Sathya77/swin-transformer-satellite

Swin Transformer — Satellite Image Classification

PyTorch implementation of Swin Transformer (Liu et al. 2021) trained on NWPU-RESISC45 satellite imagery dataset.

Model Details

Property	Value
Architecture	Swin Transformer (4 stages)
Dataset	NWPU-RESISC45
Classes	45 land use categories
Test Accuracy	82%
Input Size	224×224
Embed Dim	96
Training Hardware	RTX 4050 6GB
Framework	PyTorch (from scratch)

Classes

airplane, airport, baseball_diamond, basketball_court, beach, bridge, chaparral, church, circular_farmland, cloud, commercial_area, dense_residential, desert, forest, freeway, golf_course, ground_track_field, harbor, industrial_area, intersection, island, lake, meadow, medium_residential, mobile_home_park, mountain, overpass, palace, parking_lot, railway, railway_station, rectangular_farmland, river, roundabout, runway, sea_ice, ship, snowberg, sparse_residential, stadium, storage_tank, tennis_court, terrace, thermal_power_station, wetland

Usage

from huggingface_hub import hf_hub_download
import torch
from torchvision import transforms
from PIL import Image

checkpoint = torch.load(
    hf_hub_download("Sathya77/swin-transformer-satellite", "swin_resisc45.pth"),
    map_location='cpu'
)

model = SwinTransformer(embed_dim=96, num_classes=45)
model.load_state_dict(checkpoint['model_state_dict'])
model.eval()

Live Demo

Try it here: Sathya77/swin-transformer-satellite

References

• Swin Transformer Paper — Liu et al. 2021
• NWPU-RESISC45 Dataset