Combining imaging techniques with nonparametric modelling to predict seepage hotspots in irrigation channels

Using the Murrumbidgee Irrigation Area, Australia as a case study, we present an integrated approach for identifying seepage hotspots and predicting seepage losses from open channel. The approach is particularly important to facilitate investments for improving irrigation conveyance efficiencies, thus enabling sustainable agricultural water use. A qualitative assessment is used for capturing seepage hotspots with electromagnetic inductance (EM31) imaging techniques, followed by actual seepage measurements. Based on data from major irrigation systems in the southern Murrumbidgee Irrigation Area, a case is made for cost-effective methodology to locate seepage hotspots and quantify seepage losses in channels. In particular, a predictive model was developed based on EM31 survey and direct measured channel seepage data. The main input data for the model were EM values, soil types, water depth in channels, wetted perimeter of channels and whether water is flowing in channels. The output from the model was a seepage loss value in channels. The three different modelling techniques considered were the Generalised Linear Mixed (GLM) model, Random Forest (RF) model and Generalized Boosted Regression Model (GBM), and a best performing model for seepage prediction was identified. The RF model was found to the most reliable, explaining 60% of the variability in the data and with the least mean absolute error. The study indicated that the RF model can be used to locate seepage hotspots in channels and determine the magnitude of seepage losses.