Bridging the Gap Between Radar and Satellite: A Multi-Source Validation and Bias Correction Framework for Precipitation Estimation in Saint Lucia

Accurate precipitation monitoring is critical for Small Island Developing States (SIDS) like Saint Lucia, where complex topography and high vulnerability to tropical cyclones necessitate precise data for disaster preparedness and water resource management. While ground-based radar provides high-resolution estimates, it is spatially limited. Conversely, satellite products offer global coverage but often suffer from accuracy issues at island scales. This study presents a comprehensive evaluation and calibration framework comparing NASA’s IMERG and ERA5 reanalysis products downscaled to 2km spatial resolution against Caribbean radar observations with 1km spatial resolution. The objective is to quantify satellite performance during extreme weather events and demonstrate a robust operational workflow for bias correction.

The analysis employs a dual-framework approach. First, we conducted a spatial and temporal validation across nine major hurricane and storm events (2007-2024), including Hurricanes Dean and Tomas. This phase utilized Root Mean Square Error (RMSE) and Structural Similarity Index (SSIM) to assess agreement between radar, IMERG, and ERA5 datasets, alongside a point-based comparison at the Grace Station utilizing rain gauge data. Second, we developed a seven-step operational workflow using continuous 2021 data to implement pixel-wise Quantile Mapping bias correction. This workflow involved parallel processing of raw radar imagery, precise temporal-spatial matching, and the training of reusable correction functions.

Analysis of the storm events reveals that uncorrected satellite and reanalysis products systematically underestimate rainfall intensity, particularly during hazardous convective peaks. Satellite storm-mean values often capture only 10-60% of radar-observed totals. While ERA5 and IMERG exhibit comparable performance, both struggle to resolve fine-scale convective structures, yielding modest SSIM scores (averaging 0.15-0.40) that degrade as storm intensity increases. Point-based analysis at Grace Station highlights distinct "smoothing" effects in gridded products. For example, during Hurricane Tomas, a rain gauge recorded a peak intensity of 1516 mm/h (likely an extreme burst or anomaly) and radar recorded 33 mm/h, whereas satellite and ERA5 estimates were smoothed to 15.33 mm/h and 19.18 mm/h, respectively.

To address the identified underestimation, the Quantile Mapping method applied to the 2021 dataset yielded significant improvements. The correction reduced systematic bias by 87% (from a relative bias of -185% to -23%) and decreased the Mean Absolute Error (MAE) by 43%. Crucially, the correction dramatically improved the spatial structure of the precipitation fields, raising the temporal SSIM by 70% (from 0.490 to 0.834). The methodology successfully extended the dynamic range of satellite estimates to match radar observations, correcting the "capped" maximum values (from ~8 mm to >87 mm) and enabling a more realistic representation of extreme events. This study confirms that while raw satellite and reanalysis products underestimate intense Caribbean precipitation, they can be effectively calibrated using ground-based radar. The proposed workflow establishes a reusable framework for training bias correction functions, allowing meteorologists and hydrologists in Saint Lucia to better model flood risks and enhance climate resilience.