An application of artificial intelligence for investigating the effect of COVID-19 lockdown on three-dimensional temperature variation in equatorial Africa

Abstract

We present interesting application of artificial intelligence for investigating effect of the COVID-19 lockdown on 3-dimensional temperature variation across Nigeria (2°-15° E, 4°-14° N), in equatorial Africa. Artificial neural networks were trained to learn time-series temperature variation patterns using radio occultation measurements of atmospheric temperature from the Constellation Observing System for Meteorology, Ionosphere, and Climate (COSMIC). Data used for training, validation and testing of the neural networks covered period prior to the lockdown. There was also an investigation into the viability of solar activity indicator (represented by the sunspot number) as an input for the process. The results indicated that including the sunspot number as an input for the training did not improve the network prediction accuracy. The trained network was then used to predict values for the lockdown period. Since the network was trained using pre-lockdown dataset, predictions from the network are regarded as expected temperatures, should there have been no lockdown. By comparing with the actual COSMIC measurements during the lockdown period, effects of the lockdown on atmospheric temperatures were deduced. In overall, the mean altitudinal temperatures rose by about 1.1 °C above expected values during the lockdown. An altitudinal breakdown, at 1 km resolution, reveals that the values were typically below 0.5 °C at most of the altitudes, but exceeded 1 °C at 28 and 29 km altitudes. The temperatures were also observed to drop below expected values at altitudes of 0-2 km, and 17-20 km.

Keywords: COVID-19 lockdown; Equatorial Africa; Neural network; Sunspot number; Temperature; Time-series.

Graphical abstract

Fig. 1

Flowchart containing descriptive summary of processes employed in this study.

Fig. 2

MAEs computed for the neural networks, with and without SSN input neuron, and varying the number of hidden layer neurons from 1 to 15 in steps of 1.

Fig. 3

Altitudinal profiles of the Mean Absolute Errors (MAE), the Minimum Absolute Errors (MnAE), the Maximum Absolute Errors (MxAE), the standard deviations of the atmospheric temperatures (STD), the Coefficients of Determination (RSq), the Percentages of data points with an absolute error higher than the MAE (PAM), and the Root-Mean-Square Errors (RMSE), binned at 1 km interval.

Fig. 4

COSMIC temperature measurements versus neural network predictions for the months of April to September in year (a) 2018, and (b) 2020. The black lines are best-fit straight lines for the data points.

Fig. 5

Altitude-based distribution of differences computed between the COSMIC measurements and neural network predictions for the months of April to September of year (a) 2018, and (b) 2020. Normalized distribution of the differences for the months of April to September of year (c) 2018, and (d) 2020.

Fig. 6

(a) Altitudinal variation of the Mean Temperature Differences for April to September 2020, with the neural network mean prediction errors removed (The neural network mean prediction errors are estimated using the Mean Temperature Differences for April to September 2018), and (b) Altitudinal variation of the Mean Temperatures, computed using dataset for the months of April to September 2018 and April to September 2020. The lengths of the error bars in (a) represent the standard deviations of the differences between the predictions and measurements at each altitude bin.

Fig. 7

Seasonal variation of the Mean Absolute Error computed as a function of altitude, using the test dataset.