Pranilllllll/geonusaf-segNext-random-fold2

GeoNUSAF - SegNeXt-T - random split, fold 2

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

field	value
architecture	MSCAN-T encoder + LightHamHead decoder (SegNeXt, NeurIPS 2022)
encoder init	ImageNet-1K, 100.0% of tensors loaded
params	4.23 M
head fuse stride	8 (stages [1, 2, 3])
NMF rank / steps	R=16, 6 train / 7 eval
split mode	random
fold	2 of 3
seed	42
input	512x512, ImageNet norm, effective GSD 0.586 m/px
classes	Residential, Road, River, Forest, UnusedLand, Agricultural
lr (head/encoder)	0.0006 / 6e-05
regularization	wd 0.01, drop_path 0.1, smooth 0.05, EMA True
best epoch	191
val mIoU	0.5450
val mF1	0.6914
val OA	0.7990
val kappa	0.6848

Per-class (validation)

class	IoU	F1
Residential	0.8304	0.9073
Road	0.4453	0.6162
River	0.4574	0.6277
Forest	0.6735	0.8049
UnusedLand	0.3315	0.4979
Agricultural	0.5317	0.6943

best.pt holds model_state (EMA weights when EMA is on), arch (the dict needed to rebuild the network), the run cfg, and metrics. Rebuild with segnext_model.py from this same repo. Model code derives from Visual-Attention-Network/SegNeXt (Apache-2.0).