Enhancing the evaluation of Deep Learning Downscaling methods using Explainable Artificial Intelligence (XAI) techniques

Deep learning (DL) models have emerged as an effective statistical downscaling method for  generating local-scale climate data based on large-scale atmospheric fields following the Perfect Prognosis (PP) approach. In particular, these models have been shown to be efficient for producing regional climate change projections over continental domains from coarse Global Climate Models, with great potential in initiatives such as CORDEX [1]. However, the black box nature of DL models rises concerns regarding their ability to adhere to physical constraints and accurately learn patterns grounded in physics, which may hinder their adoption by the vulnerability impacts and adaptation community.  In this work, we focus on this problem and present two novel eXplainable Artificial Intelligence (XAI)-based diagnostics to complement the standard downscaling validation approach: Aggregated Saliency Map (ASM) and Saliency Dispersion Map (SDM). As an illustrative application, we use these diagnostics to evaluate and understand the physical consistency of different Convolutional Neural Networks (CNNs) for  downscaling temperature and precipitation over the Iberian domain within the framework of Escenarios-PNACC, the initiative responsible for the generation and provision of national climate change scenarios in Spain. Overall, we show how these novel diagnostics allow for a better assessment of the models, as well as for a physics-based intercomparison of the effect of different configurations (e.g., spatial extension of the predictors), since they allow us to take into account the physical relevance of the relationships learned.  [1] Baño-Medina, J., Manzanas, R., Cimadevilla, E., Fernández, J., González-Abad, J., Cofiño, A. S., & Gutiérrez, J. M. (2022). Downscaling multi-model climate projection ensembles with deep learning (DeepESD): contribution to CORDEX EUR-44. Geoscientific Model Development, 15(17), 6747-6758.