<p>In this paper, We formulate date palm leaf disease recognition as a hyperparameter optimization problem with three objectives, jointly maximising Macro–F1, robustness to acquisition shifts, and energy efficiency for compact CNN backbones deployed on edge devices. To address this, we propose a <i>cooperative GA–MOPSO</i> framework in which two solver components (a GA population and a MOPSO population) interact through a shared archive of non-dominated CNN configurations and periodically exchange selected solutions to balance exploration and exploitation. Experiments on a nine-class dataset of 11,420 date palm leaf images from Al–Madinah and Al–Qassim—with testing restricted to Qassim palms to mimic realistic deployment—show that the proposed cooperative GA–MOPSO improves the median hypervolume from 0.70 to 0.76 (<InlineEquation ID="IEq1"> <EquationSource Format="TEX">$\varvec{\Delta }{=}+\varvec{0.06}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mrow> <mi mathvariant="bold">Δ</mi> </mrow> <mo>=</mo> <mo>+</mo> <mrow> <mn mathvariant="bold">0.06</mn> </mrow> </mrow> </math></EquationSource> </InlineEquation> absolute; about <InlineEquation ID="IEq2"> <EquationSource Format="TEX">$\varvec{8.6}\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mrow> <mn mathvariant="bold">8.6</mn> </mrow> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>), reduces IGD from 0.108 to 0.082 (<InlineEquation ID="IEq3"> <EquationSource Format="TEX">$\varvec{\Delta }{=}-\varvec{0.026}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mrow> <mi mathvariant="bold">Δ</mi> </mrow> <mo>=</mo> <mo>-</mo> <mrow> <mn mathvariant="bold">0.026</mn> </mrow> </mrow> </math></EquationSource> </InlineEquation>; about <InlineEquation ID="IEq4"> <EquationSource Format="TEX">$\varvec{24.1}\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mrow> <mn mathvariant="bold">24.1</mn> </mrow> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>) and spacing from 0.042 to 0.031 (<InlineEquation ID="IEq5"> <EquationSource Format="TEX">$\varvec{\Delta }{=}-\varvec{0.011}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mrow> <mi mathvariant="bold">Δ</mi> </mrow> <mo>=</mo> <mo>-</mo> <mrow> <mn mathvariant="bold">0.011</mn> </mrow> </mrow> </math></EquationSource> </InlineEquation>; about <InlineEquation ID="IEq6"> <EquationSource Format="TEX">$\varvec{26.2}\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mrow> <mn mathvariant="bold">26.2</mn> </mrow> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>), and enlarges the median archive size from 20 to 25 (<InlineEquation ID="IEq7"> <EquationSource Format="TEX">$\varvec{\Delta }{=}+\varvec{5}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mrow> <mi mathvariant="bold">Δ</mi> </mrow> <mo>=</mo> <mo>+</mo> <mrow> <mn mathvariant="bold">5</mn> </mrow> </mrow> </math></EquationSource> </InlineEquation>; about <InlineEquation ID="IEq8"> <EquationSource Format="TEX">$\varvec{25}\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mrow> <mn mathvariant="bold">25</mn> </mrow> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation>) compared with the best baseline, while incurring only a small <InlineEquation ID="IEq9"> <EquationSource Format="TEX">$\varvec{\approx } \varvec{6}\%$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mrow> <mo mathvariant="bold">≈</mo> </mrow> <mrow> <mn mathvariant="bold">6</mn> </mrow> <mo>%</mo> </mrow> </math></EquationSource> </InlineEquation> extra runtime over classical MOPSO (4100 s to 4350 s; <InlineEquation ID="IEq10"> <EquationSource Format="TEX">$\varvec{\Delta }{=}+\varvec{250}$</EquationSource> <EquationSource Format="MATHML"><math> <mrow> <mrow> <mi mathvariant="bold">Δ</mi> </mrow> <mo>=</mo> <mo>+</mo> <mrow> <mn mathvariant="bold">250</mn> </mrow> </mrow> </math></EquationSource> </InlineEquation> s) and remaining faster than NSGA–II/III.</p>

Energy-aware and robust date palm disease classification in the Qassim region using a cooperative hybrid GA–MOPSO on edge devices

期刊论文
出版模式：

开放获取
发表日期：

2026年4月6日

查看全文

Issam Zidi,
Salim El Khediri,
Suliman Aladhadh

摘要

In this paper, We formulate date palm leaf disease recognition as a hyperparameter optimization problem with three objectives, jointly maximising Macro–F1, robustness to acquisition shifts, and energy efficiency for compact CNN backbones deployed on edge devices. To address this, we propose a cooperative GA–MOPSO framework in which two solver components (a GA population and a MOPSO population) interact through a shared archive of non-dominated CNN configurations and periodically exchange selected solutions to balance exploration and exploitation. Experiments on a nine-class dataset of 11,420 date palm leaf images from Al–Madinah and Al–Qassim—with testing restricted to Qassim palms to mimic realistic deployment—show that the proposed cooperative GA–MOPSO improves the median hypervolume from 0.70 to 0.76 ( $\varvec{\Delta }{=}+\varvec{0.06}$ $Δ = + 0.06$ absolute; about $\varvec{8.6}\%$ $8.6 %$ ), reduces IGD from 0.108 to 0.082 ( $\varvec{\Delta }{=}-\varvec{0.026}$ $Δ = - 0.026$ ; about $\varvec{24.1}\%$ $24.1 %$ ) and spacing from 0.042 to 0.031 ( $\varvec{\Delta }{=}-\varvec{0.011}$ $Δ = - 0.011$ ; about $\varvec{26.2}\%$ $26.2 %$ ), and enlarges the median archive size from 20 to 25 ( $\varvec{\Delta }{=}+\varvec{5}$ $Δ = + 5$ ; about $\varvec{25}\%$ $25 %$ ) compared with the best baseline, while incurring only a small $\varvec{\approx } \varvec{6}\%$ $\approx 6 %$ extra runtime over classical MOPSO (4100 s to 4350 s; $\varvec{\Delta }{=}+\varvec{250}$ $Δ = + 250$ s) and remaining faster than NSGA–II/III.