Hybrid machine learning framework for simulating biological yield in food systems using environmental and agronomic determinants

Résumé fourni par la source

Research Hypothesis This study tests the hypothesis that a hybrid machine‑learning framework integrating enhanced evolutionary feature selection (Bees Royalty Offspring Algorithm, BROA) with Gradient Boosting Regression (GBR) can accurately predict wild blueberry (Vaccinium angustifolium) yield using ecological, climatic, and agronomic determinants. A secondary hypothesis is that the feature selection process can identify the most influential yield‑driving factors, providing both high predictive accuracy and biological interpretability for precision agriculture. Data Description The dataset used in this study was generated by the Wild Blueberry Pollination Simulation Model – a spatially explicit, individual‑based model validated against 30 years of field observations from Maine, USA, and the Canadian Maritimes (Qu & Drummond, 2018). The simulation data are publicly available in the Mendeley database (DOI: 10.17632/p5hvjzsvn8.1; Qu et al., 2020). Dataset structure: The final dataset consists of 777 records, each representing the average yield (kg/ha) over 100 simulation runs for a unique combination of input parameters. Each record contains 16 predictor variables and the target variable (yield). The predictors include: Agronomic factors, Biological factors, Climatic factors. Code Description The deposited code implements the full hybrid machine‑learning pipeline: Optimized BROA feature selection: An enhanced version of the Bee Royalty Offspring Algorithm (Jamshidnezhad & Nordin, 2013) with six key modifications: (a) simplified Euclidean distance‑based mating mechanism (mating_radius = 0.3), (b) elimination of rigid algorithmic phases, (c) dynamic local search with adaptive step‑size reduction, (d) cross‑validated Mean Squared Error (MSE) fitness evaluation using Gradient Boosting Regressor, (e) simplified population management (combining current and offspring populations), and (f) improved parameterisation with configurable hyperparameters (n_royalty = 5, n_workers = 20, offspring_size = 10, max_iter = 50, feature_threshold = 0.6, local_search_iter = 5, mutation_rate = 0.1). Predictive model training: Five machine learning models are trained on the selected features: (GBR), (RFR), (SVR), (KNN), and (ANNs) with three architectures. Robustness assessment: The entire pipeline is repeated across 10 different random seeds (42, 123, 456, 789, 1024, 111, 222, 333, 444, 555) to evaluate stability. Software environment: All code was implemented in Python 3.8.8 using Spyder 4.2.5 (Anaconda 3) with the following library versions: scikit‑learn 0.24.1, numpy 1.20.1, pandas 1.2.4, matplotlib 3.3.4, seaborn 0.11.1, shap 0.41.0, openpyxl 3.0.9, and xgboost 2.0.0 (for comparison only). All random seeds are fixed to ensure full reproducibility. AI-assisted tools were used to assist in coding Python scripts.