INTRODUCTION

The need to know the personal air pollution data is vital because it is better to provide each individual with regional air quality data, which seems to be more accurate than the global data measured from far away weather stations. The problem is finding a suitable method to predict air quality data in a local area from the global data. This paper reports our solution to tackle this challenge.

To know more about this challenge and the dataset that we will use, you can refer to the overview paper of MediaEval 2020 -Insight for Wellbeing: Multimodal personal health lifelog data analysis [1].

RELATED WORK

The inverse distance weighting method [4] is used commonly in spatial interpolation [3]. This paper will apply the basic form of IDW without any modification.

APPROACH

Due to the limited time available for experimenting with algorithms requiring more time to train data, such as neural network-related algorithms, we choose the IDW. Moreover, because there are no statistical assumptions involved [2], it is simpler than Kriging or other statistical interpolation methods. The way it works is easy to understand. Based on the assumption that closer points will have similar values than further points, it will use the measured values surrounding the unknown point to predict the value. By giving each known point a weight, the predicted value will be the average of those points.

The weight 𝑤 𝑖 for a known point 𝑖 is the inverse of the distance 𝑑 from that point to the unknown point 𝑥, which is computed as:

𝑤 𝑖 = 1 𝑑 (𝑥, 𝑥 𝑖 ) 𝑝 (1)

with 𝑝 is the power value that is used to control the value of the weight. It should be noticed that the Haversine method is used to calculate the distance between the two coordinates.

The value 𝑦 of an unknown point 𝑥 is calculated as:

𝑦 (𝑥) = 𝑛 𝑖=1 𝑤 𝑖 𝑋 𝑖 𝑛 𝑖=1 𝑤 𝑖(2)

with 𝑤 𝑖 is the weight, 𝑋 𝑖 is the value of the known point 𝑖 𝑡ℎ .

Prediction

At first, all possible time frame in hour-interval is listed by grouping the training data. Then, we start to loop through the training data per time frame.

In each loop, we get the coordinates of all unknown points that need to be predicted. After that, we get the values of the known points and their respective coordinates from the public air pollution data provided by 26 weather stations surrounding the Tokyo area also in that time frame.

With all the necessary data gathered, we can use the IDW formula to make the prediction. Please note that the initial power value 𝑝 of the IDW formula is 2.

After repeating those steps for each air pollutant data (PM2.5, NO2, O3), we have the final results.

Optimization

To have the best performance, we could find the optimal value of power value p by trying different values of 𝑝 until the IDW produces acceptable values of SMAPE/RMSE/MAE.

After evaluating the 𝑝-value ranges from 0 to 5, we find that the best power values for PM2.5, NO2, and O3 are 1.5, 3.5, and 0, respectively.

RESULTS AND ANALYSIS

The evaluation of PM2.5, NO2, O3, and AQI prediction provided by MediaEval task organizers are shown in Table 1, Table 2, Table 3, and Table 4, respectively.

In general, PM2.5 prediction is acceptable, but there is a big gap in NO2 and O3 prediction results. It is mainly because the IDW formula does not have any offset parameters to compensate for the big difference between weather stations' public weather data and the one carried out by personal equipment used by volunteers. This could be because of some differences in methods and devices of those two data providers.

MediaEval'20, December 14-15 2020, Online Quan N.T. et al.

DISCUSSION AND OUTLOOK

We intend to explore more advanced algorithms in our future work, such as the advanced form of IDW [4], the combination of IDW with multiple regression. Also, we plan to utilize more weather data, such as wind direction, wind speed, temperature, to improve accuracy.