air pollution prediction deep learning for traffic and · 2019-11-21 · deep learning for traffic...
TRANSCRIPT
![Page 1: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/1.jpg)
Deep Learning for Traffic and Air Pollution Prediction
Georges Alkhouri, Moritz WilkeScaDS Leipzig
www.scads.de
![Page 2: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/2.jpg)
Goals
▪ Predict air pollution
▪ Investigate the influence of car traffic
Problems
▪ Are the sensor data reliable?
▪ Can we build a successful model of the complex system?
▪ How can we get traffic data in a similar resolution, time and location?
![Page 3: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/3.jpg)
+Public webcam
How to get traffic data?
+
Deep Learning for object recognition¹
= ?¹https://github.com/matterport/Mask_RCNN
![Page 4: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/4.jpg)
Steps
▪ Finding the Right Spot
▪ Preprocessing Sensor Data
▪ Building a Prediction Model
▪ Traffic Data
▪ Transfer Learning
▪ Results and Outlook
![Page 5: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/5.jpg)
Finding the right spot
![Page 6: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/6.jpg)
Finding the right spot
blue: camera, red: too many invalid data points, green: chosen
![Page 7: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/7.jpg)
Data Preprocessing
![Page 8: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/8.jpg)
Too much humidity
Red: Not usable data points for PM measurements Discarded locations 3123, 10574
![Page 9: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/9.jpg)
Data Preprocessing Steps
▪ remove measurements where humidity > 70%▪ outliers removal▪ average over all (3) locations
![Page 10: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/10.jpg)
Time series prediction model
![Page 11: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/11.jpg)
Model: learning, baseline, results
▪ because of possible time dependency → LSTM net usage
Models:
1. Baseline/Persistence Model (PM10t = PM10t-1)
2. TS Model (Stateful LSTM)3. Normal Model
Input as time series for Model 2
Shuffled and broken time series for Model 3
![Page 12: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/12.jpg)
Results: Models vs. Baseline
PersistenceModel
NormalModel
TSModel
MSE 0.00447 0.01781 0.02953
std 0.00236 0.00967 0.01947
![Page 13: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/13.jpg)
Data: do we have a Random Walk?
▪ Persistence Model is still the best
▪ Temporal dependence from used features is strong
→ add more features to Model
PM10 Autocorrelation
![Page 14: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/14.jpg)
Working with webcam data
![Page 15: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/15.jpg)
Webcam Data
▪ Store an image every 5 minutes▪ Process with pre-trained neural network (Mask R-CNN / COCO) ▪ Result: car timeseries for february 2019
![Page 16: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/16.jpg)
Webcam Problems
▪ Only short timespan▪ Pre-trained model not optimized on perspective and night ▪ No exact numbers, only magnitude
![Page 17: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/17.jpg)
Transfer LearningCombining both data sources
![Page 18: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/18.jpg)
Transfer Learning
▪ Car Dataset is very small compared to other data▪ Approach: train model first and use extracted features with
car data
Input
DenseCar feat. Input
Regression
Pre-trained Networks Not trainable
![Page 19: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/19.jpg)
The Influence of Traffic
New models traffic can now be parametrized manually to simulate future traffic behaviour
Blue: PM measures for actual trafficOrange: PM measures for simulated traffic (higher traffic higher traffic density)
![Page 20: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/20.jpg)
Conclusion and Outlook
![Page 21: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/21.jpg)
Conclusion and Outlook
▪ Pollution time series can not easily be forecasted from the
given data
▪ Traffic Data can be obtained from public sources
▪ To investigate the relationship between particulate matter
pollution and traffic, different features and more data are
needed
![Page 22: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/22.jpg)
https://github.com/GeorgesAlkhouri/golddust
On Github
![Page 23: Air Pollution Prediction Deep Learning for Traffic and · 2019-11-21 · Deep Learning for Traffic and Air Pollution Prediction Georges Alkhouri, Moritz Wilke ScaDS Leipzig](https://reader034.vdocuments.us/reader034/viewer/2022042308/5ed489ab7d08392cd01d1fed/html5/thumbnails/23.jpg)
Thank you for your attention.
Questions?