[Reading Science] The Eyes and Ears of Forecasts... The Power of Observational Data Supporting AI

'The Eyes of Data' Matter More Than AI
Forecast Accuracy Depends on Observation Networks and Quality Control

Even if AI can predict the sky, it is ultimately the sensors that observe it. The true variables that determine the accuracy of forecasts are not the computational power of artificial intelligence (AI), but rather the quality of the data and the precision of the observation infrastructure.

The AI-based forecasting model NowAlpha uses data from 10 weather radars across the country. The radar receives precipitation observations over a two-hour period and predicts precipitation intensity in 10-minute increments up to six hours ahead.

Approximately 700 Automated Weather Stations (AWS) nationwide undergo a correction process because their data does not align with the AI grid structure. According to the Korea Meteorological Administration, “AI forecasts operate mainly based on radar and satellite remote sensing data.”

What supports the brain of AI is an invisible quality control (QC) system. The process of removing noise from observed values and synchronizing the time intervals and coordinate systems of different datasets underpins the reliability of AI forecasts.

Woo Jingyu, a forecaster at the Korea Meteorological Administration, said, “AI forecasting is a complex system that must handle different types of data, such as radar, satellite, and marine data. The key is not the small errors in individual datasets, but how precisely these are integrated and how much errors are reduced during the quality control stage.”

He explained, “AI forecasting models are fast, but ensuring the stability and consistency of forecasts requires the physical consistency of input data. The Korea Meteorological Administration is strengthening quality control procedures such as removing non-meteorological echoes (terrain reflection, waves, chaff, radio interference, etc.), radiometric correction, and satellite position verification.”

24-Hour Quality Control System That ‘Cleans the Eyes of AI’

According to the Korea Meteorological Administration, radar data filters out non-meteorological signals by analyzing observed variables such as reflectivity, differential reflectivity, and cross-correlation coefficient from dual-polarization radar. Satellite data undergoes radiometric correction using solar observation (visible channel) and internal blackbody (infrared channel), and the performance of position correction is verified through observation residuals (errors) around the Earth. The level of noise and coordinate errors generated during this process are recorded as quality indicators (QI) and are automatically filtered out from AI training data.

The National Institute of Meteorological Sciences refers to this as a “24-hour maintenance system that cleans the eyes of AI.” Maintaining the physical consistency of observational data is essential for AI to learn without forecasting bias.

No matter how advanced the AI forecasting model becomes, there are clear limitations if the input data lacks detail. To address this, the Korea Meteorological Administration is simultaneously carrying out a project to upgrade its observation infrastructure.

According to experts in academia and the meteorological industry, in 2026, the Cheollian 2C and 2D satellites will be launched to double the resolution of visible and infrared imagery, and the network of small weather radars for local precipitation monitoring will be expanded. To cover blind spots in mountainous regions and urban areas, an AI-based drone observation network will be introduced, and more buoys equipped with real-time sensors will be deployed in the ocean.

A representative from the weather app industry said, “As the quality of public data improves, the accuracy of private forecasting services also increases. Ultimately, the reliability of AI forecasts depends on how cleanly the data is ingested.”

If AI is the brain of forecasting, observation is its senses. While AI calculates the future, countless sensors and quality control systems underpin the reliability of those predictions. Seeing the sky more accurately and refining the data more thoroughly is the most practical path toward forecasts that do not miss the mark.