sugam24/geonusaf-s4-tcsegformer-R0-block-fold1

GeoNUSAF Stage 4 - TC-SegFormer - arm R1 - block fold 1

Kathmandu Valley land-use segmentation, 6 classes, ignore_index=255, 512x512.

field	value
arm	R0 (real only)
train rows	804 (804 real + 0 synthetic)
synthetic source	sugam24/geonusaf-stage3-fakepairs-block-fold1, fake_pairs_R1
split	block fold 1, train_sha1 eb0aebdd9919
backbone	nvidia/segformer-b0-finetuned-ade-512-512
detail path	True (ch=[8, 16, 32], fuse=64)
CSA	True (tau=[0.6, 0.35, 0.35, 0.6, 0.6, 0.6], w_min=0.25, on_fake=True)
soft-clDice	True (mu=0.3 from step 1000)
balanced sampler	True (cap 8.0, freq from real)
schedule	6000 steps, warmup 150, cosine 6000
class weights	from real
seed	42
best step	6000
val mIoU	0.5608
val mF1	0.7001
val OA	0.8293
val kappa	0.6931

Per-class (validation, 136 real fold-1 tiles)

class	IoU	UA (prec)	PA (rec)
Residential	0.8456	0.9401	0.8938
Road	0.4279	0.5276	0.6936
River	0.4761	0.6154	0.6778
Forest	0.7462	0.8919	0.8204
UnusedLand	0.2966	0.5979	0.3704
Agricultural	0.5722	0.6214	0.8784

Validation contains no synthetic pixels in either arm.

Not comparable to the part-1 fold-1 run: that one used a 260-tile train split, had frozen augmentation (the persistent_workers bug), determinism on, and an epoch-based schedule.