Machine Learning for Light Sensor Calibration.

Area of Science:

• Atmospheric science
• Environmental monitoring
• Sensor technology

Background:

• Sunlight drives crucial photochemical and environmental processes, including atmospheric radiative heating.
• Accurate measurement of wavelength-resolved solar irradiance is vital for understanding these processes.
• Existing reference instruments are costly, limiting widespread deployment.

Purpose of the Study:

• To develop and validate a low-cost sensor ensemble for measuring wavelength-resolved solar irradiance spectra.
• To achieve high accuracy comparable to expensive reference instruments.
• To enable dense, neighborhood-scale sensor networks for micro-scale environmental variability studies.

Main Methods:

• Utilized an ensemble of very low-cost sensors (total cost <$20).
• Employed machine learning algorithms for sensor calibration against a reference instrument (NIST-calibrated Konica Minolta CL-500A).
• Achieved 1 nm spectral resolution between 360-780 nm.

Main Results:

• The calibrated low-cost sensor ensemble accurately reproduced reference instrument measurements with a correlation coefficient R² > 0.99.
• Demonstrated the feasibility of using machine learning to calibrate inexpensive sensors.
• Made sensor circuits and calibration code publicly available.

Conclusions:

• Low-cost sensor ensembles, when calibrated with machine learning, can provide high-fidelity solar irradiance spectra.
• This methodology allows for unprecedented spatial and temporal insights into irradiance micro-variability.
• Applications include air quality, environmental science, and agronomy.