Estimating precipation in open ocean with SWOT at sub-kilometer resolution

Precipitation is a fundamental component of the Earths hydrological cycle, with profound implications for water resource management, marine traffic, and disaster risk mitigation. Accurate rainfall estimation is critical for understanding weather patterns, forecasting flood events, and modeling climate change scenarios. While ground-based systems, such as the NEXRAD WSR-88D network, provide high-resolution data, they are geographically limited and leave vast oceanic regions unmonitored. Alternatively, satellite missions offer global coverage but often lack the necessary spatial resolution for fine-scale analysis (e.g., IMERG provides rates at ~8 km/pixel) or have limited acquisition in open oceans (e.g., Sentinel-1's constellation default observation mode in open ocean is composed of scattered imagettes).

In this context, the Surface Water and Ocean Topography (SWOT) mission presents a novel opportunity to bridge the gap between global coverage and high spatial resolution. Although primarily designed for altimetry, SWOT’s Ka-band Radar Interferometer (KaRIn) is sensitive to atmospheric hydrometeors. KaRIn offers sub-kilometer resolution (250 m/pixel), matching NEXRAD resolution in range, and provides continuous data over both coastal zones and the open ocean.

We present a machine learning framework to estimate precipitation rates using SWOT observations. We build a NEXRAD/SWOT dataset between August 2023 and February 2025, composed of 7009 patches (512×512 pixels) out of which 1090 contains precipitation (more than 1 mm/h on more than 1\% of the observation). A U-Net architecture was trained to retrieve Digital Precipitation Rates (DPR) from the WSR-88D. The input features include the backscattering coefficient (normalized to mitigate the incidence angle variability), total coherence, incidence angle, and a wind speed prior from atmospherical models. To ensure robust performance, the training loss is spatially weighted: pixels closer to NEXRAD stations are prioritized to minimize ground-truth uncertainty related to radar beam broadening and elevation in altitude, while null-DPR pixels are down-weighted to address class imbalance. Furthermore, quantile mapping was applied to align the model’s output distribution with NEXRAD's statistics, ensuring the accurate replication of heavy rainfall tails. An ensemble of independently trained models allow to compute a consensus score, providing a metric for estimating confidence.

Evaluation against NEXRAD data shows the model achieves 67\% accuracy in categorical classification (rainless, low, high intensity), a performance comparable to dual-station consistency checks. In the open ocean, validation against collocated IMERG tracks reveals strong correlations of their respective time series, reaching 95\% in the Pacific Inter-Tropical Convergence Zone (ITCZ) and 75\% in the Atlantic ITCZ. However, correlation degrades at higher latitudes, suggesting a sensitivity to convective precipitation regimes. This behavior is consistent with observations from C-Band SAR rainfall retrieval such as the future rainfall product of the Sentinel-1 constellation.

These results demonstrate the feasibility of using SWOT KaRIn high-resolution products for robust rainfall estimation, particularly in tropical and equatorial regions. By unlocking precipitation data in data-sparse regions, this approach offers a significant contribution to global precipitation monitoring and hydrological modeling.