Watershed groundwater level multistep ahead forecasts by fusing convolutional-based autoencoder and LSTM models.

Area of Science:

• Environmental Science
• Hydrology
• Artificial Intelligence

Background:

• Groundwater level forecasting is crucial for sustainable water resource management.
• High-dimensional groundwater dynamics pose challenges for traditional models.
• Accurate prediction is needed to manage groundwater resources effectively.

Purpose of the Study:

• To propose a novel deep learning model for accurate spatiotemporal groundwater level forecasting.
• To evaluate the performance of the proposed ConvAE-LSTM model against benchmark models.
• To enhance groundwater resource management strategies through improved forecasting.

Main Methods:

• Developed a hybrid Convolutional-based Autoencoder (ConvAE) and Long Short-Term Memory (LSTM) neural network model (ConvAE-LSTM).
• Utilized a dataset of groundwater levels from 33 wells in Taiwan (2000-2019), incorporating point data and derived imagery.
• Compared ConvAE-LSTM with HBV-light and LSTM models for three-month ahead forecasts.

Main Results:

• The ConvAE-LSTM model effectively extracted features from both point and imagery datasets.
• Established spatiotemporal dependencies between regional imagery and groundwater level data.
• Achieved significant improvements in R-squared values compared to HBV-light (18-49%) and outperformed LSTM.

Conclusions:

• The ConvAE-LSTM model demonstrates superior accuracy in multi-step-ahead groundwater level forecasting.
• This model offers a significant advancement in environmental modeling for groundwater resource management.
• The findings provide insights for designing effective, long-term sustainable groundwater management strategies.