sugam24/geonusaf-tcsegformer-random-fold0

GeoNUSAF TC-SegFormer -- random split, fold 0

Kathmandu Valley land-use segmentation, 6 classes, ignore_index=255, 512x512 @ 0.586 m/px.

backbone	nvidia/segformer-b0-finetuned-ade-512-512
split mode	random (sequence-block CV, export-order proxy)
fold	0 of 3
seed	42
best epoch	119
detail path	True
CSA reweighting	True (tau=[0.6, 0.35, 0.35, 0.6, 0.6, 0.6], w_min=0.25)
soft-clDice	True (mu=0.3)
balanced sampler	True

Validation

mIoU 0.5525 | mF1 0.7012 | OA 0.7935 | kappa 0.6922

class	IoU	UA (prec)	PA (rec)
Residential	0.8093	0.9216	0.8692
Road	0.4432	0.5260	0.7380
River	0.4537	0.6009	0.6494
Forest	0.6628	0.7934	0.8011
UnusedLand	0.3947	0.5449	0.5889
Agricultural	0.5512	0.6912	0.7313