Machine Learning-Driven Construction of High-Yielding Cucumber Plant Architectures in Greenhouse Environments

In the context of declining arable land, the development of plant architectures that maximise the use of finite resources is crucial for addressing food security. This study collected yield data, along with aboveground and root traits, from 263 cucumber varieties. Machine learning models and scenario simulations were utilised with the goal of identifying a high-yielding cucumber architecture suitable for greenhouse cultivation. Our findings indicate that cucumber yields can be predicted using aboveground and root phenotypes, such as the position of the first female flower node, leaf width, stem diameter, and root angle, with the combination of GBDT and SVM algorithms yielding the most accurate results (R² = 0.6155, RMSE = 0.2601). Analysis of 157 464 phenotypic combinations revealed antagonistic interactions between robust aboveground structures and fine root systems, and synergistic interactions between slender aboveground parts and broad root systems. Yields were up to 20% higher in phenotypes that combined a compact, robust aboveground structure with a narrow yet larger-diameter and shallower root system, reflecting additive effects rather than synergistic ones. Additionally, this study proposes a reference range for high-yielding phenotypes. Overall, this research provides a theoretical foundation for optimising cucumber plant structures under greenhouse environments by predicting yields and investigating phenotypic interactions through modelling.