CITRUS DISEASE CLASSIFICATION WITH TRANSFER LEARNING AND CNN BASED MODELS

Abstract

In recent years, image processing and deep learning have been widely used in the detection and classification of plant diseases. These uses offer great opportunities for the early detection of plant diseases in agriculture. Early detection of the disease is essential to prevent disease symptoms from spreading to intact leaves and to reduce crop damage. For the stated reasons, a deep learning model with three different approaches has been proposed and used for the classification of diseases that are most common in citrus leaves and affect citrus export to a great extent. Training and test data used in the proposed model are separated according to the K-fold 5 value. For this reason, the average of the performance values obtained according to the K-fold 5 value is presented in the study. As a result of the experimental studies, with the fine-tuned DenseNet201 model, which is the first model, an accuracy rate of 0.95 was achieved. In the second model, with the proposed 21-layer CNN model, an accuracy rate of 0.99 was achieved. The third model is defined to show the progress of the proposed DenseNet201 model over the basic DenseNet201 model. With the CNN method recommended for the classification of citrus grades, Blackspot (citrus black spot (CBS), canker (citrus bacterial cancer (CBC)), greening (huanglongbing (HLB)), and (healthy) Healthy) 100%, 100%, 98% and 100% rates have been reached.