农作物病虫害识别技术的发展综述

ArticleCNN architectureValidationDeep learning frameworkCropDatasetPerformance metric[8]Faster R-CNNDifferent feature extractor CNNs: AlexNet, VGG-16, GoogLeNet, ZFNet, ResNet-50, ResNet-101 and ResNetXt-101CaffeTomato5000 self-acquired field images representing 9 classes of pests and diseasesmAP (0.8306)[9]CaffeNet (AlexNet)noneCaffeApple, pear, grape,cherry, peach4483 Self-acquired field imagesAcc(96.3%)[14]LeNetnoneDeeplearning4jBananaPlantVillage (3700 field images)Acc(0.9861), precision(0.9867),recall (0.986),F1-score (0.9864)[16]GoogLeNetnoneMatlab14 crops1575 self-acquired field imagesAcc(94%)[18]AlexNetMLPCaffeApple1450 Self-acquired lab imagesAcc(96.6%)[20]Custom CNN (AlexNet precursor with cascade inception)SVM, BP ,AlexNet, GoogLeNet, ResNet-20,VGG16CaffeApple1053 self-acquired lab imagesAcc(97.62%)[22]Modified LeNetSIFT + RBF-SVM, background suppressing Gabor energy filtering + RBF-SVM,uLBP + RBF-SVMMatlabOliveSelf-acquired lab imagesAccuracy (98.6%), Matthew’s correlation coefficient(0.9798),F1-score (0.9689),Precision (0.9882),Recall (0.9718)[26]AlexNet, GoogLeNetSVM, Random ForestDIGITSTomatoPlantVillageAcc(99.185%), Macro precision (98.529%),Macro recall (98.532%), Macro F score (98.518%)[27]AlexNet, DenseNet169,Inceptionv3, ResNet34,SqueezeNet, VGG13nonepyTorch14 crops, 26 diseasesPlantVillageAcc(0.9976)[28][29]Custom CNNnoneCaffeCucumberSelf-acquired field images(7320 good condition images& 7520 bad condition images)Acc(82.3%), sensitivity (79.9%), specificity (92.6%)[31]AlexNet, GoogLeNetnoneCaffe14 crops, 26 diseasesPlantVillageF1-score (0.9934),Precision (0.9935),Recall (0.9935),Acc(0.9935)[33]Custom ensemble CNNnoneKeras on TheanoMaize1796 self acquired field images (1028 infected with NLB and 768 healthy)Acc(0.967)[35]Faster R-CNN (modified)Faster R-CNN (unmodified)MatlabSugar beet155 self-acquired images (38healthy, 20 mild, 35 severe,62 mixed mild and severe infection)Sensitivity (95.48%),Specificity (95.48%),Acc(95.48%)[69]MobileNetsnoneCaffeTomatoPlantVillageAcc(90.3%)[70]Inceptionv3, ResNet-50,VGG19, XceptionTransfer learning (TF),Feature tuning 25, 50, 75,100%, No TFKerasSoybean300 high resolution field images divided into 3000 superpixel imagesAccuracy (99.04%),Learning error (0.049)[68]DCGAN, ProGAN and StyleGAN for generation of synthetic training images Custom CNN – PDNet withYolo/AlexNet detector and 32 layer residual classifier networkAlexNet, Vgg19, Inceptionv3,DenseNet201, ResNet152Not stated12 crop species, 42 diseases18,334 PlantVillage lab images + 79,265 self-acquired field images (Plant disease dataset)Detection mAP (0.9165) Recognition acc(0.9367)[67]Faster R-CNN with VGG16 feature extractornoneCaffeTomato8927 self-acquired images representing 9 anomalies and one class for backgroundmAP (96%)[37]GPMNet – (GoogLeNet)noneMatlabGrape14,180 self-acquired lab sub-imagesAcc(94.3%),AuC (0.984)[52]INAR-SSDFaster R-CNN, SSD, R-SSDCaffeApple2029 lab and field imagesmAP (78.8%)Detection speed (23.13 FPS)[53]AlexNetnoneNot statedRice227 self-acquired images(both pest and disease)Acc(91.23%)[54]Modified Cifar10Basic Cifar10 and GoogLeNet CNNsCaffeMaize (8 diseases + healthy)500 imagesAcc(98.9%)[56]Custom CNNVGG16, VGG19, Inceptionv3,ResNet-50Keras on TheanoApplePlantVillage (healthy & black rot images)Acc(90.4%)[57]Custom CNNnoneNot statedSoybeanPlantVillageAcc(99.21%),Recall(0.99),Precision (0.99),F1-Score (0.99)[71]AlexNet, SqueezeNetnoneCaffe on Nvidia JetsonTomatoPlantvillageAcc(0.943)[72]Custom CNNBP , SVM, PSOMatlabRice500 self-acquired field images representing 10 rice diseasesAcc(95%),Missing report rate (0),False report rate (0)[73]Modified Inceptionv3Inceptionv3Keras on TensorFlowTomatoPlantVillageAcc(0.971)[74]CaffeNet up to fc layers + hand crafted features + SVM classifierHand crafted + SVM, CNN,fc6/7 + SVMCaffeTrees – 57 speciesPlantNet, Flavia and LeafSnap + self-acquired imagesAcc(99.68%)[88]Custom 3 channel CNNImage processing technology(IPT), Global-Local Singular Value Decomposition(GLSVD), SVM, Sparse Representation based Classification (SRC)MatlabTomato, cucumberPlantVillage (tomato), 500self-acquired images(cucumber)Accuracy (94.15% for cucumber and 91.16% for tomato)[89]AlexNet, ResNet-50, ResNet-101SVM, BPCaffePests550 self-acquired field images (10 species each having 55 images)Acc(98.67%)[106]VGG16, AlexNetnoneMatlabTomatoPlantVillage (6 tomato diseases+ healthy)Acc(96.19%)[120]Modified ResNet-50Shallow classifiers in [7], Re-sNet-50 with fully convolutional dense layerTensorFlowWheat8178 self-acquired images(3338 Rust, 2744 Septoria,1568 Tan Spot, 1116 healthy)AuC (0.97),Sensitivity (0.91), Specificity(0.95),Balanced acc(0.93)[107]Modified CIFAR10-quick CNNLeNet-5, AlexNet, VGG16, BP,Bayesian, SVM, KNN classifiersMatlabTea144 self-acquired images (36 images each of healthy, tea leaf blight, tea bud blight, tea red scab)Accuracy (92.5%),Loss (0.002)[123]AlexNet, VGG16Fine tuning, training from scratch, CNN feature extractors + SVM/LDA, CNN feature extractors + RNNNot stated14 crop species, 87 tree species obtained from 3 public datasetsPlantVillage (54,306 images),Flavia (1907 images),Swedish leaf dataset (1125),UCI leaf dataset (443 images)Acc Flavia –VGG16 + LDA(99.10%),Swedish leaf – VGG16(99.11%),UCI leaf – AlexNet + LDA(96.2%),PlantVillage – VGG16 (99.8%)[124]Custom CNN (MCNN)PSO, SVM, RBFNNTensorFlowMango2200 images (512 healthy mango, 558 infected mango,520 other leaves healthy, 610 other leaves infected)Accuracy (97.13%),Missing report rate (2.87),False report rate (0)