=Paper=
{{Paper
|id=Vol-3181/paper74
|storemode=property
|title=Multimodal Deep Learning for Transboundary Haze Prediction
|pdfUrl=https://ceur-ws.org/Vol-3181/paper74.pdf
|volume=Vol-3181
|authors=Thinh Nguyen,Nazmudeen Mohamed Saleem
|dblpUrl=https://dblp.org/rec/conf/mediaeval/NguyenS21
}}
==Multimodal Deep Learning for Transboundary Haze Prediction==
https://ceur-ws.org/Vol-3181/paper74.pdf
Multimodal Deep Learning for Transboundary Haze Prediction
Phuc-Thinh Nguyen1 , Nazmudeen Mohamed Saleem2
1 University of Information Technology, Vietnam
2 University Teknologi Brunei
ABSTRACT With testing data (from 2018-2019), we’ll utilize the zero-fill
Environmental pollution, particularly air pollution, has long been method to generate a mask to help the model work with missing
an issue in every major city on the planet. For many years, accurate values of data that is used for prediction.
estimation of PM2.5 and PM10[8] [12] fine dust concentration val-
ues has been a fascinating study area. This study focused on 3-Day 2.2 Features Extraction
Transboundary Air Pollution Prediction, which proposed to merge We model the PM10 index prediction problem as a regression prob-
many deep learning models and pick appropriate properties for the lem with the following features to estimate the PM10 index in the
PM10 index prediction problem by utilizing various features such as near future from a list of specified attributes.
timestamps, geographical features, and public weather data. Using
the dataset provided by MediaEval, we examined the performance 2.2.1 Timestamp features. Because outdoor air quality varies
of several learning models and features in order to investigate the greatly depending on the time of day, timestamp information can
problem. Experimental results show that combining multiple deep be valuable for PM10 estimation difficulties. In particular, each
learning models together gives a higher overall performance than country’s PM10 index is unique, as are the provinces within the
other techniques and features in RMSE, MAE, SMAPE. same country. As a result, we created a PM10 index for a country
Keywords—PM10 prediction, LSTM, BiLSTM, multimodal by averaging the PM10 scores of the provinces. We take the daily
average and then average it across provinces in provinces where
PM10 is reported hourly.
1 INTRODUCTION
2.2.2 Location features. We choose one province as a landmark,
Haze air pollution is defined as the presence in the air of particulate and then we evaluate the closest distances between provinces in
matter such as smoke, dust, and other vapours that arise from the the vicinity of the landmark as it can provide useful information
large-scale forest and land fires, factories, and automobiles. When for plot analysis and reduce noise due to air pollution in the data.
the concentration of airborne pollutants reaches dangerous levels, We calculated the distance using the Haversine formula[9], which
it causes respiratory issues and has significant consequences for is an integral equation for navigation that yields precise results
visibility, economic productivity, transportation, and tourism.[7] for calculating the great circle distance between two points on the
Transboundary haze is a recurring problem in many parts of the Earth’s surface based on their latitudes. The Haversine formula can
world, particularly in Southeast Asia, where haze pollution sources be calculated using two positions, A and B:
differ from nation to country, with varying percentages coming
from localized or transboundary sources.[7]
𝜑 − 𝜑
The goal of this article is to address the sub-task 2 of the competi- 𝑑 (𝐴, 𝐵) = 2.𝑟 .𝑎𝑟𝑐𝑠𝑖𝑛 𝑠𝑖𝑛 2
𝐵 𝐴
+
tion which is to examine the transnational PM10 estimate problem 2
𝜆 − 𝜆 1
using timestamp information, location data, and weather data using 2
𝑐𝑜𝑠 (𝜑𝐴 ).𝑐𝑜𝑠 (𝜑 𝐵 ).𝑠𝑖𝑛 2 𝐵 𝐴
, (1)
the technique of mixing multiple deep learning models.[1][11] 2
where r is the Earth’s radius, and 𝜑𝐴 , 𝜆𝐴 , 𝜑 𝐵 , 𝜆𝐵 are the latitudes
2 METHODOLOGY and longitudes of two points A and B, respectively.
We’ll start with a high-level overview of the topic and then go
over to our specific approach that we have developed for a PM10 2.2.3 Weather features. Public weather features include infor-
prediction algorithm. mation on weather, such as "temperature," "precipitation," "humid-
ity," "wind direction," and "wind speed," obtained from local stations.
2.1 Pre-processing data These characteristics can be thought of as supplementary data that
can help machine learning models become more robust and reliable.
At this stage the data for each station is separated from the original
In order to provide the best forecast results, we analyze the
data (train air quality and train weather) based on the station’s
correlation between these weather features and the PM10 index in
ID. Then we will combine the data by province as the data on air
each country and select the features with the strongest correlation.
quality and weather can be found in the same province.
The wind direction had a strong correlation with PM10 in the
We’ll utilize the interpolation method[6] for missing values in
Indonesia dataset; the temperature, rain, and humidity in the Brunei
the data. In addition, we fill in the mean for the variables that cannot
dataset; the rain, humidity, and wind speed in the Thailand dataset;
be interpolated.
and the temperature in the Singapore dataset.
Copyright 2021 for this paper by its authors. Use permitted under Creative Commons
If some of the provinces are lacking "Temperature" data We
License Attribution 4.0 International (CC BY 4.0). utilize a basic LSTM[2] approach to predict the "Temperature" using
MediaEval’21, December 13-15 2021, Online weather data from 2018-2019 (testing data). .
