Pranilllllll/geonusaf-s4-segformer-b0-R0-block-fold1

GeoNUSAF Stage 4 - SegFormer-B0 - arm R0 - block fold 1

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

field	value
arm	R0 (real only)
train pairs	804 (804 real + 0 synthetic)
synthetic source	sugam24/geonusaf-stage3-fakepairs-block-fold1, fake_pairs_R1
split	block fold 1, train_sha1 eb0aebdd9919
schedule	6000 steps, warmup 500, cosine 6000
class weights	from real
seed	42
best step	5200
val mIoU	0.4637
val mF1	0.5929
val OA	0.7786
val kappa	0.6300

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

class	IoU	F1
Residential	0.7918	0.8838
Road	0.2905	0.4502
River	0.1441	0.2518
Forest	0.7444	0.8535
UnusedLand	0.2485	0.3981
Agricultural	0.5625	0.7200

Validation contains no synthetic pixels in either arm.

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