Hybrid Machine Learning Framework for Analyzing and Optimizing Real-time Soil Moisture Sensor Arrays

Abstract:

Soil moisture sensors have become widely available for ecosystem and water resources research as well as agriculture applications. These relatively low-cost sensors are promising for building wireless sensor networks (WSN) for real-time observations and predictions, including evapotranspiration (ET) estimation based on the subsurface water budget, and predictions of when to water farmlands given weather forecasting. However, there are few methods available to take these real-time soil moisture data into the real-time estimation and prediction framework. In addition, determining how many sensors need to be installed and at which depths is challenging. In this study, we first develop a hybrid machine learning framework to couple a hydrological model and real-time soil moisture data. This framework consists of a 1D vadose-zone flow model and plant root uptake model as well as a Sampling-Importance-Resampling Particle Filter (SIR-PF). This algorithm is a sequential Monte Carlo method that can advantageously handle nonlinear dynamic models and non-Gaussian distributions. In SIR-PF parameter estimation, transpiration rate and other hydraulic parameters (e.g. saturated hydraulic conductivity) are sampled in their feasible range, then soil moisture measurements are used to estimate the posterior probability density distribution of the model states and parameters. In addition, we develop an optimization algorithm to determine the minimum-but-sufficient number of sensors and their optimal positioning depth. We use a Gaussian Process Model with sequential updating. We evaluate a different configuration depending on different target variables such as the ET estimation and soil moisture prediction. We demonstrate our approach using the soil moisture data at the East River watershed in Colorado.