Also available as PDF (figures and tables appear in the PDF).
Plant diseases significantly affect crop productivity and global food security, making early detection essential for sustainable agriculture. Traditional visual diagnosis methods are often slow, subjective, and ineffective during early infection stages. Recent advances in multispectral and hyperspectral imaging combined with machine learning and deep learning techniques enable rapid, non-destructive, and accurate detection of plant diseases before visible symptoms appear. Spectral imaging captures physiological and biochemical changes in plants across visible, near-infrared, red-edge, and thermal bands, while algorithms such as SVM, Random Forest, CNN, RNN, and Vision Transformers effectively classify healthy and diseased plants. This review discusses imaging techniques, preprocessing methods, machine learning models, UAV-based sensing, explainable AI, and edge computing applications in precision agriculture, along with challenges and future directions for smart disease detection systems.
Keywords: Machine Learning; Multispectral Imaging; Hyperspectral Imaging; Plant Disease Detection; Deep Learning; MachineLearning; MultispectralImaging; HyperspectralImaging; PlantDiseaseDetection
• NearInfrared(NIR:750–1000nm) Highlights • ShortwaveInfrared(SWIR) • Multispectralimagingenablesnon-destructive earlydiseasedetectionincrops.
Healthyvegetationstronglyabsorbsredlightdueto • Machinelearninganddeeplearningmodelsim- chlorophyllandreflectsNIRbecauseofcellularstruc- provediseaseclassificationaccuracy.
ture. Disease-induced physiological stress modifies • CNNs,SVMs,andVisionTransformersarewidely thesespectralresponses[7]. Hyperspectralimaging usedinplantdiseasediagnosis.
differsfrommultispectralimagingbycapturinghun- • UAV-based sensing systems support real-time dredsofnarrowcontiguousspectralbands.Although large-scalecropmonitoring.
hyperspectralimagingprovidesricherspectralinfor- • AI-integratedprecisionagriculturecanenhance mation,multispectralsystemsaremorecost-effective sustainablecropprotectionandfoodsecurity. andcomputationallyefficient[8].
2.2 SpectralSignaturesofDiseasedPlants 1.Introduction Diseaseinfectionsalterplantphysiologythrough: Agricultureisoneofthemostimportantsectorssup- • Chlorophylldegradation portingglobalfoodsecurity,economicdevelopment, • Reducedphotosynthesis andhumansurvival.However,plantdiseasescontinue • Waterstress toposeseriousthreatstoagriculturalproductivityand • Cellulardamage cropqualityworldwide. Variousfungal,bacterial,vi- • Pigmentchanges ral,andparasiticpathogensareresponsibleforsub- stantialreductionsincropyieldandquality,resulting Thesephysiologicalchangesaffectspectralreflectance in severe economic losses every year. According to patterns.Forexample: globalestimates,plantdiseasesaccountforapprox- • Increasedreflectanceinredwavelengths imately20–40%lossesinagriculturalproductionan- • ReducedNIRreflectance nually,therebythreateningfoodavailabilityandsus- • Alteredred-edgeposition tainable farming systems. Consequently, early and Suchspectralchangescanbedetectedbeforevisible accuratedetectionofplantdiseaseshasbecomean symptomsemerge[9].
essential requirement for modern precision agricul- ture.Conventionalmethodsofplantdiseasediagnosis mainlyrelyonvisualinspection,laboratoryculturing, 3. MachineLearningTechniquesforDisease moleculartesting, andexpertknowledge. Although Detection thesemethodsareeffectiveundercontrolledcondi- 3.1 TraditionalMachineLearningModels tions,theysufferfromseverallimitationssuchashigh 3.1.1 SupportVectorMachine(SVM) laborrequirements,longprocessingtime,highopera- SVMisoneofthemostwidelyusedclassifiersinspec- tionalcosts,anddependenceonexperiencedpathol- tralimageanalysis.Itidentifiesoptimalhyperplanes ogists.Moreimportantly,traditionalvisualdiagnosis separatingdiseaseclasseswithhigh-dimensionalspec- oftenfailstodetectdiseasesduringtheirearlyasymp- tralfeatures[10].
