Regional estimation of garlic yield using crop, satellite and climate data in Mexico

Abstract Garlic (Allivium sativum L.) has been consumed by humans since the beginning of recorded history. Despite its culinary and economical importance, little efforts are reported in the literature to forecast garlic yield. In this study, we developed the first attempt to predict garlic yield at large scale using crop, satellite (TERRA-MODIS) and climate data (ERA-5). The study area was located in Mexico (2004 – 2018). We compared the predictive capacity of three Machine Learning (ML) methods: generelised linear model (glm), support vector machine radial (svmR) and random forest (rf); as well as different combinations of predictors. The best performance was obtained by the svmR model (R2 = 0.68 and %RMSE = 20.08%) which combined YieldBaseLine, NDVI and climate data under Feature Selection scenario (FS) = 0.90. The rf method performed similarly well (R2 = 0.66), while glm achieved lower performance (R2 = 0.61). To test the predictive capacity of the models in practice, we additionally used a leave-one-year-out method in a time-wise manner. Models gradually improved over the years as more data was incorporated for training. The svmR and rf methods presented an overall performance of R2 ~0.60 (%RMSE ~20.5%), reaching the highest accuracy (R2 = 0.78, %RMSE ~16.6%) in 2014. Both approaches reveal the potentiality of the proposed models to account for the inter-annual and spatial garlic yield variability in Mexico. This procedure can be adjusted and used for other crop types or locations.