Fusion and Feature Selection of Frozen Foundation Models for In-the-Wild Plant Disease Recognition

Image-based recognition can help growers diagnose plant diseases early. Classifiers trained on laboratory leaf images, however, lose accuracy on field photographs. Frozen foundation models provide features for field photographs, but fusing several of them increases the feature dimension. In this study, we concatenated the features of ten frozen backbones from five pretraining families into a 9472-dimensional vector. We reduced this vector to the size of one backbone by feature selection or projection and classified it with a linear probe. The main results on three field datasets are means over 30 repetitions. On the official PlantWild test split, early fusion reached 81.4% accuracy and late fusion 81.7%, against 76.7% for the best single backbone. To our knowledge, these are the highest accuracies reported on this split. L1-penalized feature selection to 1536 dimensions reached 80.6%, compared with 80.2% for the previous best method at 2816 dimensions. On PlantWild, L1 selection took its largest shares from the backbones that contributed most, whereas relevance filters concentrated on two backbones that contributed less. Removing the two backbones that L1 selection used least lowered the macro-averaged F1 score by at most 0.002 on any dataset and left eight backbones at inference.