An Evaluation of Soil Temperature Predictions Based on the Long Short-Term Memory Model and Remote Sensing Data

Abstract

Soil temperature is a key variable in several fields of Earth science, and accurate predictions of soil temperature at different depths are of great significance for scientific research and agricultural production. Soil temperature data observations by meteorological stations are categorized as discrete and discontinuous, and moderate resolution imaging spectroradiometer (MODIS) remotely sensed data are used to perform routine soil temperature predictions on a large scale. In this study, data from three MODIS products, namely, normalized vegetation index, atmospheric precipitable water, and surface temperature, and daily average soil temperature measurements at depths of 40, 100, and 200 cm from the ground surface in Liaoning Province from 2017 to 2021 were used to establish a soil temperature prediction model based on the long short-term memory (LSTM) model. To improve the prediction accuracy and stability, an optimized LSTM model was established to perform comparative predictions of soil temperature concentrations based on the LSTM model, and it considered the hysteresis factor of soil temperature relative to the surface temperature. The LSTM soil temperature prediction models established based on the fusion of remote sensing data (NDVI, PWV, and LST) and soil temperature data at 40, 100, and 200 cm from the surface and the optimized LSTM models that considered hysteresis obtained R2 values of 0.86, 0.81, 0.69, and 0.90, 0.91, 0.88, respectively, with RMSE values of 0.30, 3.41, 3.74 °C, and 0.30, 3.41, 3.74 °C. Moreover, the SDRMSE of the optimized model considering hysteresis decreased compared to that of the LSTM. The LSTM models before and after optimization can achieve long-term daily temperature prediction of inter-annual soil temperature, although the prediction model that considers the hysteresis factor had a better fit and stability. Thus, the model considering hysteresis is more advantageous for obtaining accurate predictions of spatially continuous multidepth soil temperatures.