Ιnnovation - Quantifying uncertainty of satellite precipitation estimates
The overarching objective of extrAIM is to advance towards a new type of satellite-based hydroclimatic datasets that, apart from single-valued estimates, explicitly embed uncertainty, account for modelling errors, pay special attention to extremes, as well as provide multi-valued equiprobable estimates.An aspect that it is expected to significantly improve end-users’ confidence and trust on satellite-based products.
Specifically, extrAIM will develop a general-purpose framework capable of modelling and quantifying the uncertainty of satellite-based products.This will be done by combining both probabilistic and AI methods (with a focus on explainability) for multiple datasets integration, as well as uncertainty modelling. This will eventually, allow the quantification and correction of the mismatch between integrated products and in-situ observations, resulting to an adjusted product capable of reproducing the extremes’ probabilistic behavior. This will pave the way for developing uncertainty-aware (UA) datasets for hydroclimatic variables, as well as contribute towards uncertainty-informed water resources applications, such as robust estimations of irrigation needs, and development of reliable flood/drought forecasting systems driven by UA hydroclimatic data from space.
Using this general-purpose framework extrAIM will create an uncertainty-aware precipitation dataset for the Mediterranean, as well as a user-friendly data analysis and visualization platform serving multiple purposes, such as the easy data retrieval (targeting researchers and engineers), and communication of risks arising from individual and compound extreme events (targeting a broader audience, such as non-scientists citizens, policy makers and organizations).
In this respect, extrAIM intends to answer the following research questions:- Can we create a single, and improved satellite-based precipitation product (SPP) by optimally, and in a way that the end-users can understand and trust, combining multiple SPPs through AI methods?
- Is it possible to model and quantify the uncertainty and errors (arising from the mismatch with ground-based observations) of the integrated SPP, and eventually adjust it, so as it reproduces correctly the observed extremes’ behavior?
- How can we move from single-valued precipitation estimates to multi-valued, yet equally-probable, satellite precipitation estimates incorporating uncertainty?
- How can we monetize the extrAIM’s developments to increase trust and confidence on satellite data and simultaneously improve the understanding and awareness on hydroclimatic risks and extremes?
- The development of an AI-enhanced, yet explainable and operational approach capable of optimally combining multiple SPPs into a single, and improved integrated SPP.
- The development of a general probabilistic framework for the uncertainty modelling and quantification of the quantitative precipitation estimates obtained by SPPs (with focus on extremes).
- The creation of a first-of-its-kind UA satellite-based precipitation dataset for the Mediterranean region.
- The development of a user-friendly data analysis and visualization platform, which will enable the easy data retrieval and visualization, aiming to increase understanding and awareness against hydroclimatic risks arising from individual and compound extreme events.
Scientific methodology and novel aspects
Explainable AI methods for the optimum integration/merging of multiple satellite-based precipitation products
Special attention will be given to the development of an AI approach capable of accounting for the probabilistic behavior of precipitation extremes (e.g., by embedding to the AI/ML model architecture appropriate transformation/activation functions, mimicking the probabilistic behavior of precipitation– thus ensuring a proper mapping of the data during the model training/calibration procedure).
Explainable AI methods for the optimum integration/merging of multiple satellite-based precipitation products
Uncertainty modelling and quantification of the quantitative precipitation estimates
This modelling framework will build upon the notion of copulas, which in turn allows the derivation of non-Gaussian conditional distributions, capable of modelling and quantifying the error between SPPs and gauge-based measurements (conditionally on the SPPs estimates), thus adjusting them appropriately and accounting for extremes.
Uncertainty-aware daily precipitation dataset for the Mediterranean region
extrAIM, using the above-mentioned approaches, will create of a first-of-its-kind uncertainty-aware satellite-based daily precipitation dataset for the Mediterranean region. Apart from uncertainty estimates, the extrAIM dataset will provide multi-valued, equiprobable, quantitative precipitation estimates (i.e., stochastic realizations), increasing this way the users confidence and trust.
Uncertainty-aware daily precipitation dataset for the Mediterranean region
extrAIM data analysis and visualization platform
The extrAIM platform will provide a user-friendly interface facilitating easy data retrieval, and communication of risks arising from individual and compound extreme events over the Mediterranean. The extrAIM platform aims this way to: a) enhance the understanding and awareness of scientists, engineers, citizens, and policy makers against such risks, as well as b) the easy use and application of the extrAIM data product in real-world engineering studies.