Machine learning models for predicting morphological traits and optimizing genotype and planting date in roselle (Hibiscus Sabdariffa L.).

Abstract

Accurate prediction and optimization of morphological traits in Roselle are essential for enhancing crop productivity and adaptability to diverse environments. In the present study, a machine learning framework was developed using Random Forest and Multi-layer Perceptron algorithms to model and predict key morphological traits, branch number, growth period, boll number, and seed number per plant, based on genotype and planting date. The dataset was generated from a field experiment involving ten Roselle genotypes and five planting dates. Both RF and MLP exhibited robust predictive capabilities; however, RF (R² = 0.84) demonstrated superior performance compared to MLP (R² = 0.80), underscoring its efficacy in capturing the nonlinear genotype-by-environment interactions. Permutation-based feature importance analysis further revealed that planting date had a more significant impact on trait variation than genotype. To identify optimal combinations of genotype and planting date for maximizing morphological traits, the RF model was integrated with the Non-dominated Sorting Genetic Algorithm II (NSGA-II). According to the RF-NSGA-II optimization results, the optimal values, including 26 branches per plant, a growth period of 176 days, 116 bolls per plant, and 1517 seed numbers per plant, were achieved with the Qaleganj genotype planted on May 5. Collectively, these findings highlight the potential of integrating machine learning and evolutionary optimization algorithms as powerful computational tools for crop improvement and agronomic decision-making.

Plant phenotyping relevance

形態形質を予測する機械学習モデルを開発し、RFとMLPの性能比較・検証を行っているため、計算的な形質推定が研究の中心である。

Code and data availability

The paper's roselle phenotype dataset (10 genotypes × 5 planting dates, morphological traits) and analysis are not publicly deposited; the Data availability statement says datasets are available from the corresponding author on request. No public code, models, or data URLs are provided.

No evidence-backed public reproduction asset is currently recorded.