On the use of the Ocean Virtual Laboratory open tools to prepare training sets for AI deep learning of ocean surface signatures

Today when dealing with AI and deep learning to automatize recognition of ocean surface structures on the wealth of satellite observations produced every day, scientists are faced with the challenge of preparing representative training sets. These training sets consists of classes of targeted structures such as oceanic eddies with associated satellite image containing signatures of these structures. The network training process will then attempt to learn this classification from the training set. Therefore, the performance of the trained network will critically depend on the quality of the training set. However, preparing training sets of sufficient size and quality to be representative of the majority of signatures of the targeted structures requires significant amount of time from qualified people, ending up with a significant cost. We will be demonstrating some tools, such as the Ocean Virtual Laboratory (https://ovl.oceandatalab.com) that can help to significantly speed up and therefore reduce the preparation cost of these training sets by offering a mean to quickly browse, interpret and annotate satellite observations, in the context of a wide range of complementary satellite observations, together with model outputs and in-situ data.