Every year, #DigitAg grants Master 2 Internships for French and foreign students
In 2020 subjects from different disciplines are proposed in Human and Social Sciences & Agronomy , as well as in Engineering, Mathematics & Agronomy
- To apply please send your curriculum and application letter to the relevant contact persons listed below
- Allowance: #DigitAg gives internships grants to the research units: the amount of your allowance is to be ask to contact persons
Human and Social Sciences & Agronomy
|Social Sciences – Interdisciplinarity, interviews
Uses of digital artefacts for agroecological transition in Benin
Context – Agroecological transition requires creating, acquiring, exchanging information and knowledge. Digital tools take an even greater role to support this transition through the emergence of communities of practice. The digital transition is, in Benin, spurred by massive investment (for instance $100 million by the world bank in rural areas). Digital artefacts are now seen as a necessary condition for rural areas to adapt. Both transitions generate inequalities and deep structural and cognitive changes: tacit knowledge, access, financial means, available time, capacity in technology use. Users are various – farmers, technicians, extension services. Technologies and uses are multiple – information exchange (eg: WhatsApp groups, forums), building knowledge commons (eg: wiki), collective norm making (eg: specifications), decision aid tools (eg: remote sensing tools, online training videos).
Research question and objectives – What changes could emerge from such infrastructure building paving the way to larger bandwith in rural areas in Benin? What are the current uses of digital artefacts? What do they allow, how frustating are they? How practices would evolve? What are the pros and cons of digital artifacts compared to more traditional ways of information/knowledge exchange? This internship aims at exploring such questions with both a critical and open to surprises approach set in empirical grounds. We expect a typology of uses and users, a description of pros and cons of digital compared to face-to-face as a result. This preliminary investigation could lay the foundations for a thesis on the subject on the uses of digital artefacts to favor the agroecological transition.
Methods – State of the art on intellectual commons, social-ecological systems and existing collaborative tools. Field study investigation in Benin with actors who use or are willing to use digital tools. The specific focus will evolve depending on the intern interests and field observations. The internship will start by investigating small and identified communities inclined towards agroecological practices which use digital tools: a participatory guarantee system (https://amapbenin.com/2017/08/17/certification/), the Organic and Ecologic Agriculture Platform (PABE – https://www.pabebenin.org/?p=169) or other profesional organisations such as niébé based products.
|Innovation economics – Innovation studies / Agronomy of Technical change – Skills for strategic approach to farms and cooperatives, interviews, teamwork & writing
Adoption of digital technologies in the wine cooperatives: what impact on the agroecological transition?
The use of digital technologies on farms is becoming more and more frequent, but the exact nature of these uses remains little known, as are their real impacts on the greening of farming practices. This question is particularly important in the wine sector, which is at the same time integrating many technological innovations, and facing increasing environmental challenges. The strong growth of organic viticulture illustrates this trend, which also depends on the cooperative cellars strategies that can have a major influence on the use of digital and the change of practices on the farms members. The Master’s internship refers to the field of Innovation Studies. It aims i) to characterize the digital uses of grape growers that are members of cooperative cellar in Occitanie, ii) to highlight the links between these uses and commitment to organic viticulture and iii) to specify the conditions that influences the adoption of these technologies and in particular the role of the cooperative. The work will therefore focus on the entire process of adoption and use of digital in a sector, including its conditions and impacts. Several hypotheses can be tested, such as the role of the cooperative’s strategy or the viticulturist’s personal trajectory. The internship will consist in the realization of about thirty surveys with grape growers in the Occitanie region, starting from a selection of a reasoned sample of organic, non-organic and mixed cooperatives, then by a selection of 3 or 4 members per cooperative. The work will help i) Identify the digital technologies used and build a typology of their usage patterns; ii) specify their possible links with the agronomic practices implemented on the farm, in particular those involved in the agroecological transition; (iii) Identify the factors affecting the adoption and use of these technologies; iv) Identify the links between farms and cooperative in the use of digital. These results will not be sufficient for a thorough quantitative analysis, but it will bring very rich qualitative information around a study cases analysis. The originality of this work is to propose a characterization of digital usage profiles and to examine their relation with the organic or non-organic nature of farms, in a context of cooperatives. It may lead to the production of a scientific article. The internship participates in a research program built around Eléonore Schnebelin’s thesis on the links between digital and organic farming in several agricultural sectors, including viticulture. It will be carried out in conjunction with UMR Innovation and UMR AGIR, and in partnership with Coop de France Occitanie.
|Economic and social geography / Sociology of work – Knowledge on farms typology and an interest in multidisciplinary study and digital innovations.
