13/04/2024
What is the problem?
The Earth is experiencing tangible climate changes. This is proven by the dynamic of the essential climate variables (ECV) that are available from the analysis of data produced by current and next-generation missions of Earth observation. Project goal RETINA is a multidisciplinary and curiosity-driven project aiming to develop new methods for the analysis of data produced by the interaction between electromagnetic waves and the Earth surface, transferring theoretical findings to practical problems. RETINA focuses on characterizing surface soil moisture (SM) and freeze/thaw (FT) state, which are variables connected to ECVs. Current solutions and their limitations Machine learning techniques and data-driven approaches are widely used to implement retrieval. However, suitable dataset for training are needed, requiring considerable effort for annotation. In addition, often data and ancillary data are not continuously available, due to the features of the acquisition methods (e.g., the satellite orbit, the presence of disturbances like clouds). Proposed solution We propose, for the first time, the use of multivariate neural network (NN) operators in conjunction with their (approximate) inversion for the modeling and estimation of SM and FT from data delivered by space missions. The NN operators retrieval is complemented by Bayesian inversion performed using Monte Carlo methods.
To this aim, both an analytical and a probabilistic strategy will be considered:
1. Data modeling by means of well-known multivariate NN operators. Leveraging on functional analysis, the operators will be theoretically inverted. An analytical expression of the approximated model targeting the involved geophysical variables is obtained. Then, since some variables can be disturbed, the NN operators model will be extended to represent them with interval-valued fuzzy sets (IVFS).
2. Bayesian inversion via Markov Chain Monte Carlo (MCMC). Sampling from the posterior distribution via MCMC is a well-established technique. RETINA proposes to exploit a class of Markov Chains called Probabilistic Cellular Automata (PCA) characterized by a parallel updating rule. This is expected to be particularly advantageous when the physical quantity to retrieve is multi-component, e.g., when data are collected in matrix format.
