GANs-LSTM Model for Soil Temperature Estimation From Meteorological: A New Approach

Soil temperature (T
_s
) is a vital meteorological parameter for ecological, physical and biological research. T
_s
estimation is very important for a variety of fields and presents great challenges because the relevant areas have complex characteristics, human activities, and a nonlinear nature between T
_s
and its environmental factors. Hence, a novel model based on long short-term memory (LSTM), is proposed here as an alternative data-intelligence tool. The proposed model designs a novel function that combines LSTM loss with an adversarial term for enhancing the correlation between T
_s
and its environmental factors. For this purpose, we designed a novel function that combines LSTM loss with an adversarial term. The adversarial term encourages the LSTM model to estimate T
_s
, which cannot be distinguished from the observed T
_s
by an adversarial model. In this study, the proposed model is trained using meteorological information from two stations in China (Changbai Mountain and Haibei). The model is constructed using hourly input variables, including air temperature (T
_a
), wind speed (W), relative humidity (RH), solar radiation (SR), and vapor pressure deficit(VPD), while the objective variable is the T
_s
measurement at a 5 cm depth for the period 2003-2005. Different statistical evaluation criteria, including the root mean square error (RMSE), mean absolute error (MAE), Nash-Sutcliffe model efficiency coefficient (NS), Willmott index of agreement(WI) and the Legates and McCabe index (LMI), have been employed to assess the model performance. Through experimentation, the proposed model generally performs superior to the other seven state-of-the-art estimation models (autoregressive integrated moving average model, linear regression, backpropagation neural networks, support vector regression, extreme learning machine, eXtreme gradient boosting and long short-term memory) in T
_s
estimation at a 5 cm depth over the Changbai Mountain and Haibei stations. For this case, the most accurate performance is attained for T
_s
estimation at a 5 cm depth, with the highest values of NS = 0.915, WI = 0.978, and LMI = 0.686, and the lowest values of the relative RMSE and MAE being 2.276 and 1.796, respectively. In accordance with the present results, it is concluded that the proposed model can serve as an alternative approach for estimating T
_s
, while ensuring that an appropriate combination of meteorological inputs is applied to yield an optimal model.