Digital companies: for which farmers and which agricultures? The case of the dairy cattle sector
Digital farming (DF), including information and communication technologies (ICT), agricultural robot and big data leads to many hopes for transforming agricultural models. Some speak of revolution and see it as a source of innovation sometimes for a more sustainable agriculture sometimes for more competitive farms. The scientific literature is mainly focused on the technical aspects of these digital tools and their economic impacts. A wide field of research has presented itself to understand the contribution of this DF to the socio-economic transformations of farms. It is against this background that a PhD work (#DigitAg) is being taken since early September 2019 at UMR Innovation. This study aims to understand the change of farmers’ work in farms adopting digital technologies. As digital concerns society at large, beyond the farm’s level, we consider the agricultural innovation in a sociotechnical context. This study is also interested in an innovation system dedicated to the digitization of agriculture (training courses, exhibitions, research units, agricultural councils, experimental farms, etc.) and in markets.
Understanding this context involves studying the strategies of the institutions with which the farmer interacts (professional organizations, actors of agricultural council, suppliers of inputs and machines, state services, etc.). This internship will aim specially to characterize the representation of agricultural modela and farm systems held by digital companies . A focus will be made on strategies of conceptions and marketing that they implement. We formulate two hypotheses:
The internship will focus on dairy cattle farming systems, as this sector has been using digital technologies for a relatively long time. The study will concern about twenty companies selected according to qualitative sampling (size, years of existence, type of products and services such as robots for milking, the cowshed cleaning or feed supply and/or connected devices). Semi-structured interviews will be conducted to gather a free speech while addressing specific topics and previously formulated questions. These themes will be structured along three main lines:
|Economics / Stastistics – Knowledge of economics, statistics and econometric modelling, fluent English
Productive and commercial strategies: Digital technology for organic certification and short food supply chain sales
In a context where the environment is a growing concern, both public and private actors multiply initiatives to guarantee the environmental quality of production sold. As a result, certifications based in particular on the compliance of production with the Maximum Residue Limits and the use of authorized molecules are implemented. Among public, private, national and international certifications that producers must comply with, we focus on the organic certification. In 2018, respectively more than 90% and 75% of consumers reported having already consumed organic certified products or consuming them regularly. In parallel, certified production is developing. Thus, 2 million hectares are organic certified, or 7.5% of the agricultural area. At the same time, and in order to meet consumer expectations of proximity, producers are more likely encline to sell their production through short food supply chain channels (SFSC). By this way, the producer can capture a larger share of the added value, increase his income and thus develop his activity. Today, 1 in 5 producers sells through SFSC, which represents less than 10% of food consumption. On the basis of econometric models based on exhaustive data, our previous work identified brakes and levers – whatever individual, structural, financial and organizational – to SFSC and demonstrated an interdependence between the productive strategy and the commercial one. This interdependence will be reinforced by the framework of the Food and Agriculture Act adopted in 2018, which sets a target of 50% of local or organic or quality products in collective catering by 2022. In order to go further and consider the environment in which producers operate, it is essential to analyse the interaction of digital technology with commercial and productive strategy. In an increasingly interconnected world, we can wonder about the contribution of digital tools on one hand on the strengthening of the producer-consumer link and on the other hand on the enhancement of productive efforts.
The objective of the internship is to develop a typology of farms based on the triptych: use of digital technology, marketing methods and production methods. Different steps will be necessary:
|Agronomy / Geography – Student in agronomy, motivated by analysis of agricultural systems, precision farming with livestock production, and sustainability assessment
The role of digital in agroecological transitions of monogastric livestock systems
The development of agroecology in livestock production is a major challenge for sustainability of French agriculture, in which intensive livestock systems cause important emissions of greenhouse gases and nitrate pollution. Ruminant livestock systems have long been studied to identify pathways for their agroecological transitions: grass-based systems and pastoralism are two ways of reducing the use of inputs and preserving biodiversity (Dumont et al., 2018). Monogastrics, mainly pig and poultry, despite being very important production in volumes, are still at the crossroad between intensive models, based on confined systems and massive importation of grains and proteins, and extensive, full-range models answering to niche markets. A third pathway seems reachable if monogastric systems is more connected to the local territory and production systems, implementing approaches of circular economy or territory ecology (Van Zanten et al., 2018). Complementarities can then be found to use by-products of other activities (non-eatable fruits and vegetables, barley bran from breweries) and local natural resources (fruits from oak trees or chestnuts, grazing in orchards or vineyards). In turn, livestock manure can be used for fertilization of surrounding crops. Such synergies exist in small-scale diversified farming (often organic), but are poorly explored in large scale, specialized livestock systems, because of the lack of knowledge of other activities and local actors, the lack of references on value of the different feed sources, and a strong integration of livestock farms in supply chains that organize the full supply of feed. In order to identify possible complementarities for monogastric livestock systems at territory scale, digital tools can be mobilized for the diagnosis of feeding practices and for the connection with local production systems and actors.
The Master thesis will aim at (i) characterize the use of digital tools in monogastric livestock systems; (ii) identify possible options of local complementarities (for feed supply and manure management) in a circular economy. The stakeholders of precision livestock farming will be interviewed (companies, suppliers, technical advisors and farmers). Brainstorming workshops will be conducted to identify and co-design models of circular farming systems at territory level. The area of Languedoc (Pyrénées-Orientales, Lozère, Aude, Hérault and Gard districts) will be targeted, to explore around 20 livestock farms using digital tools. The potential of digital tools in precision livestock farming (animal monitoring, feed control) will be analysed and discussed to define how far such tools can be helpful to design agroecological models of monogastric livestock farms.
