Machine Learning Regression Model for Predicting Honey Harvests

Abstract

Honey yield from apiary sites varies significantly between years. This affects the beekeeper’s ability to manage hive health, as well as honey production. This also has implications for ecosystem services, such as forage availability for nectarivores or seed sets. This study investigates whether machine learning methods can develop predictive harvest models of a key nectar source for honeybees, Corymbia calophylla (marri) trees from South West Australia, using data from weather stations and remotely sensed datasets. Honey harvest data, weather and vegetation-related datasets from satellite sensors were input features for machine learning algorithms. Regression trees were able to predict the marri honey harvested per hive to a Mean Average Error (MAE) of 10.3 kg. Reducing input features based on their relative model importance achieved a MAE of 11.7 kg using the November temperature as the sole input feature, two months before marri trees typically start to produce nectar. Combining weather and satellite data and machine learning has delivered a model that quantitatively predicts harvest potential per hive. This can be used by beekeepers to adaptively manage their apiary. This approach may be readily applied to other regions or forage species, or used for the assessment of some ecosystem services.