Article | Année | Single/ Multi task | Crop | Part | Diseases | Severity Grade/ Level | Dataset | Models used | Results |
[18] | 2017 | single | apple | leaf | black rot | Healthy, Early, Middle and End | Plant Village | Lightweight CNN, VGG16, ResNet50 and Inception-V3 | Best accuracy with VGG16: 90.4% |
[10] | 2019 | Multi-task | apple, grape, cherry, peach, pepper, tomato, Strawberry, Potato, Corn curvularia, Puccinia polysora, Cercospora zeaemaydis… | Leaf | * | general and serious | Synthetic dataset from AI Challenger Global AI Contest (www.challenger.ai) and from others reseach works | PD2SE-Net50 | Accuracies of 0.99, 0.98 and 0.91, respectively for species recognition, disease classification and disease severity estimation. |
[12] | 2020 | single | Tomato | Leaf | Early Blight | healthy, mild, moderate, and severely diseased leaves | Plant Village (1000 images) | ResNet101, VGG16, VGG19, GoogLeNet, AlexNet, and ResNet50 | accuracy: 94.6% |
[22] | 2020 | single | Citrus (sweet orange) | Leaf | Huanglongbing (HLB) or Citrus Greening disease or citrus “cancer” | Early Stage, Moderate Stage, and Severely Infected Stage | Plant Village and crowdAI (5406 images) | AlexNet, DenseNet-169, Inception v3, ResNet-34, SqueezeNet-1.1, and VGG13 | accuracy: 92.60% |
[25] | 2020 | multi-task, | Coffee | Leaf | Leaf miner, rust, brown leaf spot and cercospora leaf spot | healthy (<0.1%), very low (0.1% - 5%), low (5.1% - 10%), high (10.1% - 15%) and very high (>15%). | Self-collected dataset of 1747 images | AlexNet, GoogleNet, VGG19 and ResNet50 | Accuracies of 94.05% for biotic stress classication and 84.76% for severity estimation. |
[14] | 2020 | Multi-task | corn, grap, peach, pepper, patato, strawberry, tomato | leaf | Puccinia Polysora, Curvularia Leaf Spot Fungus, Black Rot Fungus, Black Measles Fungus, Bacterial Spot, Late Blight Fungus, Early Bligh Fungus, scorch, Leaf Mold Fungus | normal, general and serious | IA Challenger (12.691 images) | BR-CNN based on DenseNet121, InceptionV3, NasNet or ResNet50 | BR-CNN based on ResNet50 obtained the best accuracy (86.70%). |