A Novel Multichannel Long Short-Term Memory Method With Time Series for Soil Temperature Modeling

• Main Authors: Qingliang Li, Yang Zhao, Fanhua Yu
• Format: Article
• Language: English
• Published: IEEE 2020-01-01
• Series: IEEE Access
• Subjects: Machine learning; soil temperature modeling; long short-term memory; autoregressive integrated moving average
• Online Access: https://ieeexplore.ieee.org/document/9214479/

Soil temperature (Ts) plays an important role in earth sciences. The temporal and spatial variations of Ts are affected by several factors. To acquire the deterministic component and the stochastic component of time series data and further improve estimation performance, a multichanne long short-term memory model (MC-LSTM) is developed to address the challenge of accurate Ts estimation. Specifically, we connect multiple channels in an LSTM structure and an autoregressive integrated moving average model (ARIMA) directly to the output layer to characterize the deterministic part and the stochastic component of time series data. Meanwhile, to improve the correlations in the time series by increasing the number of memory units, we also connect the units in the different steps directly to the output layer in the LSTM structure to learn the long-term pattern of historical Ts by connecting the unit in an earlier step directly to the output layer, to learn the features that occurred in a short-term pattern by connecting the unit to the later step captures. In this article we test the hypothesis that Ts estimation is performed at depths of 5, 10 and 15 cm over 6, 12, and 24 hours. To achieve this, the input data are composed of half-hour data from the Ts of two synoptic stations (Laegern and Fluehli) in Switzerland. Meanwhile, the estimation accuracy is verified by three performance criteria; RMSE, MAE and R². As expected, the proposed model achieves the highest relative R² values of 0.9965 and the lowest values of RMSE = 0.3414 and MAE = 0.2310 for Ts estimation over 6 hours at Fluehli Station (10 cm soil depth) when compared with other state-of-the-art machine learning methods. Consequently, the proposed model can serve as an alternative approach for Ts profile estimation.