Estimating Road Transportation Emissions Globally using CNNs and Satellite Imagery

Road transportation is one of the largest sectors of greenhouse gas emissions. To better understand road transportation emissions and make progress towards reducing them, we must develop scalable inventory capabilities. As such, our team has been experimenting with convolutional neural networks (CNNs) to estimate road transportation emissions from satellite imagery. Our initial work focused on the conterminous United States, where high-resolution gridded inventory estimates are available for training our models. We explored training CNNs to perform pixel-wise road transportation emissions regression using a variety of inputs, including 10m satellite imagery, road network data, carbon measurements, and population data. Our best-performing model used visual and road data with an MA-Net architecture and EfficientNet-B3 backbone. All models were trained using DARTE data as labels with a loss function based on root mean squared logarithmic error. To scale globally, we developed a pipeline leveraging Pangeo tools and Microsofts Planetary Computer. Dask, xarray, and StackSTAC enabled us to parallelize and accelerate the mosaicking of low-cloud satellite imagery and subsequent CNN inference of 1024x1024 chunks. Raw inference results are stored in Zarr on Azure Blob storage and capture non-country-specific geographic regions. To estimate country-specific emissions, we cropped the model outputs to country boundaries and stored them in an Azure Cosmos database for rapid summation. With our scalable processing pipeline in place and initial global emissions estimates generated, we have identified several scientific challenges associated with estimating road transportation emissions globally. For example, generalizing to drastically different climates, such as ones with substantial snow and ice, is challenging and currently causes unpredictable over-estimates from our model. We will discuss the challenges and potential solutions that we have identified.