Modeling the CO2 Sequestration Potential of Agricultural Soils with a Hybrid System

The increase of carbon dioxide concentration in the atmosphere and its consequences on climate and the environment are widely discussed topics, and in the last few years the creation of future projections in different emission scenarios has highlighted the need to take serious action against further increase in greenhouse gas content. After the Paris Agreement of 2015, the signing parties committed to specific goals to reach before 2100, to keep the temperature increase well below 2°C with respect to pre-industrial conditions, not only reducing the current emission rate but creating strategies to reach Net Zero Emissions. This is why carbon sequestration strategies have been developed: in this project, the work is focused on sequestration in cultivated soils, which could represent a way to reduce the impacts of pastures and agricultural activities. The main scope of this thesis is to create a model to describe the quantity of Soil Organic Carbon (SOC) that is contained in agricultural soils, and to simulate the future trends of this soil feature, which is seen as a possible strategy for Carbon Dioxide sequestration for climate change mitigation purposes. The model is constructed in a hybrid configuration, therefore combining a process-based relationship (theory-driven model) and the introduction of Remote Sensing data through Machine Learning techniques, more specifically using Artificial Neural Networks (data-driven model). The choice of a hybrid model was made to preserve the physical consistency and the interpretability of the process-based model, while integrating a large quantity of data in complex patterns to be able to create generalized models which can ideally be upscaled through the Machine Learning algorithm. The data used to train the model are collected in 9 different farms, 8 of which are in Portugal and the last is in Spain, that are all based on the cultivation of Sown Biodiverse Permanent Pastures Rich in Legumes. The process-based model that was used as base for the study is a simple equation describing the relationship between SOC content at a given instant and the value after a defined time interval, depending only on two parameters (carbon inputs and mineralization rate). Also, Neural Networks were used to estimate the best values of such parameters. The research has been focused on finding the model hyperparameters which allowed to obtain the best possible fit between measurements and modelled values. The results obtained highlighted that the simplicity of the model is not ideal to describe such complex relationships as well as the ones that occur between soil variables. On the other hand acceptable fits were obtained, so the model can be used to make predictions of SOC content and, therefore, carbon sequestration potential, by inserting projections of soil temperature and moisture.