Engineering, Mathematics & Agronomy
|Agronomy & ICT – Informatics
Specialisation of a location and classification neuronal network for the implementation of a mango maturity assessment tool
Keywords: Deep learning, neural network, mango, maturity, Senegal
|Informatics – Semantic Web
Taking into account uncertainties in a decision support tool for the AOP cheese farmer and producer
Keywords: Uncertainty management, decision support system, imperfect knowledge management<
Cheese making chains valorizing their terroir represent an important economical and agricultural activity in France, with around 17900 milk producers, 1290 farm producers and 432 transformation companies. Cheese making chains enjoying an “Indication Géographique” (AOP/IGP) base their product differentiation strategy on the valorization of local resources in connection with their terroir and on the expression of the know-how of experience as well at the level of the production as the transformation. Internal evolutions to appellations, especially in terms of turn-over and training of operators, strongly weaken the preservation and transmission of these know-how. The development of artificial intelligence methods allowing the exploitation of knowledge bases opens new perspectives in terms of preservation and data management of operational experience, by proposing complex modes of reasoning going well beyond the description and formalization of standard processes. As part of the CASDAR Docamex project (2017-2020), the knowledge engineering team (ICO team) of UMR IATE is developing a method and a tool for decision support (DSS) for the farmer who helps to control a defect or quality of manufacture by recommending the most relevant technological actions to undertake. The DSS also allows, for a given action, to determine all the defects and qualities impacted. These recommendations are based on the formal representation of the causal relationships linking default / quality to actions by explanatory mechanisms. In Docamex, the ICO team works with several AOP chains of different characteristics (Comté, Reblochon, Emmental of Savoy, Salers, Cantal) in order to develop a generic and adaptable DSS. The knowledge manipulated by the ADO is formalized with the Semantic Web languages well adapted to integrate knowledge from heterogeneous sources in the chains. In the master’s project, the methodological lock aimed at modeling the uncertainties associated with the causal relationships linking the defects / qualities to the actions. The first objective is to be able to propose in the ADO a prioritization of actions taking into account these uncertainties. The second objective is to be able to update the uncertainties by taking into account the result of the new cheese experiments recorded in the field. The proposed modeling will be tested by the creation of a software prototype that will be integrated into the DSS.
|Statistics – biostatistics – Modelling
Assimilation of proximal and remote sensing data to improve the forecast capacity of the crop growth model SiriusQuality
Keywords: Decision support, Data assimilation, Crop model, High-throughput phenotyping, Remote sensing
Crop models simulate the interaction between plants and the environment to estimate yield, harvest quality and the environmental impact of the crop. They represent the daily crop growth as a function of meteorological conditions, water and nitrogen availability and varietal characteristics. Crop models can be used to optimize (strategical and tactical decisions) or to drive (operational decisions) crop management practices, to maximize crop yield and quality and minimize unwanted nitrogen losses (e.g. leaching, N2O emissions). Data assimilation improves the forecast capacity of models and reduces uncertainty of simulations. By collecting spatial information for large areas at low resolution (>10 m), remote sensing provides the means of monitoring agricultural systems at intra-plot to regional scale, to evaluate the effect of different management strategies on yields and the environment. Proximal remote sensing methods (terrestrial or aerial) provide complementary information with higher temporal and spatial resolution. In addition to crop management, proximal remote sensing methods are developed for high throughput phenotyping for varietal selection. Assimilation of fast phenotyping data allows to determine genotype x environment interactions in multi environment experiments, or to analyze the genetic variability of not or punctually measurable characters. The objective of this Master project is to evaluate data assimilation methods to be coupled with a crop model (Sirius Quality; http://www1.clermont.inra.fr/siriusquality/). The student will use leaf area index, soil moisture and chlorophyll concentration data from different sources (Sentinel 2, drone, Phenomobile), and will evaluate a hybrid data assimilation method that combines parameter and initial conditions estimation and an original particle filter method (Chen, Trevezas, Cournede, 2014 ; doi.10.1137/1.9781611973273.10) developed in the framework of a collaboration with the Cybeletech agricultural start up. This will contribute to improve the representation of crop growth and reduce uncertainty. The student will be involved in the choice of the parameter and initial condition estimation method, and in the coupling with the data assimilation method. He/She will implement and test the methods for the Sirius Quality model. At the end the developed method will be validated for two additional crop models (STICS and Monica) and will be integrated in the decision support tool developed by Cybeletech.
