Nitrogen prediction in the Great Barrier Reef using finite element analysis with deep neural networks

Abstract

The corals of the Great Barrier Reef (GBR) in Australia are under pressure from contaminants including nitrogen entering the sea. To provide decision support in reaching target water quality outcomes, development of a nitrogen forecasting model may be useful. Here, we propose a new technique that considers the whole GBR as a frame and treats forecasting of nitrogen as a next-frame prediction task, to produce spatial maps of nitrogen over the whole GBR at forecast time-steps. To achieve this, we design an innovative Deep Neural Network (DNN) inspired by the Finite Element (FE) analysis concept. In our proposed method, the GBR area is meshed into small elements with pre-calculated stiffness matrices first. Next, both the stiffness matrices and the nitrogen values of each element are fed into the designed DNN for element-wise nitrogen prediction. The final result is then gained by attaching separate outputs of each element. Unlike other next-frame prediction models, our FE-DNN model generates accurate forecasts with unblurred prediction frames. We demonstrate that our model is the first to provide nitrogen forecasts for the entire GBR with low Mean Square Error (MSE), while generating a high-resolution prediction frame. The proposed model is applicable to other environmental modelling applications that are governed by Partial Differential Equations (PDE), e.g., sea temperature prediction and sediment distribution forecasting. Nonetheless, no knowledge of the underlying PDEs is required to use our DNN-based model. Our method can produce accurate forecasting predictions by leveraging existing hindcasting simulation models.