Exploration of Machine Learning Methods from Decision Trees to Long-Short Term Memory Operators to Quality Control Soil Moisture Observations.

Soil moisture monitoring sensors measure the quanity of water filling the empty pore spaces the in the soil matrix. These measurements are often expressed in terms of volumetric water fraction. Since the typical pore size and other relevant hydrological properties (i.e., bulk density, hydraulic conductivity) are highly sensitive to soil type, the range of volumetric soil moisture observations and the response (wet up and dry down rates) to incoming precipitation varies greatly even over short spatial scales. This can make efforts to developing automated quality control methods that work well across diverse climate regions of the U.S. challenging. The purpose of this investigation is to explore the utility of machine learning methods at identifying and flagging anomalous soil moisture behavior. Several machine learning techniques were applied to soil moisture observations from the U.S. Climate Reference Network (USCRN). USCRN is a set of high-quality climate stations that monitor both above ground and below ground conditions. Soil moisture observations are monitored at either 2 (5 and 10 cm) or 5 (5, 10, 20, 50, and 100cm) depth configurations. At each depth, there are three separate sensors monitoring soil moisture conditions redundantly for a total of over 1500 sensors across 113 stations. A goal of this exploration is to identify machine learning techniques that have proved useful in detecting anomalous soil sensors behavior and can be applied within a human-in-the-loop implementation strategy.