UC Davis researchers use synthetic data to fix AI crop detection drops

A team led by Hamid Kamangir at UC Davis quantified how genotype-by-environment (GxE) shifts degrade AI detection of cowpea flowers and pods, then tested whether synthetic data can overcome the gap. Their paper, posted on arXiv, shows that flower detection mAP@50 fell from 76.3% to as low as 50.6% under unseen conditions across two California locations and two growing seasons, with pod detection even more sensitive. Diagnostics confirmed these losses track measurable distributional shifts in feature space and image quality, making the case that real-data annotation alone is impractical for every GxE combination.

To replace that annotation burden, the team rendered procedural 3D cowpea models into synthetic imagery. Synthetic supervision alone improved over pretraining but hit a domain gap driven by camera image formation rather than scene content. They then applied a domain-gap-aware camera-realism augmentation, optimized against real-image statistics via Wasserstein distance, and found that a linear HDR representation converted a smaller measured gap into a larger detection gain than 8-bit RGB. Critically, optimized HDR synthetic data combined with as few as five real images matched or exceeded the real-data baseline for spatial generalization, with pod detection benefiting most at the lowest shot counts. The results show synthetic data can work—but only when the domain gap is explicitly measured and optimized.