Estimating Power Plant Greenhouse Gas Emissions from Satellite Imagery
Applied Research Earth Observation Deep Learning Multi-task Learning Power Plants CO2 Sentinel-2
This project forms the logical extension of our previous project on the characterization of industrial plumes from remote sensing data. Instead of simply identifying and characterizing plumes, we utilize this information to estimate Greenhouse Gas emissions from fossil fuel-firing power plants.
The idea is rather simple: we know that we can segment plumes and robustly distinguish them from natural clouds. For all European power plants, we know for any given time what their power generation rates are. We can therefore train a regression model to estimate the power generation rate from the extent of the observed plume for any given time. However, the apparent size of the plume is not only a function of the power generation process. Instead, it is a highly complex process that relies on environmental variables. To approximate these variables, we provide concurrent weather information, including ambient temperature, air pressure and wind speed to our regression model.
To improve our power generation rate predictions, we utilize a multi-task learning approach: instead of simply training a model that predicts power generation, we learn additional tasks, namely a power plant classification model (see here and a plume segmentation model (see here). By learning additional tasks, the model generates richer representations of the underlying data, resulting in a more robust training and better performance. This is reflected in our results: the combined multi-task learning approach performs better on each learned tasks than a separately trained model. We can also show the benefit of adding weather data to our model input.
So far, our model is able to estimate power generation rates (with an MAE of 139 MW) - how do we get CO2 emission rates from this result? Our model is also able to identify power plant types; depending on the fuel they use, they emit different amounts of CO2 per MWh into the atmosphere. By approximating these emission factors, we are able to convert our power generation rate estimates into emission rate estimates.
Our approach enables the estimation of power generation rates and CO2 emission rates of power plants based on freely available Sentinel-2 satellite imagery. Results have been published at the Tackling Climate Change with Machine Learning Workshop at NeurIPS 2021 and a journal paper with all details is currently in review.
Hanna, J., Mommert, M., Scheibenreif, L. Borth, D., “Multitask Learning for Estimating Power Plant Greenhouse Gas Emissions from Satellite Imagery”, NeurIPS 2021 Tackling Climate Change with Machine Learning Workshop; publication (open access).
Hanna J. et al. 2021, code
Hanna J. et al. 2021, dataset