Machine Learning (ML) techniques for the exploitation of hyperspectral satellite data by the enhancement of High-Priority Products (HPPs) detection capacity, identification of potential new ones and improvement in expanding applications in different hyperspectral domains.

The Copernicus Hyperspectral Imaging Mission for the Environment (CHIME) by ESA, together with other hyperspectral upcoming mission such as NASA’s Surface Biology and Geology (SBG) and IRIDE- PLATiNO, promises high spectral resolution across the 400-250 nm range. These missions target diverse applications in environmental monitoring and resource management, addressing the needs of both Copernicus and SBG communities. The applicability of CHIME data for detecting raw materials and plant functional traits, along with high spectral resolution hyperspectral satellite data such as PRISMA and EnMAP, depends on the ability to refine and improve the currently used algorithms and consequently the resulting products to support natural and anthropic risk management, agriculture, water resource monitoring, forestry management, and human activity control. The availability of hundreds of contiguous bands with high spectral resolution provides unique information through hyperspectral sensing. The Mission Requirement Document for CHIME identifies the HPPs that are variables or indicators that are physically meaningful, spatial and temporally consistent and suitable to supporting decision-making processes. In particular, the mission outcomes will support food security, in the domains of soil and vegetation variables assessment, and mineral resources mapping. However, existing algorithms do not fully meet user demands across key domains. Thus, the development of innovative processing chains for the hyperspectral exploitation is needed. This research aims to investigate how Machine Learning can bridge gaps in hyperspectral satellite data, enhancing detection and mapping capabilities in two primary application domains: raw materials and vegetation. In particular, this is focus on selected HPPs: mineral and methane mapping for raw materials and canopy water and chlorophyll content in vegetation application domain. This work uses hyperspectral satellite data, such as PRISMA and EnMAP, to explore hybrid algorithms (Machine Learning+Physics-based) capable of delivering enhanced hyperspectral products for future multi-mission Earth Observation services. This work is organized in four parts: Firstly, an analysis of state of art was carried out to assess the maturity of algorithms for the exploitation of hyperspectral data in the selected application domains and assess the existing gaps in these applications. Then three study cases were developed to test the application of hybrid models for the retrieval of the selected HPPs: (1) Rare Earth Element detection by hybrid models and construction of prototypal user-friendly tool; (2) Methane plume mapping improvement trough hybrid models considering climate variables; (3) Vegetation stress index assessment by the extraction of spectral features from multi-scale hyperspectral data. The findings indicate that tailoring machine learning through physically informed hybridization is emerging as a robust approach to strengthen High-Priority Products and support the development of new hyperspectral products under real mission constraints. The results of this work can support the achievement of major policy priorities, including the European Green Deal, the EU Raw Materials Act and the Global Methane Pledge.