{"status":"ok","message-type":"work","message-version":"1.0.0","message":{"indexed":{"date-parts":[[2026,10,22]],"date-time":"2026-10-22T15:35:03Z","timestamp":1792683303764,"version":"4.5.0"},"reference-count":27,"publisher":"Frontiers Media SA","license":[{"start":{"date-parts":[[2026,8,11]],"date-time":"2026-08-11T00:00:00Z","timestamp":1786406400000},"content-version":"vor","delay-in-days":0,"URL":"https:\/\/creativecommons.org\/licenses\/by\/4.0\/"}],"content-domain":{"domain":["frontiersin.org"],"crossmark-restriction":true},"short-container-title":["Front. Plant Sci."],"abstract":"<jats:p>\n                    Automated plant disease recognition from field photographs remains challenging: models trained on laboratory datasets collapse on in-the-wild images, and the current state of the art on PlantWild\u2014the largest open in-the-wild benchmark (18,542 images, 89 classes)\u2014relies on text prototypes derived from a language model to reach 76.18% top 1 accuracy. We ask whether text supervision is truly necessary or whether the bottleneck is the\n                    <jats:italic>diversity<\/jats:italic>\n                    of the visual representation. Our central methodological finding is that\n                    <jats:italic>backbone selection and fusion matter far more than classifier-head engineering<\/jats:italic>\n                    : across 118 experiments, a simple fixed-weight linear\u2013prototype combination on top of three complementary frozen backbones yields larger and more reproducible gains than any head-level adaptive routing mechanism we test. Specifically, through a systematic study of six frozen vision backbones (three CLIP, two DINOv3, and one DINOv2), three classifier heads, and four fusion configurations, completed in a single day on one consumer GPU, we establish three findings. (i) A single self-supervised backbone (DINOv2 ViT-L\/14) already surpasses text-augmented MVPDR (77.56% vs. 76.18%). (ii) Concatenating three complementary backbones (DINOv2 + DINOv3 + CLIP) reaches 80.23% \u00b1 0.41% (five seeds), exceeding the\n                    <jats:italic>published<\/jats:italic>\n                    MVPDR accuracy by +4.05 points and our\n                    <jats:italic>own reproduction<\/jats:italic>\n                    of MVPDR under an identical evaluation protocol by +7.96 points, without any language supervision. (The difference between the two deltas reflects evaluation\u2013protocol differences\u2014our split, model-selection criterion, and training schedule\u2014rather than any discrepancy in the reported numbers; see Section 4.6.6 for a full reconciliation.) (iii) Every form of learned routing we test\u2014per-class gating, backbone gating, sample-wise gating\u2014is inessential; the gain is entirely attributable to backbone diversity and a simple linear\u2013prototype scoring combination. On the smaller PlantDoc benchmark, the same principle transfers but with substantially higher seed variance: the best configuration reaches 80.09% at a favourable seed but 76.97% \u00b1 1.48% over five seeds\u2014a suggestive rather than robust gain. Beyond the accuracy headline, we provide a pathology-aware per-class analysis showing that DINOv2\/v3 dominate on fine-texture lesion classes (rusts, mildews, and leaf spots) whilst CLIP\u2019s narrow advantage concentrates on organ\/species-level identification (rice leaf and potato late blight). All primary claims are validated over five seeds, and all code, feature caches, and result files are released for full reproducibility.\n                  <\/jats:p>","DOI":"10.3389\/fpls.2026.1860665","type":"journal-article","created":{"date-parts":[[2026,8,11]],"date-time":"2026-08-11T05:37:02Z","timestamp":1786426622000},"update-policy":"https:\/\/doi.org\/10.3389\/crossmark-policy","source":"Crossref","is-referenced-by-count":1,"title":["Backbone diversity beats text supervision: a systematic study of frozen multi-foundation model fusion for in-the-wild plant disease recognition"],"prefix":"10.3389","volume":"17","author":[{"given":"Hajar","family":"El Karch","sequence":"first","affiliation":[{"name":"Laboratory of Advanced Systems Engineering, National School of Applied Engineering, Ibn Tofail University","place":["Kenitra, Morocco"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Youssef","family":"Natij","sequence":"additional","affiliation":[{"name":"Scientific Research and Innovation Laboratory, Higher School of Technology, Ibn Tofail University","place":["Kenitra, Morocco"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Mohamed","family":"Benaly","sequence":"additional","affiliation":[{"name":"Innovative Systems Engineering Laboratory, National School of Applied Sciences of Tetouan, Abdelmalek Essaadi University","place":["Tetouan, Morocco"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Rachid","family":"El Gouri","sequence":"additional","affiliation":[{"name":"Laboratory of Advanced Systems Engineering, National School of Applied Engineering, Ibn Tofail University","place":["Kenitra, Morocco"]}],"role":[{"vocabulary":"crossref","role":"author"}]},{"given":"Abdelkader","family":"Mezouari","sequence":"additional","affiliation":[{"name":"Scientific Research and Innovation Laboratory, Higher School of Technology, Ibn Tofail University","place":["Kenitra, Morocco"]}],"role":[{"vocabulary":"crossref","role":"author"}]}],"member":"1965","published-online":{"date-parts":[[2026,8,11]]},"reference":[{"key":"B1","doi-asserted-by":"publisher","first-page":"46","DOI":"10.1016\/j.compag.2018.08.013","article-title":"Impact of dataset size and variety on the effectiveness of deep learning and transfer learning for plant disease classification","volume":"153","author":"Barbedo","year":"2018","journal-title":"Comput. Electron. Agric."