The “ADELE4RAIN” Project
Revelis has developed ADELE4RAIN (A DEep LEarning-based framework for RAINfall estimation and forecasting), a comprehensive platform for rainfall monitoring and prediction that integrates various data analysis methods and machine learning techniques to be used to accurately assess, predict, and mitigate the impacts of rainfall-related hazards on both urban infrastructure and agricultural landscapes, enabling the extraction and integration of information from a wide range of data to provide more precise and timely forecasts of extreme weather events, including heavy rainfall, storms, and flash floods.
The overall objective of the ADELE4RAIN project is to provide communities, governments, and stakeholders with the necessary tools and knowledge to proactively manage and mitigate the impacts of heavy rainfall events, thereby safeguarding critical infrastructure, protecting livelihoods, and promoting sustainable development.
Project Type: Public call for proposals aimed at granting funding for activities consistent with the program under the resources of the National Recovery and Resilience Plan (PNRR) Mission 4, “Education and Research” – Component 2, “From Research to Business” – Investment Line 1 1.4, financed by the European Union – NextGenerationEU”, project “ICSC” “National Centre for HPC, Big Data and Quantum Computing (HPC)” Project Code CN00000013, CUP C83C22000560007.
Project Objectives
The main objectives of the ADELE4RAIN project are as follows:
– Management of extreme weather events;
– Protection of natural resources;
– Improvement of community resilience;
– Prevention of environmental disasters.
Distinctive Features
The distinctive features of the ADELE4RAIN project include:
– The use of a Machine Learning module that will include a suite of models with different complexities and capabilities;
– The monitoring and forecasting of rainfall;
– The use of deep learning techniques, such as multi-layer neural networks;
– The use of an evaluation module that will allow for continuous monitoring and validation of predictive models.
The Application Case
The ADELE4RAIN platform is based on Mixture of Experts (MoE) models and Snapshot Ensembles and Soup Models.
A case study in Calabria, with its climatic variability and vulnerability to floods, offers the opportunity to test and optimize this innovative system, thus contributing to a more effective management of hydrogeological risk in the region.
The state-of-the-art techniques used in the project fall within the areas of: Machine Learning and Deep Learning.
It’s interesting that this case study is focused on Calabria! Given that I’m currently in Reggio Calabria, it makes this application particularly relevant to the region. The challenges posed by the local climate and the risk of flooding are significant, so a tool like ADELE4RAIN could be very beneficial.
Mixture of Experts (MoE) Models
These deep ensemble models combine the predictions of multiple specialized models, or “experts,” into a single framework.
Each expert is trained on a specific subset of the data or specializes in modeling a particular aspect of the problem.
A gating network is then used to dynamically weight the contributions of each expert based on the input data.
Ensemble Snapshot e i Soup Model
Lightweight deep ensemble methods, such as Snapshot Ensembles and Soup Models, are designed to overcome the computational challenges associated with traditional deep ensemble methods.
Snapshot Ensembles involve training multiple neural networks with different initializations or training procedures and then averaging their predictions to obtain a final ensemble prediction. This approach can be computationally efficient, as it does not require training multiple models from scratch.
Innovation Plan
The activities for the creation of the ADELE4RAIN platform are organized into 3 macro-phases.
Critical Analysis of the State of the Art
In this activity, a comprehensive analysis of the existing state of the art in rainfall monitoring and forecasting is conducted, identifying the main challenges and issues faced by current solutions.
Evaluation of Current Solutions
This activity includes:
– An assessment of the capabilities and limitations of current solutions for rainfall monitoring and forecasting;
– The design and definition of a Deep Learning-based framework for rainfall estimation and forecasting;
– The design of a data extraction and integration module;
– The design of methods for rainfall estimation.
Experimental Development
This activity includes:
– Framework development;
– Definition of the case study and framework evaluation.
