Marine ecological information prediction by using adjacent location spatiotemporal deep learning model with ensemble learning techniques
Rising sea temperatures and shifting tidal patterns, fuelled by climate change, cause formidable threats to marine ecosystems. Accurate prediction of sea surface temperature (SST) and tidal height (TH) is essential to address these challenges. Recently, many studies have used machine learning or dee...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-03-01
|
| Series: | Ecological Informatics |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1574954124005065 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Rising sea temperatures and shifting tidal patterns, fuelled by climate change, cause formidable threats to marine ecosystems. Accurate prediction of sea surface temperature (SST) and tidal height (TH) is essential to address these challenges. Recently, many studies have used machine learning or deep learning to predict SST and TH. However, achieving more precise predictions for SST and TH remains crucial. In this paper, we introduce an innovative deep learning approach that combines ensemble learning techniques with multi-featured spatiotemporal deep learning models to accurately predict SST and TH. First, we train a temporal model using LSTM for the nearby sea areas of the target prediction location. Then, we use various ensemble learning (EL) techniques to integrate the models of these nearby sea areas, forming a multi-featured spatiotemporal model (STH-LSTM). We test our model in the Penghu Sea area near Taiwan (Penghu, Dongji, Qimei, Jibei, Jiangjun, and Wangang) using data from the Central Weather Bureau (CBW) from 2020 to 2022. The ensemble techniques employed include STH-MLR-LSTM, STH-AGG-LSTM, STH-MAE-LSTM, and STH-SAM-LSTM. In this study, we evaluate the proposed model's performance using metrics such as Mean Absolute Error (MAE), Root Mean Square Error (RMSE), and the coefficient of determination (R2), alongside comparative analyses against SVR (Support Vector Regression), AdaBoost, and RF (Random Forest) models. The results show that STH-MLR-LSTM achieves the best average prediction results across the six locations. Using the Average Improvement Percentage (AIP) for evaluation, STH-MLR-LSTM improves SST MAE by 13.7 % to 69 % compared to other models, with specific improvements of 62.91 % over SVR, 58.73 % over AdaBoost, and 44.85 % over RF. The worst-performing model was STH-SAM-LSTM. For TH MAE, the improvement ranges from 5.8 % to 78 %, with specific improvements of 67.62 % over SVR, 78.04 % over AdaBoost, and 66.54 % over RF. The worst-performing model was AdaBoost. These findings indicate that our innovative STH-MLR-LSTM, which combines ensemble learning techniques and a multi-featured spatiotemporal deep learning model, to achieve the best results in tidal height prediction. |
|---|---|
| ISSN: | 1574-9541 |