Image-Based Identification of Crop and Weed Species at the Seedling Stage: Deep Learning Classifiers Rely on Acquisition Context Beyond Plant Morphology
TL;DR
Reported accuracy partly measures acquisition context, not morphology; morphological grounding should be measured and reported alongside accuracy.
Abstract
Distinguishing crop from weed species at the seedling stage is a fine-grained morphological discrimination problem. We study it on the public Plant Seedlings benchmark, which contains twelve species (three crops and nine weeds) imaged at the seedling stage. Automated classifiers report near-ceiling accuracy on this benchmark. Yet its images are acquired in controlled trays containing soil, gravel, rulers, barcodes, and printed labels, so a model may identify a species from its acquisition context rather than its morphology. We audit two architectures, EfficientNet-B7 and ViT-B/16, trained on the V2 dataset (5539 images) and probed with plant-only, background-only, and background-swapped inputs. Near-ceiling models (96.1% and 96.4% over three seeds) recover the correct species for up to 47.7% of samples from the background alone (chance 8.3%) and lose about 60 points under background swapping. Context reliance is therefore a property of the benchmark, not any single architecture. Tracing this to its source, an independent learned representation of the plant-free background alone identifies the species at 72.4%. The reliance is correctable end to end: a consistency-regularisation scheme retains 92.1% full-image accuracy for the transformer with no segmentation at inference, at an architecture-dependent cost. Reported accuracy thus partly measures acquisition context, not morphology; morphological grounding should be measured and reported alongside accuracy.