On-Device Continual Learning for Identifying Invasive Plants of Biosecurity Concern to the United States from Smartphone Imagery Captured in Vietnam and Australia

A field tool for recognising invasive plants must work offline, learn local species from its user’s26 own photos, and run on the phones people have. We evaluate on-device continual learning27 with the Med Herb Lens training engine, which runs the same Kotlin code on iOS, Android28 and the web: a frozen pretrained backbone embeds each photo, and only a small classifier29 is learnt. We release 7,269 geotagged smartphone photos of 21 invasive plant species, each30 photographed with an iPhone and an Android phone in Vietnam and Australia. Models trained31 on one phone’s photos were tested on the other’s. Across 4,860 runs with three backbones,32 three scenarios, nine strategies and ten seeds per platform, the backbone mattered most: with33 a trained classifier, a frozen, 8-bit DINOv2 vision transformer reached about 96% accuracy,34 17 to 25 points above MobileNetV3, matching fully fine-tuned networks. Streaming linear35 discriminant analysis, which calculates the classifier instead of training it, did better on every36 platform: 97 to 98% with DINOv2, well calibrated. Replaying 50 stored embeddings per class37 came within 0.4 points of retraining with DINOv2, while fine-tuning without stored data forgot38 almost everything. Platforms agreed within about 2 points on average; embedding a photo took39 129 to 935 milliseconds. A random split of one phone’s photos inflated the smaller backbones’40 accuracy by 7.5 to 15 points. The contribution to artificial intelligence, this evaluation and its41 cross-device test, applies to any application adding classes on the device from a frozen backbone;42 ours identifies invasive plants.