#DigitAg Master2 Internship Offers 2023

Every year, #DigitAg grants Master 2 Internships for French and foreign students

In 2023, new subjects from different disciplines are suggested in Human&Economic Sciences, Environment&Life sciences, as well as in Engineering and Mathematics.

  • To apply please send your curriculum and application letter to the relevant contact from the list below
  • Allowance: #DigitAg gives internships grants to the research units:  the amount of your allowance is to be asked to the contact person

Environnement and life sciences

Engineering sciences

Human and Economic sciences

Maths and its applications

Environmental and life sciences

Better understanding of water use in agro-ecological farms in Occitania: evaluation of existing low-tech sensors and application to monitoring irrigation practices

Keywords: low cost, low tech, sensors, soil moisture, water meter, irrigation

  • Additional discipline: Physics and sciences of the universe
  • Duration (nb of months): 6
  • Desired start date: March 1st, 2023
  • Research unit: G-eau, Institut Agro
  • Contact: Gilles Belaud – gilles.belaud@supagro.fr
  • #DigitAg: Axis 3, Axis 2, Challenge 1, Challenge 6

In Occitania, water is a critical resource for agriculture. Climate change reinforces this fact. By allowing to reduce the water stress for crops, irrigation is part of the possible means of adaptation and can constitute an important lever of agroecological transition. On farms already committed to agroecology, a better knowledge of water and irrigation management at the farm level is necessary. The TAI-OC project aims to characterize the irrigated agroecological systems of Occitania, to understand the factors of the agroecological transition and to accompany this transition. To understand these systems, particularly in terms of water use, we wish to rely on the developments, initiated within the framework of a European project (www.prima-hubis.org), of a set of technical solutions based on low-tech and low cost sensors. These solutions allow a better understanding of the irrigation management within the farms.

The trainee’s objective will be to evaluate the interest of sensors of this type for the monitoring and control of irrigation in agroecological farms. For this, several sub-objectives are defined:

1/ to take stock of the low-cost low-tech sensors available on the market, concerning the capacities of measurement of the volume of irrigation and pressure in irrigated systems (conduits, canals).

2/ to evaluate their usability (on economic and end-user criteria: cost, availability, ease of use but also technical: reliability in terms of measurements and uncertainties, number and type of failures, possibility of construction and repair).

3/ to analyze the interest and the operationality of their use in agroecological farms (positioning, number to be used, etc.) thanks to the modeling of irrigation networks.

Modelling collective strategies of crop-livestock cooperations in landscapes with mobile ovine herds

Keywords : multi-agent system ; viti-pastoralism ; crop-livestock ; landless shepherd ; agroecology

  • Additional discipline: Engineering sciences
  • Duration: 6 months
  • Desired start date: Mar 1, 2023
  • Research unit: Agir, Inrae
  • Contact: Grillot Myriam – myriam.grillot@inrae.fr
  • #DigitAg: Axis 6, Axis 5, Challenge 6, Challenge 4

Within the Sagiterres research project, innovative agroecological models have been identified in the Minervois area. They integrate agriculture and livestock in organic field crops, wine and livestock farming systems. Livestock plays a central role (maintenance of space, supply of organic matter, etc.). However, maintaining livestock raises questions about access to fodder and grazing resources, and about animal movements in the landscape. The hypothesis made is that the landscape has resources that could be opened up to herds and that it is possible to better coordinate the shepherds with various stakeholders in the landscape (wine growers, cereal growers, municipalities, forest managers, etc.) to adjust the level of stocking density with the available food resources. A spatialized multi-agent model is being developed specifically within the research project to represent the landscape in its current functioning. It integrates involved stakeholders’ logic, potential resources (plant cover, vineyard inter-rows, wasteland, etc.), herds’ feeding needs and logistical elements related to space management.

The intern will explore the model and use it to simulate scenarios that will allow the identified initiatives in the landscape to be strengthened, perpetuated and even multiplied. The scenarios could be based on participative workshops conducted during the use of the serious game Dynamix.

Using the main results of the simulations, the intern will be able to evaluate scenarios of the initiatives’ functioning and access to resources in terms of services provided in the landscape. The results will serve as a support for discussions on perspectives of a real deployment of such initiatives on the landscape and as a basis to involve territorial actors and public services.