tomaticstagesbecausevisiblesymptomsusuallyap- Advantages: pear only after severe physiological damage has al- readyoccurred. Delayeddiseasedetectioncanlead • Effectiveinhigh-dimensionaldata torapidpathogenspread,excessivepesticideapplica- • Goodgeneralization tion,reducedcropquality,andsignificanteconomic • Workswithsmalldatasets losses.
Limitations: • Computationallyexpensiveforlargedatasets 2.MultispectralImaginginPlantDiseaseDe- • Sensitivetoparameterselection tection SeveralstudiesreportedSVMaccuraciesabove90%for 2.1 PrinciplesofMultispectralImaging diseaseclassificationusinghyperspectraldata[11]. Multispectralimagingcapturesreflectanceinformation atdiscretewavelengthbands. Commonspectralre- 3.1.2 RandomForest(RF) gionsinclude: RandomForestusesensembledecisiontreesforclas- sification.RFperformswellwithnoisyandnonlinear • Blue(450–495nm) agriculturaldatasets[12].
• Green(495–570nm) • Red(620–750nm) Advantages:
• GoogLeNet • ResNet RFhasbeenextensivelyusedinwheatrustdetection, • DenseNet citrus canker diagnosis, and rice blast classification • EfficientNet [13].
CNNseliminatemanualfeatureengineeringandsignifi- 3.1.3 k-NearestNeighbor(kNN) cantlyoutperformtraditionalMLmethods[16].Studies kNNclassifiessamplesbasedonsimilaritywithneigh- demonstratedclassificationaccuraciesexceeding98% boringdatapoints[14].
usingCNNswithhyperspectralimagery[17]. Advantages: • Simpleimplementation 4.2 3D-CNNModels 3D-CNNssimultaneouslyprocessspectralandspatial • Non-parametric dimensionsofhyperspectralcubes[18].
• Limitations: • Highmemoryrequirement Advantages: • Sensitivetodimensionality • Betterspectral-spatialfeatureextraction Althoughsimplerthandeeplearningmethods, kNN • Improveddiseasediscrimination remainsusefulinsmall-scalediseasedetectiontasks. However, 3D-CNNsrequireextensivecomputational resourcesandlargeannotateddatasets[19]. 4.DeepLearningModels 4.3 TransferLearning 4.1 ConvolutionalNeuralNetworks(CNN) Transferlearningutilizespretrainedmodelstrainedon CNNshavebecomedominantinimage-basedplant largedatasetssuchasImageNet[20].
diseasedetectionduetotheirabilitytoautomatically learnhierarchicalfeatures[15].
Benefitsinclude: PopularCNNarchitecturesinclude: • Reducedtrainingtime • Improvedperformancewithlimiteddatasets 6.ApplicationsinDifferentCrops • Lowercomputationalcost Theintegrationofmultispectralimaging,hyperspec- Transferlearninghasshownpromisingresultsincrop tral sensing, and machine learning techniques has diseasedetectionusingmultispectralimagery[21]. beenextensivelyinvestigatedacrossawiderangeof agriculturalcropsforearlydiseasedetectionandstress 4.4 VisionTransformers(ViTs) monitoring.Differentcropsexhibituniquephysiologi- Transformerarchitecturescapturelong-rangedepen- calandspectralresponsestopathogeninfections,en- denciesusingattentionmechanisms[22]. ablingmachinelearningmodelstoidentifydisease- specificspectralsignatureswithhighaccuracy[27][28]. ViTsdemonstrate: Researchersworldwidehavedemonstratedthatcom- • Stronggeneralizationcapability biningspectralimagingtechnologieswithadvanced • Improvedfeaturerepresentation artificialintelligencemodelssignificantlyimprovesdis- • Enhancedscalability easediagnosis,cropmonitoring,andprecisiondisease Recentstudiesindicatetransformersmayoutperform managementpractices[29][30].
