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

## TL;DR

This paper shows how combining weather data from multiple sources improves temperature forecasts for agriculture, enabling better field-level decisions.

## Contribution

A novel data fusion workflow using TabPFN-KNN achieves significant temperature forecast improvements for precision agriculture.

## Key findings

- Multi-source data fusion improves 24-hour 2m air temperature forecasts compared to raw GFS data.
- TabPFN-KNN achieves a 24% lower MAE (1.26°C) than GFS in the most demanding validation regime.
- The hybrid model supports generating high-resolution temperature fields compatible with sensor infrastructures.

## Abstract

What are the main findings?
Multi-source data fusion (GFS predictors + station observations + static physiography) consistently improves 24 h 2 m air temperature forecasts relative to raw GFS across all spatiotemporal splits.The best operational configuration is TabPFN-KNN, achieving MAE = 1.26 °C in the most demanding regime (time = validation, space = validation), i.e., ≈24% lower error than GFS (1.66 °C).

Multi-source data fusion (GFS predictors + station observations + static physiography) consistently improves 24 h 2 m air temperature forecasts relative to raw GFS across all spatiotemporal splits.

The best operational configuration is TabPFN-KNN, achieving MAE = 1.26 °C in the most demanding regime (time = validation, space = validation), i.e., ≈24% lower error than GFS (1.66 °C).

What are the implications of the main findings?
High-resolution, spatially continuous near-surface temperature fields can be generated from routinely available forecast inputs and regional station networks, supporting field-scale agricultural decisions.The hybrid design (station-level learning + physiography-conditioned KNN propagation) provides a deployable pathway for superresolution services integrated into sensor infrastructures (e.g., SensLog/ALIANCE).

High-resolution, spatially continuous near-surface temperature fields can be generated from routinely available forecast inputs and regional station networks, supporting field-scale agricultural decisions.

The hybrid design (station-level learning + physiography-conditioned KNN propagation) provides a deployable pathway for superresolution services integrated into sensor infrastructures (e.g., SensLog/ALIANCE).

Accurate field-scale meteorological information is required for precision agriculture, but operational numerical weather prediction products remain spatially coarse and cannot resolve local microclimate variability. This study proposes a data fusion superresolution workflow that combines global GFS predictors (0.25°), regional station observations from Southern Moravia (Czech Republic), and static physiographic descriptors (elevation and terrain gradients) to predict the 2 m air temperature 24 h ahead and to generate spatially continuous high-resolution temperature fields. Several model families (LightGBM, TabPFN, Transformer, and Bayesian neural fields) are evaluated under spatiotemporal splits designed to test generalization to unseen time periods and unseen stations; spatial mapping is implemented via a KNN interpolation layer in the physiographic feature space. All learned configurations reduce the mean absolute error relative to raw GFS across splits. In the most operationally relevant regime (unseen stations and unseen future period), TabPFN-KNN achieves the lowest MAE (1.26 °C), corresponding to an ≈24% reduction versus GFS (1.66 °C). The results support the feasibility of an operational, sensor-infrastructure-compatible pipeline for high-resolution temperature superresolution in agricultural landscapes.