work 2.1 Super-Resolution in Computer Vision Super-resolution has a rich history in computer vision, where the goal is to reconstruct high-resolution images from low-resolution observations. Classical deep SR methods achieved rapid progress through CNN-based residual architectures such as EDSR [ 11 ] and attention-based designs such as RCAN [ 20 ] . For perceptual quality, adversarial and perceptual-loss formulations, exemplified by ESRGAN [ 18 ] , can generate visually sharp textures. However, the high-frequency content produced by these methods is not uniquely determined by the input; multiple plausible completions exist. This ambiguity poses problems for geophysical applications. Downscaled atmospheric or oceanic fields are evaluated not only by pixel-wise error but also by their spectral characteristics, representation of extremes, and spatiotemporal consistency. A sharp gradient that looks convincing in isolation may be physically meaningless if it appears at the wrong location, moves in the wrong direction, or violates conservation principles. These requirements motivate methods that incorporate dynamics and physical constraints rather than purely appearance-driven texture synthesis. Recent generative approaches (e.g., diffusion-based nowcasting and downscaling) have improved perceptual fidelity and can represent forecast uncertainty, but they still require strong constraints to ensure transport-consistent evolution and physically meaningful extremes [ 8 , 12 , 19 , 4 ] . 2.2 Deep Learning for Climate and Weather Downscaling In climate and weather applications, deep learning has been widely adopted as a form of statistical downscaling. Early work explicitly adapted image SR architectures to climate data, with DeepSD [ 16 ] among the pioneering examples. Since then, a substantial body of research has explored CNNs, U-Nets, GANs, and transformers for meteorological SR across different variables, regions, and resolutions. Systematic reviews have identified recurring challenges including scale-factor dependence, multivariate consistency, uncertainty quantification, and out-of-distribution generalization [ 5 ] . Recent probabilistic approaches reframe downscaling as distribution matching, using generative models to explicitly target tail behavior and multivariate dependencies [ 17 ] . These advances are complementary to CASCADE. While probabilistic models focus on capturing the full conditional distribution of fine-scale fields, CASCADE emphasizes