Pranilllllll/geonusaf-s4-unet-r18-R0-block-fold1

GeoNUSAF Stage 4 - UNet-ResNet18 - arm R0 - block fold 1

Kathmandu Valley land-use segmentation, 6 classes, ignore_index=255. Weights are the EMA weights (decay 0.999), not the raw final weights.

field	value
architecture	smp.Unet, encoder resnet18 (ImageNet), decoder [128, 64, 32, 16, 8]
params	12.46 M
arch version	unet-r18-v1
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	0.4572
val mF1	0.5914
val OA	0.7855
val kappa	0.6289

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

class	IoU	F1
Residential	0.8260	0.9047
Road	0.3617	0.5313
River	0.1587	0.2739
Forest	0.6510	0.7886
UnusedLand	0.2202	0.3609
Agricultural	0.5255	0.6889

Validation contains no synthetic pixels in either arm.

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