Deep Convolutional Neural Network for Cattle Disease Identification

Abstract

Most of the Ethiopian population resides in rural regions, and the foundation of its economy relies on agriculture. Ethiopia is famous in Africa for being a prominent producer of cattle. Despite this fact, cattle's economic contribution is not satisfactory due to several factors, among which animal diseases trouble the cattle business productivity. Therefore, Early and accurate identification of cattle diseases could be among the key measures required for preventing disease in its early stages. Several efforts have been made to advance cattle disease identification; In this thesis, a deep residual CNN model, ResNet18, with channel and spatial attention mechanisms (CCSAM) added at each residual network basic block is proposed. The CCSAM is integrated into the basic blocks of the ResNet18 to extract disease-related features from spatial and channel dimensions. These features are combined via elementwise multiplication to amplify the disease feature and again added to the original cattle disease image feature map via skip connection to minimize the impact of model overfitting. Apart from the aforementioned model development, the proposed method also applied several geometrical and photometric augmentation techniques for improving the diversity of the dataset, increasing class distribution, and minimizing model overfitting. The proposed method was extensively trained on two datasets. The first dataset was collected from Haramaya University Veterinary Medicine, Chiro Woreda Veterinary Clinic, and West Hararghe Zone Veterinary Office. The second was extracted from the Kaggle open cattle disease dataset. To measure the overall performance of the proposed method, the mean aggregate was calculated for values of precision, recall, F1 score, and accuracy for each class. To further strengthen our evaluation, the confusion matrix and classification report metrics were used. In general, the fine-tuned proposed ResNet18-CCSAM method on a geometrically augmented dataset has obtained a precision of 99%, a recall of 98%, F1 score of 98%, and an accuracy of 98.97%. Compared to the other related research works, our proposed method has demonstrated better cattle disease identification performance.