PhD position (IA, vegetation, ecohydrology)
Fabrice Vinatier (PRO)
Contribution of artificial intelligence to the characterisation of plant biodiversity in agricultural landscapes in response to climate change and the agroecological transition
Supervision :
Director: Fabrice Vinatier - Co-director: Alexis Joly
Keywords :
diversity; deep learning; classification; botany; agro-ecological infrastructures; Pl@ntNet; CNN; ViT
Abstract:
New methods are needed to assess the success of biodiversification programmes and the impacts of climate change on plant biodiversity. The classical methods consisting in carrying out floristic surveys with the help of botanical experts often prove insufficient to establish a diagnosis that is both spatially and temporally dense. The aim here is to test the potential of artificial intelligence associated with imagery in order to propose monitoring protocols at unmatched spatial and temporal frequencies to meet the challenges of biodiversity conservation in landscapes. The thesis will be based on proven methods, i.e. using Pl@ntNet, associated with large datasets from labelized observatories, i.e. OMERE, which will be completed during the thesis. In addition to the development and validation of a new floristic monitoring protocol, the thesis will provide new elements of knowledge on the relationships between plant biodiversity and water resources.
Context :
Plant biodiversity is a major component of agroecosystem resilience, providing a significant number of ecosystem services for agricultural production (resource for crop auxiliaries, pollination) and for the preservation of soil (vegetation cover) and water resources (e.g. phytoremediation) (Garcia et al 2018). Hardly impacted by intensive agriculture and climate change (Lichtenberg et al 2017), it is gradually being reintroduced into agricultural landscapes via biodiversification and agroecology, based on the observation that more complex landscapes, i.e. richer in semi-natural elements, facilitate the establishment of associated biodiversity in the plots (Estrada-Carmona et al 2022).
This biodiversity, better characterized in space and time, attests to the success or not of agroecological programmes in terms of restoring ecological connectivity and provides information on the health of agroecosystems. Plant communities translated into functional properties provide information both on the response of plants to climate change and on the contribution of biodiversity to the restoration of ecosystem services (Rudi et al. 2020). It is therefore important to be able to characterize this diversity in all the elements of the landscape, and in particular the interstitial spaces of crops, i.e. hedges, ditches, edges, which are likely to be biodiverse because they are not productive.
However, the characterisation of biodiversity at a specific scale, i.e. the identification of plant species and their coverage per unit area, is a task reserved for a certain number of specialists, professional or amateur botanists, according to specific protocols established according to the type of environment and the type of diversity (alpha, beta, gamma), sometimes with a significant observation bias (Archaux et al 2006, Morrison 2016). These limitations reduce the sampling strength required to meet current resource preservation challenges at the landscape level (cf. agro-ecological infrastructures). This sampling must be both spatially dense in order to finely characterize the biodiversity of all the landscape elements impacted by the agroecological transition, and temporally dense to allow plant biodiversity to play the role of sentinel of the impacts of climate change.
The recent arrival of artificial intelligence applied to imagery, and the dissemination of applications dedicated to automatic species recognition, such as Pl@ntNet (Affouard et al., 2017), has turned the situation upside down, adjusting the imbalances by allowing anyone to recognise plant species from an image taken in the field. This method is particularly effective at recognising a species from one or more land cover images (Bonnet et al 2016), but its potential to recognise and position multiple specimens from a single image is yet to be explored further.
Although significant improvements have been made on multi-specimen images of cultivated plants (van der Velde, 2022), a number of methodological hurdles remain to link this type of application to biodiversity sampling protocols in agricultural environments. Indeed, the latter require exhaustive analyses of plant cover (e.g. quadrats or transects) for which it is necessary to recognise several species in a mixture per unit area, at various phenological stages. Carrying out such surveys by photographing and identifying individuals one by one with an application such as Pl@ntNet would be a very long and tedious task that would limit the acceptability and implementation of such approaches. One solution would be to develop new tools to automate this type of survey directly from images of a multispecies plant cover.
