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

Abstract

The measurement of air temperature is strongly influenced by environmental factors such as solar radiation, humidity, wind speed and rainfall. This is problematic in low-cost air temperature sensors, which lack a radiation shield or a forced aspiration system, exposing them to direct sunlight and condensation. In this study, we developed a machine learning-based calibration method for air temperature measurement by a low-cost sensor. An artificial neural network (ANN) was used to balance the effect of multiple environmental factors on the measurements. Data were collected over 305 days, at three different locations in Japan, and used to evaluate the performance of the approach. Data collected at the same location and at different locations were used for training and testing, and the former was also used for k-fold cross-validation, demonstrating an average improvement in mean absolute error (MAE) from 1.62 to 0.67 by applying our method. Some calibration failures were noted, due to abrupt changes in environmental conditions such as solar radiation or rainfall. The MAE was shown to decrease even when the data collected in different nearby locations were used for training and testing. However, the results also showed that negative effects arose when data obtained from widely-separated locations were used, because of the significant environmental differences between them.

Keywords: agriculture; artificial neural network; low-cost sensor; machine learning; sensor calibration.

Figure 1

Appearance of the air temperature sensor of the Automated Meteorological Data Acquisition System (AMeDAS). The sensor is covered by a radiation shield made of stainless steel. There is 5-${\mathrm{ms}}^{-1}$ air flow in the radiation shield generated by a forced aspiration system.

Figure 2

Appearance of the low-cost air temperature sensor.

Figure 3

Location map of experimental sites and sensor setups (Pin1: tanashi-117; Pin2: tanashi-118; Pin3: sensor measuring reference air temperature for Tokyo site; Pin4: sagatea-111; Pin5: AMeDAS in Saga site). (a) Experimental sites in Japan; (b) sensor setup in Saga; (c) sensor setup in Tokyo.

Figure 3

Location map of experimental sites and sensor setups (Pin1: tanashi-117; Pin2: tanashi-118; Pin3: sensor measuring reference air temperature for Tokyo site; Pin4: sagatea-111; Pin5: AMeDAS in Saga site). (a) Experimental sites in Japan; (b) sensor setup in Saga; (c) sensor setup in Tokyo.

Figure 4

Appearance of Stevenson screens (photographed by Nachoman-au, distributed under a CC BY-SA3.0 license).

Figure 5

Structure of the neural network.

Figure 6

Indoor air temperatures measured by low-cost and high-accuracy (reference) sensors. The indoor experiment was conducted for 15 days, and the air temperatures were very close in the whole period.

Figure 7

MAE between air temperatures from high-accuracy sensor and low-cost sensor (original), linear regression-based calibration (linear) and ANN-based calibration (ANN).

Figure 8

R-squared values between air temperatures from high-accuracy sensor and low-cost sensor (original), linear regression-based calibration (linear) and ANN-based calibration (ANN).

Figure 9

Examples of successful calibration. (a) 25 January 2015; (b) 28 September 2015.

Figure 10

Average and diurnal temperatures before and after calibration. “Low-cost sensor”, “linear” and “ANN” indicate original values, linear regression-based calibration values and ANN-based calibration values of the observations by the low-cost sensors respectively. (a) Daily average temperature; (b) diurnal range of temperature.

Figure 11

Histogram of relative error between reference and calibrated air temperatures.

Figure 12

Examples of calibration failure. (a) 24 March 2015; (b) 14 April 2015; (c) 4 September 2015.

Figure 12

Examples of calibration failure. (a) 24 March 2015; (b) 14 April 2015; (c) 4 September 2015.

Figure 13

Box plots of differences between air temperatures measured by the high-accuracy and low-cost sensors at different hours. Bar plots represent mean solar radiation in each hour. (a) sagatea-111; (b) tanashi-117; (c) tanashi-118.