COASTAL HABITAT MAPPING WITH UAV MULTI-SENSOR DATA: AN EXPERIMENT AMONG DCNN-BASED APPROACHES

We assessed the predictive skill of forecasting tools over the next season for water discharge and lake temperature. Tools were forced with seasonal weather predictions; however, most of the prediction skill originates from legacy effects and not from seasonal weather predictions. Yet, when skills from seasonal weather predictions are present, additional skill comes from interaction effects. Skilful lake seasonal predictions require better weather predictions and realistic antecedent conditions.

We develop a Gaussian Bayesian network (GBN) for seasonal forecasting of lake water quality and algal bloom risk in a nutrient-impacted lake in southern Norway. Bayesian networks are powerful tools for environmental modelling but are almost exclusively discrete. We demonstrate that a continuous GBN is a promising alternative approach. Predictive performance of the GBN was similar or improved compared to a discrete network, and it was substantially less time-consuming and subjective to develop.

In order to allow researchers to quickly prototype and build models of natural systems, we have created the Mobius framework. Such models can, for instance, be used to ask questions about what the impacts of land-use changes are to water quality in a river or lake, or the response of biogeochemical systems to climate change. The Mobius framework makes it quick to build models that run fast, which enables the user to explore many different scenarios and model formulations.