Deep transfer learning in remote sensing with large pretrained model: study on crop type classification by Quentin Deffense

Louvain-La-Neuve

March 07, 2023

13h

Salle Ocean

Recently, large self-supervised pre-trained models such as GPT-3 (ChatGPT), Bert (Google), InceptionV3, … have shown that they can outperform previous state-of-the-art models and transfer to different tasks and domains. In remote sensing, a large amount of unlabelled data are freely available but gathering in situ dataset is costly. Pre-trained models could leverage the features in the unlabelled images to both outperformed the current models and reduce the need of manual labels.

In this seminar, I will explain how GPT-3 and Bert use unlabelled data to train and I will show how their mechanisms can be used in remote sensing. Finally, I will show the first results on a crop type classification.