sugam24/geonusaf-unet-r18-random-fold2

GeoNUSAF - UNet-ResNet18 - random split, fold 2

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
split mode	random
fold	2 of 3
seed	42
input	512x512, ImageNet norm, effective GSD 0.586 m/px
regularization	wd 0.01 (norm/bias exempt), ls 0.05, drop 0.1, EMA 0.999
classes	Residential, Road, River, Forest, UnusedLand, Agricultural
best epoch	77
val mIoU	0.4484
val mF1	0.5859
val OA	0.7560
val kappa	0.6248

Per-class (validation)

class	IoU	F1
Residential	0.8104	0.8953
Road	0.3750	0.5455
River	0.1349	0.2377
Forest	0.6457	0.7847
UnusedLand	0.2777	0.4347
Agricultural	0.4469	0.6177

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