A Segmentation-driven Dual-branch Deep Learning Framework with Attention Fusion for Plant Disease Phenotyping

Abstract

Plant phenotyping plays a critical role in understanding plant health and improving agricultural productivity by enabling quantitative analysis of disease-related physiological characteristics.Among these, leaf diseases significantly impact crop yield and quality, necessitating accurate and automated phenotyping approaches.Traditional phenotyping methods rely on manual inspection or handcrafted feature extraction, which are time-consuming, prone to human error, and lack scalability under diverse environmental conditions.This study introduces a deep learning (DL)-based approach for image-based plant phenotyping, focusing on the classification of disease-affected traits.The developed method integrates segmentation-based region extraction, dual-branch feature learning, and attention-based feature fusion.Initially, input images are preprocessed and passed through a TransUNet-based segmentation module to isolate phenotypically relevant leaf regions while suppressing background interference.Both the original image and the segmented region are then processed using a RegNet-based feature extraction network to capture global structural information and localized disease-specific characteristics.The extracted features are fused using an attention-based mechanism, followed by fully connected layers for multiclass classification.Experimental results obtained on the controlled PlantVillage grape leaf dataset, which serves as a standardized benchmark for plant disease classification, demonstrate an overall classification accuracy of 97.8%, with precision, recall, and F1-score values of 97.7%, 97.9%, and 97.8%, respectively.In addition, the segmentation module achieves an Intersection over Union (IoU) of 94.1% and a Dice score of 96.8%, confirming its effectiveness in isolating relevant phenotypic regions.

Plant phenotyping relevance

植物病害形質の画像取得・領域抽出・分類を中核とする深層学習フェノタイピング手法の開発と性能評価であり、方法が中心的です。

Code and data availability

The paper uses the PlantVillage grape leaf subset (accessed via TensorFlow Datasets) but provides no authors' public dataset deposit, code repository, trained model, or supplement with an explicit availability URL. PlantVillage is a cited prior public dataset, not a paper-specific asset, and no author code availability

No evidence-backed public reproduction asset is currently recorded.