Ten-Layered Deep Convolution Neural Network-Based Tea Leaf Disease Prediction

Abstract

Pests and pathogens have a significant impact on overall food production systems, resulting in massive losses in terms of volumes, reliability, survivability, and economy. Investigational evaluations and geometric experiments are two typical appropriate diagnostic methodologies designed to automatically reduce plant economic loss. Plant pathogens have quite a worldwide effect on the agricultural industry, leading to significant industrial and post-harvest damages. Timely identification of plant and diseases is essential to maintaining long-time survivability, and it makes it very difficult for research scientists. Deep learning techniques could be used to identify the tea leaf ailment using this overview. In this paper, the tea leaf images from the Mendeley data repository were used for implementation with 160 images under four classes as healthy, red scab, red leaf spot, and leaf blight. The augmentation process is done ending with 1040 augmented images, thereby forming input dataset with 1200 images. The training dataset is formed with 920 images, with validation dataset of 80 images, and system is tested with 200 images. The dataset is preprocessed starting with the train-equipped layers of ten-layered Conv2D base model. The base model is then pretrained with ImageNet with categorical cross-entropy as training loss function along with Adam optimizer and the trainable parameters as 150,477,897. To obtain the high-performance key characteristics of tea leaf image data, the base model is added to develop custom layers using a transfer deep learning approach including several CNN models like VGG-19, 10-DCNN, Inception V3, ResNet-50, MobileNetV2, and DenseNet20. Performance metrics such as RoC curve, model loss, sensitivity, precision, specificity, accuracy, and F1-score are used to assess the effectiveness of ten-layered CNN frameworks with transfer learning. The project is implemented in Python and runs on NVidia Tesla V100 GPU server with 1000 training iteration and batch size of 32. Experimental results show that 10-DCNN model is found to exhibit 99.25% accuracy, precision and sensitivity of 98.5, and specificity and F1-score of 99.5.

Keywords