A Machine Learning Hybrid Approach for Diagnosing Plants Bacterial and Fungal Diseases

Abstract

Bacterial and Fungal diseases may affect the yield of stone fruit and cause damage to the Chlorophyll synthesis process, which is crucial for tree growth and fruiting. However, due to their similar visual shot-hole symptoms, novice agriculturalists and ordinary farmers usually cannot identify and differentiate these two diseases. This work investigates and evaluates the use of machine learning for diagnosing these two diseases. It aims at paving the way toward creating a generic deep learning-based model that can be embedded in a mobile phone application or in a web service to provide a fast, reliable, and cheap diagnosis for plant diseases which help reduce the excessive, unnecessary, or improper use of pesticides, which can harm public health and the environment. The dataset consists of hundreds of samples collected from stone fruit farms in the north of Jordan under normal field conditions. The image features were extracted using a CNN algorithm that was pre-trained with millions of images, and the diseases were identified using three machine learning classification algorithms: (1) K-nearest neighbour (KNN); (2) Stochastic Gradient Descent (SGD); and (3) Random Forests (RF). The resulting models were evaluated using 10-fold crossvalidation, with CNN-KNN achieving the best AUC performance with a score of 98.5%. On the other hand, the CNN-SGD model performed best in Classification Accuracy (CA) with a score of 93.7%. The results shown in the Confusion Matrix, ROC, Lift, and Calibration curves also confirmed the validity and robustness of the constructed models.

Enzyme dosage detection to degrade feathers in edible bird’s nests: A comparative convolutional neural networks study

“…The 70:30 dataset split ratio was chosen as it has been proven to demonstrate better performance of CNN compared to the ratios of 80:20 and 60:40 in Distributed Denial of Service classification (Gadze et al, 2021). The training data was used to train the model to classify images into 30 enzyme classes, while the validation data was used to evaluate the accuracy of the trained model in classifying unseen images during training (Ho et al, 2020;BaniMustafa et al, 2023).…”