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

GeoNUSAF Stage 4 - UNetFormer (ResNet-18) - arm R0 - block fold 1

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

field	value
architecture	UNetFormer, timm resnet18 encoder (ImageNet), global-local attention decoder
arch version	unetformer-r18-v2 (sig 29554ac657c8)
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
input	512x512, ImageNet norm, working GSD 1.374 m/px
lr (dec/enc)	0.0003 / 3e-05, AdamW wd 0.01
schedule	6000 steps, warmup 500, cosine 6000
regularization	EMA 0.999, label smoothing 0.05, drop path 0.1, dropout 0.1
aux head weight	0.4
class weights	from real
weights	EMA
seed	42
best step	6000
val mIoU	0.5055
val mF1	0.6470
val OA	0.8054
val kappa	0.6597

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

class	IoU	F1
Residential	0.8358	0.9106
Road	0.3662	0.5361
River	0.3290	0.4951
Forest	0.6959	0.8207
UnusedLand	0.2613	0.4143
Agricultural	0.5450	0.7055

Validation contains no synthetic pixels in either arm.

Not comparable to the part-1 fold-1 UNetFormer run: that one used an epoch ceiling, epoch-granularity early stopping and epoch-seeded augmentation.

Architecture is an independent implementation of Wang et al. (2022), ISPRS J. Photogramm. Remote Sens. 190:196-214 (the reference repo is GPL-3.0).

LR deviates from the UNetFormer paper's 6e-4 (reduced to 3e-4 for parity with the other baselines at this dataset scale).

Checkpoint best.pt holds model_state (EMA weights) plus cfg, metrics, arch_sig.