DeepRun
DeepRun – Deep Learning for Reunion Energy Autonomy
The intelligent specialisation strategy for research and innovation S3 2021-2027 of the Réunion Region seeks to achieve the objective, among many others, of electrical autonomy by 2023 (Siby, 2020). And the regional programme “Green Revolution, La Réunion île solaire et terre d’innovation” proposes to further encourage renewable sources (solar, wind, thermal, biomass) and to store unused energy by sustainable means. Priority areas 1 and 4 of the ERDF programme are to invest in the levers of growth and to move towards energy transition and electrical autonomy. Two levers are available: the first is the increase of renewable energies (biomass, solar, etc.) and the second is the control of intermittent energies (solar and wind) through hydrogen storage.
By 2023, coal-fired power stations will have to be converted to biomass power stations (ALBIOMA, 2020). However, to date there is no detailed mapping of the agricultural area on Reunion Island, making it difficult to characterise the agricultural areas on the scale of the island. Remote sensing using satellite images offers a means of producing large-scale knowledge of soils at a reasonable cost. This image capture not only makes it possible to reach areas that are difficult to access, but also opens up the possibility of periodic revisits for project monitoring. Vegetation is easily distinguished from concrete (Gaetano et al., 2018). However, differentiating sugarcane from maize, or wooded Creole orchards with satellite resolutions is a much more difficult task. Here, recent Deep Learning approaches promise to efficiently analyse the phenomenal amounts of data (Watanabe et al., 2020) and improve land use classifications while being easily exploitable by a user (Ayhan et al., 2020). These tools will thus be implemented in the DeepRun project to obtain fine maps of land use.
Among all the means of sustainable storage, hydrogen represents an energy vector of the future, very promising for the territory. In this sense, France has just launched the hydrogen recovery programme for 2030. Despite this, many technological and societal obstacles remain to be resolved. To reduce production costs, Japan, for example, has just inaugurated the largest electrolyser plant in the world. One of the many challenges is the lack of reliability through the occurrence of malfunctions (Dijoux et al., 2017). Conventional diagnostic tools only integrate external variables which are difficult to interpret with little sensitivity. The DeepRun project aims to improve the understanding and detection of these faults by integrating internal observations with classical tools. And, recent Deep Learning methods show excellent results in this kind of application (Haas et al., 2020).
The postdoctoral project DeepRun investigates in a transversal way artificial intelligence tools for image recognition in order to support a transition towards energy independence.
Postdoctoral project duration: 18 months (October 2021 – March 2022)
Main objective
The main objective is the development of cross-sectional multi-scale image recognition tools, using Deep Learning algorithms, applied to the estimation of biomass resources for bioenergy production, and to the reliability of hydrogen converters for energy storage optimisation.
Expected outcomes
- Development and research of a deep learning tool that meets the operating constraints of the Reunionese territory.
- Multi-scale application of the developed method, fine detection of crop types on a plot scale, and recognition of bubble/drop regimes.
- Availability of research results to the scientific community and society.
The DeepRun project consists of three actions:
Action 1: Design of a multi-scale image recognition tool
- Bibliography on multi-scale Deep Learning image recognition tools
- Development of a scientific tool specific to the conditions of the Reunionese territory
- Operation of hydrogen converters in a tropical environment (humidity and temperature)
- Diversity of Reunionese cultures and steep relief
- Creation of databases with satellite images (Pléiades) and new truth maps from the DEAL REUNION / IGN
- Creation of databases for hydrogen converters
Action 2: Application to the recognition of land use on a landscape scale
- Detection of crop types and generation of fine maps of land use (1 per year)
- Cross-referencing of maps (risks, deposits, uses) for photovoltaic decision making
Action 3: Application on a microbubble scale at the electrochemical cell scale
- Creation of databases and detection of bubbles/droplets, then determination of operating regimes
- Coupling of the obtained spatial distributions to models and design of a multimodal diagnostic tool
To complete the project, the postdoctoral researcher will be provided with computing machines at different scales:
- Local computing stations at ENERGY-lab and CIRAD
- GPU: 2 x Nvidia Quadro RTX 4000 (8 Go) + 1 x Nvidia RTX 5500 (24 Go)
- RAM: 128 Go
- Storage: 4 To
- Mésocentre HPC et données Meso@LR
- GPU: 2 modified Nvidia RTX 6000 (48 GB) visualisation nodes
- RAM post: 3 To
- Storage: 15 Po
- Supercalculateur HPE CNRS IDRIS Jean-Zay
- GPU: 7 accelerated partitions, Nvidia V100 (16-32 Go), Nvidia A100 (40-80 Go), max 1024 GPUs/job
- RAM post: 3 To
- Storage: 30 Po
The hydrogen test bench of the SYSPACREVERS project is also available for experimentations:
Scientific partners
- CIRAD
- Pierre Todoroff (pierre.todoroff@cirad.fr)
- Lionel Le Mézo (lionel.le_mezo@cirad.fr)
- Mickaël Mezino (mickael.mezino@cirad.fr)
- Bertrand pitollat (bertrand.pitollat@cirad.fr)
- CNRS IDRIS (advanced support)
- Maxime Song (maxime.song@idris.fr)
- Pierre Cornette (pierre.cornette@idris.fr)
Financial partners
The DeepRun postdoctoral project is co-financed by the EU, the Réunion Region and the University of La Réunion, with the support of the DRARI.
