A Novel Deep Feature Extraction Approach Based on DenseNet201 and ResNet50 for Cotton Leaf Disease Detection

In this study, a new deep feature extraction approach is proposed for automatic detection of diseases observed on cotton plant leaves. In the proposed approach, feature extraction is performed using DenseNet201 and ResNet50 deep learning architectures. and the obtained feature vectors are combined. Then, the most informative features are selected with the Iterative Chi2 algorithm, and disease detection is performed using the Support Vector Machine (SVM) classifier. The developed model is tested on an open access dataset consisting of 2,137 cotton leaf images and 7 different classes (1 healthy, 6 diseased). 10-fold cross-validation and 80:20 hold-out cross-validation strategies are applied in the testing phase. As a result of the tests performed without using any data augmentation technique, 97.29% and 96.96% classification accuracies are obtained, respectively. The proposed approach makes significant contributions to the literature in terms of showing high success on the imbalanced dataset and providing a computationally lightweight architecture.