Applying ensemble learning in ecophysiological models to predict spring phenology

Abstract

Accurate and transparent phenological models have become a vital tool for reflecting the feedbacks and interactions between the biosphere and atmosphere and accurately predicting future phenological responses to climate change. With the rapid accumulation of ground-observed phenological data, an increasing number of studies have used process-based ecophysiological (Eco) models to predict future phenological changes. Many algorithms have been used to optimize the parameters of Eco models, but there is a lack of evaluation of different algorithms. Although no single Eco model can show obvious advantages, ensemble learning can improve model performance by combining different trained models. In this study, based on the historical observation data (leaf unfolding date and first flowering date) of more than 100 woody plants from 1962 to 2018 in Heilongjiang Forest Botanical Garden, we evaluated the performance of five Eco model parameter optimization algorithms, and compared the performance of 20 Eco models in phenological observation data prediction. Most importantly, based on the idea of ensemble learning in machine learning, we proposed improving the prediction accuracy of Eco models by applying ensemble learning methods in the trained Eco models. Our research results show that among the five optimization algorithms involved in this study, the generalized simulated annealing algorithm is more recommended for Eco model parameter optimization. Compared with the more complex three-phase and four-phase models, the simpler the model structure, the better the generalization performance of one-phase and two-phase models. The RMSEs of Eco models of many species on the test set were greater than 4 days, which indicates that the ability of Eco models to predict phenological data based on ground observations of some specific species is relatively limited. Our results highlight that the prediction accuracy of Eco models can be significantly improved by the Voting ensemble learning method. In the future, we can improve the accuracy of phenological predictions by using ensemble learning methods, so as to more accurately detect future phenological changes and their responses to climate change.