Oil Palm Disease Lesion-Level Segmentation Using YOLO based Framework

This article reports on a machine learning based methodology for identifying lesions associated with oil palm disease from RGB images. The research has significantly expanded upon the Oil Palm Ganoderma Detection Dataset through converting COCO formatted annotations to segmentations in polygon format to be used to train a segmentation-based YOLO model. As opposed to other methodologies that typically detect objects rather than segments diseases at the lesion-level, the developed methodology has the capability to provide higher interpretive value than traditional classification or detection methodologies. It was shown in experimental trials that the methodology can perform well (Accuracy = 84.6%, Precision = 86.3%, Recall = 84.6%, F1-Score = 85.4%, Segmentation mAP @ 50 = 0.735) which demonstrates its ability to capture spatial properties of disease.