|Applied Mathematics – Agronomy & ICT
Numerical optimization for precision spraying in viticulture: obtaining intervention maps maximizing the protection level associated to reduced doses
Keywords: Optimization, spray efficiency, disease risk, map segmentation
The wide access to spatialized data at the scale of an agricultural plot as well as technical innovations make it possible to adjust spatially cultural interventions by taking into account spatial heterogeneity. Precision agriculture opens up the possibility of reducing the use of phytosanitary products while preserving agronomic performance, which is a particularly important challenge in viticulture (10 to 24 treatments per year depending on the production context). This Master internship proposal is part of the research action “Precision spraying in viticulture” developed at IFV within the UMT EcoTechViti, and at UMR ITAP and aims to mobilize optimization methods developed at UMR MISTEA dedicated to map zoning (R package ‘Geozoning’). Contrary to the usual scheme of defining action zones by processing vegetation maps independently of cultural actions, we propose here to optimize the very definition of action maps by guiding it with a function representing the level of protection actually achieved and with an estimation of a risk accepted by the vinegrower. This approach would also allow to take into account in-season observations (typically observation of outbreaks of cryptogamic diseases) for the generation of maps. The selected student will implement and test an optimization algorithm defining an action map for precision spraying that i) is based on the modeling of the grapevine protection level and ii) meets the objectives of the vinegrower. The first step will consist in formalizing with crop scientists the problem of optimizing the level of protection from the sprayed dose, the spray efficiency and an the plant density obtained using Lidar data. In a second step, the student will study with mathematicians the use of the optimization method ‘geozoning’ to solve the problem numerically. This step may lead to implement in R or Python a “1d” version of the initial algorithm by considering the curvilinear abscissa of the sprayer path. Finally, scenarios for evaluating the obtained maps will be constructed using data acquired by the UMT EcotechViti. The student will be hosted at UMR MISTEA, supervised by Patrice Loisel and Sébastien Roux and co-supervised by Xavier Delpuech (IFV, UMT EcotechViti) and Olivier Naud (UMR ITAP, UMT EcotechViti). This project will also provide an opportunity for interdisciplinary collaboration on management issues and the economic context at the farm scale. Indeed new questions are raised when using concepts as “phyto budget” (dose to be distributed) and localized risk management. This will be done with colleagues from UMR MOISA (in particular Isabelle Piot-Lepetit and Karine Gauche).
|Remote Sensing / Agronomy – Knowledge in optical remote sensing & solid background in programming and statistics. QGIS, R, Python) would be a plus
Landscape scale estimations of agricultural performance of smallholder farming systems using satellites and UAVs: Case study of an agroforestry parkland of Senegal
Incorporation and management of isolated trees as an integrated part of smallholder farming systems has long been a key food security and livelihood strategy while improving farmers’ resilience to climate change in Africa (Garrity et al., 2010; Mbow et al., 2014). Hence, parklands are considered as a good option to foster the sustainable intensification of African agriculture. Methods for the monitoring of agronomical performance of parklands are therefore necessary to optimize farmers’ practices. With the democratization of high spatial and temporal resolution satellite imagery (Sentinel-2, Venus or PlanetScope), an accurate estimation of cereal crop yields in complex agricultural landscapes is now possible. Recently, Leroux et al (2019) showed that integrating information on the parkland structuring into a statistical remote sensing based model for millet yields estimation improved the agronomical performance assessment in a Faidherbia albida parkland, Senegal. However, this type of approach based on observed yields aggregated to the plot scale doesn’t take into account the intra-plot variability of yields due to environmental micro-variabilities, farmers’ practices or the presence of trees. UAVs are a tangible alternative for crop monitoring and their intra-plot variability in smallholder family farming (Blaes et al., 2016; Roupsard et al., 2019). They can also be used to extend the observed data over a limited number of plots to a larger area. The objective of this internship is therefore to study the complementarities between high spatial and temporal resolution remote sensing images and UAV images to estimate the spatial variability of yields in the Faidherbia albida parkland in the Senegalese groundnut basin. Supported by CIRAD’s AÏDA RU (L. Leroux), the Eco&Sol RU (O. Roupsard) and AGAP RU (A. Audebert), it is part of the EU-RAMSES2 project (https://josianeseghieri.wixsite.com/ramsesii). The intern will work from a set of remote sensing/UAV images acquired in 2018-2019. More precisely, the intern will first have to test different spectral and/or textural indices to estimate millet yields for a set of plots from drone imagery. In a second step, based on a land use map previously produced over the area (Ndao et al., 2019), these estimates will be extended by establishing a drone/satellite relationship to integrate intra-plot variability into the landscape scale estimates. To this end, data mining and geostatistical methods will be used to exploit the complementarities between the different sources of information. The results of this internship will contribute to the spatialized assessment of ecosystem services provided by trees in the RAMSES2 project where the proposed approach will be validated on other types of agroforestry parks.
