Pranilllllll/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
arch version	unetformer-r18-v2 (sig 29554ac657c8)
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)	0.0003 / 3e-05, AdamW wd 0.01
schedule	500-step warmup, cosine over 120 epochs
regularization	EMA 0.999, label smoothing 0.05, drop path 0.1, dropout 0.1
aux head weight	0.4
weights	EMA
best epoch	80
val mIoU	0.5007
val mF1	0.6505
val OA	0.7755
val kappa	0.6465

Per-class (validation)

class	IoU	F1
Residential	0.8208	0.9016
Road	0.3525	0.5213
River	0.3918	0.5630
Forest	0.6285	0.7719
UnusedLand	0.3306	0.4969
Agricultural	0.4797	0.6484

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

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

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 (EMA weights) plus cfg, metrics, arch_sig.