Semantic segmentation of plant roots from RGB (mini-) rhizotron images—generalisation potential and false positives of established methods and advanced deep-learning models

Table 3 The performance of different techniques/models (Table 4) trained on mixed data (incl. different species, soil types and imaging devices), predicting a different, unseen Chicory/RootPainter dataset

Technique/Model (+ aug)	SSIM	DSC	IoU	FPR
Dummy classifier^b	0.9732	0.4103	0.4103	–
Frangi Vesselness	0.3361	0.0054	0.0042	0.9984
Adaptive thresholding	0.9408	0.1506	0.1199	0.8814
SVM	0.8260	0.0845	0.0775	0.8862
SegRoot^b	0.9732	0.4103	0.4103	–
SegRoot + aug^a,b	0.9732	0.4103	0.4103	–
UNetGNRes	0.9611	0.2489	0.2276	0.5769
UNetGNRes + aug	0.9623	0.3707	0.3022	0.6858
U-Net SE-ResNeXt-101 (32 × 4d)	0.9764	0.4750	0.4156	0.3125
U-Net SE-ResNeXt-101 (32 × 4d) + aug	0.9780	0.5676	0.5043	0.1442
U-Net EfficientNet-b6	0.9771	0.5350	0.4739	0.1827
U-Net EfficientNet-b6 + aug	0.9784	0.6103	0.5411	0.1233

^aNo differences in model performance were found using + aug in the SegRoot model
^bExcluded from FPR determination as not predicting roots. Indicates methods tghat do not predict roots. The best scores for the evaluation metrics average structural similarity index (SSIM), average Sørensen-Dice similarity coefficient (DSC), and average Jaccard index/intersection over union (IoU), and the lowest false positive rate (FPR) are shown in bold. Number of images: Training: 557/augmented Training: 2228, Validation: 62, Testing: 1537. Models trained with augmented data are indicated by + aug

ISSN: 1746-4811