|Computer Sciences, Embedded application, geolocalized data, sampling, statistics / Agronomy – Skills in development under Android environnement, statistics and teamwork
Building the “Coffe Health Diagnosis” application
The coffee farmer looks at a diseased tree and wonders if he should treat his coffee trees. He takes his smartphone, activates the ChDiag application “Coffee health diagnosis”, and reports the damage to a virtual coffee tree. ChDiag indicates other coffee trees to sample and then provides him with an estimate of the level of damage on the plot. He can make a decision. Thus, the objective of this project is to create a tool to diagnose the severity of pest attacks on a plant and on a plot scale. The project aims to contribute to the reduction of chemical inputs in crop protection. Indeed, Liebig et al (2016) find that many producers trust the effectiveness of chemical treatments while other types of treatments, such as cultural methods, exist. In their study, producers using chemical treatments have no better results than those using alternative methods. ChDiag will also allow a more accurate assessment of attacks: Ribeyre & Cilas (1998) showed that, in the case of the coffee berry borer, the “random” sampling of visible damage resulted in a significant overestimation of the damage. Diagnosis on smartphones will be carried out via guided damage entry. A damage scale will be represented on a typical plant with variable numbers of branches and fruits. The user will choose the type of plant that most closely resembles the actual plant and report the observed damage in situ. A sequential sampling plan will guide the producer in the choice of plants to be sampled and will stop it when the required precision is reached. ChDiag will then provide an estimate of the damage to the plot. This tool intended for coffee producers in areas where the pest infestation occurs will be specialized for damage caused by leaf disease, rust and a bark beetle, the black twig borer (branch attacks). According to the assessment of the level of damage caused, ChDiag will recommend that the producer switch to farming methods when infestations are below the epidemic threshold. ChDiag will therefore help to limit the use of chemical treatments. A procedure for control sampling, spaced over time, will be recommended. This smartphone tool will integrate 1) a statistical sampling procedure based on literature, past or acquired experimental data, and validated by partnerships (coffee growers’ associations) established as part of ongoing projects (including SWITCH Africa Green / EuropAid in Uganda and PROCAGICA UE in Costa Rica); 2) a geo-located data entry interface; 3) action advice based on the literature. A tool for predicting the impact of damage on production (coupling a bio-aggressor model with the FSPM GreenLab model) is planned to be integrated later.
|Image analysis / Agronomy & et Ecology – Agronomist interrest into numeric tools or computer scientist interrest by agronomy and ecology
Image segmentation in order to improve rangeland biomass rangeland using photogrammetry
Quick estimation of grass biomass in rangeland and grassland is an important goal for the management of grass-based livestock. Indeed, based on the available biomass quantity, livestock management can be implemented. Indirect measurement could be very useful to simplify the yield estimation. The photogrammetry allows from images toke with different points of view, to produce 3D model of the object such as an herbaceous layer. The first trial in Senegal show that 3D model obtain from simple camera is related to the grass volume and so to the grass biomass. However, the link between grass volume and grass biomass change with the species, phenological state and Herbaceous covers structure. The internship goal will be to used images from 3D model obtain from camera and segmentation images approach to evaluate the percentage of the different species and of the different phenological states. These information could be reused to improve prediction of biomasses for the grass volume estimated by photogrammetry? This internship will work on data produce during the summer 2019 along a monitoring experiment where data were collected every 10 days and/ or data that will be collected in southern France during the spring 2020 . Segmentation technique based on deep learning will be used on these images. The possibility of predicting biomass with a simple camera open numerous perspectives for the monitoring of experimentation but also used for a participatory observatory of grass growth. These observatory could contributed to pastoral crisis warning system.
|Computer Science : Text Mining – Good knowledge of data acquisition systems (including Web APIs) and data base systems
Acquisition and analysis of Youtube video transcripts – Application to the problem of food security in West Africa
The internship is part of an interdisciplinary project on food safety risk management. The project focuses on the case of West Africa, where agricultural risks are always increasing as national monitoring and follow-up services may fail due to lack of technical and financial resources. The overall goals of the project are twofold: (i) to show how remote sensing data can be enriched by other data sources to make them more suitable for the analysis of food security conditions and (ii) to define original data mining techniques. The analysis and interpretation of agro-climatic data (e.g. satellite imagery, climate data) could be supported by the joint use of independent data from textual sources (social media, press, market analysis), which would help to correctly identify agricultural risks on a regional scale in near-real time. Nevertheless, obtaining quality textual data, being able to analyze it and link it to the agro-climatic sources is a challenging task
The focus of this internship is on the acquisition and analysis of textual data on the theme of food security coming from Youtube videos transcripts (i.e., textual transcription of the audio content). The geographical area of study is Burkina Faso. The idea is to process a source of information representing an unexplored alternative to the ones classically exploited while building textual corpora and perfoming text mining tasks (e.g., newspapers, scientific articles, classic social media platforms). A Youtube channel has already been targeted for this analysis, and it is the one managed by the RTB – Radiodiffusion Télévision du Burkina, containing nearly 12000 videos. By choosing such an official channel, we are targeting an ideal compromise between the dynamical aspects of social media content and the information quality of official sources. The hypothesis is that the videos being issued by an official news channel are more likely to contain valuable information (i.e., news, documentaries, interviews, filming of official events, and so on). Moreover, the standard and clear language used in this kind of videos guarantees a good quality of the textual transcripts. The data acquisition and analysis processes will be based on the use of web APIs and python libraries.
