A Deep Learning Approach For Anthracnose Infected Trees Classification In Walnut Orchards

This paper presents a novel approach for the detection of disease-infected leaves on trees with the use of deep learning. Focus of this study was to build an accurate and fast object detection system that can identify anthracnose-infected leaves on walnut trees, in order to be used in real agricultural environments. Similar studies in the literature address the disease identification issue; however, so far, the detection was performed on single leaves which had been removed from trees, using images taken in controlled environment with clear background. A gap has been identified in the detection of infected leaves on tree-level in real-field conditions, an issue which is tackled in our study. Deep learning is an area of machine learning that can be proved particularly useful in the development of such systems. The latest developments in deep learning and object detection, points us towards utilizing and adapting the state-of-the-art single shot detector (SSD) algorithm. An object detector was trained to recognize anthracnose-infected walnut leaves and the trained model was applied to detect diseased trees in a 4 ha commercial walnut orchard. The orchard was initially inspected by domain experts identifying the infected trees to be used as ground truth information. Out of the 379 trees of the orchard, 100 were randomly selected to train the object detector and the remaining 279 trees were used to examine the effectiveness and robustness of the detector comparing the experts? classification with the predicted classes of the system. The best inputs and hyper parameter configuration for the trained model provided an average precision of 63% for the object detector, which correctly classified 87% of the validation tree dataset. These encouraging results indicate that the detector shows great potential for direct application in commercial orchards, to detect infected leaves on tree level in real field conditions, and categorize the trees into infected or healthy in real time. Thus, this system can consist an applicable solution for real-time scouting, monitoring, and decision making.