Classifying Glyphosate Resistance in Lolium rigidum Using Smartphone RGB Imaging and Deep Learning

Glyphosate-resistant annual ryegrass (Lolium rigidum) is a major problem in Australian grain production. Early detection of resistant populations is important for effective management, yet current bioassay-based detection takes 21–28 days. Here, this proof-of-concept study tested whether consumer-grade smartphone imaging, combined with machine learning and deep learning, can compress that timeline. One confirmed glyphosate-resistant (GR) and one susceptible (GS) L. rigidum population were sprayed with glyphosate at 540 g ae ha−1. RGB images were captured with a Samsung Galaxy S21 at 0, 3 and 9 days post-application (DPA) and split at the plant level using stratified 5-fold cross-validation. A pretrained EfficientNet-B0, fine-tuned via transfer learning, achieved 95.45% binary accuracy at 3 DPA, outperforming a custom CNN (90.91%) and Random Forest (81.82%). By 9 DPA, EfficientNet-B0 reached 100%. For four-class classification (biotype × treatment), EfficientNet-B0 scored 60.98% at 3 DPA and 94.44% at 9 DPA, exceeding both CNN and RF. Baseline accuracy of 54.55–72.73% before spraying points to intrinsic population differences; the post-glyphosate gains suggest the herbicide sharpens the contrast between biotypes. Because only one population pair was tested, these results should be treated as proof-of-concept; validation across multiple independent populations in the field is the obvious next step.