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

Table 2 Performance of different techniques/models (Table 4) on a test data subset of the mixed data set, which was not used during the training on the mixed data set

Technique/model (+ aug)	SSIM	DSC	IoU	FPR
Dummy classifier^b	0.9173	0.3250	0.3250	-
Frangi Vesselness	0.3665	0.1009	0.0626	1.0
Adaptive thresholding	0.8348	0.2367	0.1804	0.8636
SVM	0.7617	0.1744	0.1341	0.9090
SegRoot^ab	0.9173	0.3250	0.3250	-
SegRoot + aug^ab	0.9173	0.3250	0.3250	-
UNetGNRes	0.9246	0.4399	0.3585	0.6363
UNetGNRes + aug	0.9313	0.5326	0.4452	0.4545
U-Net SE-ResNeXt-101 (32 × 4d)	0.9352	0.5708	0.4800	0.3636
U-Net SE-ResNeXt-101 (32 × 4d) + aug	0.9360	0.6217	0.5299	0.2272
U-Net EfficientNet-b6	0.9375	0.6418	0.5498	0.1363
U-Net EfficientNet-b6 + aug	0.9381	0.6848	0.5920	0.0454

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 used: Training: 557/augmented Training: 2228, Validation: 62, Testing: 69. Models trained with augmented data are indicated by + aug
^aNo differences in model performance were found using + aug in the SegRoot model
^bEXCLUDED from FPR determination as not predicting roots. indicate methods where the labels are zero or do not predict roots

ISSN: 1746-4811