Climate change is expected to increase the frequency and intensity of extreme weather events, leading to significant socioeconomic and human losses. A crucial aspect of this change is the alteration of precipitation patterns, especially in tropical regions. While the Clausius-Clapeyron (CC) relation suggests a 7% increase in atmospheric moisture per Kelvin of warming, projections from global climate models (GCMs) vary considerably, ranging from 1.5%K⁻¹ to 11%K⁻¹, with some high-resolution models suggesting even higher sensitivities (13–17%K⁻¹). Observation-based studies, however, show a contrasting trend, with decreasing inner-core TC precipitation despite increased outer rainband precipitation. This discrepancy stems from the complex convective dynamics within TCs, requiring kilometer-scale resolution simulations to accurately capture the interplay of slantwise and buoyant convective cells. The computational cost of such high-resolution global simulations has hindered their widespread use, making it difficult to assess the precise mechanisms behind TC precipitation changes. High-resolution forecasting of TC extreme precipitation is critical for building climate-resilient cities and communities. The potential intensification of precipitation significantly impacts flood and landslide risks, yet comprehensive assessments are lacking due to limitations in high-resolution climate simulations and uncertainties in process models. This study uses deep learning (DL) to overcome these challenges, focusing on the South China region, a zone particularly vulnerable to TC-induced precipitation extremes and subsequent landslides.

Existing literature displays significant uncertainty regarding the sensitivity of tropical precipitation extremes to warming. Studies using global climate models (GCMs) show a wide range of projections, with some suggesting a 7% increase per Kelvin based on the Clausius-Clapeyron (CC) relationship, while others report a wider range (1.5%K⁻¹ to 17%K⁻¹). This disparity arises from the limitations of coarse-resolution GCMs in accurately resolving convective processes within tropical cyclones (TCs). High-resolution studies are limited due to computational constraints. Observation-based studies have also presented conflicting results, with some indicating a decrease in inner-core TC precipitation while others show increases in outer rainband precipitation. The lack of consensus highlights the need for high-resolution simulations capable of resolving the intricate convective dynamics within TCs to better understand their response to warming.

This study employs a novel approach combining deep learning (DL) and dynamical downscaling to simulate high-resolution TC precipitation under global warming. First, a deep learning model is trained using coarse-resolution global climate simulation data (Community Earth System Model 2, CESM2) and high-resolution gridded precipitation data (APHRODITE) to identify time slices associated with potential extreme precipitation events in South China. This model predicts daily subgrid maximum rainfall based on atmospheric circulation patterns. The model incorporates five atmospheric variables (geopotential, specific humidity, temperature, u and v wind components) at six pressure levels, trained on ERA5 reanalysis data. The trained model is then applied to CESM2 data (SSP5-8.5 scenario, 2015-2100) to select extreme events for higher-resolution simulation. The top 16 most intense events from both present (2015-2034) and future (2081-2100) periods are selected. The Weather Research and Forecast (WRF) model is used to perform targeted dynamical downscaling of these identified events at a 1-km resolution, capturing the detailed convective processes. This deep learning-assisted method, termed smart dynamical downscaling (SDD), is compared against a direct dynamical downscaling (DDD) method, which directly selects extreme events based on CESM2 precipitation outputs. The WRF simulations utilize four nested domains with a 1-km grid spacing in the innermost domain. The resulting high-resolution precipitation data are then used to assess the impact on landslide risks in Hong Kong using the Hong Kong Landslip Warning System (LWS), a statistical model based on historical observations. A modified LWS is used for future scenarios where rainfall exceeds historical maxima, assuming that the number of landslides plateaus beyond a certain rainfall threshold. The analysis examines precipitation sensitivity to warming at various spatial and temporal scales and investigates changes in TC structure, including updraft intensity, convective core size, and TC movement speed.

The study reveals a scale-dependent sensitivity of extreme precipitation to warming. At hourly and kilometer scales, the sensitivity is close to the CC rate (4-8%K⁻¹). However, at larger spatial scales (25-50 km) and longer time scales (12-18 h), the sensitivity significantly exceeds the CC rate, reaching approximately 18%K⁻¹. This super-CC scaling is observed across various temporal and spatial accumulation scales, emphasizing its significance for urban flood management and landslide risk assessment. The analysis shows that while the maximum upward velocity within convective cores decreases slightly in the future climate, the size of these cores increases significantly. This expansion of deep convective cores, accompanied by a reduction of shallow cumulus and cumulus congestus clouds, leads to increased spatial and temporal accumulation of rainfall, exceeding the effects of slightly weaker updrafts. This is more pronounced at spatial scales of 25-50 km and temporal scales of 12-18 h, consistent with the characteristic scales of convection clusters and TC rainbands. High-resolution simulations of the most intense events in the present and future periods demonstrate the importance of finer spatial resolution for accurately assessing rainfall patterns. The 1-km resolution data reveals significantly higher 24-h maximum rolling rainfall (MRR) values compared to coarser resolutions, with a 54% increase in MRR maxima in the future case due to a 4-K warming. This corresponds to a 13%K⁻¹ sensitivity, consistent with the super-CC scaling observed in the analysis of various spatial and temporal scales. The landslide risk assessment based on the modified LWS shows a potential increase in the number of landslides in Hong Kong ranging from 92% to 215% with a 4-K warming, underscoring the significant threat of compound disasters from intensified TC precipitation.

The findings of this study challenge the simplistic application of the CC scaling to project the intensification of TC precipitation under global warming. The observed scale-dependent sensitivity emphasizes the critical role of resolving convective processes at high resolution. The super-CC scaling observed at larger scales is directly linked to changes in TC structure, specifically the expansion of deep convective cores. This mechanistic understanding is crucial for accurate risk assessment. The significant increase in projected landslide risks highlights the compounding effects of extreme precipitation under climate change. The study's results have significant implications for infrastructure planning and risk management in coastal cities vulnerable to TCs, particularly in South China. The study’s findings emphasize the need to consider scale-dependent responses when assessing the impact of climate change on extreme precipitation and subsequent hazards. The integration of deep learning and high-resolution simulations offers a powerful approach to improve the accuracy of climate change impact projections.

This study demonstrates that the intensification of TC precipitation in South China under global warming is more substantial than previously anticipated, exceeding the Clausius-Clapeyron scaling at scales relevant to urban infrastructure and landslide risk. The scale-dependent sensitivity, driven by changes in TC structure, necessitates high-resolution simulations for accurate impact assessments. The projected increase in landslide risk underscores the urgent need for enhanced climate adaptation strategies. Future research should investigate the mechanisms behind the changing organization of convective cores within TCs, explore the application of the methodology to other regions, and refine the landslide risk assessment models to account for other relevant factors.

While the kilometer-scale resolution simulations represent a significant advancement, they still involve turbulence parameterizations, an area of ongoing model development. The limited distribution of observational networks for model validation also poses a challenge. The landslide risk assessment is specific to Hong Kong's man-made slopes and may not be directly generalizable to other regions with different geological and geotechnical characteristics. The study focuses on a specific region and a limited number of extreme events, limiting the generalizability of the findings to other geographic areas and TC types.