},{"key":"B2","doi-asserted-by":"crossref","DOI":"10.1109\/ICCV48922.2021.00951","article-title":"Emerging properties in self-supervised vision transformers","author":"Caron","year":"2021"},{"key":"B3","doi-asserted-by":"publisher","first-page":"311","DOI":"10.1016\/j.compag.2018.01.009","article-title":"Deep learning models for plant disease detection and diagnosis","volume":"145","author":"Ferentinos","year":"2018","journal-title":"Comput. Electron. Agric."},{"key":"B4","doi-asserted-by":"publisher","DOI":"10.1007\/s11263-023-01891-x","article-title":"CLIP-Adapter: Better vision-language models with feature adapters","author":"Gao","year":"2023","journal-title":"Int. J. Comput. Vision (IJCV)"},{"key":"B5","doi-asserted-by":"crossref","DOI":"10.1109\/CVPR52688.2022.01553","article-title":"Masked autoencoders are scalable vision learners","author":"He","year":"2022"},{"key":"B6","article-title":"LoRA: Low-rank adaptation of large language models","author":"Hu","year":"2022"},{"key":"B7","article-title":"An open access repository of images on plant health to enable the development of mobile disease diagnostics","author":"Hughes","year":"2015","journal-title":"arXiv preprint arXiv:1511.08060"},{"key":"B8","doi-asserted-by":"crossref","DOI":"10.1007\/978-3-031-19827-4_41","article-title":"Visual prompt tuning","author":"Jia","year":"2022"},{"key":"B9","doi-asserted-by":"publisher","DOI":"10.1109\/cvpr.2019.00277","article-title":"Do better ImageNet models transfer better","author":"Kornblith","year":"2019","journal-title":"IEEE\/CVF CVPR"},{"key":"B10","doi-asserted-by":"crossref","DOI":"10.1109\/ICCV.2017.324","article-title":"Focal loss for dense object detection","author":"Lin","year":"2017"},{"key":"B11","article-title":"Decoupled weight decay regularization","author":"Loshchilov","year":"2019"},{"key":"B12","first-page":"281","article-title":"Some methods for classification and analysis of multivariate observations","author":"MacQueen","year":"1967"},{"key":"B13","doi-asserted-by":"publisher","DOI":"10.3389\/fpls.2016.01419","article-title":"Using deep learning for imagebased plant disease detection","volume":"7","author":"Mohanty","year":"2016","journal-title":"Front. Plant Sci."},{"key":"B14","article-title":"Uncovering bias in the PlantVillage dataset","author":"Noyan","year":"2022","journal-title":"arXiv preprint arXiv:2206.04374"},{"key":"B15","article-title":"DINOv2: Learning robust visual features without supervision","author":"Oquab","year":"2024","journal-title":"Trans. Mach. Learn. Res. (TMLR)"},{"key":"B16","first-page":"8748","article-title":"Learning transferable visual models from natural language supervision","author":"Radford","year":"2021"},{"key":"B17","doi-asserted-by":"publisher","first-page":"430","DOI":"10.1038\/s41559-018-0793-y","article-title":"The global burden of pathogens and pests on major food crops","volume":"3","author":"Savary","year":"2019","journal-title":"Nat. Ecol. Evol."},{"key":"B18","article-title":"DINOv3: A versatile vision foundation model","author":"Sim\u00e9oni","year":"2026","journal-title":"Trans. Mach. Learn. Res. (TMLR)"},{"key":"B19","first-page":"249","article-title":"PlantDoc: A dataset for visual plant disease detection","author":"Singh","year":"2020"},{"key":"B20","article-title":"Prototypical networks for few-shot learning","author":"Snell","year":"2017"},{"key":"B21","doi-asserted-by":"crossref","DOI":"10.1109\/CVPR.2016.308","article-title":"Rethinking the inception architecture for computer vision","author":"Szegedy","year":"2016"},{"key":"B22","doi-asserted-by":"publisher","first-page":"272","DOI":"10.1016\/j.compag.2018.03.032","article-title":"A comparative study of finetuning deep learning models for plant disease identification","volume":"161","author":"Too","year":"2019","journal-title":"Comput. Electron. Agric."},{"key":"B23","first-page":"1593","article-title":"Benchmarking in-the-wild multimodal disease recognition and a versatile baseline","author":"Wei","year":"2024"},{"key":"B24","article-title":"Plant disease recognition in the wild","author":"Wei","year":"2024"},{"key":"B25","article-title":"PlantCaFo: Few-shot plant disease recognition by combining clip, dino and dinov2 with a weight-decomposed adapter","author":"Jiang","year":"2025","journal-title":"Plant Phenomics"},{"key":"B26","article-title":"Tip-Adapter: Training-free CLIP-adapter for better vision-language modeling","author":"Zhang","year":"2022"},{"key":"B27","article-title":"Combining self-supervised and contrastive foundation models for robust visual recognition","author":"Zhao","year":"2024","journal-title":"NeurIPS Workshop Self-Supervised Learn"}],"container-title":["Frontiers in Plant Science"],"original-title":[],"link":[{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/fpls.2026.1860665\/full","content-type":"unspecified","content-version":"vor","intended-application":"similarity-checking"}],"deposited":{"date-parts":[[2026,8,11]],"date-time":"2026-08-11T05:37:04Z","timestamp":1786426624000},"score":1,"resource":{"primary":{"URL":"https:\/\/www.frontiersin.org\/articles\/10.3389\/fpls.2026.1860665\/full"}},"subtitle":[],"short-title":[],"issued":{"date-parts":[[2026,8,11]]},"references-count":27,"alternative-id":["10.3389\/fpls.2026.1860665"],"URL":"https:\/\/doi.org\/10.3389\/fpls.2026.1860665","relation":{},"ISSN":["1664-462X"],"issn-type":[{"value":"1664-462X","type":"electronic"}],"subject":[],"published":{"date-parts":[[2026,8,11]]},"article-number":"1860665"}}