Grouping together during hot weather? A study of the collective behaviour of aggregation in sheep

Keywords : Collective behaviour, aggregation, heat stress, animal welfare, embedded sensors

  • Additional discipline: Engineering sciences
  • Duration: 6 months
  • Desired start date: Feb 20, 2023
  • Research unit: CRCA – Centre de Recherches sur la Cognition Animale, Université Toulouse III – Paul Sabatier École doctorale
  • Contact: Bon Richard – richard.bon@univ-tlse3.fr
  • #DigitAg: Axis 3, Axis 2, Challenge 4, Challenge 1

In grazing livestock farming systems, sheep express a unique collective behaviour of aggregation linked with environmental heat. During the expression of this behaviour, which can last several hours in the event of prolonged heat, individuals no longer ingest. The collective dynamics associated with the expression of this behaviour, their links with environmental temperature and the individual or collective fitness associated with the expression of this behaviour are poorly determined.

However, this behaviour well known by shepherds can be a hindrance to the herd’s performance and as well as an issue for the management of animal welfare. The global evolution of climate and the multiplication of heat waves in the Mediterranean area accentuate the need to study this behaviour in order to understand it and propose ways of managing the grazing environment.

During this project, using digital tools, we will focus on the characterisation of (i) the collective dynamics of grouping (using embedded radio transmitters and ultra wideband sensors) and (ii) the environmental conditions within these animal groups (using sensors for temperature and gas content such as CO2 and O2), in relation to the local climatic conditions (temperature, humidity and solar radiation sensors). This monitoring will be carried out in different groups of ewes (5 to 15 individuals) at the Domaine du Merle (Institut Agro Montpellier, Salon-de-Provence). This monitoring will be completed by behavioural and health monitoring of the flock, notably by monitoring the level of oxidative stress with the assay of plasma metabolites.

This project represents a great opportunity to formalise research collaboration between the CRCA and UMR SELMET teams.

Engineering sciences

Characterisation of the potential of Mediterranean pastoral areas by remote sensing and multi-use of territories

Keywords: pastoralism, multi-use, remote sensing, diversity

  • Additional discipline: Life end environmental sciences
  • Duration (nb of months): 6
  • Desired start date: Feb 1st, 2022
  • Research unit: UMR Selmet, Inrae
  • Contact: Fabien Stark – fabien.stark [AT] inrae.fr
  • #DigitAg: Axis 3, Axis 6, Challenge 4, Challenge 1, Challenge 6

Pastoral areas represent 2.2 million hectares in France (RPG, 2018). Behind this classification, there is a wide variety of environments and plant cover (Nozières et al., 2021). These pastoral areas are used for different purposes: agricultural and forestry production, opening up of environments, preservation of biodiversity, and recreational activities. To ensure the multifunctionality offered by these natural areas, it is necessary to be able to characterise them in detail, in space and time.

Remote sensing appears to be a promising tool for addressing this issue. Indeed, satellite imagery data would make it possible to identify pastoral spatial units and their arrangement in diversified landscape mosaics. In addition, multi-spectral satellite images such as SPOT 6/7 and Sentinel could be used to characterise the biomass potential of the different types of vegetation cover present (Dusseux et al., 2014). Finally, the characterisation of pastoral areas and their potential would make it possible to promote a whole range of uses of these environments by stakeholders, depending on their characteristics.

The proposed internship aims to develop a method for characterising available pastoral resources (spatial, temporal and quantitative) using remote sensing. To do this, it will be based on a case study (Minervois) where the issue of reintroducing pastoral-type livestock farming is seen as an agroecological lever of interest, in terms of recycling nutrients and transferring fertility, opening up environments, maintaining wine-growing plots, and fighting fires. On this basis, it will be necessary to calibrate a generic model for estimating the potential offered by pastoral areas and their possible uses, which can be used in other territories and at other scales to support stakeholders.

How does an animal’s movement and/or its posture influence measurements in a data capture system based on 3D imaging?

