Machine Learning-Based Calibration of Low-Cost Air Temperature Sensors Using Environmental Data

Area of Science:

• Environmental Science
• Sensor Technology
• Machine Learning

Background:

• Low-cost air temperature sensors are susceptible to environmental influences like solar radiation and humidity.
• Lack of radiation shields or forced aspiration in these sensors leads to inaccurate readings.

Purpose of the Study:

• To develop a machine learning-based calibration method for improving air temperature measurements from low-cost sensors.
• To mitigate the impact of environmental variables on sensor accuracy.

Main Methods:

• An artificial neural network (ANN) was employed to calibrate temperature measurements.
• Data collected over 305 days across three Japanese locations were used for training and validation.
• k-fold cross-validation was utilized to assess model performance.

Main Results:

• The machine learning calibration reduced the mean absolute error (MAE) from 1.62 to 0.67.
• Calibration effectiveness was demonstrated even when using data from nearby, different locations for training.
• Calibration failures occurred during abrupt environmental shifts (e.g., rainfall).

Conclusions:

• The ANN-based calibration method effectively enhances the accuracy of low-cost air temperature sensors.
• Sensor data accuracy can be improved using data from nearby locations, but significant environmental differences between distant locations negatively impact results.