STARLAB BARCELONA SL (ES)
Traditional empirical and analytical Earth Observation (EO) algorithms retrieving physical parameters are getting to a fundamental change where learning algorithms without any prior background will be able to set themselves through the ingestion of Inputs/Outputs training datasets. Nowadays, Deep Learning (DL) networks among many other Machine Learning (ML) techniques are accurate enough, and computation technology is available to run such models. One of the key issue of such approach is the availability of massive or, large enough, reference datasets to train the models. As the models learn from the available data within the training datasets, if the size of such dataset is relatively small, the models learn very specific features that do not allow o generalize to any input data due to the lack of representativeness of the training dataset.
This project addresses this issue in the context of a specific ML application, ie target/feature detection. The main goals are (1) to develop a PLATFORM to build and share collaborative training datasets for combined EO/ML communities, and (2) to implement a generic ML algorithm to detect targets in EO scenes for expert and/or non-expert users online