Objective and Method :
The objective of the thesis is to study the impact of agroecological transition and climate change in agricultural landscapes on the structuring of plant communities. In order to ensure a spatio-temporal sampling consistent with the studied phenomena, it will be necessary to define, implement and test an analysis method applicable to vegetation cover images. The method will be based essentially on deep learning algorithms such as convolutional neural networks or vision transformers. It will be tested in the field to cover the diversity of conditions inherent to sampling (weather, access to survey areas, diversity of species phenological stages).
The PhD student will collect and format the data sets already acquired, annotate and label the images already produced, and set up new acquisition protocols if necessary. Two types of data sets already acquired and to be completed will be mobilized: multi-site snapshot sampling according to environmental gradients and single-site temporal monitoring with a regular frequency.
Various classification methods will be studied: tiling-based, multi-label or instance segmentation, in order to evaluate the accuracy, systematic bias and associated uncertainties by taxonomic level, in comparison with the observation datasets.
Once validated, the method will be used to address specific issues of the host laboratory, LISAH, with respect to ecohydrology, an interdisciplinary science studying the dual relationships between vegetation and water (Vinatier et al 2017, Rudi et al 2018). The first step will be to map the response of plant communities to a variation in water resources, according to the hydrological niche hypothesis (Silvertown et al 2015). The sampled floristic diversity will be transformed into average effect traits on water transfers in landscapes, via databases internal to the laboratory or external (LEDA, TRY). Examples of functions to be tested are the infiltration capacity of roots, its potential evapotranspiration, and its impact on the water pathway, by linking the identified species to functional traits: root length, water use efficiency, and plant porosity. Finally, intra-specific (phenological stages) and inter-specific (species) successions will be studied over time to assess the impact of seasonality and climate change on the functional properties of the cover crops.
Within this framework, specific sampling will be carried out at unprecedented spatial and temporal frequencies to study the variation of beta diversity along hydrological gradients in particular, and the succession of plant species and their associated traits over time. They will be based on the observation systems of the LISAH laboratory (ORE OMERE) and the other observatories (Pech Rouge, Basse Vallée Durance) coordinated via the MOMAC project (Labex Agro 2021-2024 funding).
Expected results:
From a methodological point of view, several results are expected during this thesis, on the one hand annotated multi-specimen image datasets, FAIRised and uploaded to the Pl@ntNet platform, on the other hand deep learning models for the analysis of vegetation cover images. From a cognitive point of view, the results of the specific sampling will bring a gain of knowledge on the role of hydrological connectivity of landscapes on the functional structuring of plant communities, on the feedback impact of these communities on hydrological fluxes, and on the potential to use the dynamic monitoring of plant successions as markers of climate change.
References:
Affouard, A., Goëau, H., Bonnet, P., Lombardo, J.C. and Joly, A., 2017, April. Pl@ntnet app in the era of deep learning. In ICLR: International Conference on Learning Representations.
Archaux, Frédéric, Frédéric Gosselin, Laurent Bergès, et Richard Chevalier. 2006. « Effects of Sampling Time, Species Richness and Observer on the Exhaustiveness of Plant Censuses ». Journal of Vegetation Science 17 (3): 299‑306. https://doi.org/10.1111/j.1654-1103.2006.tb02449.x.
Bonnet, Pierre, Alexis Joly, Hervé Goëau, Julien Champ, Christel Vignau, Jean-François Molino, Daniel Barthélémy, et Nozha Boujemaa. 2016. « Plant Identification: Man vs. Machine ». Multimedia Tools and Applications 75 (3): 1647‑65. https://doi.org/10.1007/s11042-015-2607-4.
Estrada-Carmona, Natalia, Andrea C. Sánchez, Roseline Remans, et Sarah K. Jones. 2022. « Complex Agricultural Landscapes Host More Biodiversity than Simple Ones: A Global Meta-Analysis ». Proceedings of the National Academy of Sciences 119 (38): e2203385119. https://doi.org/10.1073/pnas.2203385119.
