针对来自异构数据的 OLCI 图像中海冰通道检测的弱监督深度学习方法

本研究旨在通过整合高度计提供的1D数据与光学传感器捕获的2D图像，开发半监督及弱监督学习方法，以有效识别海冰结构。该项目致力于解决准确描述冰体结构分布的问题，从而理解北极气候下的热交换机制，这涉及重要的环境与经济问题。

Job description: Project description: Weakly supervised learning approaches have been developed to solve tasks for which little or no labeled data is available while taking advantage of the performance gains of deep learning. Semi-supervised or weakly supervised methods construct a structured representation of the data from which information can be propagated. In this thesis, the objective is to develop semi-supervised or weakly supervised methods to propagate 1D information obtained through altimeters to 2D image sensors for the purpose of mapping ice floe leads. Sea ice leads, which are fractures between ice floes, are the main sites of heat and moisture transmission from the ocean to the atmosphere. To assess this heat transmission and model the evolution of the sea ice, which extent has been decreasing since the 1990s, it is important to accurately characterize the distribution of the leads. Such characterization is possible from an accurate 2D altimetric map. Altimeters allow for floe/lead classification and accurate elevation measurement, but they only acquire pointwise measurements along their trajectory. On the other hand, optical imagers enable floe/lead differentiation on a large scale but this discrimination by adaptive thresholding methods alone is still very noisy [Muchow2021] because the appearance of sea ice varies greatly depending on external conditions. Using the OLCI (a multispectral imager) and SRAL (a SAR altimeter) sensors, both on board the Sentinel-3 satellite, a method for floe/channel patch classification has already been proposed [Chen2024]. An OLCI patch classification database was built in the team [Orgambide2025] to train a method for classifying sea ice superpixels obtained by SAM [Kirillov2023]. These two methods are promising but do not solve a semantic segmentation task and do not offer pixel-level classification. Moreover, optical sensors cannot measure directly sea-ice elevation. To obtain a reliable elevation and sea ice lead map, several research questions must be addressed: - How can imprecise labels from altimeter sensors and a small number of manually labeled patch be best used to train a deep learning semantic segmentation method for lead detection? - The OLCI sensor has a resolution of 300 m, which does not allow the detection of thin leads, which nevertheless involve significant thermal exchanges [Marcq2012]. The Sentinel-2 MSI sensor has a less frequent revisit, but a resolution of 10 m. Is it possible to extend the channel detection capabilities of the OLCI sensor to the MSI sensor? - Finally, can the methodology developed for detecting leads in optical images using labels from the SRAL altimeter be extended the elevation measurements? Altimeter measures the elevation of the sea ice through the measurement of its freeboard, which necessitate a higher precision than on the floe/lead classification. Methodologically, this thesis is part of the theme of weakly learning, a theme found in many scientific and industrial problems. The objective is to use deep learning to construct latent representations of image data that can solve a variety of tasks that are not necessarily defined at the time of learning this representation. On the application side, this thesis is part of the study of the ice pack and, more generally, of the Arctic climate, the understanding of which is primarily an environmental issue, but also an economic and even military one.

Profile and skills required: We welcome candidates with a Master 2 including courses in computer vision, image processing and remote sensing. Niveau de français requis: Intermédiaire supérieur: Vous pouvez utiliser la langue de manière efficace et vous exprimer précisément.