Air Quality Data Time Series Modeling Using Deep Recurrent Neural Networks on Sentinel-5p Products

Abstract:

The rise in atmospheric greenhouse gases is known to be the primary driver of global warming. Air pollution is a huge public health issue, whose mitigation along with the restriction of Green-House Gas (GHG) emissions are critical factors in strategic planning, management, and decision-making at worldwide level. Therefore, closely monitoring the chemical composition and physical properties of the atmosphere is crucial to understanding climate change. Remote Sensing has lately led to mature assessments of atmospheric composition from space. Recently, the Sentinel-5 Precursor (S5p) mission, provides information and services on air quality, climate, and the ozone layer. In particular, the TROPOspheric Monitoring Instrument (TROPOMI) is an ultraviolet, visible, near and short-wavelength infrared spectrometer that provides accurate and timely observations of key atmospheric species for air quality monitoring.

In this work, a number of geophysical (L2) products collected by the TROPOMI instrument, are processed and used for air quality prediction as well as for monitoring the concentrations of atmospheric constituents. The main purpose of this study is to accurately detect recurring patterns of GHG emissions and other pollutants, using advanced machine learning algorithms. To better estimate temporal trends in the troposphere related to air quality and climate at regional scale, a special class of artificial neural networks are used, recurrent neural networks (RNNs). RNNs form connections between nodes along a temporal sequence, allowing to exhibit temporal dynamic behavior. Derived from feedforward neural networks, RNNs can use their internal state (memory) to process variable length sequences of inputs, whereas a special variation of RNNs, called Gated Recurrent Unit (GRU), can also successfully deal with temporal dependencies. Special structures of these networks, called gates, help to regulate the flow of information to the network, thus permitting to learn which data in a sequence should be considered significant and thus retained, and which not. The proposed model will also be adaptable, able to capture extreme conditions and adjust its behavior under new circumstances (e.g. regions with increased anthropogenic GHG emissions).