Convective Environments over the Arabian Peninsula in Current and Future Climate

Convective precipitation is the primary source of freshwater and groundwater recharge in the Arabian Peninsula (AP), occurring as sporadic, localised events. Understanding the Lagrangian properties of these convective systems and how they are influenced by land surface characteristics, topography, and climate change requires convection-permitting modelling, which can be computationally inefficient in such an arid region. However, we trained three machine learning (ML) models -Random Forest (RF), Extreme Gradient Boosting (XGB), and Deep Learning (DL)- to establish relationships between convective environments and precipitation. These models can be applied to any numerical model output (e.g., CMIP6) to infer the probability of convective precipitation, thereby identifying when convection-permitting simulations should be performed. Constrained by CMIP6 temporal (6h) and spatial (~1°) resolutions, we aggregated IMERG V07 precipitation to 6h and averaged over 1 °. We derived 102 features from ERA5 describing moisture, lift, instability, and location to characterise the atmospheric profile. A profile is labelled convective if the accumulated precipitation exceeds the local climatological median, thereby reducing the problem to a binary classification task. The ML classifiers show high skill in identifying convective environments over the northern AP during cold months (Oct-Apr), with a Heidke Skill Score (HSS) of approximately 0.65. However, over the southern AP during warm months (May-Sep), the HSS values drop to around 0.35. These results support the established finding that convective systems over the AP in cold months are linked to large-scale atmospheric patterns. In contrast, in warm months they are localised and/or orographically influenced. Findings also demonstrate that the ML models learned convective environment patterns across the AP. Furthermore, the dirnual performance of ML models remains comparable (HSS: ~ 0.55), except at 12 UTC (HSS: ~ 0.48), when the convection is relatively localised. The SHapley Additive exPlanations (SHAP) method enables the interpretation of each feature’s contribution to the ML models’ prediction. The top three features identified by SHAP differ across ML models. Equivalent potential temperature at 850 hPa, humidity index, and lightning potential index are most important for RF. XGB emphasizes precipitable water vapour, relative humidity at 1000 hPa, and most unstable CAPE. In DL, precipitable water vapour, relative humidity at 700 hPa, and 2m dew-point temperature are the key contributors. The current warming over the AP is +1.22 °C relative to pre-industrial levels. We aim to analyse periods with comparable warming across 10 CMIP6 historical simulations to evaluate the models' ability to reproduce the spatiotemporal distribution and characteristics of convective environments. To assess the effect of climate change, we will analyse two future periods with changes of +2.22 and +3.22 °C from the SSP5-8.5 scenario. SHAP can help evaluate whether the model‑inferred importance ranking of mechanisms controlling convective precipitation changes between present and future simulations. In-depth analysis is undergoing.