Forecasting and assessing pollutants values accurately is always an attractive topic in the research community during the last decade. It will help to provide a good approach to problems related to health efects associated with current air quality conditions. In this paper, we will provide 2 approaches utilizing machine learning to solve 2 subtasks. In subtask 1, we use multi-models to forecast AQI value and in subtask 2, we use VGG16 model which was customized to predict AQI level using only pictures. We evaluated the performance of two models using the dataset from sensors stations installed across Dalat City, Vietnam. The experiment results show that our proposed models yield high performance in terms of MAE, MSE, and F1-score.
2.1. Subtask 1 After some visualization stages, there are several issues with provided data: untrusted data, missing data, and duplicated data. With untrusted data and missing data, we just simply remove all of them from our dataset. With some sensors (such as sensor 1), the amount of data missing was not too much to afect the result but with others, there are approximately 80% amount of data was missing so we calculated a ”base” value and applied it to our missing-timestamp data.
Because of missing data and the time of making the prediction was the time of intersection between the rainy and dry seasons can cause rapid weather condition changes so we decided to use a window time of 1 week.
Our used method contains 3 steps: Forecast weather components, predict pollutants’ value, and predict AQI based on the value of pollutants.
Because temperature is a regular feature, we used Fourier-transform to forecast the temperature values for the forecasting week, and after that used it to predict the humidity feature because of a strong correlation (Figure 2 -right) between them.
In the second stage, we used features forecasted and predicted in the last step to predict the
value of pollutants. Because several models cause negative predictions, after all, we choose
XBGBoost[
In the last stage, we predict AQI for each pollutant based on [
According to the requirements of the organizers, the pollutant indicators are calculated through AQI according to EPA standards. However, the EPA does not determine the AQI value for O3_1H when the index is less than 0.124, so the reference becomes unreasonable. So we propose a slight modification to the calculation of the AQI value for O3_1H with a value less than 0.124 according to the following rule: AQI 50, 100 corresponding to 0.08 and 0.124 O3_1H respectively. 2.2. Subtask 2 We approach this subtask using the image only. Firstly, we think about solving the problem with features extracted by hand including a number of vehicles and pedestrians, but we realize that this work does not ensure whether these features are good and we believe inexact features will get us the wrong way to approach this task and make the model potentially lead to inaccurate predictions. Therefore, we decide to solve this task using Multiple inputs and Multiple models such as VGG16 to predict each pollutant level because we would like features to be extracted by robust models through learning from data.
In the model, we use VGG16[
For each AQI pollutant level, we use this model architecture and train on a dataset including the AQI level of each pollutant of the sensor and CCTV nearest sensor (Figure 2 -left). We generate 6 models in the same architecture for predicting AQI pollutants - PM2.5, PM10, CO, O3, SO2, and NO2.
This model was based on Classification problems, we classify 5 levels of AQI pollutants PM2.5, PM10, CO, O3, SO2, and NO2. Before adding the vector time feature, we tested it on only VGG16. The loss and accuracy of this baseline model are not good. After adding the vector time feature, this new model predicts better than.
We train it on 20 epochs and we realize, the original VGG16 does not increase accuracy or loss, but our new model can approach 0.9 while the epoch comes over 100. The optimizer is Adam, we evaluate in sparse categorical cross-entropy loss. The activation function is Relu.
With each sensor, we use images from its nearest camera and these images we use are from 7 days ago or days near the time which we need to predict.
According to table 1, chemical pollutants have higher prediction performance than particulate matter values. This can be explained by some problems related to missing data and the lack of some necessary features (such as trafic being constantly updated because the request for the subtask is not allowed). Moreover, values of 2, 2, and 2 in Dalat as we know when visualizing them are always within a fixed range of AQI level such as good level, so in terms of ideas, we do not have to predict it.
In subtask 2, We evaluated the VGG16 on image features only and saw that the performance is not good. We thought that because image features are not enough and the time we need to predict AQI level is an important factor. Therefore, we tried to add additional encoded timestamp features and chose a small but comprehensive model, VGG16, that can capture useful features. Adding the time feature, in combination with our experimental evidence, has shown that our hyperthesis is correct. The addition of these timestamp features has led to an increase in accuracy and F1 score.
Model
VGG16 VGG16+ Time
Accuracy 0.71 0.80
Loss 0.66 0.47
Instead of using VGG16, with our conclusion, it can be replaced by more powerful models such as ResNet, DenseNet, ..., however the results will not be much better, the important thing here is the hand-features added.
In this competition, we used multi-models to forecast AQI for each pollutant. Because of several reasons, our performance was not as good as we had hoped but we have gained a lot of experience by participating in this task.
When building the correlation hypothesis between trafic, weather, and air pollution, both people’s experience and the model’s knowledge played important roles in the process. The diference between them is the source of the information being used and the level of subjectivity involved. People’s experience is based on their observations, learning and sometimes, ideas come from hunch. In fact, we are Vietnamese and have been to Dalat city many times, so that, we know which human activities often cause air pollution, which areas are crowded, as well as geographical, weather, and seasonal factors, but the prediction can be influenced by our own preconceptions. While a model’s knowledge is based on data and information that has been input into the model. When the data was collected incorrectly, no matter how good the model is, it cannot make the right prediction. However, when the data is “acceptable”, the model can identify patterns and trends that may not be immediately apparent to the human observer, just based on data and following logical rules and algorithms. The similarity between people’s experience and model’s knowledge is that both can be used to identify patterns and trends, test and validate the hypothesis. Therefore, the two are often used together to aid in the hypothesis-building process.
In the future, we think that we can have some improvements to our methods and maybe can get higher performance. For subtask 1, we can combine it with real trafic features (not now because we still cannot get the exact trafic from CCTV) to calculate the trafic flow over the city.