The goals of this internship include the production of a public corpus and of a series of analysis tasks based on the use of state of the art text mining techniques (e.g., LDA, word2vec). The deliverable will consist in a research paper presenting the knowledge about food security that can be discovery in such an information source.
The provisional planning is structured as follows: study of the specifications of the corpus to be built, definition and implementation of the data collection process, constitution of the corpus on the study area, analysis of the corpus, and writing of the deliverables.
|Artificial intelligence, Image processing / Agronomy, Phenotyping – very good computing skills, Python language, fluent English, writing and teamwork skills
Comparison of shallow and Deep Learning methods to estimate the vegetation green fraction from RGB imagery. Impact of the spatial resolution
High-throughput phenotyping is developing rapidly for plant breeding and smart farming applications. The green fraction, i.e. the fraction of green pixels in an image, is one of the most useful trait to monitor vegetation development that can be extracted from high resolution imagery taken from a range of systems including UAVs, ground robotic rovers and fixed cameras for continuous crop monitoring. When the spatial resolution is fine enough to minimize the fraction of mixed pixels, segmentation techniques appear very efficient to estimate the green fraction. They include either pixel based methods such as random-forest classification, and deep-learning approaches such as the Unet model. However, when the spatial resolution degrades, mixed pixels represent a significant fraction of the image. In these conditions the segmentation techniques have difficulties to segment the mixed pixels. The application of super-resolution techniques based on generative adversarial networks (GAN) could recover the spatial resolution so that the previous segmentation techniques can be still performing. However, when the resolution is too degraded, segmentation and super-resolution techniques failed. In these cases, other approaches either based on vegetation indices or on deep-learning models could be developed to yield estimates of the green fraction.
The objectives of the proposed study are to evaluate the performances of the three types of approaches described above as a function of the image resolution relative to the size of the vegetation elements. For this purpose, a collection of annotated high resolution RGB imagery (resolution better than 0.5 mm) taken over wheat crops will be used as a reference. The spatial resolution of these images will be degraded by binning adjacent pixels up to the point where the image texture disappeared (resolution of few cm). The performances of the three approaches described above (1: segmentation, 2: super-resolution + segmentation, 3: vegetation index or deep-learning) will be evaluated over this range of spatial resolutions for the reference images. The techniques may be adapted to better suit the actual spatial resolution using independent datasets for the training.
|Automatics / Phenotyping – solid bases in robotics and programming with predisposal for image analysis
High-throughput phenotyping of developmental heterogeneity of berries in grape clusters
In order to adapt Mediterranean viticulture to climate change, high expectations are weighing on the creation of varieties and the design of new viticultural practices. This questions the genetic and physiological bases of the vine’s response to fluctuations in the environment. In particular, ongoing research within the UMR LEPSE in collaboration with UMR AGAP highlights the need to take into account the developmental heterogeneity of berries in the cluster both to improve the accuracy of phenotypic characterizations but also because this heterogeneity of development is often increased in response to abiotic stresses and strongly impacts the quality of the harvest. Unfortunately, the implementation of such characterizations faces the heaviness of phenotyping methodologies while genotyping today benefits from very high speed technologies. The UMR LEPSE is piloting large national projects to advance these methodologies by focusing more particularly on automated devices in a controlled environment. The objective of the proposed internship is to evolve the automatisms dedicated to image acquisition and pretreatment in the PhénoArch phenotyping platform to analyze the developmental heterogeneity of berries in the cluster. The project will rely on the equipment and mechanisms in place in the platform that allow: positioning the plant in an imaging cabin with respect to the cameras shooting side and zenith; to position a side camera (as regards height and distance to the object) with respect to the plant; pretreat the image to retrocontrol if necessary the positioning of the plant and the lateral camera. The project will use robotics and machine learning methods. It will develop an automation algorithm for the placement of the camera and the plant that allows to optimize (quality and speed) image capture of the cluster and identification of a maximum number of berries in the cluster from the set of views taken by different cameras. The trainee will be able to rely on a similar project that has made it possible to automate the detection and measurement of early ear development on maize plants in the same platform (Brichet et al., 2017. Plant Methods 13 : 96). He will benefit from the expertise of the head of the platform (access to code and knowledge of robotics), an engineer specialized in pattern recognition in plant images and possibly LIRMM colleagues who have contributed to the success from the previous project on the ear of maize. Potted vine plants of various genotypes will be prepared in the greenhouse in the winter of 2019-2020 to be available for the development of the method.
