Plant disease affects agricultural productivity, food security, and, in turn, economic stability; thus, early-stage, accurate detection of plant diseases is necessary. This research proposes Hybrid Feature-based Plant Disease Classification Network (HyFPlantNet), a hybrid feature based deep learning framework that integrates both spatial and spectral descriptors for robust plant disease classification through visualization. The feature extraction modules include Scale Invariant Feature Transform (SiFT) for keypoints detection, Wavelet based SFTA (WSFTA) for fractal texture analysis, Gray Level Co occurrence Matrix (GLcM) for statistical texture characterization, and HSV color space analysis for color based discrimination. The resultant feature vectors are concatenated and filtered using Principal Component Analysis (PCA), and finally fed into a fully connected neural network for classification. HyFPlantNet was evaluated on three benchmark datasets - <Emphasis FontCategory="NonProportional">PlantVillage</Emphasis>, <Emphasis FontCategory="NonProportional">AI Challenger 2018 Crop Disease Detection</Emphasis>, and <Emphasis FontCategory="NonProportional">Plant Pathology 2021</Emphasis> (cf. FGVC8) where it achieved accuracies of <InlineEquation ID="IEq1"> <EquationSource Format="TEX">$96.41\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>96.41</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>, <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$96.41\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>96.41</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>, and <InlineEquation ID="IEq3"> <EquationSource Format="TEX">$96.41\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>96.41</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>, respectively, which accounts for <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$0.1663\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>0.1663</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> improvements from Vision Transformer on <Emphasis FontCategory="NonProportional">PlantVillage</Emphasis>, <InlineEquation ID="IEq5"> <EquationSource Format="TEX">$0.0952\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>0.0952</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> improvements from Xception on <Emphasis FontCategory="NonProportional">AI Challenger</Emphasis>, and <InlineEquation ID="IEq6"> <EquationSource Format="TEX">$0.9741\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mn>0.9741</mn> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> from ResNet 50 on <Emphasis FontCategory="NonProportional">Plant Pathology 2021</Emphasis>, serving as state of the art.

HyFPlantNet: hybrid feature-based plant disease classification network

期刊论文
发表日期：

2025年10月3日

查看全文

Deepkiran Munjal,
Mrinal Pandey,
Laxman Singh

摘要

Plant disease affects agricultural productivity, food security, and, in turn, economic stability; thus, early-stage, accurate detection of plant diseases is necessary. This research proposes Hybrid Feature-based Plant Disease Classification Network (HyFPlantNet), a hybrid feature based deep learning framework that integrates both spatial and spectral descriptors for robust plant disease classification through visualization. The feature extraction modules include Scale Invariant Feature Transform (SiFT) for keypoints detection, Wavelet based SFTA (WSFTA) for fractal texture analysis, Gray Level Co occurrence Matrix (GLcM) for statistical texture characterization, and HSV color space analysis for color based discrimination. The resultant feature vectors are concatenated and filtered using Principal Component Analysis (PCA), and finally fed into a fully connected neural network for classification. HyFPlantNet was evaluated on three benchmark datasets - PlantVillage, AI Challenger 2018 Crop Disease Detection, and Plant Pathology 2021 (cf. FGVC8) where it achieved accuracies of $96.41\%$ $96.41 %$ , $96.41\%$ $96.41 %$ , and $96.41\%$ $96.41 %$ , respectively, which accounts for $0.1663\%$ $0.1663 %$ improvements from Vision Transformer on PlantVillage, $0.0952\%$ $0.0952 %$ improvements from Xception on AI Challenger, and $0.9741\%$ $0.9741 %$ from ResNet 50 on Plant Pathology 2021, serving as state of the art.