sugam24/geonusaf-unetformer-r18-random-fold1

GeoNUSAF - UNetFormer (ResNet-18) - random split, fold 1

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

field	value
architecture	UNetFormer, timm resnet18 encoder (ImageNet), global-local attention decoder
split mode	random
fold	1 of 3
seed	42
input	512x512, ImageNet norm, effective GSD 0.586 m/px
classes	Residential, Road, River, Forest, UnusedLand, Agricultural
lr (dec/enc)	0.0006 / 6e-05, AdamW wd 0.0001
aux head weight	0.4
best epoch	33
val mIoU	0.5148
val mF1	0.6640
val OA	0.7774
val kappa	0.6427

Per-class (validation)

class	IoU	F1
Residential	0.8252	0.9042
Road	0.3881	0.5592
River	0.4677	0.6373
Forest	0.6327	0.7750
UnusedLand	0.3101	0.4734
Agricultural	0.4652	0.6350

Split caveat: block is sequence-block CV using an export-order proxy, not spatial CV.

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).

Checkpoint best.pt holds model_state plus cfg and metrics.