|Computer Science, web-based AR ; location-based AR ; SIG / Modeling – IT engineer (knowledge of an AR SDK) interested in agroecology, or agronomist, with an experience in agroecological system design, in particular in design workshops, willing to get his/her hands in the code
Exploring the possibilities of augmented reality to support agroforestry systems design
Agroforestry is recognized nationally and internationally as a way of developing a sustainable agriculture that is both resilient to and mitigates climate change. However, its adoption by farmers in developed countries remains low. Beyond the technical issues (technical feasibility, lack of skills in tree management) and economic issues (lack of economic references on the costs and benefits associated with agroforestry), a major obstacle to the adoption of this practice is the necessary shift in the temporal scale: farmers are used to make decisions on an annual time step or over a few years, not over the several decades of tree growth. In addition, because of the complexity of these systems, the number of combinations of species, spatial configurations, tree and crop management options is huge and the choice of a particular system must be specific to each farm, depending on its pedo-climatic constraints, the local value chains and the objectives of the farmer. Farmers therefore need support for the design of their system, in the form of system design workshops or individual coaching. New technologies can facilitate this process and thus promote the adoption of agroforestry. In particular, augmented reality (superimposition of digital objects on real world images) offers the possibility of visualizing different options and thus to immerse oneself in different scenarios in order to choose the most desirable one and to initiate a process of change. Agroforestry is an ideal case study for augmented reality. First, the introduction of trees profoundly changes the appearance of the plots and landscapes, so visualization tools would be very effective. Second, the growth of trees is slow and the consequences of the choices farmers make now will only be felt several decades later, hence the utility of accelerated-time visualization tools.
This internship will define the methodology and technical environment allowing building virtual agroforestry systems, describing them and visualising them in situ. We intend to explore two complementary uses of augmented reality (AR): marker-based AR to enhance user experience during design workshops (e.g. after placing markers on a map of the farm, one could visualize the growth of trees on a digital elevation model superimposed on the map) and geolocation-based AR (following the design of his/her agroforestry system, the farmer could visit the real fields and visualize their appearance in 10, 20 or 40 years). The applications will be first developed for temperate agroforestry systems, for which 3D models of tree growth already exist within AMAP unit (e.g. walnut, poplar, wild cherry…), and then, tropical agroforestry systems (there is an on-going PhD thesis on the architecture of forest trees associated with coffee plantations in Côte d’Ivoire).
|Crop modeling / Statistics – Spatial analysis
Comparing common sensitivity metrics for assessing a multi-scale crop model
Crop models are being adjusted to accept high-resolution data to provide in-season predictions at farm, field and sub-field scales i.e. the models and the input data are becoming increasingly scalar in nature. A recent project (involving Dr Taylor) successful incorporated high-resolution UAV imagery (and derivatives) into a well-known potato crop model. Using very basic modelling criteria, such as the RMSE, the project demonstrated improved predictions when then UAV imagery input was used (compared to the original model). However, several limitations were noted with this approach. Only predictions at the same scale could be compared, whereas a key issue is identifying the optimal scale (spatial footprint) for running the model. As the modelling footprint changes, the influences of stochastic effects on model behaviour will change (both in the potential additional information gain and the potential increase in ‘noise’). Accounting for this is crucial in identifying a preferred modelling scale (output) in the presence of inputs at varying scales. From the Tuberzone project, a large database of crop data at point, zone and field scale is available to permit crop modelling at various scales. Within this Master’s project, the existing model will be run at various scales and with various scales of input (that can be adjusted in scale by interpolation or sub-sampling). Variance-based sensitivity analysis, using primarily Sobol-based approaches, will then be tested on permutations of scaled model inputs and outputs to understand how well these existing sensitivity analyses perform in assessing spatial crop model behaviour. In particular, a key aim is to start to understand how spatial variance structures, which exist in these data, are accounted for (and could potentially be better accounted for). The lead company on the original project has given permission for these data to be used in this project. This Master’s project will be focussed on the application and performance of sensitivity analysis indices to a known spatialized crop model. There is no need to understand the mechanics of the crop model or the underlying phenology/agronomy. The key criteria is for the student to be able to work in the (geo-) statistical space. A PhD project that expands these ideas was accepted by #DigitAg in 2019, but unfortunately the selected candidate declined the offer. It is hoped that this Master’s project could identify a student to potentially continue onto further post-graduate work in this area.
|Computer Science, Artificial Intelligence / Agronomy – Computer or Data Scientist with skills related to Artificial Intelligence, Deep Learning and Computer Vision/Image Analysis
Multi-temporal and multi-scale satellite data fusion through deep learning methods for land cover mapping
Nowadays, more and more data from satellite missions like the European Sentinel program are produced and made available, offering the possibility to monitor the same geographical area continuously over time thanks to high revisit period. Indeed, the Sentinel-1 (A / B) and Sentinel-2 (A / B) satellites acquire respectively radar and optical images of the Earth with a time frequency of about 5 days over the same geographical area and a spatial resolution up to 10 meters. The time series of satellite images thus generated represent a non-negligible source of information for efficiently managing our agriculture and adapting our farming practices to the major challenge of already noticeable changes in the climate. However, the need to put in place methods to efficiently and robustly manage and analyze this large and heterogeneous amount of data is still relevant. It is with this in mind that this internship proposal is included. In the continuity of the past work of Ms. Paola Benedetti (DigitAg master’s internship in 2018) and current work of Mr. Jean Eudes Gbodjo (DigitAg thesis), we want to evaluate innovative methods based on deep learning and especially neural networks to map land cover as a fundamental input to crop monitoring systems, by coupling radar (Sentinel-1) and optical (Sentinel-2) time series to 10 meters of spatial resolution with SPOT 6/7 satellite images very high spatial resolution (THRS – 1.5 meters) whose annual acquisition repeatability is more modest. The SENTINEL satellite images will be made available through the THEIA and PEPS distribution platforms while the SPOT6/7 images will be obtained via the GEOSUD Equipex.