CNNsincomplexagriculturalimagingtasks[23]. 6.1 RiceDiseaseDetection Rice (Oryza sativa L.) is one of the most important 5. Image Preprocessing and Feature Extrac- staplefoodcropsglobally,feedingmorethanhalfof tion theworld’spopulation. However,riceproductionis 5.1 NoiseReduction severelyaffectedbydiseasessuchasriceblast,bac- Hyperspectralimagesoftencontainsensornoiseand terialblight, sheathblight, andbrownspot[31][32]. illuminationvariations.Commonpreprocessingtech- Earlydetectionofthesediseasesisessentialtoprevent niquesinclude: majoryieldlosses. Severalresearchershaveapplied • Savitzky–Golayfiltering multispectralandhyperspectralimagingtechniques • Waveletdenoising forricediseasediagnosis.Hyperspectralimaginghas • Gaussiansmoothing beenshowntodetectphysiologicalchangesassoci- • Medianfiltering atedwithfungalinfectionsbeforevisiblesymptomde- velopment[33].Hyperspectralimagingcombinedwith Thesemethodsimprovespectralqualityandmodel SVMmodelscanclassifyriceblastdiseasewithhigh performance[24].
accuracy[34].Spectralbandsinthenear-infraredre- gionareparticularlysensitivetodisease-inducedstress 5.2 DimensionalityReduction [27][28].CNNshavebeenappliedforautomatedrice Hyperspectraldatacontainhundredsofbands,lead- diseaseclassification,achievingaccuraciesabove95% ing to the “curse of dimensionality.” Dimensionality [35]. Deep CNN architectures integrated with multi- reductiontechniquesinclude: spectralimagerycanoutperformtraditionalmachine • PrincipalComponentAnalysis(PCA) learningmodels[36].TransferlearningusingResNet • LinearDiscriminantAnalysis(LDA) andDenseNethasalsoachievedremarkableaccuracy • IndependentComponentAnalysis(ICA) forricediseaseidentification[37]. UAV-basedmulti- • Autoencoders spectralimagingfurtherallowslarge-scalefield-level Thesetechniquesreduceredundancywhilepreserving diseasesurveillance[38][39].
6.2 WheatDiseaseDetection 5.3 VegetationIndices Wheat (TriticumaestivumL.) sufferssignificantlosses Vegetationindicesderivedfromspectralbandsareuse- duetodiseasessuchasleafrust,stemrust,striperust, fulindicatorsofplanthealth.
Fusariumheadblight,andpowderymildew[40][41]. Hyperspectralandmultispectralimagingtechnologies Commonindicesinclude: havebeenwidelyexploredforearlydetectionofwheat • NDVI diseases.Hyperspectralimagingcombinedwithneu- • PRI ralnetworkclassifierscandistinguishhealthyandin- • SAVI fectedwheatleaves[42][43].SVMclassifierseffectively • MCARI identify wheat diseases under field conditions [44]. Theseindicesimprovediseasediscriminationcapabil- RandomForestmodelsusinghyperspectralreflectance ity[26].
