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. »
