DSA-net: a lightweight and efficient deep learning-based model for pea leaf disease identification.

Abstract

Introduction Pea is a nutrient-dense, functionally diversified vegetable. However, its leaf diseases have a direct impact on yield and quality. Most approaches for identifying pea leaf diseases exhibit low feature extraction efficiency, significant environmental sensitivity, and limited large-scale applications, making it impossible to meet the expectations of modern agriculture for accuracy, real-time processing, and low cost. Methods Therefore, we propose a deep learning model for pea leaf disease identification based on an improved MobileNet-V3_small, deformable convolution strategy, self-attention, and additive attention mechanisms (DSA-Net). First, a deformable convolution is added to MobileNet-V3-small to increase the modeling skills for geometric changes in disease features. Second, a self-attention mechanism is integrated to improve the ability to recognize global features of complex diseases. Finally, an additive attention strategy to enhance the feature channel and spatial position response relationship in edge-blurred lesion areas. The experimental pea leaf data set consists of 7915 samples divided into five categories. It includes one healthy leaf and four diseases: brown spot, leaf miner, powdery mildew, and root rot. Results The experimental results indicate that the suggested DSA-Net has an average recognition accuracy of 99.12%. It has a parameter size of 1.48M. Discussion The proposed approach will help with future edge device deployments. The current proposed technique considerably enhances the diagnostic accuracy of pea leaf diseases and has significant promotion and application potential in agriculture.

Plant phenotyping relevance

エンドツーエンドの画像ベース深層学習モデルを開発し、エンドウ葉の病徴・病害状態を直接識別する手法が研究の中心であるため、植物フェノタイピング手法として収録する。

Code and data availability

The supplied blocks describe a self-built pea leaf disease image dataset (7,915 samples, five classes) and a modified MobileNet-V3 model, but contain no data availability statement, no public dataset deposit, and no author code/model release with a URL. No paper-specific public asset is actionable.

No evidence-backed public reproduction asset is currently recorded.