Hyperspectral foundation model trained by spectral reconstruction for greenhouse gas emission estimation

Abstract

Remote sensed satellite data, in particular, the data that has been sensed very finely in the electromagentic spectrum lends itself to learn the signatures of many objects on the surface and in the atmosphere of Earth. Very large amounts of such data are fed in to artificial intelligence models that are carry out learning general representations from the data in a self supervised manner. On top of this learning of general representations, the model can be made to see labeled or ground truth data to associate data patterns with specific signatures of objects. In this work, the objects are the greenhouse gas such as methane and carbon dioxide concentrations in the atmospheres and models learn the signatures of gas concentrations. We will use such a model to estimate the greenhouse gas concentrations in the atmosphere.