Summer Drought Prediction in Europe combining Climate Simulations and Remote Sensing

The occurrence of droughts is ruled by the interplay of complex processes with very different natures and spatio-temporal scales. Different modes of climate variability, like the North Atlantic Oscillation or ENSO (El Niño-Southern Oscillation), set the prevalence of distinct weather regimes providing sources of predictability at large-scale. On the other hand,  land-atmosphere feedbacks play a crucial role in climate extremes, and particularly, in the evolution and amplification of droughts. However, the weak predictability of the former large-scale variability in the extratropics together with the poor representation of these feedbacks in current seasonal predictive systems lead to a limited capability of predicting droughts months in advance. In this study (part of the AI4Drought project, funded by ESA), we aim to enhance summer drought prediction in Europe from spring conditions by the combination of state-of-the-art climate simulations and remote sensing. A hybrid model combining climate simulations and high-resolution remote sensing data is proposed to boost the predictability signal at seasonal time-scale through the integration of two machine learning (ML) models. The first model (model-A) enhances large-scale predictability. It consists of a generative model (conditional variational auto-encoder, based on Pan et al., 2022), which is trained with 10.000s years of CMIP6 climate simulations to empirically learn the probability distributions between global spring fields; e.g., sea surface temperatures and 500 hPa geopotential height; and summer drought conditions (SPEI3). A local-scale model for extremes amplification is developed (model-B). A pixel-based (multi-layer neural network) model aims to capture land-atmosphere feedbacks; integrating local conditions from satellite-based products and reanalysis data, e.g. soil moisture (SM), temperatures and NDVI together with information from the large-scale predictions from model-A in order to predict SM anomalies for the whole summer season. Preliminary results highlight the significance of local conditions in enhancing drought predictions, particularly in the Mediterranean region, where land-atmosphere feedbacks are pronounced. Experiments conducted under ideal conditions, knowing the future large-scale conditions in advance, demonstrate improved prediction skill when local conditions (e.g., soil moisture, NDVI) are included as predictors. Moreover, a DeepSHAP analysis (eXplainableAI-based method) is performed to understand which are the most important drivers for the local-scale model prediction of summer SM anomalies. As expected, the spring’s SM anomalies are the most important input features; together with the large-scale conditions described by August SPEI-3. Additionally, temperature anomalies have a relatively high importance when predicting summer drought conditions. This research underscores the potential of a hybrid approach integrating climate simulations and remote sensing data to advance the understanding and prediction of summer droughts in Europe.