andtexturefeatureshavebeenshowntoimprovedis- MachineLearning/ Imaging Crop MajorDiseases DeepLearning References Technique Models Blast,BacterialBlight, Multispectral, SVM,CNN, Rice [31][32][33][34] BrownSpot Hyperspectral ResNet,DenseNet Rust,PowderyMildew, Hyperspectral, Wheat RF,SVM,3D-CNN [35][36][37][38] FusariumHeadBlight UAVImaging NorthernLeafBlight, Multispectral, CNN,ViT, Maize [39][40][41][42] GrayLeafSpot Hyperspectral TransferLearning CNN,SVM, Potato LateBlight,EarlyBlight Hyperspectral [43][44][45][46] 3D-CNN EarlyBlight, CNN,AlexNet, Tomato SeptoriaSpot, Multispectral [47][48][49] VGGNet MosaicVirus CitrusCanker, UAV Citrus RF,CNN,SVM [50][51][52] GreeningDisease Multispectral PowderyMildew, UAV CNN,Thermal Grapevine [53][54][55] DownyMildew Hyperspectral FusionModels Rust,SuddenDeath Soybean Hyperspectral CNN,EfficientNet [56][57][58] Syndrome Multispectral, Apple AppleScab,BlackRot DeepCNN,RF [59][60][61] UAVImaging easeclassificationperformance[45]. Recentstudies 6.5 TomatoDiseaseDetection demonstrate that 3D-CNN architectures can extract Tomato (SolanumlycopersicumL.) ishighlyproneto spectral-spatialfeaturessimultaneouslyfromhyper- fungal,bacterial,andviraldiseases[58].Multispectral spectralcubes,achievinghighlyaccuratewheatdis- imagingcombinedwithCNN-baseddeeplearningmod- easedetection[46].UAV-basedmultispectralsensors elsachieveshighclassificationaccuracyfortomatodis- enablelarge-scalemonitoringofwheatrustseverity eases[59][60].AlexNet,VGGNet,andotherCNNarchi- [47].
tecturesarewidelyusedforautomaticdiseaserecog- nition[61]. Hyperspectralspectralsignaturesinthe visibleandnear-infraredregionscandetectdiseases 6.3 MaizeDiseaseDetection atasymptomaticstages[27][28].
Maize (ZeamaysL.) isvulnerabletodiseasessuchas northernleafblight,grayleafspot,commonrust,and southernrust[48]. Earlydiseasedetectioniscrucial 6.6 CitrusDiseaseDetection topreventrapidspread. Hyperspectralimagingcan Citrus crops are threatened by diseases such as cit- identifymaizefungalinfectionsbeforevisiblesymp- rus canker and Huanglongbing (HLB) [62][63]. UAV- tomsappear[49].Machinelearningclassifierssuchas mountedmultispectralimagingcombinedwithSVM CNNshavebeeneffectivelyusedformaizediseaseclas- orCNNmodelsprovidesearlyandaccuratediseasede- sification[50]. DenseNetandtransferlearningmod- tection[50][51].Hyperspectralimagingisalsohighly elsfurtherimproveclassificationaccuracy[51].Vision effectiveforidentifyingcitrusgreeningdisease[52]. Transformers(ViTs)haverecentlyoutperformedcon- ventionalCNNsincomplexmaizediseaserecognition 6.7 GrapevineDiseaseDetection tasks[52].
Grapevinesareaffectedbypowderymildew,downy mildew,andEscadisease[64][65].Hyperspectralimag- ing can detect disease-induced spectral variations 6.4 PotatoDiseaseDetection Potato (SolanumtuberosumL.) issusceptibletolate [53].UAV-basedmultispectralimagingcombinedwith blightandearlyblight[53].Earlydetectioniscriticalfor CNNmodelsenablesrapidvineyarddiseasemapping minimizingyieldloss.Hyperspectralimaginghasiden- [54][55].
tifiedpathogen-inducedstressinpotatoleaves[54]. SVMandCNNclassifiersachievehighaccuracyinearly 6.8 SoybeanDiseaseDetection disease detection [55][56]. Spectral-spatial 3D-CNN Soybeandiseases,includingsoybeanrustandsudden modelsprovideenhancedclassificationperformance deathsyndrome,severelyreduceyield[66][67].Hyper- overconventionalmethods[57].
