Non-parametric Crop Yield Forecasting, a Didactic Case Study for Zimbabwe

Operational crop yield forecasting is mostly achieved with empirical statistical regression equations relating regional yield with predictor variables, termed “factors”. Regional yield (the “dependent variable”) refers to average yield over districts, provinces or, more rarely, whole countries; they are provided by national statistical services. The factors can be any combination of raw environmental variables such as weather variables or indices, satellite indices such as Normalised Difference Vegetation Indices (NDVI), farm inputs (fertiliser use) or outputs from simulation models, for instance water transpired over a given phenological phase, maximum leaf area index (LAI), average soil moisture, etc. The approach above is termed “parametric” for two reasons: (1) it derives or requires a number of parameters, for instance regression coefficients and the parameters characterise crop simulation models and (2) it attempts to identify the factors that condition yields and to understand their action. The difference between “parametric” and “non-parametric” methods is not clear-cut; it is mostly operational. Parametric forecasting approaches derive a “model” (through a process known as “calibration”) based on historical yield and climatic data. The model is subsequently applied to current crops and within season data to issue a forecast of yields. A number of calculations are performed; they are basically the same in the calibration and in the forecasting phases. Non-parametric crop yield forecasting techniques attempt to establish a typology (qualitative description) of the environmental conditions that occur during the growing season, assuming that similar types of seasons lead to similar yields. Similar years are grouped in classes. During the calibration phase, the types of seasons are defined in such a way as to minimize the variability of yields within classes and maximise between-classes variance. The forecast proper is done by categorizing the current year into one of the classes, and by assigning the class yield to the current forecast. Depending on the actual method, the forecast itself may require little more than comparing some variables with reference values, e.g. a threshold. This paper offers a rough comparison of simple yet classical parametric approaches with two different non-parametric methods, applied to national maize yields in Zimbabwe. The conclusion suggests that the simple non-parametric approaches are not inferior, in terms of accuracy and ease of use to the more complex parametric models