Augmented Grape Leaf Disease Dataset for Realistic Vineyard Environment using Background Blending

Plant disease detection models achieve >99% accuracy on datasets with near-perfect conditions but fail catastrophically in real agricultural environments due to spurious correlations with homogeneous backgrounds. We introduce an augmented grape disease dataset that incorporates realistic vineyard environments, doubling PlantVillage to 18,054 images through principled background integration and geometric transformations. Models trained on laboratory-only data suffer severe degradation (62-87% accuracy) on realistic conditions, while our augmented dataset enables robust performance (>97%) across both scenarios. This work challenges current agricultural AI benchmarking and enables deployment-ready disease detection systems. Dataset is publicly available on Huggingface.<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">1</sup>