Are Neural Networks Really the Holy Grail? A Comparison of Multivariate Calibration for Low-Cost Environmental Sensors

Data obtained from low-cost environmental sensors can have various issues such as low precision and accuracy and incompleteness. A calibration process is often applied to address such issues. With the recent advances in artificial intelligence, we have seen an increased number of applications that starts to use an artificial neural network (ANN) to calibrate the sensors, and their results are promising. In this work, we used a six-months worth of real hourly data to demonstrate that the ANN may not always be the best choice of a calibration method. Our evaluation compares an ANN-based method with a simple regression-based method in various aspects. The result shows that the ANN-based method does not consistently outperform the regression-based method. More interestingly, in the comparison, our results suggest that the performance of a calibration can be more sensitive to some of the factors (e.g. training and testing data, model parameters) than the use of different calibration methods. Even though the results may not be generalised in other sensors or datasets, our evaluation provides evidence showing that inappropriate use of a calibration method can compromise the calibration result, and the use of the ANN will not magically solve that problem.