Optimized classification of potato leaf disease using EfficientNet-LITE and KE-SVM in diverse environments.

Abstract

Introduction Potatoes are a vital global product, and prompt identification of foliar diseases is imperative for sustaining healthy yields. Computer vision is essential in precision agriculture, facilitating automated disease diagnosis and decision-making through real-time data. Inconsistent data in uncontrolled contexts undermines classic image classification techniques, hindering precise illness detection. Methods We present a novel model that integrates EfficientNet-LITE for enhanced feature extraction with KE-SVM Optimization for effective classification. KE-SVM Optimization cross-references misclassified instances with correct classifications across kernels, iteratively refining the confusion matrix to improve accuracy across all classes. EfficientNet-LITE improves the model's emphasis on pertinent features through Channel Attention (CA) and 1-D Local Binary Pattern (LBP), while preserving computational economy with a reduced model size of 12.46 MB, fewer parameters at 3.11M, and a diminished FLOP count of 359.69 MFLOPs. Results Before optimization, the SVM classifier attained an accuracy of 79.38% on uncontrolled data and 99.07% on laboratory-controlled data. Following the implementation of KE-SVM Optimization, accuracy increased to 87.82% for uncontrolled data and 99.54% for laboratory-controlled data. Discussion The model's efficiency and improved accuracy render it especially appropriate for settings with constrained computational resources, such as mobile or edge devices, offering substantial practical advantages for precision agriculture.

Plant phenotyping relevance

ジャガイモ葉の病徴を画像から分類する新規コンピュータビジョン手法を開発・評価しており、植物状態の推定が中心的な方法論的貢献である。

Code and data availability

The paper uses two pre-existing public datasets (an uncontrolled potato leaf disease dataset from Shabrina et al. and PlantVillage potato), but these are cited prior-work resources, not paper-specific assets. No author code, models, checkpoints, or data deposit with a public URL is mentioned; no data availability URL/`

No evidence-backed public reproduction asset is currently recorded.