[Defended thesis] Yawogan Jean Eudes Gbodjo: Leverage Multi-Source Remote Sensing data via machine learning to improve Crop Monitoring Systems

Jean Eudes  is one of the #DigitAg co-funded PhDs

Jean-Eudes defended his thesis on 8th November at 1.30 PM at Maison de la Télédétection, Bâtiment ADRET salle ASIE 1er étage. 

Leverage Multi-Source Remote Sensing data via machine learning to improve Crop Monitoring Systems

  • Start Date:  November 2018
  • University:  University of Montpellier
  • PhD School: I2S Information Structures Systèmes
  • Field(s): Computer Science
  • Doctoral Thesis Advisor: Dino Ienco, Inrae, UMR Tetis
  • Co-supervisors : Dino Ienco, Inrae, UMR Tetis & Louise Leroux, Cirad, UMR Aida
  • Funding: #DigitAg – Inrae
  • #DigitAg: Axis 5 – Challenges 8 & 6

Keywords: Crop production, Land cover mapping, Yield estimation and forecasting, Multi-source, Multi-temporal, Multi-scale remote sensing, SAR and optical images, Supervised learning, Deep learning, Recurrent neural networks, Convolutional neural networks.


Abstract: Crop monitoring systems play a key role in the assessment of crop production worldwide. Nowadays, a plethora of earth observation systems providing large scale multi-source information with high spatial and temporal resolutions, as well as the breakthrough induced by the deep learning have opened up new opportunities for crop monitoring systems in crop production assessment. This thesis investigates methodological trails in order to enhance crop production monitoring through the use of multi-source remote sensing and deep learning. We propose two methods for land cover mapping and cropland identification. The first approach is based on recurrent neural networks equipped with attention strategies which employ multi-source radar and optical time series as well as specific domain knowledge. The second approach is built on convolutional neural networks and further explores the combination of multi-source and multi-scale information especially, thanks to the integration of a very high spatial resolution optical source. We assess our proposals on territorial and local scales through a range of study sites with various landscapes, agro-climatic conditions and agricultural practices, which are located in smallholder agriculture systems and more conventional ones. We also investigate the estimation and forecasting of crop yields at the local scale of smallholder agriculture, using multi-source radar and optical time series. In this context, also characterized by a limited amount of ground reference data, we assess the potential of deep learning methods compared to commonly used modeling approaches. The evaluation of our proposals for the setting of land cover mapping and cropland identification shows that deep learning techniques seem well suited than common machine learning approaches to leverage the complementarity of multi-source, multi-temporal and multi-scale information, as there is sufficient amount of data for the training stage. On the other hand, the investigation carried out for crop yield estimation and forecasting did not show significant contributions from these methods. In this latter setting, the limited amount of ground reference data seems to be the main explanation.

Jury compound

Bertrand LE SAUX,    Rapporteur, Chargé de recherche, HDR  lab ESA

Clément MALLET, Rapporteur,  Cadre scientifique des EPIC, HDR IGN, LASTIG Université Gustave Eiffel Paris France

Thomas CORPETTI,  Examinateur, Directeur de recherche, LETG Université Rennes 2

Germain FORESTIER, Examinateur,  Professeur, IRIMAS, Université Haute-Alsace

Mme Carmen GERVET, Examinateur, Professeur, Espace-Dev, Université de Montpellier

Mme Laurence HUBERT-MOY,  Examinateur, Professeur, LETG Université Rennes

Dino IENCO, Directeur de these , Chargé de recherche, HDR, Université de Montpellier

Mme Louise LEROUX, Co-encadrante, Chargé de recherche, PhD , UPR AIDA CIRAD

International papers

Censi A.M., Ienco D., Gbodjo Y.J.E., Pensa R.G., Interdonato R., Gaetano R. (2021) Attentive Spatial Temporal Graph CNN for Land Cover Mapping from Multi Temporal Remote Sensing Data, IEEE Access

Ienco D., Eudes Gbodjo Y.J., Gaetano R., Interdonato R. (2020) Weakly supervised learning for land cover mapping of satellite image time series via attention-based CNN, IEEE Access

Gbodjo, Y.J.E.; Ienco, D.; Leroux, L.; Interdonato, R.; Gaetano, R.; Ndao, B., 2020, Object-Based Multi-Temporal and Multi-Source Land Cover Mapping Leveraging Hierarchical Class Relationships

Yawogan Jean Eudes Gbodjo; Dino Ienco; Louise Leroux, 2019, Toward Spatio–Spectral Analysis of Sentinel-2 Time Series Data for Land Cover Mapping

Eudes Gbodjo Y.J., Leroux L., Gaetano R., Ndao B., 2019, RNN-based multi-source land cover mapping: An application to West African landscape

Contact :   jean-eudes.gbodjo [AT] irstea.fr​ – Tél : 04 67 54 87 54

Social Networks:   GitHubLinkedInTwitter