Self Supervised Vision for Climate Downscaling

TL;DR

The paper introduces a self-supervised, instance-specific climate downscaling model that operates without high-resolution ground truth data. It combines self-supervised pre-training, channel segregation, and topoclimatic attention to learn from a single data instance at runtime, effectively handling multi-channel climate variables such as temperature, precipitation, and topography gradient. Evaluations on CESM data show substantial RMSE improvements for 2x–4x downscaling of TS and PRECT with favorable runtimes, demonstrating potential for larger climate ensembles and historic data downscaling. The approach offers a data-efficient, adaptable framework with practical impact on climate research and policy-relevant regional projections, and opens avenues for extending to more variables and temporal dimensions.

Abstract

Climate change is one of the most critical challenges that our planet is facing today. Rising global temperatures are already bringing noticeable changes to Earth's weather and climate patterns with an increased frequency of unpredictable and extreme weather events. Future projections for climate change research are based on Earth System Models (ESMs), the computer models that simulate the Earth's climate system. ESMs provide a framework to integrate various physical systems, but their output is bound by the enormous computational resources required for running and archiving higher-resolution simulations. For a given resource budget, the ESMs are generally run on a coarser grid, followed by a computationally lighter $downscaling$ process to obtain a finer-resolution output. In this work, we present a deep-learning model for downscaling ESM simulation data that does not require high-resolution ground truth data for model optimization. This is realized by leveraging salient data distribution patterns and the hidden dependencies between weather variables for an $\textit{individual}$ data point at $\textit{runtime}$. Extensive evaluation with $2$x, $3$x, and $4$x scaling factors demonstrates that the proposed model consistently obtains superior performance over that of various baselines. The improved downscaling performance and no dependence on high-resolution ground truth data make the proposed method a valuable tool for climate research and mark it as a promising direction for future research.

Figures (7)

• Figure 1: A snapshot of a time step representing total precipitation (PRECT) and surface temperature (TS). Both variables exhibit remarkably different behaviors.
• Figure 2: Information repetition across scales in climate data. The similarity between the 10 x 10 patches in the source image is measured across multiple scales (1.5x--4.0x). y-axis represents levels L1 to L6 - the increasing RMSE values for the cutoff threshold that determines the level of similarity. The cutoff range is 1E-2 to 1E-1 for TS and 0.5E-14 to 1E-12 for PRECT. the x-axis represents the number of patches that have RMSE values lower than the cutoff value at that level.
• Figure 3: Core architecture of the proposed model. A low-resolution input ${X}$ is downscaled to ${X}$$\uparrow_{s}$ ($s>1$) and (${X}$, ${X}$)$\uparrow_{s}$$\in \mathbb{R}^{C \times W \times H}$ where $C=3$, represent three weather variables: temperature (TS), total precipitation (PRECT), and gradient of topography (dPHIS). ${X}$ is resized to target size before feeding into the model. The upper half of the image shows the channel segregation with three pipelines, with respective filters for each input variable. The topoclimatic attention mechanism is the lower half parallel to the main model flow. The attention mechanism re-weights the intermediate feature maps at multiple stages and enforces learning of topoclimatic relationships.
• Figure 4: Ground truth high-resolution data (a), compared to downscaled output without channel segregation (b) downscaled output with channel segregation (c). Regions with noticeable artifacts are marked with pink rectangles. Region A is zoomed-in Figures (d), (e), and (f). RMSE values for (b) and (c) are smaller than bicubic interpolation, but model output without channel segregation suffers from undesirable artifacts.
• Figure 5: Performance comparison for temperature (TS) Figure (a) and precipitation (PRECT) Figure (b) over scaling factors ranging from 2x until 8x with 0.1 increments.