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.
