Estimating soil water and salt contents from field measurements with time domain reflectometry using machine learning algorithms

Soil water and salt contents are key soil physical parameters that play a crucial role in soil-related hydrological, ecological, environmental, and agricultural processes. Time domain reflectometry (TDR) is commonly used to measure in-situ soil water and salt contents, and provide possible solutions to quickly obtain soil bulk density (BD). However, the measurement accuracy is greatly influenced by the interaction of soil water and salt contents on the measured soil dielectric constant and electrical conductivity, especially for salinized soils. To accurately estimate the soil gravimetric (GWC) and volumetric (VWC) water contents, soil salt content (TS), and BD based on the TDR measurements, we designed different model input schemes to quantify the effect of different soil factors, and applied eight machine learning algorithms to map the non-linear relationship between model inputs and each target soil property. Results of a case study in Hetao Irrigation District in Northwest China indicated that soil particle-size fractions (psfs) are important inputs to predict all the above soil properties. Furthermore, BD mainly contributes to the prediction of soil GWC, and soil surface temperature (T) is effective in improving the GWC and TS estimations. Among eight machine learning algorithms used, extreme gradient boosting (XGB) and gradient boosting regression tree (GBRT) showed good robustness and strong learning capacity. It is rec-ommended to apply XGB to precisely estimate GWC and BD, which resulted in the coefficients of determination (R2) of 0.80 and 0.69, respectively. On the other hand, GBRT precisely estimated the VWC and TS with R2 of 0.71 and 0.84, respectively. The evaluation of spatial distribution characteristic indicated that it is reliable to obtain the spatial distributions of the above soil properties from the TDR measurements based on the recommended model input schemes and machine learning algorithms.

Automated summary

Soil water and salt contents are important soil physical parameters that have significant impacts on hydrological, ecological, environmental, and agricultural processes. Time domain reflectometry (TDR) is commonly used to measure in-situ soil water and salt contents, and provide possible solutions to quickly obtain soil bulk density (BD). This study designed different model input schemes and applied machine learning algorithms to accurately estimate various soil properties based on TDR measurements. The results showed the importance of soil particle-size fractions (psfs) in predicting soil properties, and extreme gradient boosting (XGB) and gradient boosting regression tree (GBRT) algorithms demonstrated good robustness and strong learning capacity.