The objective of this work package is to develop a new metamodeling framework of peatland’s Critical Zone based on a smart combination of a distributed-physics-based model and a series of deep neural network fed by geophysical data and geochemical analyses. The developped methodologies will be applied to the two study sites to predict the spatial and temporal variability of carbon, nutrients, water and heat fluxes and stocks of the Critical Zone.
Physics-informed neural networks (PINNs)1 represent a class of versatile function approximators capable of incorporating the physics principles, e.g., partial differential equations (PDEs), governing a specific dataset within the learning process. In LandSense, we use PINNs as an alternative to conventional numerical inversion methods for soil hydraulic parameter estimation.
We use a soil respiration model2 coupling a physics-based 1D model describing soil water, heat and CO2 transport, and a pool-based model of soil carbon decomposition to predict water, heat and carbon fluxes at specific spatial locations across the study sites. An in-depth parametrization will be conducted to adapt the model to peat soils.
Predicting environmental data at both high spatial and temporal resolution remains a challenge, especially when using point observations in an irregular grid. Novel approaches using Deep Learning3 showed promising results as an alternative to geostatistics methodologies. In LandSense, UAV remote sensing data with high spatial resolution and physical model outputs (integrating depth) will be used as input to such models to predict fluxes across the landscape.