Keywords: 3D imaging; posture, squeletal modelling; correction

  • Additional discipline: Engineering sciences
  • Duration (nb of months): 6 + 2 months (short-term period contract)
  • Desired start date: Feb 15th, 2023
  • Research unit: UMR PEGASE, Inrae
  • Contact: Le Cozler Yannick – yannick.lecozler [AT] agrocampus-ouest.fr
  • #DigitAg: Axis 2, Axis 3, Challenge 1, Challenge 2, Challenge 4

3D imaging allows precise measurements of morphology on animals and to follow their evolution over time, while minimizing the risks related to handling, both for people and for animals. The Morpho3D device had shown that it was also possible to estimate volume, surface, weight or chemical composition from 3D images on live animals. The repeatability and reproducibility of the measurements were good, but strong variations were sometimes observed, linked probably to the position of the animal. However, correct positioning of the animal would require human intervention or the implementation of constraining devices, which would be contrary to the objectives of high-throughput phenotyping. This reference posture would ultimately allow to limit, or even correct, errors during acquisitions. This work will be carried out from data collected within the experimental facilities of UMR PEGASE. It will aim to determine the reference posture of the animals and the approaches to consider when the image obtained is different from the desired one. One of our hypotheses is that it is possible, from morphological landmarks, to model an animal as a skeleton with a landmark tree. The joints seen as ball-and-socket joints and the length of the skeleton segments being constant for a given animal, we can thus define its position and its global orientation in the 3D world and determine the reference posture. For this, the morphological landmarks of interest will be specified, to define the static reference position, validated by expertise. We assume that it will then be possible, if necessary, to correct the recordings, from a reference database (reference and associated corrections), created from a limited number of animals, of different development and size, with very different postures.

Temporal transfer of land cover mapping models from multi-source/multi-scale remote sensing imagery by reusing ground truth data acquired in the past

Keywords : Remote Sensing, Agriculture, Machine Learning, Land cover mapping

  • Additional discipline: Environmental and life sciences
  • Duration: 6 months
  • Desired start date: Apr 3, 2023
  • Research unit: Tetis, Inrae
  • Contact: Ienco Dino – dino.ienco@inrae.fr
  • #DigitAg: Axis 5, Axis 4, Challenge 6, Challenge 8

Nowadays, more and more remote sensing data are available, offering the possibility to follow a geographical area over time. The time series thus generated represent an essential source of information to efficiently manage our agriculture on a territorial scale.

To this end, remote sensing data is used as input to machine learning (ML) methods to provide updated land cover maps (LCM). To do this, ML methods require a large amount of ground truth, which poses challenges for their use where little or no reference data is available. For example, when a LCM needs to be updated from year to year.

Re-using ground truth data acquired in the past to transfer a model to a successive period will avoid new costs and take advantage of previous investments. Unfortunately, directly transferring a model from one year to the successive one can be inefficient as the two periods are affected by different environmental and/or climatic conditions. This results in differences in the distribution of the acquired satellite data.

This internship proposal aims at developing an innovative deep learning/transfer learning method to produce a LCM for a year T using remote sensing data on year T as well as previous ground truth data (i.e. year T-1). We will tackle a multi-source framework where the input will be multi-temporal Sentinel-2 imagery and SPOT6/7 imagery with the question of how to fuse these heterogeneous data (in spatial resolution and spectral content). The satellite images will be obtained through the THEIA and PEPS platforms and the Equipex GEOSUD (DINAMIS). The method developed will be evaluated on a study site in West Africa, Burkina Faso, featured by contrasted agricultural landscape.

Automatic normalization of variables from agroecology databases

Keywords : Data science, text mining, databse, agroecology, semantics

  • Additional discipline: Environmental and life sciences
  • Duration: 6 months
  • Desired start date: Sep 4, 2022
  • Research unit: Aïda, Cirad
  • Contact: Sandrine AUZOUX – sandrine.auzoux@cirad.fr
  • #DigitAg: Axis 5, Axis 4, Axis 6, Challenge 1, Challenge 3, Challenge 8

Agro-ecological studies generate many heterogeneous databases in terms of structure and content. They are difficult to exploit and require curation to be used in statistical or modeling approaches. Curation consists in selecting the most relevant data and enriching them with the metadata necessary to understand them, in order to make them accessible, shareable and reusable. To annotate the data and increase the precision of the terms used, an interdisciplinary group of CIRAD researchers has built a dictionary of variables. A variable consists of semantic terms derived from expert knowledge and reference ontologies. A list of usual variables has been defined to facilitate data comparison and analysis, and links with crop models.

