Soil Moisture Prediction Using Meteorological Data, Satellite Imagery, and Machine Learning in the Red River Valley of the North

Weather stations provide key information related to soil moisture and have been used by farmers to decide various field operations. We first evaluated the discrepancies in soil moisture between a weather stations and nearby field; due to soil texture, crop residue cover, crop type, growth stage and duration of temporal dependency to recent rainfall and evaporation rates using regression analysis. The regression analysis showed strong relationship between soil moisture at the weather station and the nearby field at the late vegetative and early reproductive stages. The correlation thereafter declines at later growth stages for corn and wheat. We can adduce that the regression coefficient of soil moisture with four-day cumulative rainfall slightly increased with an increase in the crop residue resulting in a low root mean square error (RMSE) value. We then investigated the effectiveness of machine learning techniques such as random forest regression (RFR), boosted regression trees (BRT), support vector regression, and artificial neural network to predict soil moisture in nearby fields based on RMSE of a 30% validation dataset and to determine the relative importance of predictor variables. The RFR and BRT performed best over other machine learning algorithms based on the lower RMSE values of 0.045 and 0.048 m3 m-3, respectively. The Classification and Regression Trees (CART), RFR and BRT models showed soil moisture at nearby weather stations had the highest relative influence for moisture prediction, followed by the four-day cumulative rainfall and Potential Evapotranspiration (PET), and subsequently followed by bulk density and Saturated Hydraulic Conductivity (Ksat). We then evaluated the integration of weather station data, RFR machine learning, and remotely sensed satellite imagery to predict soil moisture in nearby fields. Soil moisture predicted with an RFR algorithm using OPtical TRApezoidal Model (OPTRAM) moisture values, rainfall, standardized precipitation index (SPI) and percent clay showed high goodness of fit (r2=0.69) and low RMSE (0.053 m3 m-3). This research shows that the integration of weather station data, machine learning, and remote sensing tools can be used to effectively predict soil moisture in the Red River Valley of the North among a large diversity of cropping systems