Update! HASSA Brown Bag Lunch Series Spring 2025

Omid Zandi Talk

Maria Castro Talk

Developing a Novel Long-Term IR-based Precipitation Product from AVHRR in High Latitudes Through a Double Machine Learning Strategy

Accurate long-term precipitation estimation is critical to studying Earth's hydrologic cycle and energy budget. Geostationary infrared (IR) observations have made it possible to produce long-term precipitation records within ~ 60^o N/S. Still, in higher latitudes, geostationary images are too oblique, and Passive Microwave (PMW) precipitation products have shorter records and face challenges with precipitation retrieval, especially over snow-ice surfaces. Therefore, in high latitudes, the Global Precipitation Climatology Project (GPCP) has so far used IR-based precipitation retrievals from the TIROS Operational Vertical Sounder (TOVS) and the Atmospheric Infrared Sounder (AIRS). Besides some inconsistency issues between the two sensors, precipitation estimation from TOVS and AIRS is based on an old retrieval method. Here, we present a novel double Machine Learning (ML) based precipitation retrieval method that uses brightness temperature and cloud properties from the Advanced Very High-Resolution Radiometer (AVHRR) observations via the PATMOS-X product, environmental information from the MERRA-2, and surface type data from AutoSnow as major input features. The AVHRR observations have higher spatial resolution compared to TOVS and AIRS. Precipitation estimates from coincident PMW-based precipitation retrievals, CloudSat, and ERA5 were used for training where needed. To mitigate the overestimation of light precipitation intensities and occurrence, a post processing step is used to limit the raining areas, in which each precipitating cloud patch is filtered and corrected based on the local statistical characteristics of the cloud. 

Leveraging the continuous and consistent AVHRR observations across multiple platforms over more than four decades, this product provides a long-term and consistent precipitation data record in high latitudes that will be helpful to GPCP, other long-term products, or merged products that may not trust PMW estimates over certain conditions in high latitudes. Quantitative comparison with IMERG using independent data in 2010 shows that the product outperforms AIRS precipitation estimates, achieving daily Kling-Gupta Efficiency (KGE) skill scores of 0.62 and 0.54 over 55^oN-90^oN and 55^oS-90^oS, respectively, where AIRS  retrievals yielded KGE of 0.26 and -0.3.

Probabilistic Dam Break Flood Mapping via Monte-Carlo Simulations using a 2D Local-Inertial Model

Dam-break flood hazard assessment is essential to enhance preparedness and safeguard downstream areas in case of failure. Deterministic event-based approaches typically do not consider the inherent uncertainty arising from the effects governing a dam-break scenario. A probabilistic dam-break model was developed based on the Monte-Carlo method, coupling a 2D local-inertial hydrodynamic model (HydroPol2D) with Bayesian-generated governing parameters to the dam-break flow propagation problem to capture a set of scenarios with different reservoir initial volumes, breach hydrographs, and terrain roughness. The Tapacurá dam in Pernambuco, Brazil, was assessed under ensemble probabilistic scenarios. To validate the numerical approach, the local-inertial model (HydroPol2D) was benchmarked with the HEC-RAS 2D full momentum model for spatial resolutions of 10 (LiDAR) and 30 meters (Copernicus DEM). Results obtained from the probabilistic maps point out that the expected inundation area tends to increase 6% for all probabilities, on average, as model spatial resolution decreases from 10 m to 30 m (threefold). By this increase in the pixel size, the computational times are reduced, on average, by a factor of three. For both resolutions, several internal points of the domain were assessed, with flood inundation probability results for the 10 m resolution including (i) a school (Prob = 0.4 ± 0.4), (ii) an emergency care unit (Prob = 0.3 ± 0.5), (iii) a public market (Prob = 0.6 ± 0.5), (iv) a historical cinema (Prob = 0.2 ± 0.4), (v) a hospital (Prob = 0.4 ± 0.5), and (vi) a soccer stadium (Prob = 0.05 ± 0.2). The large standard deviations arise from the uncertain nature of the dam-break characteristics, emulated by different fitted probability distribution functions to represent HydroPol2D parameters and boundary conditions.