The objective of this internship is to automate the labelling of variables from agro-ecology databases from the list of usual variables.

The intern will use data from agro-ecological trials set up in Reunion by CIRAD and its partners (eRcane and CTIS).

Several methods will be mobilized and combined to propose the dictionary variables that are most in line with the database variables:

– lexical proximities measures,

– contextual proximity methods based on the description of the variables given by the experts,

– contextual proximity methods based on corpora. Contexts will be constituted from textual corpora and word embedding methods from deep learning methods.

Beyond an extension of the method by proposing original text mining methods, an important objective of this internship is to propose a generic approach to label data and facilitate the interoperability of databases in agroecology

Visualization of ecosystem services to support the design of agroforestry systems

Keywords : Human-Machine Interaction, Visualization, Agroforestry

  • Additional discipline: Environmental and life sciences, Human and social sciences
  • Duration: 6 months
  • Desired start date: Mar 1, 2023
  • Research unit: AbSyS, Inrae
  • Contact: Laëtitia LEMIERE – laetitia.lemiere@inrae.fr
  • #DigitAg: Axis 2, Axis 6, Challenge 1, Challenge 5

Agroforestry is a promising farming method as part of the agroecological transition. However, due to their great diversity, the design of these systems is complex and requires the combination of multiple areas of expertise. To facilitate the choice of species and their spatial arrangement during collective design workshops, mock-ups are often used to represent the plot to be designed. However, it is difficult to project oneself in time and space using a simple mock-up. Augmented reality visualizations could be used to visualize the ecosystem functions and services produced by the agroforestry plot. Several visualizations are possible (3D objects, layers with transparency, text…) and they each have their own advantages and disadvantages.

The goal of the internship is to determine the visualizations that are the most adapted to the technological context (augmented reality), application (agroforestry mock-up) and use case (co-design workshops). This problem of human-machine interaction will be addressed by implementing several examples of visualizations and evaluating them with the users targeted by the tool. A prototype allowing to visualize agroforestry systems in augmented reality has already been developed; at the moment, it allows to visualize the growth of trees, the instantaneous shade, geometrical structures such as the tree lines or the area occupied by the crops. The student will complete this prototype by adding a module for visualizing ecosystem services. The tool will then be evaluated with agroforestry advisors.

Estimating traits related to tree vigor, photosynthesis and phenology using image analysis for the screening of genetic resources in fruit trees

Keywords : RGB imagery, deep learning, trait estimation, stone fruit

  • Additional discipline: Life and environmental sciences
  • Duration: 6 months
  • Desired start date: Feb 1, 2023
  • Research unit: GAFL, Inrae
  • Contact: ROTH Morgane – morgane.roth@inrae.fr
  • #DigitAg: Axis 5, Axis 2, Axis 3, Challenge 2, Challenge 1

Fruit production is a key sector for French agriculture that contributes to a healthy and local diet while diversifying crops production and storing carbon in the soil. However, this sector is also fragile and strongly in need for modern varieties able to cope with different sources of biotic and abiotic stress. To this aim, it is crucial to phenotype fast and accurately large genetic resources to identify useful breeding candidates. This project ambitions to tackle this challenge with RGB (Red-Green-Blue) imaging. It aims at setting up image analysis protocols combining deep learning, mathematical morphology and stereovision to extract traits related to vigor, photosynthetic activity and phenology in peach and apricot tree. Images will be obtained in orchards with high genetic diversity (150 to 206 genotypes) thanks to two cameras carried by a light phenotyping device called PHENOMAN, along with routine measurement of target traits. These data will feed an existing database which the student will use to develop algorithms quantifying:

– increase of trunk volume and chlorophyll content, functions of vigor and health status,

– flowering stage and flower density, determining local adaptation and yield

The student’s goal will be to improve existing estimation methods for vigor and flowering traits, and to set up the very first protocols (SPAD measurements, RGB images) and image processing methods for estimating chlorophyll content. This project is expected to yield reliable tools to phenotype more efficiently the targeted traits in research teams or even in breeding companies.