Scientific articles
- Christophe Lin-Kwong-Chon, Cedric Damour, Michel Benne, Jean-Jacques Amangoua Kadjo, et Brigitte Grondin-Pérez, « Adaptive neural control of PEMFC system based on data-driven and reinforcement learning approaches », Control Engineering Practice, vol. 120, p. 105022, mars 2022, doi: 10.1016/j.conengprac.2021.105022.
- Christophe Lin-Kwong-Chon, Pierre Todoroff, Lionel Le Mezo, Michel Benne et Jean-Jacques Amangoua Kadjo, « Multi-level deep-based classification of land use and land cover: A case study on Réunion Island », Proceedings of the International Society of Sugar Cane Technologists, volume 31, xx–xx, 2023 (à paraitre)
Congregational acts
- Christophe Lin-Kwong-Chon, Kenza Benlamlih, Pierre Todoroff, Jean-Jacques Amangoua Kadjo, « Deep learning et imagerie satellitaire pour cartographier l’occupation du sol : performances et perspectives », RGR2021, 17 novembre 2021, sciencesconf.org:rgr2021:374451.
- Idriss Sinapan, Christophe Lin-Kwong-Chon, Cédric Damour, Michel Benne, Jean-Jacques Amangoua Kadjo, Optimisation des interfaces fluidiques dans un système de production d’hydrogène par électrolyse à partir des outils de reconnaissance d’images Deep Learning, CNRS Aussois FRH2, 2 juin 2022
- Christophe Lin-Kwong-Chon, Idriss Sinapan, Dominique Grondin, Michel Benne, Segmentation d’images pour la classification de données massives, TEMERGIE ValoREn, 1er decembre 2022
- Christophe Lin-Kwong-Chon, Pierre Todoroff, Lionel Le Mezo, Michel Benne et Jean-Jacques Amangoua Kadjo, « Multi-level deep-based classification of land use and land cover: A case study on Réunion Island », ISSCT2023, Hyderabad, 2023
Posters
- Pierre Todoroff, Christophe Lin-Kwong-Chon, Kenza Benlamlih, Deep Learning et imagerie satellitaire pour cartographier l’occupation du sol : performances et perspectives, CST SIAAM, 15 novembre 2021
- Pierre Todoroff, Christophe Lin-Kwong-Chon et Kenza Benlamlih, Modèle d’apprentissage profond pour cartographier l’occupation du sol en zone tropicale à partir d’images THRS, SEAS-OI, 13 juin 2022
- Christophe Lin-Kwong-Chon, Pierre Todoroff, Mickaël Mezino, Lionel Le Mezo, Cartographie de l’occupation du sol par imagerie satellitaire et apprentissage profond : Premiers résultats, CST CapTerre, 24 novembre 2022
Land Usage and Land Cover mapping
Tools and interfaces
- DeepRun-GUI : A GUI application for deep learning model generation, model training and images inference. Source code.
Outreach activity
- Fête de la Sciences, 17 novembre 2022, 30ème édition, Université de La Réunion
Laboratory Director and Scientific Director
Pr. Michel Benne (michel.benne@univ-reunion.fr )
DeepRun project leader
MCF HDR Jean-Jacques Amangoua Kadjo (amangoua.kadjo@univ-reunion.fr)
Scientific partner
Dr HDR Pierre Todoroff (pierre.todoroff@cirad.fr)
Postdoctoral student
Dr Christophe Lin-Kwong-Chon (christophe.lin-kwong-chon@univ-reunion.fr)
Contact :
Laboratoire ENERGY-lab
Jean-Jacques Amangoua Kadjo
amangoua.kadjo@univ-reunion.fr
Tel : +262(0)262 938216
15, Avenue
René Cassin
CS 92003
97744 Saint-Denis Cedex 9
La Réunion