Low-cost environmental sensors encounter challenges of reliability and accuracy. This thesis aims at increasing the readings accuracy of low-cost air quality sensors, particularly using three Air Quality Egg (AQE) version 2 to measure the CO and NO₂ concentrations in the air, in a supervised manner, by proposing the outlier and adjustment modules. The outlier module detects and eliminates noise in the sensor readings, while the adjustment module aims to increase the sensors’ reading accuracy. Four proposed detection schemes and three mathematical algorithms were trained and tested in the outlier module and adjustment module, respectively.

The temperature, humidity, and CO sensors on the AQEs had good readings agreement based on the index of agreement (d-value), except for the NO₂. Thus, the selection of the schemes for the outlier module can be based on the sensors’ characteristic. The scheme that verifies the adjacent nodes before marking an outlier has better ability to classify the outlier than a scheme that does not. Artificial neural network (ANN) outperformed the other two mathematical techniques in the adjustment module because the accuracy performance is influenced by the sensors readings agreement and the performance of the outlier module.

The CO, temperature, and humidity sensors on the three AQEs can fit the CO reference from Environment Canterbury’s data by at least 80% with or without the existence of the outlier module. On the other hand, the NO₂, temperature and humidity sensors can fit the NO2 reference by 40% and at least 80%, respectively, without and with the filter from the outlier module.