Detection of Cotton Leaf Diseases Using Transfer Learning Techniques

Abstract

Cotton crop is among the leading source of income globally through its production is harmed by many diseases which attack the leaves. Identification of these diseases is a tedious, cumbersome, and prone to a lot of errors hence the need to come up with automatic systems to detect them. This study presents a transfer learning-based approach for the detection and classification of cotton leaf diseases, focusing on four primary categories: bacterial blight, fusarium wilt, curl virus and apparently healthy one. To support training and evaluation, 3418 images were gathered and preprocessed; this involved resizing and normalization and augmentation to ensure generalization by the developed model. The following five modern transfer learning models: ResNet50, VGG16, DenseNet201, InceptionV3, and InceptionResNetV2 were adjusted to the dataset. Accuracy was used to determine the ability of the models; Densenet201 recorded the highest value of 99.71%, with VGG16 recording 99.41%, while InceptionResNetV2 recorded 98.98%. InceptionV3 made improved performance compared to the other models in the test with the test accuracy 98.10 %, however Resnet50 had lower accuracy at 81.14%. Accordingly, owing to the employment transfer learning the worksheets incorporated proficient at feature extraction and disease pattern discovery in cotton leaves. It is proved that the proposed system is not only effective for diseased and healthy leaves and also for the prediction that may be uncertain at the end. This work gives a solution in which cotton farmers will gain from as it can highlight on the disease and perhaps reduce crop losses as they can be determined early.