A spatial machine-learning model for predicting crop water stress index for precision irrigation of vineyards

Optimization of water inputs is possible through precision irrigation based on prescription maps. The crop water stress index (CWSI) is an indicator of spatial and dynamic changes in plant water status that can serve irrigation management decision-making. The driving hypothesis was that in-season CWSI maps based on combined static and spatial-dynamic variables could be used to delineate irrigation MZs. A primary incentive was to minimize thermal-imaging campaigns and to complement CWSI maps between campaigns with cost-effective multi-spectral imaging campaigns producing normalized difference vegetative index (NDVI) maps. A spatial machine-learning model based on a random-forest (RF) algorithm combined with spatial statistical methods was developed to predict the spatial and temporal variability in CWSI of single vines in a vineyard. Model criteria and objectives included the reduction of sample data and input variables to a minimum without impacting prediction accuracy, consideration of only variables readily available to farmers, and accounting for spatial location and spatial processes.

The model was developed and tested on data from a ‘Cabernet Sauvignon’ vineyard in Israel over two years. Prediction of CWSI was driven by terrain parameters, slope, aspect and topographical wetness index, soil apparent electrical conductivity (ECa), and NDVI.

Spatial models based on RF were found to support CWSI prediction. Adding a geospatial component significantly improved model performance and accuracy, particularly when raw data was represented as z-scores or when z-scores were used as weights. NDVI, followed by ECa, aspect, or slope, was the most important variable predicting CWSI in the non-spatial models. The stronger the variable importance of NDVI, the better the model performed. The weaker the effect of NDVI in predicting CWSI, the stronger the effect of terrain and soil variables. In the spatial models, based on z-transformed values or on weighted values, the most important variable in predicting CWSI was either NDVI or location.

The model, based on a limited and readily accessible number of variables, can serve as the basis for user-friendly decision support tools for precision irrigation. Additional research is needed to evaluate alternative prediction variables and to account for case studies in more geographical locations to address overfitting specific input data. Socio-economic and cost-benefit considerations should be integrated to examine whether precision irrigation management based on such models has the desired effects on water consumption and yield.