Navigating the multidimensional implications of agroecology for animal and plant health decision-making

Keywords : Agroecology, plant use, Formal Concept Analysis, Multidimensional implication rule, Visual analytics

  • Additional discipline: Life and environmental sciences
  • Duration: 6 months
  • Desired start date: Feb 1, 2023
  • Research unit: Lirmm, Université de Montpellier
  • Contact: Marianne Huchard – huchard@lirmm.fr
  • #DigitAg: Axis 5, Axis 1, Challenge 1, Challenge 3

For a farmer, deciding on a practice requires considering those already in place to avoid disturbing the balance of the system through the introduction of new interactions. He therefore has to know the diversity of available practices. In the case of crop protection, for example, the literature presents various plant-based solutions (in aqueous or essential oil form) to control the infestation by a pest. For example, choosing a solution that repels a pest population may result in the population moving to a neighboring crop, which is usually not attacked.

The Knomana database [Silvie et al., 2021], with over 48,000 descriptions of pesticidal and antibiotic plant use, can enable this choice. The RCAviz [Muller et al., 2022] and RCAvizIR software platforms can be used to navigate through this knowledge base, whose knowledge has been previously classified using Relational Concept Analysis. In order to accurately represent the data of various dimension and to facilitate their interpretation by the farmer, a promising solution is to express them in the form of multidimensional implication rules, a new method derived from Formal Concept Analysis. For a ternary relation (3-D data) relating pests, plants, and protected crop, this method makes it possible, for example, to express knowledge in the form “when Pest1 is controlled by plant1 on crop1, then Pest1 is also controlled by plant2 on crop1, and by plant3 on crop2”. This method can be applied to relations of dimension greater than 3.

The objective of the internship is to develop a software prototype for visualizing knowledge, expressed in the form of multidimensional implication rules. These rules are produced by an algorithm implemented in Python. We will also develop a strategy so that the rules are presented to the user according to his interests and according to the semantics of the rules’ content.

Characterization of the vulnerability of coffee plantations using satellite images and neural networks (VULNCAF)

Keywords : Coffee, biodiversity, neural network, satellite images, diseases, pests, landscape indicators, landscape, texture

  • Additional discipline: Life and environmental sciences
  • Duration: 6 months
  • Desired start date: Mar 1, 2023
  • Research unit: Amap, Cirad
  • Contact: BORNE Frédéric – frederic.borne@cirad.fr
  • #DigitAg: Axis 5, Challenge 1, Challenge 3, Challenge 6

The objective of this study is to identify indicators helping the characterization of coffee landscape vulnerability to pest and disease (P&D), based on remote sensing images at very high spatial resolution (Pleiades, WV2) combined with exogenic data (soil, climate…). We thus propose to test the Swin Transformer Neural Network approach. The attended deliverables are (1) a map of the level of resistance to P&D spread, (2) a typology formalizing the P&D resistance in terms of landscape structure components.

The student’s main mission will be to build an image processing algorithm adapted to this problem. His.her work will rely on reference data priorly collected in the fields in Uganda during the DESIRA Robust project, thanks to its large pluri-disciplinary research team (ex : UMR PHIM). This context will allow to produce a large and accurate data base, geographically extended, and with a good representativity, such factors being a common limit for the network’s approaches.

A large number of plots (>50 per type) will thus be identified in the fields, being characteristic and representative of the coffee plantation structure and density diversity (shading type, diversity, organisation…), and showing various degrees of P&D vulnerability. These reference plots will be divided in three datasets: first to train and fine-tune a neural network priorly trained on standard remote sensing data, second to test and improve the quality of training results while avoiding overtraining effects, and third to map the whole study zone and evaluate the final result.

Various input combinations will be studied to determine the most performing networks depending on the scale of results: recognition and typology at the coffee orchard scale, and then at lower levels (intraplot structure), with adaptation of the input data resolutions. The most significative parameters will then be yielded for each working scale.

Human and Economic sciences

No subject available yet 

Maths and its applications

No subject available yet