GreenMind: An Explainable Plant Identification and Disease Screening System, with a Temporal and Lab-to-Field Evaluation

Plant identification and plant disease applications are usually evaluated on test images drawn from the same collection as their training images, which says little about how they perform on new photographs taken by users. We built GreenMind, an open-source web application that combines a local species classifier with the Plant.id and Pl@ntNet services, screens leaves for 15 disease classes, and explains each prediction with a class-activation map. We then evaluate its models under two kinds of distribution shift. For species identification we rebuild the data from research-grade, Creative-Commons-licensed iNaturalist observations and split them by upload date around the deployment of the application. Retrained EfficientNet-B0 and ResNet-18 models reach 76.0% and 64.9% top-1 accuracy on 958 species in an in-distribution test split, and 69.0% and 58.2% on photographs uploaded after deployment, while the originally deployed classifier reaches only 22.0%. For disease screening, classifiers trained on PlantVillage reach 99.7% to 99.9% accuracy on a held-out PlantVillage split but only 28% to 29% (mean of three seeds) on matching PlantDoc field photographs, with confidence well above their accuracy. Adding 842 PlantDoc field photographs to training raises field accuracy from 29.4% to 49.0% without loss on PlantVillage, classifying leaf crops from a detector trained on PlantDoc boxes raises it further to 52.0%, and temperature scaling fitted on field images reduces the expected calibration error from 0.27 to 0.07, whereas fitting it on lab images increases it to 0.61. All code, notebooks and per-image predictions are released so that the results can be reproduced on free cloud GPUs.