Healthy- Disease Mustard Leaf Set: A Dataset for Mustard Disease Detection

Introduction: Agriculture is vital to the economy, particularly in developing countries in which it may serve as the primary source of employment as well as revenue. Plant diseases cause India to lose 10-35% of its yearly crop yield. The absence of lab equipment and experience makes early plant disease identification challenging. One of the biggest obstacles to enabling vision-based plant disease identification is the lack of a suitable large-scale non-lab dataset. Objectives: The main objective of this study is to introduce Healthy-Disease Mustard Leaf Set (HDMLS), a dataset consisting of both healthy and diseased mustard leaves to promote deep- learning based identification of Alternaria Blight and White Rust disease, evaluating its effectiveness in terms of Accuracy, Precision, F1-Score, mIoU, mAP, Specificity and Sensitivity. Methods: Approximately 350 person hours were spent clicking photos for our dataset, which has 2889 images of mustard plant leaves containing 1 healthy and 4 diseased classes. To increase the diversity of the images and model’s capacity for generalization, data augmentation is applied on the dataset increasing its size to 5,663. Pre-processing is done to remove unneseccary features. With an 80-20 split, dataset is separated into training, validation and test sets. Three different state-of-art CNN models- ResNet-50, Xception and DenseNet are applied to both our self-created dataset and PlantDoc dataset which is freely available. Performances are then evaluated and compared. Results: Confusion matrices have been developed for every model trained on the dataset in order to assess the models. Model Performance is compared using metrics such as precision, recall, accuracy, FPR, F1-Score, Specificity, Sensitivity, m AP, mIoU. 20 epochs are used to evaluate results. Xception model and ResNet-50 model achieve same highest accuracy of 0.88. ResNet-50 model also achieve high mAP (0.70) and mIoU (0.52). DenseNet model also achieves good accuracy (0.84) and mAP (0.63) but it has low mIoU (0.35). Conclusions: ResNet-50 performs exceptionally well in terms of mAp, , mIoU, and accuracy indicating that it can recognize, classify, and segment objects with superior performance.