Embeddings from standardized sorghum leaf images capture variation in disease response that human scoring misses

Human raters scoring plant disease severity using ordinal scales fails to capture the full spectrum of variation in lesion color, size, and number. This inconsistency complicates comparisons across different environments. To overcome this challenge, researchers created a portable imaging chamber capable of capturing high-quality images of thousands of sorghum leaves under consistent lighting, orientation, and backgrounds. The leaves were captured in large numbers across three different states.

The team employed vision encoders to generate embeddings from the leaf images, which were then used to predict the human-assigned disease severity scores. Remarkably, the embeddings were able to accurately capture the disease severity information that was missed by traditional ordinal scoring methods. In a genome-wide association study (GWAS), no significant hits were found using traditional disease severity scores, either human-assigned or derived from vegetation indices.

However, when applied to the embeddings, the GWAS identified twelve genomic hotspots that control leaf appearance, nine of which were linked to variation in disease symptom severity.

Five of these hotspots corresponded to previously known sorghum genes, with all three hotspots not linked to disease and two of the nine that were. Roughly one-third of the embedding-hotspot associations replicated across at least two states, and twenty replicated across all three states. By combining scalable leaf imaging with pretrained image encoders, researchers demonstrated the potential of this approach to capture genetically controlled variation in diverse disease symptoms that traditional human scoring methods often overlook.