Room 109, Building 9, St Eloi campus Machine Learning in Montpellier, Theory & Practice The field of remote sensing (RS) has witnessed remarkable advancements, but the scarcity of labeled data remains a challenge. This talk focuses on leveraging transfer learning and self-supervised learning (SSL) methods to benefit downstream tasks with limited data by learning representations from large datasets. However, most pre-trained models in RS are based on ImageNet or MS COCO, which may not capture the spectral and spatial characteristics of RS data. To address this, SSL provides a solution by learning useful feature representations from unlabelled RS data. By considering multimodal data, particularly optical and radar images, SSL models can improve accuracy and robustness in RS analysis. The talk aims to shed light on the potential of transfer learning and SSL methods in RS, emphasising the importance of considering the multi-modal nature of RS data. Machine Learning in Montpellier, Theory & Practice

Room 109, Building 9, St Eloi campus Machine Learning in Montpellier, Theory & Practice Registration was required for this session.,, Machine Learning in Montpellier, Theory & Practice

Room 109, Building 9, St Eloi campus Machine Learning in Montpellier, Theory & Practice Food sovereignty challenges in the South are occurring in an increasingly uncertain environment of geopolitical instability and climate change threat. This situation calls for the development of sustainable and resilient food production systems. In this context, modelling, image analysis, and high-throughput observation systems have become ubiquitous in agronomy to monitor and simulate the performance of agricultural systems. In this presentation, I will present the activities of the Phenomen team in image analysis of plants observed at scales ranging from nanometers to kilometers, using registration, geostatistics, segmentation, and time-lapse tracking techniques. The core of the presentation will focus on segmentation and time-lapse tracking of root architectures. With its simple geometry but complex topology, this problem efficiently resists deep-learning approaches; recently, we studied this subject and contributed to the state of the art by rethinking "traditional"" pipelines based on deep learning [1]. The talk will conclude with a discussion to suggest ways to go beyond the state of the art by considering a new formalization of the problem leveraging recent deep-learning techniques for image, graphs, and time-lapse tracking." Machine Learning in Montpellier, Theory & Practice