ParaLeafNet-Severity: A Parallel Deep CNN with Squeezeand-Excitation Attention for Automated Plant Disease Severity Assessment and Agricultural Health Monitoring

Abstract

Accurate and timely detection of plant diseases is a critical component of modern agricultural biosecurity, directly impacting crop yield, food safety, and the health of farming communities. Plant diseases cause annual agricultural losses that exceed $220 billion worldwide, yet accurately quantifying how severely a plant is infected remains one of the least-addressed problems in automated crop monitoring. This paper presents ParaLeafNet-Severity, a deep parallel convolutional neural network designed to perform disease identification and four-level severity grading simultaneously within a single forward pass. The architecture draws complementary feature representations from two lightweight backbones — MobileNetV2 and MobileNetV3Small — running in parallel, fuses their outputs through channel-wise Squeeze-and-Excitation (SE) attention, and routes the resulting shared embedding to two task-specific output heads. An optional K-Means clustering branch operates on the shared feature space to discover natural severity groupings without requiring additional expert labels. Experimental evaluation on the adapted PlantVillage dataset demonstrates that the proposed framework achieves strong performance in disease identification and severity classification, outperforming existing baseline approaches while maintaining robustness across multiple classes. The integration of an unsupervised clustering module further confirms that the learned feature representations preserve meaningful severity-related structure consistent with expert annotations. In addition, the model maintains a compact design and efficient inference characteristics, making it suitable for deployment on resource-constrained devices such as smartphones and edge platforms. Interpretability analysis using Grad-CAM indicates that the model focuses on pathologically relevant regions of the leaf, supporting transparent and reliable decision-making. Furthermore, the extended ParaLeafNet framework incorporates deployment-oriented optimization techniques that enhance performance without requiring modifications to the core architecture.

Plant phenotyping relevance

葉画像から植物病害の重症度を自動推定するCNN手法が研究の中心であり、植物状態の表現型計測に該当する。

Code and data availability

植物フェノタイピング解析を再現する公開資産であることを、入力本文と直接リンクから確認できなかったため保留しました。

No evidence-backed public reproduction asset is currently recorded.