Abstract
With the development of artificial intelligence (AI) in recent years, meteorological departments have also begun to improve algorithms and revise short-term forecasts via AI, expecting to timely capture meteorological clues in massive weather data, to "prevent meteorological disasters", and "calculate precipitation faster and more accurately". At present, AI has been initially applied to the meteorological field, especially to the analysis of massive meteorological data. For instance, the AI-based data analysis technology can rapidly judge the cloud type and the meteorological prototype in satellite images. The AI-based data fusion technology contributes to more three-dimensional and refined atmosphere data, which improves the temporal and spatial resolutions of precipitation data. During data fusion, the severe convective weather characteristics reflected by massive historical precipitation data can be learned through machine learning methods to predict the evolution trend of disastrous weather within the future 1 to 2 h. Furthermore, precipitation data errors are corrected through AI data analysis, and a daily precipitation fusion dataset with a spatial resolution of 1 km is obtained.
Introduction
Accurate precipitation data is essential for hydrological modeling and meteorological forecasting. However, current measurement techniques—ground-based stations, satellite-borne sensors, and radar—each face significant challenges. Ground stations provide high-precision point measurements but are often too sparse to capture localized variability. Satellites offer extensive spatial coverage but are limited by the accuracy of atmospheric retrieval algorithms. Radar provides high spatial resolution but is frequently degraded by complex terrain, such as mountain ranges. This research addresses the challenge of multi-source data fusion, aiming to integrate these disparate sources into a single, reliable, high-resolution (1 km) precipitation product, with a specific focus on the topographically complex Qilian Mountain region in China.
Methods and Data
The study proposes a multi-source fusion framework primarily utilizing Gaussian Process Regression (GPR). GPR is a non-parametric, Bayesian approach well-suited for handling high-dimensional data and small sample sizes with inherent uncertainty. The performance of GPR was benchmarked against Random Forest (RF) and Multivariate Adaptive Regression Splines (MARS).
Two primary fusion strategies were evaluated:
- Point-Surface Fusion (Scheme 1): Direct fusion of satellite and ground data using auxiliary variables like Digital Elevation Models (DEM), longitude, and latitude.
- Station Bias Correction (Scheme 2): Using the model to first calculate the bias at specific stations and then interpolating that error across the spatial domain.
The dataset for the Qilian Mountain region (2019–2020) included hourly ground observations from 241 stations, IMERG daily satellite data (0.25° resolution), and quantitative precipitation estimations from radar (3 km resolution). Topographic context was provided by SRTM v41 DEM data.
Results and Discussion
Analysis revealed that topography is a primary driver of measurement error, with precipitation bias following a distinct spatial pattern: minimal on north and south slopes but maximal along mountain ridges. This underscores the necessity of incorporating high-resolution terrain data into fusion models.
The GPR model demonstrated superior performance over RF and MARS. While RF often produced spatial artifacts ("massive traces"), GPR generated smoother, more physically consistent transitions across the landscape. In Scheme 1, GPR achieved a verification correlation coefficient (R) of 0.79, outperforming RF (0.77) and MARS (0.73). Furthermore, GPR maintained a lower Root Mean Square Error (RMSE) of 30.21.
The final output achieved a spatial resolution of 1 km, a significant refinement from the original 0.25° satellite data. The model also demonstrated the ability to adapt to dynamic correlations between precipitation and auxiliary variables, which can fluctuate significantly between different weather events.
Conclusion
This research demonstrates that machine learning frameworks, particularly those based on Gaussian Process Regression, can effectively fuse multi-source meteorological data to produce high-resolution precipitation products. By accounting for complex topography, the model significantly improves the accuracy of rainfall mapping in mountainous regions. Future work will focus on extending this framework to real-time "nowcasting" systems and evaluating performance across different seasons and precipitation types, such as snowfall.
Summary updated