New AI-biological model predicts cherry flowering with 40% less error

A team of international researchers has developed a new hybrid phenological model to better predict cherry blossom flowering, combining biological knowledge with machine learning. Tested in Japan, South Korea, and Switzerland, the model consistently outperformed both traditional bioclimatic models and deep learning approaches.

Unlike conventional models that often require frequent recalibration, the hybrid design replaces the chilling accumulation component with a multilayer perceptron (MLP) while retaining the thermal forcing module. This allows the system to learn how plants respond to winter temperatures directly from data, while preserving the biological structure of the process.

Using over 9,000 bloom observations paired with hourly temperature data, the model reduced prediction errors by up to 30–40% compared to classical approaches. It also proved robust in data-scarce conditions, such as in South Korea, and could even generalize to cherry varieties and climatic scenarios not included in the training set.

However, researchers noted some differences between the learned responses and expected biological patterns — for instance, the model indicated chilling contributions at temperatures above 12.5 °C, which traditional frameworks do not. Future improvements may include regularization techniques to guide the learning process toward biologically plausible outcomes.

The study demonstrates that hybrid models can bridge the gap between interpretability and predictive power, offering a valuable tool for understanding climate impacts on fruit tree phenology and supporting growers worldwide.

source: frutasdechile.cl

photo: epicgardening.com