Using Convolutional Neural Network Models with TROPOMI
Using machine learning to identify plumes on a global scale
In this letter, we review a research study titled “Automated detection of atmospheric NO2 plumes from satellite data: a tool to help infer anthropogenic combustion emissions.” This study uses machine learning models to understand the distribution of nitrogen dioxide on a global scale. The full copy of the paper (currently under review for preprint) can be found here.
Convolutional Neural Network Models
Background
A 2021 study published in Atmospheric Measurement Techniques by Finch et. al (2021) uses convolutional neural network (CNN) models to identify anthropogenic sources of plumes on a global scale using nitrogen dioxide columns from the Tropospheric Ozone Monitoring Instrument (TROPOMI).
The motivation for this study was to use nitrogen dioxide plumes as a tracer for carbon dioxide (Finch et. al, 2021). This is considered a suitable proxy for multiple reasons. Firstly, nitrogen dioxide is a product of incomplete combustion. Secondly, nitrogen dioxide has a short life span and does not disperse far from the source location, which allows researchers to use it as a method to identify plumes. Finch et. al (2021) acknowledge that carbon dioxide is already being remotely measured using the NASA Orbiting Carbon Observatory-2 and state that the dilution that takes place across the 3 kilometre grid cell from the instrument provides further motivation to use nitrogen dioxide as a trace gas (Finch et. al, 2021).
When identifying plumes, the researchers wanted to remove the influence of biomass burning, and so they harnessed data from the Visible Infrared Imaging Radiometer Suite (VIIRS) to discern sources of biomass burning from the identified plumes (Finch et. al, 2021). Therefore, any plume within a 15 kilometer radius of a known biomass burning site was removed.
Identifying Plumes
This study downloaded and combined TROPOMI nitrogen dioxide columns from and created more than six thousand images of the pollutant from July 2018 to June 2020. The researchers used an online platform to engage public support in identifying plumes (Finch et. al, 2021). There was a lack of consensus for many sites, and the researchers ended up using their judgement for plume detection. Once detected, the images were normalized to remove the influence of higher values among the individual plume images.
Deep Learning Model
Once the plumes were identified, the CNN model was developed. The ratio to train and test the model was 80:20 (Finch et. al, 2021). On a high level, CNN models use multiple layers with an increasing number of filters to detect patterns in the input image. An example image is provided below. The end product of this model is a series of categorized images with a confidence level attributed to each image.
Conclusions and Results
Over the two-year study period, more than 247,000 plumes were identified, with the highest percentage (>20%) being from China, and the second-highest number from India (Finch et. al, 2021). The model had a success rate of more than 90%. Interestingly, the plumes from shipping tracks were also identified. A small percentage (around 8%) of plumes were from small sources of combustion, and a high density of plumes were identified over large cities (Finch et. al, 2021).
Takeaway
This paper demonstrates the value that machine learning models bring when analyzing satellite remote sensing data. Further summaries of research harnessing machine learning models will be posted over the upcoming months.