The methods developed will be evaluated on two study sites: a first site located in metropolitan France in the Gard department and characterized by a conventional agriculture and a second site (Koumbia) located in West Africa in the Burkina Faso country and characterized by familiar agriculture. In addition, the developed framework may be reused to other study sites such as Reunion Island. The reference data for the Gard site will be built through the Graphical Terrain Register (RPG) and IGN BD TOPO while for the sites of Koumbia and Reunion Island, the UMR TETIS already has the ground truth data acquired via previous field campaigns.
|Mathematics / Computer Science – strong competence in mathematical programming
Robustness for precision farming
This internship comes in support of a thesis co-funded by DigitAg and led by Gabriel Volte on optimization for digital services in connection with precision agriculture. The management (Rodolphe Giroudeau, director, with Eric Bourreau of LIRMM, Olivier Naud, IRSTEA) will be completed for the internship by Mickael Poss, CNRS researcher at LIRMM recognized in the field of robust combinatorial optimization. The thesis is part of the ultimate goal of interactive optimization in the context of digital agriculture. In the early works, we considered a problem of optimization of the selective harvest with supposedly precise and exact data. The problem is illustrative of complex site organizations made necessary to exploit the local specificities of the crops. An innovative approach using constraint programming in a general branch and price scheme is under development. With this internship, we want to address the issue of robustness against uncertain data, which is a step towards interactivity. In fact, the most appropriate decision support involves a priori robust optimization and then interactivity in the field involving re-optimization according to the latest available data and user preferences. The example problem is similar to a problem of vehicle tours with business constraints very specific to the agricultural field. However, various sources of uncertainty can really affect the optimal solutions obtained in the deterministic case. Notable uncertainties include the high local variability of the crop and the quantitative errors introduced by statistical estimation methods (eg mass quantity or harvest volume). Ignoring these uncertainties can lead to inaccurate or even unacceptable solutions. To take them into account in optimization, the notion of uncertainty can be characterized either by random variables (stochastic approach), or with a worst case approach, with parameters respecting inequality constraints (robust approach). In this internship, we will privilege the second approach, for its great practical interest. We want to measure the impact of a robust approach on the quality and practical acceptability of solutions by comparing them with solutions obtained without uncertainty. The interest of coupling this internship with an ongoing thesis is the availability of realistic data and the mastery of a body of methods (column generation, branch and price, constraint programming) with related tools. The trainee will be able to focus on defining the robust problem, implementing the methods, and benchmarking. We will carry out a battery of tests on real and simulated data in order to measure the efficiency of the proposed methods, from the point of view of the quality of the solution and the calculation time. We will also carry out a parametric study (for example variation of the number of ranks, distribution of the quantities harvested and uncertainties on the parcel, …) in order to evaluate those which have the most important impact on the combinatory and the computation times .
|Electronics / Aquaculture, Marine Biology – background in electrical engineering and an experience or at least an interest in applications related to biology or ecology.
Intrabody and submarine communication for sensor network implantable in fish – benchmarking of communication techniques regarding power efficiency
This work is part of a long-term project aiming at developing the first implantable sensor network for holistic monitoring of fish state of health. This approach relies on the need for measuring several physiological parameters to gather useful information about various biological processes. For instance, a Ph or temperature sensor in stomach for feeding event identification, a sensor in fat storage to estimate the amount of available energy for a fasting period, a sensor in gonads to monitor oocyte cycle are interesting and useful solutions that should be set in a sensor network to provide useful dataset on the fish. Setting an implantable sensor network is of interest for every exploited animal species. Studying fishes in the context of aquaculture or fishing, is the targeted application. In the context of fishes living in salted water, a major issue for the deployment of such a sensor network is the wireless communication needed for data gathering. At first biological tissue is a very contraining environment for communicaiton. In addition, salted water strongly attenuates electromagnetic signals, which limits the range of communications usually used for sensor networks.
The objective of this internship is to set-up and test several intrabody communication techniques in order to evaluate their potential deployment in marine context. Intrabody communication techniques are mainly classified in two categories
Preliminary results have shown that it is possible to setup an intrabody communication through immersed fishes using LoRa protocol with a range of at least 1m.
As the application framework is the deployment in potentially small fishes for a long time (6 month – 1 year), the main criteria are the volume and the battery life. As a consequence, the design of the prototypes will focus on reducing the volume especially considering antenna and battery. The test objectives will be:
|Computer Science, knowledge engineering / Life & Environment Sciences – skills : data science knowledge and web technologies – Plus: semantic web
Development and alignment of semantic resources for agricultural data
Keywords: Data interoperability, semantic resources, vocabularies, ontologies, repositories, Agroportal, web technologies