Weather Based Strawberry Yield Forecasts at Field Scale Using Statistical and Machine Learning Models

Abstract

Strawberry is a high value and labor-intensive specialty crop in California. The three major fruit production areas on the Central Coast complement each other in producing fruits almost throughout the year. Forecasting strawberry yield with some lead time can help growers plan for required and often limited human resources and aid in making strategic business decisions. The objectives of this paper were to investigate the correlation among various weather parameters related with strawberry yield at the field level and to evaluate yield forecasts using the predictive principal component regression (PPCR) and two machine-learning techniques: (a) a single layer neural network (NN) and (b) generic random forest (RF). The meteorological parameters were a combination of the sensor data measured in the strawberry field, meteorological data obtained from the nearest weather station, and calculated agroclimatic indices such as chill hours. The correlation analysis showed that all of the parameters were significantly correlated with strawberry yield and provided the potential to develop weekly yield forecast models. In general, the machine learning technique showed better skills in predicting strawberry yields when compared to the principal component regression. More specifically, the NN provided the most skills in forecasting strawberry yield. While observations of one growing season are capable of forecasting crop yield with reasonable skills, more efforts are needed to validate this approach in various fields in the region.