Reducing Latency in Satellite-Based Precipitation Estimates Using GOES-16 and Machine Learning

Accurate representation of spatiotemporal precipitation patterns is essential for developing hydrological applications, particularly in operational hydrology and early warning systems. In regions with scarce in situ precipitation data, freely available satellite precipitation products (SPPs) provide an effective solution. However, equally important as SPPs' resolution is their latency, the time elapsed between satellite data collection and its availability to users. In this study, we developed a machine learning based on the random forest (RF) algorithm for predicting Integrated Multi-satellite Retrievals for Global Precipitation Measurement (IMERG) precipitation by leveraging low-latency GOES-16 Advanced Baseline Imager (ABI) bands with reduced latency; GOES-16 is one of the Geostationary Operational Environmental Satellites (GOES). The predictive model was applied to a mountain basin (3,391 km2) in southern Ecuador. The model was trained using hourly data over a 5-year period, and its performance was evaluated using quantitative and qualitative metrics across multiple temporal scales. The results indicated a progressive improvement in model accuracy with increasing temporal aggregation. For temporal scales ranging from hourly to monthly, RMS error (RMSE) values decreased from 0.48 to 0.05 mm/h, and Pearson's cross-correlation (PCC) improved from 0.59 to 0.87. Qualitative metrics, including probability of detection (POD), false alarm ratio (FAR), and critical success index (CSI), supported these findings by indicating the influence of temporal scale variability on model performance. These findings demonstrate the potential of the proposed approach for real-time hydrological applications and pave the way for more-advanced machine learning models capable of achieving greater accuracy at finer temporal resolutions.