Using Explainability to Inform Statistical Downscaling Based on Deep Learning Beyond Standard Validation Approaches.

Deep learning (DL) has emerged as a promising tool to downscale climate projections at regional‐to‐local scales from large‐scale atmospheric fields following the perfect‐prognosis approach. Given their complexity, it is crucial to properly evaluate these methods, especially when applied to changing climatic conditions where the ability to extrapolate/generalize is key. In this work, we intercompare several DL models extracted from the literature for the same challenging use‐case (downscaling temperature in the CORDEX North America domain) and expand standard evaluation methods building on eXplainable Artificial Intelligence (XAI) techniques. Specifically, we introduce two novel XAI‐based diagnostics—Aggregated Saliency Map and Saliency Dispersion Maps—and show how they can be used to unravel the internal behavior of these models, aiding in their design and evaluation. This work advocates for the introduction of XAI techniques into deep downscaling evaluation frameworks, especially when working with large regions and/or under climate change conditions. Plain Language Summary: Due to limitations in the computational resources available, General Circulation Models (GCMs) are often used to simulate the climate system over coarse resolution grids. This hampers the applicability of GCM products in the regional‐to‐local scale, highly demanded by different socio‐economic sectors. Statistical downscaling aims to solve this problem by generating high‐resolution climate fields. Recently, machine learning techniques—particularly deep learning (DL) models—have shown promising results in this task. These models are first trained in a historical period through observational data sets, and then applied to the GCM outputs of plausible future scenarios, thus generating high‐resolution climate change products. To assess the performance of these methods, a number of evaluation metrics have been proposed considering both the skill to reproduce present climate conditions and the ability to generalize changing conditions. Here, we illustrate the possibilities of eXplainable Artificial Intelligence (XAI) techniques to expand the evaluation framework for deep downscaling methods, introducing new XAI‐derived diagnostics to unravel their internal behavior. The results show the usefulness of incorporating XAI techniques into statistical downscaling evaluation frameworks, especially when working with large regions and/or under climate change conditions. Key Points: Explainable artificial intelligence (XAI) facilitates the evaluation of deep downscaling models by unraveling their internal behaviorXAI techniques can detect structural problems not revealed by standard evaluation, which may be relevant for understanding model differencesXAI techniques can assess the relevance and locality of the predictors of deep learning models, helping to assess their physical consistency [ABSTRACT FROM AUTHOR]