Irrigation Management using Machine Learning Regressors of Aggregated Soil Moisture Data

Soil moisture data is highly valuable for irrigation management, however, soil data can often be difficult for farmers to interpret for making informed irrigation decisions. Subsurface drip irrigation targets the root zone of crops. It is commonly used and highly efficient at minimizing evaporative loss. Factors, such as long irrigation lines and hilly terrain, influence the timing and duration of irrigation events, which makes arrival time and duration of crop irrigation water unpredictable even if there is a well-managed schedule. Also, deficit irrigation is a practice where high value crops are intentionally water stressed after the fruiting stage to improve their quality and value. In this study, we propose a new modeling method for predicting soil moisture that addresses the randomness of one of the primary boundary conditions, the irrigation event. Through machine learning regression, we aim to predict near surface soil moisture values in a subsurface drip irrigated crop in a silt loam soil texture. Our model focuses specifically on the dewatering portion of the time series soil moisture data at two depths, the soil textural data, and the evapotranspiration (ET) as the only boundary condition. By predicting future soil moisture values or stress conditions in the absence of irrigation, our model provides valuable insights for farmers making irrigation management decisions. This presentation serves as a feasibility study and reports the results of the first attempt to apply machine learning regressors to time series soil moisture data to predict future near surface soil moisture values. In our experiment, we placed two HydraProbe Soil Sensors in the root zone of a blueberry crop located near Wilsonville Oregon in the United States. Soil moisture was logged every five minutes at depths of 15 and 30 cm. The ET and soil moisture data were aggregated and parameterized into the input features for machine learning regressors. To create a training data set, algorithms were developed to isolate only the dewatering portions of the soil moisture time series data for a single growing season. The machine learning input features include: 1) the sum of ET for a specific duration interval, 2) soil moisture percentage, 3) the sum of the ET for the prior 24 hours, 4) the sum of forward-looking ET and 5) capillary features derived from soil texture pedotransfer functions (PFTs) that are part of the Richards Equation. The predicted near future soil moisture values are the output target of the model. Using the Skikitlearn machine learning regressors, we evaluated random forest, support vector machine, ridge, and LASSO regressions. Each regressor underwent regularization through a grid-search of the hyper-parameters using the training data set. To measure potential overfitting of the models, a 15% holdout was examined using r2 and the RMSE (root mean square error). Additionally, a validation data set was created using seasonal low soil moisture values and time intervals not included in the training set. Among the regressors evaluated, ridge regression performed well with an r2=0.98, RMSE=0.5% on the 15% holdout, and an r2=0.93, RMSE=1.13% on the validation data set.