Virtual Sensors for Air Quality (AQ) Measurements in Urban Environments

Air Quality (AQ) is currently measured using ground-based stations, which require funding and infrastructure to support. In contrast, high-resolution satellite imagery (~2 m/pixel) can be produced for almost any location on earth and is readily available. The goal of this project is to develop a pipeline that uses this imagery to produce maps of AQ, along with social and environmental metrics for developing cities around the globe (e.g. Accra, Dhaka) with greater resolution and coverage than existing methods. To achieve this, we engineer different types of features that are fed to a Deep Neural Network (DNN).

To extract object-based features, we feed the satellite images into the convolutional layers of a pretrained VGG-16 network. To engineer atmospheric features, we use Landsat 8 ground reflectance products to subtract ground reflectance from the satellite images. This produces an atmospheric reflectance image, which is then fed into convolutional layers. The outputs of the object-based and the atmospheric features are flattened into fully-connected layers, along with features from external sources like meteorological, demographic, and elevation data. The entire DNN is then trained to predict AQ metrics. The developed model is tested and evaluated using both ground monitoring and modeled Particulate Matter and NO 2 concentrations. We obtained images for over a decade from 35 urban locations around the globe; a case study with a model trained on images of Greater London will be presented.

Current breakthroughs in DNNs will enable, for the first time, AQ estimation using high resolution satellite imagery in the complete absence of ground measurements and help address a global challenge in Earth Science.