THREE ROADS TO DIAGNOSIS: BENCHMARKING CONVNEXT-TINY, DENSENET121, AND YOLOV8N-CLS FOR ROBUST TOMATO LEAF DISEASE RECOGNITION

Abstract

Tomato leaf disease is still diagnosed manually across most producing regions, a process too slow and subjective to scale to modern field sizes. This paper narrows the question asked by a companion seven-architecture benchmark to a sharper, three-way comparison: how do a lightweight single-stage detector repurposed for classification (YOLOv8n-cls), a densely-connected convolutional network (DenseNet121), and a modernised pure-convolutional backbone (ConvNeXt-Tiny) compare, not only on curated laboratory imagery but on their ability to survive the transition to field-captured photography? Using a class-balanced, 23,727-image, ten-class tomato leaf dataset, all three architectures were fine-tuned under an identical protocol and evaluated on a held-out 1,500-image in-distribution test split. YOLOv8n-cls and DenseNet121 tied for the best controlled-condition result at 96.20% accuracy and 96.19% weighted F1-score, with ConvNeXt-Tiny close behind at 96.13% accuracy. The same three architectures were then independently retrained and evaluated on a five-class, field-captured subset of the CCMT dataset to probe generalisation. Here the ranking inverted: ConvNeXt-Tiny, the weakest of the three under laboratory conditions, generalised best by a wide margin (85.32% field accuracy, 85.01% F1-score), followed by DenseNet121 (81.88%, 81.62%), while YOLOv8n-cls — despite tying for first place on curated imagery — generalised worst of the three (77.64%, 76.60%). The gap between laboratory and field accuracy therefore ranged from 10.81 points (ConvNeXt-Tiny) to 18.56 points (YOLOv8n-cls), a reversal that a laboratory-only comparison of these three models would never have revealed. We argue that architecture family, not parameter count or laboratory accuracy, is the more reliable signal of field robustness among the three backbones studied, and we discuss what this means for choosing a deployment architecture in practice.