spectralimagingcombinedwithCNNandEfficientNet ML/DL Reported Crop Disease ImagingType References Model Accuracy Rice RiceBlast CNN Hyperspectral 95–98% [35][36] Fusarium Wheat RF Hyperspectral 93–96% [45] HeadBlight Wheat StripeRust 3D-CNN Hyperspectral 97% [46] Maize GrayLeafSpot DenseNet Multispectral 98% [51] Potato LateBlight CNN Hyperspectral 99% [55][56] Tomato EarlyBlight VGGNet Multispectral 99% [59][60] Citrus CitrusCanker SVM UAVMultispectral 94–96% [50][51] Soybean SoybeanRust EfficientNet Hyperspectral 97% [57][58] Apple AppleScab DeepCNN Multispectral 98% [59][60] modelsenablesaccurateearlydetection[56][57][58]. traldataandlimitedtrainingsamples. However, re- centadvancementsindeeplearninghavefurtheren- 6.9 AppleDiseaseDetection hanced disease detection capabilities by automati- Appleorchardssufferfromdiseasessuchasapplescab callyextractinghierarchicalspectral-spatialfeatures andblackrot[68][69].Multispectralimaginganddeep from multispectral images. Deep learning architec- CNNsprovidehighclassificationaccuracy[59][60].UAV turesincludingConvolutionalNeuralNetworks(CNNs), imagingfacilitateslarge-scaleorcharddiseasemoni- 3D-CNNs,DenseNet,ResNet,EfficientNet,andVision toring[61].
Transformers(ViTs)haveachievedstate-of-the-artper- formanceacrossmultiplecropssuchasrice, wheat, maize,potato,tomato,citrus,soybean,grapevine,and 7.Conclusion apple.Thesemodelshavedemonstratedremarkable Machinelearningmodelsintegratedwithmultispectral classificationaccuraciesexceeding95–99%inseveral andhyperspectralimagingtechnologieshaveemerged diseasedetectionstudies.Furthermore,transferlearn- as highly effective tools for the early detection and ingapproacheshavereducedcomputationalrequire- diagnosisofplantdiseases. Thecombinationofad- ments and improved model generalization, particu- vanced imaging systems with artificial intelligence larlyinsituationswithlimitedannotateddatasets.The hassignificantlytransformedtraditionalagricultural integrationofUAVs,drone-basedmultispectralimag- disease monitoring into an automated, rapid, non- ing systems, and edge computing technologies has destructive,andhighlyaccurateprocess.Thesetech- furtherexpandedthepracticalapplicationsofintelli- nologiesenabletheidentificationofsubtlephysiolog- gentdiseasemonitoringsystemsinprecisionagricul- icalandbiochemicalchangesinplantsbeforevisible ture. UAV-mountedsensorsenablerapidlarge-scale symptoms appear, thereby allowing timely disease fieldsurveillancewithhighspatialandtemporalresolu- management and reducing major crop losses. Con- tion,facilitatingreal-timecrophealthassessmentand ventionaldiseasediagnosismethodsbasedonmanual site-specific disease management. Additionally, ad- observation and laboratory analysis are often time- vancesinexplainableartificialintelligence(XAI)have consuming, subjective, labor-intensive, and unsuit- improvedtheinterpretabilityandtransparencyofdeep able for large-scale agricultural monitoring. In con- learning models, increasing farmer confidence and trast, multispectral imaging provides detailed spec- adoptionpotential.Despitesubstantialprogress,sev- tralinformationassociatedwithchlorophylldegrada- eralchallengesremaininthepracticaldeploymentof tion,moisturestress,pigmentalteration,andcellular machinelearning-basedmultispectraldiseasedetec- damagecausedbypathogens.Machinelearningalgo- tionsystems.Majorlimitationsincludethescarcityof rithmscaneffectivelyanalyzethesecomplexspectral largeannotatedhyperspectraldatasets,environmen- datasetsandaccuratelyclassifyhealthyandinfected