We propose a method to discover periodic trafic-pollution patterns in Dalat City, Vietnam. From the air pollution, weather, and CCTV stations data recorded in Dalat City, Vietnam, we convert them into an hourly uncertain temporal database and apply a periodic frequent pattern discovery algorithm to explore the most occurring patterns in the dataset, as well as finding some insights from the found patterns. Our source code and results can be found at https://github.com/TranDucHuyVnuHcmUs/ MediaEval2022-UrbanAir-HCMUS-public.
Research has recently been trying to discover the feasibility of various types of data to predict
or understand the mutual relationship between human activities and air pollution. Mike et al
[
After collecting CCTV data, we discover that the most continuous period is from 13 August 2022
to 31 October 2022. As a result, we download all CCTV images during this time period for training
purposes (220463 images). We use the Yolo-V5 [
After we have the table from 3.3, for each row, we apply the fuzzy negation method mentioned
in [
Taking ideas from [
Each transaction will also be linked with a timestamp. A default time is chosen as the beginning of time. We simply subtract all datetime values with the beginning datetime (converted to hour) to get the timestamp.
After this step, we’ll have an uncertain temporal database. We reimplement the algorithm
from [
In this attempt, we use all cameras for all sensors (each sensor will be linked with images from all cameras, that are closest in sensor data’s time).
We make some changes in our implementation. Aside from fixing errors, we change the way we calculate the prefixed item cap into using the minimal probability of previous items in each transaction: _(, ) = (, ) * ( (1, ), (2, ), ... (− 1, )).
We also change our fuzzy negation mapping into as follow:
Furthermore, for each transactions, we delete the items with probability below 0.5. This means no two items from the same feature (for example, Temperature_LOW and Temperature_MED) can appear in one transaction. By doing this, all transactions will have statistically independent items, assuming that the features are statistically independent.
We also split a subset of data with only trafic-related features and AQI, to explore the
interesting patterns between trafic and air pollution. We customized our search to use each
and only one AQI item in one mining, and do the same for each trafic item. Please refer to our
source code [
There are also a few temporal gaps in our database, which can be at most 120 hours (5 days). This gap afects the periodicity of those patterns because of their high values (since the periodicity is calculated by the maximum diference between occurring timestamps). So we delete a part of our database to lower the efect of these gaps on our results.
The results reported here are only a notable part of our results.
Because of the changes in our implementations, we provide our updated results for the full data mining (using all features) for clarity. We do not think, however, that the new result will change our insight. So we also provide new results from mining patterns with each AQI-related item the only AQI-related item present in the database, to see which most frequent factors contribute to each AQI level.
Our patterns shows that the AQI values are mostly low, even with high trafic . And the trafic is generally low as well. However, with high trafic signal, especially for motorbike, car and person, the AQI can still rise and thus afect the environment. Interestingly, in most cases, highest expected support are found within patterns with no high trafic item included, while the most periodic patterns are the patterns with at least one high trafic item present.
On a side note, using the min function to calculate the prefixed item cap increase performance of our implementation. (Appendix D).
Our approach are prone to error, as we don’t exclude non-moving trafic objects, especially parking vehicles. We also find that applying the same fuzzy negation to all non-AQI columns cause the distribution of some items to be imbalanced (for example, Humidity_HIGH appear in the entire database!). It may be more suitable to have diferent mappings for diferent columns.
Further steps for improvement are: optimizing the algorithm, adding environmental and infrastructure factors, parking vehicles detection,...
Da Lat’s air quality is overall good, even when the trafic is high. High trafic, however, can still worsen air pollution.
Duc-Huy Tran, Vinh-Thuyen Nguyen-Truong, Xuan-Cuong Le were funded by Vingroup JSC and supported by the Master, PhD Scholarship Programme of Vingroup Innovation Foundation (VINIF), Institute of Big Data, codes VINIF.2021.ThS.JVN.09, VINIF.2021.ThS.JVN.10, VINIF.2021.ThS.JVN.08.
We execute the pre-trained YOLO V5 on 20463 images from 16 CCTV stations. Afterward, we accumulate all the results by CamereID of person, motorbike, car, truck, and bus. We choose these five types of objects because CCTV mostly covers them (the number of other types is extremely small compared to these 5 main objects).
According to the result of YOLO-V5 on our CCTV data (Table 4), Camera16 doesn’t find too many trafic objects (to our own knowledge, Camera16 mostly points to natural scenes from above). On the other hand, Camera 01 takes the most crowded pictures. We also have a broken camera - Camera14.
Thís approach are prone to error: Idle objects, especially parking vehicles, though not emitting chemicals, can still be included.
There are two attributes that need to consider before merging: time and space.
For the space, we calculate the geographical distances between all pairs of camera-sensor,
and are calculated using GeoPy [
For the time, we only merge data rows of the camera-sensor with their time diference not larger than a defined threshold. In our initial setting, this threshold is 329.015(7) = (/) * 3600, with = 32(/ℎ), and is the maximal value of each nearest camera-sensor pair. The intuition is that vehicles are mostly moving, and since the sensors recorded their values at a diferent time compared to the cameras, we assume that these vehicles would have moved a distance by the time each sensor’s respective camera took another picture.
In our second attempt, we merged all cameras with all sensor data, and use 300 (second) as tolerance constant to merge the data by time (which means camera rows can only match with sensor rows if their time diference is no more than 300 seconds).
Two tables below explain our initial mappings (see table 1) used for our submission. For values in one column X (not related to AQI) that are in 0% - 12.5% of its [min, max] range, these transactions will be given an item as X_LOW(1). For values in 18.75% (average of 12.5% and 25%) of its range, those transactions will have 2 items X_LOW(0.5) and X_MED(0.5). Note that both items will exist in those transactions.
We process the AQI-related items diferently - instead of calculating the percentage of [min, max] range first, we directly use its values to convert. Because the meaning and purpose of AQI is a unified metric for measuring the concentration of (and health afect caused by) air pollutants, and each air pollutant has a diferent [min, max] range, we don’t convert the values to percentages to not lose the unified meaning.
It’s worth noticing that our mapping is not the same as the AQI conversion table provided for us: lv1, lv2, and lv3 are in the ’GOOD’ category. Our reasoning to divide the first 3 levels into such small gaps is because the average AQI values of each air pollutant are below 50 (’GOOD’ category). Especially for NO2, SO2 and PM10 air pollutant, their mean values are lower than 14, and for SO2, its max values is 14.285714.
Our experiments with both max and min functions (’mode’) to calculate the prefixed item cap show that the second algorithm run faster than the previous implementation, and with fewer potential patterns (patterns that may meet the requirements, but not ensured). Due to stronger bounds to the actual probabilities, more recursive calls are pruned, thus results in less time and fewer potential patterns to be considered, and will be included or filtered out in the final step.
Our results also show that two modes can produce diferent longest patterns, both potential and final.
High diference (more than 1) of the longest length between potential patterns and final patterns can be a problem, as too many potential patterns are not satisfying the requirements, thus increasing runtime while not gaining better results. Therefore, it’s our future works to reduce this diference. An improvement that can be made is to introduce closer bounds to prune more recursive calls.