Prediction of hysteretic matric potential dynamics using artificial intelligence: application of autoencoder neural networks

Abstract

Information on soil water potential is essential to assessing the soil moisture state, to prevent soil compaction in weak soils, and to optimize crop management. When there is a lack of direct measurements, the soil water potential values must be deduced from soil water content dynamics that can be monitored at the plot scale or obtained at a larger scale from remote sensing information. Because the relationship between water content and soil water potential in natural field soils is highly ambiguous, the prediction of soil water potential from water content data is a big challenge. The hysteretic relationship observed in nine soil profiles in the region of Solothurn (Switzerland) is not a simple function of texture or wetting-drainage cycles but depends on seasonal patterns that may be related to soil structural dynamics. Because the physical mechanisms governing seasonal hysteresis are unclear, we developed a deep neural network model that predicts water potential changes using rainfall, potential evapotranspiration, and water content time series as inputs. To adapt the model for multiple locations, we incorporated a deep autoencoder neural network as a classifier. The autoencoder compresses the water content time series into a site-specific feature that is highly representative of the underlying water content dynamics of each site and quantifies the similarity of dynamic patterns. By adding the autoencoder's output as an additional input and training the neural network model with three stations located in three major classes established by the autoencoder, we predict matric potential for other sites. This method has the potential to deduce the dynamics of matric potential from water content data (including satellite data) despite strong seasonal effects that cannot be captured by standard methods.