Mining Spatiotemporal Information for Harmful Algal Bloom Forecasting and Mechanism Interpreting

Abstract

A multistep spatiotemporal forecasting (MSTF) network is developed through incorporating the graph convolutional network (GCN) and the long short-term memory (LSTM) network within a sequence-to-sequence (seq2seq) framework. The MSTF method can not only extract spatial and temporal information from the input data but also make multistep-ahead and continuous predictions. An MSTF-based harmful algal bloom (HAB) forecasting model is then formulated to predict the chlorophyll-a (Chl-a) concentration of the Dianchi Lake (China). The integrated gradients (IG) method is employed to interpret the trained MSTF model and quantify the attribution of each input dimension to the Chl-a prediction. Results indicate that (i) the coefficient of determination (R 2) of the MSTF model in 24-h-ahead Chl-a prediction reaches 0.926, 28.4% higher than that of the traditional LSTM model; (ii) the ammonia nitrogen (12.3%), the total phosphorus (10.2%), the total nitrogen (9.9%), and the temperature (8.6%) are significant variables for Chl-a prediction; (iii) the spatial information from neighbor lake and river stations plays an important role in the HAB forecasting, with an average contribution of 35.0%; (iv) the proposed MSTF model is also skillful in the 72-h-ahead Chl-a prediction. Results presented highlight the importance of considering both spatial and temporal dependency of monitoring data in HAB forecasting and mechanism interpreting.