talvariability,illuminationeffects,spectralnoise,com- plantsundervaryingenvironmentalconditions.Tradi- putationalcomplexity,overfittingissues,andlimited tionalmachinelearningtechniquessuchasSupport generalizationunderreal-fieldconditions. Highsen- VectorMachine(SVM),RandomForest(RF),k-Nearest sorcostsandtheneedforpowerfulcomputationalin- Neighbor(kNN),ArtificialNeuralNetworks(ANN),and frastructurealsorestrictwidespreadadoptionamong DecisionTreeshavedemonstratedconsiderablesuc- small-scalefarmers.Futureresearchshouldfocuson cessinplantdiseaseclassificationtasks.Amongthese developinglightweightandcomputationallyefficient methods,SVMandRFmodelshaveshownexcellent deeplearningmodelssuitableforreal-timefieldap- performanceinhandlinghigh-dimensionalhyperspec- Technology/Approach MajorAdvantages Limitations References Low-cost,rapidimaging, Limitedspectral MultispectralImaging [27][28] fieldapplicability resolution Highspectraldetail, Highcomputational HyperspectralImaging [33][42] earlystressdetection complexity SupportVectorMachine Effectiveforsmalldatasets, Sensitiveto [34][44] (SVM) high-dimensionaldata parametertuning Handlesnoisydata, Largeensemble RandomForest(RF) [45] featureimportanceanalysis complexity Automaticfeatureextraction, CNNModels Requireslargedatasets [55][60] highaccuracy Reducedtrainingtime, Domainadaptation TransferLearning [51][58] bettergeneralization issues Large-scalefieldcoverage, UAV-basedMonitoring Weatherdependency [38][47][54] real-timemonitoring Captureslong-range Computationally VisionTransformers [52] dependencies intensive plications.Emergingtechnologiessuchasfederated Disease,100(2):241–251,2016.
learning, self-supervised learning, multimodal sen- 4. C. H. Bock, G. H. Poole, P. E. Parker, and T. R. sorfusion,IoT-integratedsmartfarmingsystems,and Gottwald. Plantdiseaseseverityestimatedvisu- transformer-basedarchitecturesareexpectedtofur- ally,bydigitalphotographyandimageanalysis, therimprovediseasedetectionperformanceandscal- andbyhyperspectralimaging.CriticalReviews ability. Additionally, integratingmultispectralimag- inPlantSciences,29(2):59–107,2020.
ingwiththermalimaging,fluorescencesensing,LiDAR, 5. G.N.Agrios. PlantPathology.5thedition,Else- andgenomicinformationmayprovidemorecompre- vierAcademicPress,2005.
hensivecrophealthassessmentsystemsinthefuture. 6. N.W.Schaad,J.B.Jones,andW.Chun. Labora- Overall,machinelearningmodelscombinedwithmul- toryGuideforIdentificationofPlantPathogenic tispectralimagingrepresentapromisingandtransfor- Bacteria.APSPress,2003.
mativeapproachforsustainablecropprotectionand 7. S.Sankaran,A.Mishra,R.Ehsani,andC.Davis. A precisionagriculture.Thesetechnologieshavethepo- reviewofadvancedtechniquesfordetectingplant tentialtosignificantlyimproveagriculturalproductiv- diseases.ComputersandElectronicsinAgricul- ity,reduceexcessivepesticideapplication,minimize ture,72(1):1–13,2010.
economiclosses,andenhanceglobalfoodsecurityun- 8. S.J.PethybridgeandS.C.Nelson. LeafDoctor:A derchangingclimaticconditions.Continuedadvance- new portable application for quantifying plant ments in artificial intelligence, remote sensing, and disease severity. Plant Disease, 99(10):1310– sensortechnologiesareexpectedtoacceleratethede- 1316,2015.
velopmentofintelligent,affordable,andscalableplant 9. J. S. West, C. Bravo, R. Oberti, D. Lemaire, D. diseasedetectionsystemsfornext-generationsmart Moshou,andH.A.McCartney. Thepotentialof agriculture.
optical canopy measurement for targeted con- troloffieldcropdiseases.AnnualReviewofPhy- topathology,41:593–614,2003.