Table 2.

Quantitative results under ID shift, where $\mathcal{X}$ = {jitter, mirroring, rotation} applied to evaluation set. The first rows for both datasets with no $\mathcal{X}$ (under no-shift scenario) serve as the baselines. Results are reported for 1-way 20-shot tasks. The best performance metrics are shown in bold, and the second-best are underlined. Only the model with the lowest predictive uncertainty is highlighted in bold.

Evaluation Set	$\mathcal{X}$	mIoU	OA	MCC	$\mathit{mH}$
Infrabel-5	–	89.10	92.48	0.8250	0.0938	⇒ baseline
Infrabel-5	jitter	89.07	92.47	0.8247	0.0951
Infrabel-5	mirroring	89.02	92.43	0.8237	0.0952
Infrabel-5	rotation	89.01	92.42	0.8236	0.0938
WHU	–	83.04	89.02	0.7714	0.3254	⇒ baseline
WHU	jitter	82.89	88.86	0.7682	0.3261
WHU	mirroring	82.95	88.97	0.7701	0.3255
WHU	rotation	82.94	88.96	0.7700	0.3256