Advanced methodologies for atmospheric remote sensing using artificial intelligence

Atmospheric monitoring by means of Earth Observation data is receiving increasing interest worldwide, from the scientific community to governments, international organizations, public authorities and companies. This is certainly due to the more growing attention to environmental issues related to Earth pollution, the effects of which are significantly associated with the health of the planet and of living beings. Moreover, the development of the space industry with ever more advanced technologies in terms of satellite platforms and sensor systems leads to a constant improvement in remote sensing techniques and capabilities. The Earth’s atmosphere hosts many phenomena related to the environmental impacts of climate change, such as severe storms and cyclones, dispersion of aerosols generated from volcanic eruptions and wildfires, transportation of noxious chemical agents produced by anthropic activities (i.e. plants, vehicle traffic, waste treatment), and so on. The study of atmospheric parameters (such as precipitation, volcanic ash, particulate matter, CO2, SO2, clouds, ozone, temperature, various aerosol types, etc) is crucial for a better understanding of the complex phenomena occurring in the atmosphere and different algorithms and techniques dealing with detection and retrieval of atmospheric species have been developed during the years. However, actions need to be still carried out in order to improve the accuracy of the estimates of atmospheric parameters. The Earth’s atmosphere is indeed a challenging domain to simulate, it’s not straightforward to obtain accurate quantifications, and the level of complexity of the algorithms for its modeling is very high. In the retrieval of atmospheric parameters from space a key role is played by the surface underneath the atmospheric layer, i.e. the surface characteristics influence the radiation measured by satellite sensors and thus, it affects the estimation of the atmospheric parameter under investigation. The radiance measured by a satellite sensor is indeed constituted by different parts which can be summarized with an atmospheric contribution and a surface contribution, and when there is the need to derive an atmospheric parameter from space sensors, such as aerosol content in the atmosphere, the problem of separating the aerosol signal from the surface must be addressed. The simplest case in remote sensing is when the background surface is black, i.e. the quantity measured by the sensor is the top of atmosphere (TOA) radiance which is related only to the photons scattered by the atmospheric layer, while when the surface is not black the sensor measures not only the TOA radiance but also the radiance coming from the photons reflected by the surface. The decoupling of these two contributions is not straightforward, especially in the case where the background signal comes from land surface (remote sensing over oceans is less difficult given the highest degree of homogeneity of the water surface as compared to land surfaces). This highlights the importance of surface characterization in atmospheric remote sensing 3 4 Introduction applications. This work deals with surface reflectance characterization by means of the Bidirectional Reflectance Distribution Function (BRDF) and atmospheric parameter retrieval with a focus on natural hazard monitoring, in particular volcanic eruptions and tropical cyclones. With the rapid and increasing growth of AI-based algorithms, the satellite Earth Observation sector has benefited from their ability to solve complex non-linear problems, such as estimating parameters with high level of interdependence with other variables. The rapid and automatic processing of these models has made them highly suitable for situations where a prompt response is mandatory, such as the case of emergency management services. In this context, neural networks (NNs) have shown remarkable efficiency in extracting atmospheric parameters. In light of the above, this work makes use of neural network models for volcanic ash detection and tropical cyclone precipitation retrieval. In particular, given the very recent data used for the latter, i.e., satellite observations from the NASA (National Aeronautics and Space Administration) TROPICS (Time-Resolved Observations of Precipitation structure and storm Intensity with a Constellation of Smallsats) constellation, precipitation retrievals with neural networks has not yet been explored for this new space mission. Moreover, the task of deriving multi-angular surface reflectance patterns for BRDF estimation is also addressed by means of a neural network retrieval scheme that utilizes UAS-based measurements, in addition to standard methods, which is a topic not fully covered in literature.