Deep learning based multiclass classification for citrus anomaly detection in agriculture

Abstract

In regions where citrus crops are threatened by diseases caused by fungi, bacteria, pests and viruses, growers are actively seeking automated technologies that can accurately detect citrus anomalies to minimize economic losses. Recent advances in deep learning techniques have shown potential in automating and improving the accuracy of citrus anomaly categorization. This research explores the use of deep learning methods, specifically DenseNet, to construct robust models capable of accurately distinguishing between different types of citrus anomalies. The dataset consists of high-resolution images of different orange leaves of the species Citrus sinensis osbeck, collected from orange groves in the states of Tamaulipas and San Luis Potosi in northeastern Mexico was used in study. Experimental results demonstrated the effectiveness of the proposed deep learning models in simultaneously identifying 12 different classes of citrus anomalies. Evaluation metrics, including accuracy, recall, precision and the confusion matrix, underscore the discriminative power of the models. Among the convolutional neural network architectures used, DenseNet achieved the highest classification accuracy at 99.50%. The study concluded by highlighting the potential for scalable and effective citrus anomaly classification and management using deep learning-based systems.