Applying deep learning to satellite imagery for emission inventory development
File(s)
Author(s)
Sheehan, Annalisa
Type
Thesis
Abstract
Accurate traffic data are crucial for a range of different applications such as quantifying vehicle emissions, transportation planning and management, and road accident statistics. However, globally, traffic data are at best fragmented and at worst non-existent. This study had two aims. Firstly, to investigate whether an accurate assessment of traffic activity can be built from satellite imagery. Secondly, to understand whether these satellite-derived traffic data can be used to quantify vehicle emissions in order to produce an emissions inventory. Although vehicle data can be obtained locally, the use of satellite data could provide a globally standardised approach. To this end, a deep-learning object detection model was trained to identify vehicle activity using satellite imagery of Barcelona. Uniquely, this research exploited the spectral configuration of the satellite to determine vehicle speed. The resulting model classified the speed of more than 600,000 vehicles in ten satellite images, which were validated using Google Directions API data and Barcelona Council Traffic Counter data. Model accuracy was optimised at 71%, demonstrating that satellite object detection models can be used to identify vehicles and their speed. To investigate whether satellite-derived traffic data can be used to produce an emissions inventory, this study undertook an atmospheric measurement campaign in Barcelona focusing on non-exhaust emissions which are becoming increasingly important. Non-exhaust concentrations were calculated using a source-specific tracer approach and combined with exhaust gas dilution to provide emission factors. However, these were higher than literature values, and approximately twice those calculated using a similar approach for London. To ensure their accuracy, they were validated by combining complementary chemical composition measurements and comparing to independently measured PM10 mass concentration. The satellite-derived Non-Exhaust emissions inventory developed was half of the existing Barcelona municipal inventory. Highlighting that urban topography and noise in imagery such as shadow, vegetation and buildings were important characteristics when estimating traffic on all roads in a city from satellite imagery. Nevertheless, the consistency with the traffic validation data sets demonstrates that this approach can be extended to model cities around the world.
Version
Open Access
Date Issued
2022-12-23
Date Awarded
01/09/2023
License URL
Advisor
Green, David
Beevers, Sean
Sponsor
Natural Environment Research Council (Great Britain)
Grant Number
NE/L002485/1
Publisher Department
School of Public Health
Publisher Institution
Imperial College London
Qualification Level
Doctoral
Qualification Name
Doctor of Philosophy (PhD)