Noise-Resilient Hybrid EfficientNet–Vision Transformer Framework with Adaptive Symmetric Cross-Entropy Loss for Robust Plant Disease Detection

Abstract

The errors of human annotation and the noise of the environment such as lighting changes, occlusions and cluttered backdrop limit the correct detection of the plant diseases in the field condition. The research hypothesis is to present a robust deep learning model that can withstand noise and be interpretable in controlled and noisy environments to achieve high plant disease classification. The hybrid EfficientNet-Vision Transformer (ViT) network proposed is based on an EfficientNet-B4 branch of CNN and a branch of Vision Transformer (ViT) network, which focuses on capturing fine-grained lesion features and global contexts information. A data augmentation pipeline based on CycleGAN is used to introduce field-style distortions (e.g., (lighting shifts, shadowing, debris and partial occlusions), to be more robust to environmental noise, and an Adaptive Symmetric Cross-Entropy (ASCE) loss identifies and down-weights uncertain samples with normalized prediction entropy. The training is done in two phases, Stage 1 pretraining with clean images of PlantVillage and Stage 2 with increasingly noisy samples. The framework is tested in two different noise conditions, and these include the controlled synthetic label noise with PlantVillage and the real environmental noise with PlantDoc. The proposed model has an accuracy of 94.5% on the clean PlantVillage test set. It achieves 85.0% accuracy on the PlantVillage dataset under the 20% synthetic label noise protocol, outperforming ResNet-50V2 (76.5%), DenseNet-121 (78.9%), and Co-Teaching (79.5%). Macro-precision, macro-recall and macro-F1 of the model on the external PlantDoc field dataset are 0.718, 0.681, 0.681, respectively with a top-1 accuracy of 72.0, which is a manifestation of cross-domain generalization. The lesion-centric Grad-CAM images indicate that the model places emphasis on symptomatic areas of leaves and represses reactions of background soil, shadows, and clutters. The suggested hybrid EfficientNet-ViT architecture offers, in general, a robust and explainable solution to precision agriculture and intelligent crop tracking systems that are resistant to noise.

Plant phenotyping relevance

植物の病徴画像から病害状態を推定する深層学習手法を開発し、ノイズ条件・外部データセットで性能検証しているため、植物フェノタイピング手法が中心である。

Code and data availability

The supplied blocks describe a hybrid EfficientNet–ViT plant disease detection framework evaluated on PlantVillage and PlantDoc, but contain no author code release, trained model checkpoint, or paper-specific dataset deposit. PlantVillage and PlantDoc are pre-existing public datasets cited as prior work (refs [36], [13

No evidence-backed public reproduction asset is currently recorded.