Innovative, state-of-the-art methodology to reduce the size and complexity of Copernicus Sentinel1 & 2 EO data making it more accessible to core downstreamusers.
Given the rapidly growing archive of Copernicus EO data (10TB/day), new methods need to be deployed to optimise their size while maintaining their validity. Our consortium is rolling out an innovative, state-of-the-art methodology that will reduce the size of Copernicus Sentinel 1 and 2 data, making it more accessible to downstream users.
The methodology leverages Weakly Supervised Learning (WSL) which is a subcategory of machine learning and artificial intelligence. WSL is an umbrella term describing the attempt to construct predictive models by learning with weak supervision. Supervised learning techniques construct predictive models by learning from a large number of training examples, where each training example has a label indicating its ground-truth output.
Although WSL may train models for particular tasks (biomass estimation, land-cover mapping, etc.), training models from scratch per application is still inefficient. This limitation led to the development of new methodologies to build intermediate representations (embeddings) of multi-modal information from, e.g., optical, SAR, and hyperspectral satellite imagery.
Recent successes in self-supervised learning (SSL) for computer vision and natural language models motivate our objective to transfer this approach to the domain of remote sensing, as recently proposed in the literature. The generation of generic embeddings from multi-sensor data such as served by the Sentinel 1 and 2 missions facilitates downstream tasks. For example, these embeddings may be used to represent time-series for change detection application with little to no need for supervised model training.
ResponsivePics errors