Modeling Soil Temperature for Different Days Using Novel Quadruplet Loss-Guided LSTM

Area of Science:

• Geosciences
• Environmental Science
• Data Science

Background:

• Soil temperature is a critical variable in geosciences with complex spatiotemporal variations.
• Accurate soil temperature estimation presents significant challenges due to numerous influencing factors.
• Existing models require enhanced data processing for improved estimation performance.

Purpose of the Study:

• To develop and validate a novel Long Short-Term Memory (LSTM) model integrated with a quadruplet loss function (QL-LSTM) for precise soil temperature estimation.
• To enhance the accuracy of soil temperature prediction by optimizing the loss function in the estimation model.
• To assess the model's performance against established estimation techniques using meteorological data.

Main Methods:

• A Quadruplet Loss Long Short-Term Memory (QL-LSTM) model was designed, combining squared-error loss with distance metric learning.
• The model processed meteorological data (radiation, air temperature, vapor pressure deficit, wind speed, air pressure, past soil temperature) from Laegern and Fluehli stations.
• Performance was evaluated using metrics such as RMSE, MAE, NS, WI, and LMI across different soil depths and days.

Main Results:

• The QL-LSTM model demonstrated superior performance compared to backpropagation neural networks, extreme learning machines, support vector regression, and standard LSTM.
• The model achieved the highest accuracy at a 15 cm soil depth on the first day at Laegern station, with exceptional WI, NS, and LMI values.
• Low RMSE and MAE values further confirmed the model's effectiveness in estimating soil temperature.

Conclusions:

• The proposed QL-LSTM model offers a significant advancement in soil temperature estimation accuracy.
• This model is highly recommended for estimating daily soil temperature profiles, particularly on the 1st, 5th, and 15th days.
• The integration of quadruplet loss function effectively optimizes the estimation process for geoscientific applications.