AgroYOLOvNext: A Multi-Scale Transformer-Enhanced Framework with Domain-Adaptive Contrastive Learning for Plant Disease Classification on the Plant Village Benchmark

Abstract

Automated plant disease detection in a real-world agricultural setting remains a difficult problem because of variation in illumination, intra-class diversity, domain shift between laboratory and field data, and occlusions. While recent deep learning-based object detection models, such as YOLO variants and transformer-based models, have demonstrated impressive accuracy on curated datasets, their ability to perform well under open field conditions remains limited. To address this, we introduce Agro YOLO v Next, a multi-scale Transformer-enhanced YOLO framework with domain-adaptive learning that combines three complementary components rather than a single new mechanism. The proposed architecture consists of three main components: (1) a Multi-Scale Depth wise Asymmetric (MSDA) backbone for extracting fine-grained features from multiple receptive fields, (2) a Lightweight Cross-Stage Transformer Neck (LCST-Neck) for improved feature fusion and class separability, and (3) a Domain-Adaptive Contrastive Head (DACH)