The soil water retention curve (SWRC) represents the relationship between soil water content and matric potential, explaining how soil retains and releases water under various moisture conditions. Information on matric potential is vital for assessing soil's ability to store water for plant growth, and for quantification of its mechanical stability to prevent compaction damage. However, when direct measurements of matric potential are unavailable, and only soil water content data is accessible (e.g., from satellite observations), estimating matric potential becomes particularly challenging and relies on knowledge of SWRC. This complexity is further compounded by the highly variable, site-specific nature of the relationship between soil water content and matric potential, influenced by factors such as soil structure, seasonal fluctuations, and environmental stressors like drought. As a result, conventional methods based on an unambiguous pressure-saturation relationship often fail in capturing the dynamic behavior of soil moisture over time.
In this study, we address these challenges by employing a combined approach involving Seasonal-Trend decomposition using Locally estimated smoothing (STL) and an autoencoder neural network to monitor and predict changes in SWRC. STL is utilized to isolate seasonal patterns and long-term trends in water content data, capturing how environmental factors, especially prolonged drought, impact SWRC across multiple sites in Germany. The autoencoder neural network then compresses this information into a site- and period-specific feature called the "AUV" (Autoencoder Value), which represents the soil's water-holding properties and its response to changing environmental conditions. This AUV value is subsequently used to predict shifts in SWRC by modeling changes in matric potential resulting from significant alterations in the seasonal amplitude of water content or shifts in long-term trends following dry events.
Our approach was tested across several sites, where, in some locations, prolonged drought caused a noticeable reduction in seasonal amplitude and a decrease in trend values, which led to a corresponding decline in the AUV value. This decline in AUV was found to be indicative of reduced water retention capacity and decreased soil resilience. Overall, this method offers a practical solution to (i) predict dynamic changes in matric potential using only soil water content measurements, (ii) monitor shifts in SWRC over time to reflect changes in soil health and (iii) provide a scalable tool for assessing soil resilience to climate variability, making it particularly useful in regions where direct measurements of soil matric potential are not available.