Root Detection Methods

Exciting new research by Vienna Scientifics' Pavel Baykalov and Boris Rewald on root detection methodology in Minirhizotron, Rhizobox and Rhizotron / Root Window RGB root and rhizosphere images.

Manual analysis of mini-rhizotron images is time-consuming. While several methods exist for root segmentation in single-source datasets, recent advances in deep learning allow for automated feature extraction. This study compared six methods and proposes two improved deep learning models for automated, semantic root segmentation. The evaluation included performance on maize and various other species in minirhizotron and rhizotron /root window datasets of several experiments (Fig. 1), demonstrating the generalisation potential of the models on an unseen dataset. The following methodologies were compared:

• Frangi Vesselness
• Adaptive Threshold
• Support Vector Machine (SVM)
• SegRoot
• UNetGNRes
• U-Net

The focus of the research was on accuracy, false positives and transferability of these different MR root image analysis approaches, and how to improve root detection with advanced U-Net models.

U-net models outperformed others methods, with increased encoder complexity improving root segmentation accuracy and generalisation potential (Fig. 2). The mixed MR dataset, composed of different image qualities, roots and soil conditions, challenged non-U-Net techniques, but data augmentation improved performance. Deep meta-architectures and feature extractors showed improved results and reduced false positives.