Development and Implementation of a Hybrid Non-dominated Sorting Genetic Algorithm (H-NSGA-II-III) for Greenhouse Pesticide Routing on Embedded Platforms

Abstract

Greenhouse pesticide application requires route optimization that balances operational efficiency with sapling-level infection severity. Classical formulations of the Capacitated Vehicle Routing Problem (CVRP) construct routes subject to vehicle capacity constraints. In agricultural applications, such as spraying robots with limited tank capacity, these formulations do not account for infection severity, which is critical for timely treatment of highly affected saplings. This limitation necessitates a multi-objective optimization (MOO) approach to capture competing routing and treatment priorities. However, achieving high-quality Pareto fronts with algorithms such as the Non-dominated Sorting Genetic Algorithm II (NSGA-II) and NSGA-III often requires long runtimes, limiting their practicality for real-time or embedded agricultural systems. To address these challenges, this study proposes H-NSGA-II-III, a hybrid algorithm integrating the diversity-preserving mechanism of NSGA-II with the reference-point selection strategy of NSGA-III. The approach extends the classical CVRP into a bi-objective formulation that minimizes total travel distance while incorporating sapling-level infection severity as a second objective. The algorithm was benchmarked through multiple independent runs on a standard computing platform (SCP) and a lightweight embedded platform (LEP). Results show that H-NSGA-II-III achieves 11-13% higher hypervolume than NSGA-II and 20-21% higher than NSGA-III, while reducing runtime by 9-32% on the SCP and 10-17% on the LEP. These findings indicate that H-NSGA-II-III provides an efficient and lightweight framework for infection-aware routing, feasible for embedded deployment in greenhouse robots.