Garcia, Léo, Florian Celette, Christian Gary, Aude Ripoche, Hector Valdés-Gómez, et Aurélie Metay. 2018. « Management of service crops for the provision of ecosystem services in vineyards: A review ». Agriculture, Ecosystems and Environment 251 (January): 158‑70. https://doi.org/10.1016/j.agee.2017.09.030.
Lichtenberg, Elinor M., Christina M. Kennedy, Claire Kremen, Péter Batáry, Frank Berendse, Riccardo Bommarco, Nilsa A. Bosque‐Pérez, et al. 2017. « A Global Synthesis of the Effects of Diversified Farming Systems on Arthropod Diversity within Fields and across Agricultural Landscapes ». Global Change Biology 23 (11): 4946‑57. https://doi.org/10.1111/gcb.13714.
Morrison, Lloyd W. 2016. « Observer Error in Vegetation Surveys: A Review ». Journal of Plant Ecology 9 (4): 367‑79. https://doi.org/10.1093/jpe/rtv077.
Rudi, Gabrielle, Jean-Stéphane Bailly, Gilles Belaud, et Fabrice Vinatier. 2018. « Characterization of the long-distance dispersal of Johnsongrass ( Sorghum halepense ) in a vegetated irrigation channel ». River Research and Applications 34 (9): 1219‑28. https://doi.org/10.1002/rra.3356.
Rudi, Gabrielle, Jean Stéphane Bailly, Gilles Belaud, Cécile Dages, Philippe Lagacherie, et Fabrice Vinatier. 2020. « Multifunctionality of agricultural channel vegetation : A review based on community functional parameters and properties to support ecosystem function modeling ». Ecohydrology and Hydrobiology 20 (3): 397‑412. https://doi.org/10.1016/j.ecohyd.2020.03.004.
Silvertown, Jonathan, Yoseph Araya, et David Gowing. 2015. « Hydrological niches in terrestrial plant communities: A review ». Journal of Ecology 103 (1): 93‑108. https://doi.org/10.1111/1365-2745.12332.
van der Velde, M., Goeau, H., Bonnet, P., d'Andrimont, R., Yordanov, M., Affouard, A., Claverie, M., Czucz, B., Elvekjaer, N., Martinez-Sanchez, L. and Rotllan-Puig, X., 2022. Pl@ntNet Crops: merging citizen science observations and structured survey data to improve crop recognition for agri-food-environment applications. Environmental Research Letters. DOI.10.1088/1748-9326/acadf3
Vinatier, Fabrice, J-S. Bailly, et G. Belaud. 2017. « From 3D grassy vegetation point cloud to hydraulic resistance: Application to close-range estimation of Manning coefficients for intermittent open channels ». Ecohydrology 10 (8): e1885. https://doi.org/10.1002/eco.1885.
Specific material and financial conditions :
Bi-location of the PhD student half-time at LISAH (main affiliation) and half-time at LIRMM or AMAP (to be defined).
Thesis grant acquired via the PEPR Agroecology and Digital programme, Pl@ntAgroEco project.
Integration in LISAH observatories (OMERE Roujan) and in the MOMAC project observatories (Pech Rouge, Basse Vallée Durance)
Envisaged International collaborations:
Participation in GBIF, collaborations in the framework of the European projects GUARDEN (2022-2025) and MAMBO (2022-2026).
Envisaged national collaborations :
UMR IGEPP: deep learning, UMR PSH: agroecology, UMR AbSys: agroecology.
Profile required:
Master's degree in computer science/image processing. Necessary skills in programming (python) and deep learning (tensorflow, pytorch).
Interest in botany and fieldwork.
Contact persons:
Fabrice Vinatier fabrice....@inrae.fr and Alexis Joly alexi...@inria.fr