HabLSTM: A Nonstationary Feature Focusing LSTM for Spatiotemporal Prediction of Harmful Algal Bloom

Abstract

Harmful algal bloom (HAB) has long been one of the most formidable environmental problems in the world. HAB is influenced by multi-factors, its dynamic is highly non-stationary, making its prediction challenging. The existing machine learning based HAB prediction methods mainly utilize time series data, which ignores the intrinsic relationship between spatial and temporal variations of HAB. To achieve more accurate HAB spatiotemporal prediction, a novel long-short term memory (LSTM) based non-stationary focusing prediction model (HabLSTM) is proposed in this paper. The HabLSTM network is constructed by stacking HabLSTM units consisting of the hidden states spatial differential block (HSSD) and the combined states temporal differential block (CSTD). The HSSD block utilizes the gating mechanism and the difference of hidden states to generate differential features between adjacent frames, guides the network to learn short-term non-stationary features by controlling the feature update of the hidden state in the HabLSTM unit. The CSTD block utilizes the gating mechanism and the difference of combined states to generate the differential features of the current input sequence, guides the network to learn long-term non-stationary features by controlling the feature update of the memory state in the HabLSTM unit. These two differential features guide the HabLSTM network to focus on learning non-stationary spatiotemporal features and boost HAB spatiotemporal prediction accuracy. In addition, two new spatiotemporal datasets of HAB named as Taihu HAB A and Taihu HAB B are established by using the year-A and year-B Normalized Difference Vegetation Index (NDVI) images collected by Himawari-8 satellite, respectively. The experimental results on the two HAB datasets and spatiotemporal prediction learning (ST-PL) benchmark dataset MovingMNIST++ validate the outstanding HAB prediction and non-stationary spatiotemporal features learning capability of HabLSTM.