Parameterization and Calibration of Wild Blueberry Machine Learning Models to Predict Fruit-Set in the Northeast China Bog Blueberry Agroecosystem

Abstract

Data deficiency prevents the development of reliable machine learning models for many agroecosystems, especially those characterized by a dearth of knowledge derived from field data. However, other similar agroecosystems with extensive data resources can be of use. We propose a new predictive modeling approach based upon the concept of transfer learning to solve the problem of data deficiency in predicting productivity of agroecosystems, where productivity is a nonlinear function of various interacting biotic and abiotic factors. We describe the process of building metamodels (machine learning models built and trained on simulation data) from simulations built for one agroecosystem (US wild blueberry) as the source domain, where the data resource is abundant. Metamodels are evaluated and the best metamodel representing the system dynamics is selected. The best metamodel is re-parameterized and calibrated to another agroecosystem (Northeast China bog blueberry) as the target domain where field collected data are lacking. Experimental results showed that our metamodel developed for wild blueberry achieved 78% accuracy in fruit-set prediction for bog blueberry. To demonstrate its usefulness, we applied this calibrated metamodel to investigate the response of bog blueberry to various weather conditions. We found that an 8% reduction in fruit-set of bog blueberry is likely to happen if weather becomes warmer and wetter as predicted by climate models. In addition, southern and eastern production regions will suffer more severe fruit-set decline than the other growing regions. Predictions also suggest that increasing commercially available honeybee densities to 18 bees/m2/min, or bumble bee densities to 0.6 bees/m2/min, is a viable way to compensate for the predicted 8% climate induced fruit-set decline in the future.