AI-Driven Weather Data Superresolution via Data Fusion for Precision Agriculture.

Area of Science:

• Agricultural Meteorology
• Geospatial Data Science
• Machine Learning Applications

Background:

• Precision agriculture demands localized, high-resolution meteorological data, which current coarse-scale numerical weather prediction (NWP) models fail to provide.
• Microclimate variability significantly impacts crop yields and requires more granular weather information than operational NWP products offer.

Purpose of the Study:

• To develop and evaluate a data fusion superresolution workflow for generating high-resolution 2 m air temperature fields 24 hours ahead.
• To combine Global Forecast System (GFS) predictors, regional station data, and physiographic descriptors for improved temperature prediction.

Main Methods:

• A data fusion superresolution workflow integrating GFS (0.25°), Southern Moravia station data, and static physiographic descriptors (elevation, terrain gradients).
• Evaluation of multiple machine learning models (LightGBM, TabPFN, Transformer, Bayesian neural fields) using spatiotemporal cross-validation.
• Implementation of a KNN interpolation layer in the physiographic feature space for spatial mapping.

Main Results:

• All evaluated models demonstrated a reduction in mean absolute error (MAE) compared to raw GFS predictors.
• The TabPFN-KNN model achieved the lowest MAE of 1.26 °C in the operational regime (unseen stations and future periods).
• This represents an approximate 24% improvement over the GFS baseline (MAE of 1.66 °C).

Conclusions:

• The proposed data fusion superresolution workflow is feasible for operational deployment in agricultural landscapes.
• The approach is compatible with existing sensor infrastructure, enabling high-resolution temperature forecasting.
• This advancement supports more accurate, localized weather information crucial for precision agriculture.