MediaEval’21, 13-15 December 2021, Online T.Nguyen
2.3 Training and Testing Setup 3.2 Model settings
To train a model that is appropriate for the problem, we split the raw To improve the performance of the presented approaches, we use a
training dataset (2010-2017) into training and test datasets, using random search method to select ideal hyperparameters based on
80% of them for training and the remaining 20% for model testing. performance indicators. For each model, this method entails scan-
After that, we will predict PM10 on testing dataset (2018-2019). This ning a predefined parameter space and picking the best performing
study uses a regression model to forecast PM10 for the next three hyperparameters which is shown in Table 1.
days.
2.4 Multimodal Method
Process the data for each station in section 2.1, then build a
dataset for each country by averaging the weather and air quality 3.3 Results
features of the stations in that country; for example, data of Brunei In this study, we compare the proposed methodologies’ perfor-
would be obtained by averaging the weather and air quality features mance to the performance measures listed in Section 2.5.
of the stations in Brunei. The MAE and RMSE results of multimodal are better than the rest
For each country, we use data from the previous three days to of the models, as shown in Table 2. However, because multimodal
forecast the future three days. employs the average PM10 value of countries, the forecast results
We employ the concept of merging different models[5][3] to will be skewed for countries with excessively low or excessively
increase the performance of the PM10 estimator. The fundamental high PM10.
idea is to connect the outputs of each country’s PM10 prediction The MAE and RMSE values in Table 2 are better than those in
deep learning model, then feed that information into a final deep Table 3, but the SMAPE results are worse.
learning model to get the final PM10 result.
We’ll need three branches to construct our multi-input network:
The first two forks will be a simple BiLSTM[2] that will handle
Brunei and Thailand data repectively. A simple LSTM will handle
Indonesian data inputs in the last fork. Finally, concatenate these
branches to produce the final multi-input deep learning model. It’s
random; you’re free to use Indonesian data in the BiLSTM model.
We replaced it throughout the experiment, and the MAE and RMSE
results are nearly similar.
We must fill in the mean from 2010 to 2015 because Singapore
weather data is only available from 2016 to 2017, hence we do not
recommend using Singapore data to train the multimodal model
to avoid overfitting the model. We’ll run the BiLSTM model for
4 CONCLUSIONS AND FUTURE WORKS
Singapore data separately.
After three months for analyzing, we illustrated the benefits of com-
2.5 Performance metrics bining numerous models with generalization and deep learning
methods to address the PM10 index estimation problem utilizing
We employ root mean square error (RMSE), mean absolute er-
various types of features such as timestamps, location, and public
ror (MAE)[10], and Symmetric mean absolute percentage error
weather data. By incorporating a variety of characteristics, the test
(SMAPE)[4] to evaluate the performance of the proposed approaches.
results reveal that PM10 level prediction is fairly accurate when
3 EXPERIMENTS compared to ground-truth. Transnational air pollution can be pre-
dicted using the strategy of merging multiple models.
3.1 Data sets We’re excited to continue our research by examining additional
The organizer has provided us with a data set. Weather and air qual- forms of data, such as image data, video, and new deep learning
ity data are included in the dataset, with the training set spanning models. One of the new approachs is multivariate transformer
2010-2017 and the test set spanning 2018-2019. learning because this learning can learn across long timespans.
3-Day Transboundary Air Pollution Prediction MediaEval’21, 13-15 December 2021, Online
REFERENCES
[1] AR Varkonyi-Koczy. A Mosavi, S Ardabili. 2019. List of Deep Learning
Models. International Conference on Global Research and Education
(2019).
[2] Jason Brownlee. 2018. Deep Learning for Time Series Forecasting.
[3] Quang M.; Nguyen-Tai Tan-Loc; Bo Dong; Nguyen Dat; Dao Minh-
Son; Nguyen Binh T. Duong, Dat Q.; Le. 2020. Multi-source Machine
Learning for AQI Estimation. 5 (2020).
[4] Wida Susanty Haji Suhailia-Peijiang Zhaob. Effa Nabilla Aziza,
Asem Kasema. 2021. Convolution Recurrent Neural Network for
Daily Forecast of PM10 Concentrations in Brunei Darussalam. AIDIC
13 (2021).
[5] D Roggen. FJ Ordóñez. 2016. Deep Convolutional and LSTM Recurrent
Neural Networks for Multimodal Wearable Activity Recognition. mdpi
5 (2016).
[6] Y. T. GU. G. R. LIU. 2001. A POINT INTERPOLATION METHOD
FOR TWO-DIMESIONAL SOLIDS. INTERNATIONAL JOURNAL FOR
NUMERICAL METHODS IN ENGINEERING (2001).
[7] Asem Kasem, Minh-Son Dao, Effa Nabilla Aziz, Duc-Tien Dang-
Nguyen, Cathal Gurrin , Minh-Triet Tran, Thanh-Binh Nguyen, and
Wida Suhaili. Overview of MediaEval 2021: Insights for Wellbeing
Task Cross-Data Analytics for Transboundary Haze Prediction.
[8] Pingqing Fu Xiangdong Li Ling Jin, Xiaosan Luo. 2016. Airborne
particulate matter pollution in urban China: a chemical mixture per-
spective from sources to impacts. National Science Review 4 (2016),
593–610.
[9] S Suryati Widdha Mellyssa. M Basyir, M Nasir. 2017. Determination
of Nearest Emergency Service Office using Haversine Formula Based
on Android Platform. EMITTER 5 (2017).
[10] H Smith NR Draper. 1998. Applied regression analysis.
[11] G Hinton. Y LeCun, Y Bengio. 2015. Deep learning. Nature (2015).
[12] Min-Hua Shi-Yi-Xin Lian. Yu-Fei Xing, Yue-Hua Xu. 2016. The impact
of PM2.5 on the human respiratory system. Journal of Thoracic Disease
(Jan 2016).