Optimization of deep learning model for coastal chlorophyll a dynamic forecast

• Published in: Ecological modelling Vol. 467; p. 109913
• Main Authors: Wenxiang, Ding, Caiyun, Zhang, Shaoping, Shang, Xueding, Li
• Format: Journal Article
• Language: English
• Published: Elsevier B.V 01-05-2022
• Subjects: Chlorophyll dynamic forecast; Deep learning; Relative change rate; Xiamen Bay; Deep learning; Xiamen Bay; Chlorophyll dynamic forecast; Relative change rate
• ISSN: 0304-3800; 1872-7026
• DOI: 10.1016/j.ecolmodel.2022.109913

•A deep learning model was optimized using the change rate (ΔChl) and the relative change rate (ΔRChl) of Chl to replace the absolute concentration (Chl) as the output of the LSTM model.•The Chl concentration forecast result obtained through ΔChl and ΔRChl was much better than that obtained through Chl.•Combining the Chl forecast results obtained through ΔChl and ΔRChl can prevent overestimation of the Chl peak.
Chlorophyll a is an important factor in characterizing algal biomass. Its dynamic forecast model is considered to be one of the best early warning methods to prevent or alleviate the occurrence of algal blooms. In this study, the absolute concentration of Chlorophyll (Chl), the change rate of Chl (ΔChl), and the relative change rate of Chl (ΔRChl) were used as the output of a long short-term memory (LSTM) model. The model was used to carry out Chl dynamic forecasts for different seasons in Xiamen Bay. The results show that the Chl forecast result obtained using ΔChl and ΔRChl is much better than the forecast using Chl. Combining the Chl forecast results obtained using ΔChl and ΔRChl can solve the problem of overestimating the Chl high value, thereby improving the forecasting accuracy. Effectively applying our understanding of the mechanisms of deep learning forecasting models can improve forecasting capabilities.