Online machine learning for collaborative biophysical modelling

Abstract

Many initiatives try to integrate data from different parties to solve problems that could not be addressed by a sole participant. Despite the well-known benefits of collaboration, concerns of data privacy and confidentiality are still an obstacle that impedes progress in collaborative global research. This work tackles this issue using an online-learning algorithm to generate a single model where the data remains with each party and there is no need to integrate it to a single source. This approach is demonstrated in building a global soil organic carbon model based on databases of field observations held by 65 different countries. The model is trained by visiting each country, one at a time. Only knowledge and parameters of the model are transferred between countries. The results show that it is possible that the proposed approach yields a similar prediction accuracy compared with a model that is trained with all the data.