Enhanced Automatic Root Recognition and Localization in GPR Images Through a YOLOv4-Based Deep Learning Approach

Abstract

In recent years, ground penetrating radar (GPR) has become increasingly important as a nondestructive way to explore plant roots. Automatic recognition and localization of root objects from GPR images presents a significant challenge. GPR images for the root system contain complicated hyperbolic signals that appear deformation depending on root size, orientation, aggregation degree and soil background. This paper presents a new deep learning approach, YOLOv4-hyperbola, that provides fully automatic recognition and localization of root objects from GPR images. YOLOv4-hyperbola improves the YOLOv4 (You Only Look Once v4) architecture by introducing keypoints detection branch in order to accurately locate roots while identifying them. The YOLOv4-hyperbola model was trained by combining field datasets and simulated datasets to simultaneously identify and locate hyperbolic features representing potential root objects across GPR images, and evaluated on datasets of root detection from two experiments in the field. Compared with Randomized Hough transform (RHT) method, the proposed approach demonstrated higher accuracy and efficiency in root object detection on GPR image. YOLOv4-hyperbola was able to accurately recognize and locate abnormal hyperbolic signals caused by the complexity of root system in nature. The validation on the two independent datasets showed that the proposed approach had good generalization and great application potential for real-time detection and location of roots over large areas in the field.