AgroBench and ARI: A reproducible method for robustness evaluation of agricultural object detectors

Robustness evaluation of agricultural object detectors is often limited to clean-set metrics, which can conceal substantial performance degradation under realistic field perturbations such as haze, motion blur, compression artifacts, chromatic shifts, and canopy-related occlusions. This article presents a reproducible method that combines an agriculture-calibrated corruption benchmark, termed AgroBench, with a normalized robustness summary metric, the Agro-Robustness Index (ARI). The method is designed to quantify how much of a detector's clean performance is retained under a fixed, domain-grounded corruption protocol while preserving per-family diagnostic interpretability. AgroBench specifies explicit corruption families, parameter grids, and global modulation settings to ensure deterministic and comparable evaluation across runs. ARI summarizes corrupted performance as a normalized retention score relative to each detector's clean reference, thereby reducing the influence of baseline clean accuracy on robustness interpretation. The method is demonstrated on agricultural object detection using YOLOv8n and produces both a global robustness score and per-family retention components that support diagnosis and reporting. This framework is intended as a reusable evaluation method for standardized robustness assessment in agricultural vision.