MIIDAPS-AI: An Explainable Machine-Learning Algorithm for Infrared and Microwave Remote Sensing and Data Assimilation Preprocessing - Application to LEO and GEO Sensors
Eric S. Maddy, Sid‐Ahmed Boukabara
Abstract
In this paper we leverage and apply state-of-theart AI techniques to satellite remote sensing of temperature, moisture, surface, and cloud parameters in all-weather, allsurface conditions, from both microwave and infrared sensors. The Multi-Instrument Inversion and Data Assimilation Preprocessing System, Artificial Intelligence version, or MIIDAPS-AI for short, is valid for both polar and geostationary microwave and infrared sounders and imagers as well as for pairs of combined infrared and microwave sounders. The algorithm produces vertical profiles of temperature and moisture as well as surface temperature, surface emissivity and cloud parameters. Additional products from hyperspectral infrared sensors include selected trace gases. From microwave sensors, additional products such as rainfall rate, first year/multi-year sea ice concentration, and soil moisture can be derived from primary products. The MIIDAPS-AI algorithm is highly efficient with no noticeable decrease in accuracy compared to traditional operational sounding algorithms. The automatically generated Jacobians from this deep-learning algorithm could provide an explainability mechanism to build trustworthiness in the algorithm, and to quantify uncertainties of the algorithm's outputs. The computation gain is estimated to be two orders of magnitude, which opens the door to either (1) process massively larger amounts of satellite data, or to (2) offer improvements in timeliness and significant saving in computing power (and therefore cost) if the same amount of data is processed. Here we present an overview of the MIIDAPS-AI implementation, discuss its applicability to various sensors and provide an initial performance assessment for a select number of sensors and geophysical parameters.