An automated adaptative framework for weather data interpolation

Abstract

Spatially continuous estimates of reference evapotranspiration and rainfall are crucial for environmental management applications. However, their estimation often relies on weather station measurements characterized by incomplete temporal records. Spatial interpolation methods (SIMs) address this issue but are variable and data-specific, making their selection challenging due to dynamic, unharmonized networks and missing data. Advances in computing capabilities enable automation of selection routines, allowing development of automatic SIMs selection process. Several authors proposed employing combined metrics to optimize model selection. Accordingly, this paper developed a new composite metric (CM), based on random and systematic error, integrated into an automated process, called JWgrid (Judicious Widget for Geospatial Robust Interpolation Design or JW), to identify optimal interpolation methods across various contexts. Using Leave-One-Out Cross-Validation (LOOCV), the SIM with the lowest CM value is selected from popular interpolation techniques. The methodology was assessed on weather station networks in Luxembourg, evaluating rainfall and reference evapotranspiration related variables. Results showed low error values with SIM selection frequencies being variable dependent. The methodology’s robustness was assessed by (1) reducing data to 25, 50, and 75% of the original network size, by (2) using synthetic datasets with varying data variability, network configurations, and presence or absence of an external drift and (3) applying to a mountainous study area. Network density minimally affected overall accuracy, though SIM selection frequencies varied with no universally preferable method. In synthetic tests and the mountainous area, the JW methodology performed at least equally to the best single method by selecting the most appropriate method at every timestep. Those results highlight the proposed methodology’s advantages, producing reliable gridded datasets across diverse climatic conditions. The approach enables users with limited SIMs experience to easily generate gridded data, providing an effective solution for spatial interpolation challenges in environmental applications.