Description du sujet de thèse

Changes in an animal's behavior can be indicative of its intrinsic evolution (e.g. age, reproductive state, health status) or its response to the environment and its changes (e.g. responses to human activities). For this reason, the study of animal behavior has become central to many studies in scientific ecology and wildlife management and conservation. These studies have been greatly facilitated in recent decades by the development of biologging, i.e. the deployment of sensors directly on the animal. These biologgers have become smaller and smaller, more energy-efficient, and carry more and more sensors capable of recording a variety of variables of interest: GPS location, activity measured at high frequency using accelerometers, sometimes sound or image. The analysis of collected data has also benefited from significant developments in machine learning models of increasing complexity, from decision tree models (e.g. xgboost) to the deep learning models of artificial intelligence, for example. However, at present, data analysis can often only be carried out once the sensors have been retrieved. The need to retrieve the biologger to access the data generates a long delay between its installation and the possibility of analyzing the data. The current situation is problematic: frugal, energy-efficient sensors with long lifetimes are desirable, but they also delay data analysis. Faced with this state of affairs, ecologists would like to be able to work with sensors that operate over a long period of time and transmit data remotely, enabling them to work with these data rapidly, and even enabling them, when useful, to have a 'reactive' approach, enabling monitoring or experimentation to be adjusted, and in the context of management/conservation, the implementation of intervention measures if necessary. Recent technical and statistical developments point to the possibility of creating 'intelligent' sensors that can analyze in real time the large volumes of data collected by a biologger's various sensors and produce relevant metrics, or behavioral classifications directly, information that can be transmitted more easily than raw data, enabling analysis long before the sensor is recovered. This data processing at the periphery (egde computing) would enable analysis and studies to be carried out closer to the time of data collection by the sensors, to adjust monitoring to observations, and also to develop surveillance approaches based on 'sentinel' animals enabling the detection of anomalies or worrying developments. The context of the thesis proposed here is to address this issue of 'intelligent' communicating sensors.
There are methodological hurdles to overcome: embedded data processing models must be as uncomplicated as possible to limit the biologger's computation and energy consumption, but sufficiently efficient to be useful. This usefulness is linked to the application-specific cost of obtaining false positives or negatives. There is therefore no single solution - which could have been appropriated independently by electronics and software engineering - to the problems outlined in the state of the art. Interdisciplinary work is therefore essential if realistic solutions are to be found.
The aim of the thesis is therefore to examine the added value of edge computing, and if necessary to go beyond proof of concept to offer new approaches for monitoring animal behavior. The PhD student will be responsible for implementing solutions of increasing complexity: (1) A solution corresponding to some current biologgers which transmit raw, or 'summarized', data for analysis away from the sensor; the work may focus on data compression approaches in a context of long-distance transmission efficiency. This step will enable us to establish a basis of comparison for the study of solutions 2 and 3; (2) A solution in which the data is analyzed in the sensor using a pre-trained model (i.e. the sensor makes an inference from the model and the just-acquired data), corresponding to the most classical approach in edge computing; the behavioral classification is then transmitted ; (3) A solution in which the sensor itself learns to detect anomalies or evolutions, thus performing 'self-learning', potentially based on unsupervised classification to dispense with the need to obtain training data, a strong constraint for model development for solutions (1) and (2). These different solutions will be implemented through concerted developments in electronics and on-board software, and will include consideration of the choice of data analysis models and the operation of the entire animal - biologist - user platform chain. Frugality of the entire chain will be considered. This phase of the study will also be an opportunity to put emerging embedded AI technologies to the test, notably with the very recent appearance of low-power microcontrollers that embed hardware neural gas pedals (Neural Processing Units) and which have not yet been investigated in the context of biologging. Prototype testing phases, both in the laboratory and in the field, will naturally take place during this period. Once a solution has led to a reliable prototype, we will initiate pilot studies to test the envisaged solution in real-life situations.

Profile required:
- Expertise in embedded machine learning
- Good command of C and Python languages
- Expertise in electronics will be considered favourably
- Interest for the ecology topic of the thesis

Références : [1] Berger-Tal & Saltz, eds. 2016. Conservation behavior: applying behavioral ecology to wildlife conservation and management. Cambridge University Press. [2] Candolin et al. 2023. Animal behaviour in a changing world. Trends in Ecology & Evolution 38: 313-315. [3] Nickel et al. 2021. Energetics and fear of humans constrain the spatial ecology of pumas. PNAS 118 :e2004592118. [4] Whitford & Klimley. 2019. An overview of behavioral, physiological, and environmental sensors used in animal biotelemetry and biologging studies. Animal Biotelemetry 7:1-24. [5] Nathan et al. 2022. Big-data approaches lead to an increased understanding of the ecology of animal movement. Science 375:eabg1780. [6] Mao et al. 2023. Deep learning-based animal activity recognition with wearable sensors: Overview, challenges, and future directions. Computers and Electronics in Agriculture 211: 108043. [7] Nuijten et al. 2020. Less is more: On‐board lossy compression of accelerometer data increases biologging capacity. Journal of Animal Ecology 89 :237-247. [8] Yu et al. 2024. Edge computing in wildlife behavior and ecology. Trends in Ecology & Evolution 39:128-130. [9] Morelle et al. 2023. Accelerometer-based detection of African swine fever infection in wild boar. Proceedings of the Royal Society B 290:20231396. Méthode Résultats attendus - Expected results Références bibliographiques [10] Rast et al. 2024. Death detector: Using vultures as sentinels to detect carcasses by combining bio‐logging and machine learning. Journal of Applied Ecology 61:2936-2945. [11] Arablouei et al. 2023. Multimodal sensor data fusion for in-situ classification of animal behavior using accelerometry and GNSS data. Smart Agricultural Technology 4:100163. [12] Trotta et al. 2023. Optimizing IoT-based Human Activity Recognition on Extreme Edge Devices. IEEE International Conference on Smart Computing (SMARTCOMP) (pp. 41-48).

Contexte de travail

This PhD is part of the SENTINEL project (Vers des animaux 'sentinelles' : le suivi continu et à long-terme du comportement animal par biologging est-il crédible ?) supported by the Mission pour les Inititiatives Transverses et Interdisciplinaire (MITI) of the CNRS. The PhD student will be provided with a workspace, a laptop and access to the software required for his or her work. Funding is available to cover short field missions in France or abroad to gain first-hand experience of the use of the sensors developed in the thesis.
The PhD student will be assigned to the Centre d'Ecologie Fonctionnelle et Evolutive (CEFE, Montpellier), but will be hosted at the beginning of the thesis mainly at the Laboratoire d'Informatique, Robotique et Microélectronique de Montpellier (LIRMM, Montpellier), located 10 min from the CEFE. L. Latorre (LIRMM), professor at the University of Montpellier, HDR, specialist in embedded systems, will be the thesis supervisor. S. Chamaillé-Jammes (CEFE), CNRS research director, HDR, will be co-director of the thesis. The doctoral student will be registered with the I2S (Information, Structures, Systems) doctoral school, specializing in Computer Science, at the University of Montpellier.

Contraintes et risques

No specific constraints or risks.