Explaining Parameter-Free and Global Attention in Custom YOLOv11 Architectures for Plant Phenotyping

To ensure robust, safe awareness of the surroundings in robotic applications, improve generalization across different varieties of the same crop species, and achieve trustworthiness among non-technical users and stakeholders, explaining the predictions of detection models in plant phenotyping scenarios is becoming crucial. Moreover, as attention mechanisms are being embedded in lightweight detection models like You Only Look Once (YOLO), interpreting their predictions is becoming even more complicated than before. We use the Explainable AI (xAI) framework, Eigen-CAM, to uncover the motifs behind the predictions of custom YOLOv11 architectures, in which Global Attention Mechanism (GAM) and Simple and parameter-free Attention Module (SimAM) blocks were added at key points in the feature fusion. After training both the baseline and custom models on a lettuce detection dataset, localization performance across all configurations was compared, observing nonuniform gains between low and high densities. For example, in small models, the introduction of an attention module at the end of the backbone resulted in a precision variation of +1.81% with SimAM and +1.18% with GAM. Then, Eigen-CAM was used to analyze the models’ predictions on a separate, more challenging dataset where lettuce instances are heavily occluded by weeds, and observe changes in behavior with the custom configurations: they become more focused on the vegetables and less on the background noise compared to the baseline.