Integrated Response Surface Methodology with Gradient Boosting in Modelling Potato Post-Harvest Losses
Erick Kirui *
Murang’a University of Technology, Murang'a, Kenya.
*Author to whom correspondence should be addressed.
Abstract
Post-harvest losses remain a persistent constraint to potato value chains, particularly where cold-chain systems and storage facilities are limited. This study presents a simulation-based demonstration of an integrated hybrid modelling framework that combines Response Surface Methodology (RSM) with Gradient Boosting Machines (GBM) to model potato post-harvest deterioration under different storage conditions. The objective was to assess the methodological feasibility of combining an interpretable polynomial response surface with a machine-learning residual-correction model. A Central Composite Design was developed using six storage-related variables: temperature, relative humidity, storage duration, light exposure, mechanical damage and curing duration. Synthetic observations were generated through Monte Carlo simulation using post-harvest biological parameter assumptions from the literature. Weight loss, sprouting index and rotting loss were standardised and combined through Principal Component Analysis to develop a Composite Loss Index (CLI). A second-order RSM model was first fitted to the CLI to estimate linear, quadratic and interaction effects among the predictors. The unexplained residual variation from the RSM model was then modelled using GBM to capture higher-order nonlinearities and complex interactions not represented by the polynomial surface. In the simulated environment, the hybrid RSM-GBM model achieved better predictive performance than the standalone RSM and GBM models. Storage duration, relative humidity, temperature and variety were the dominant contributors to simulated post-harvest deterioration patterns. The findings indicate that combining interpretable response surface modelling with non-parametric residual learning can improve prediction while retaining useful explanatory structure. However, all results are based on simulated data and should therefore be interpreted as preliminary computational evidence rather than empirical agricultural findings. Further validation using controlled and field-based potato storage observations is required before practical application.
Keywords: Response surface methodology, gradient boosting machines, composite loss index, potato storage, post-harvest loss, central composite design, monte Carlo simulation, principal component analysis, residual correction, predictive modelling