Models Overview¶

The Pansharpening Toolkit provides a comprehensive set of models ranging from classic signal processing methods to state-of-the-art deep learning architectures.

Model Categories¶

Classic Methods¶

Traditional pansharpening algorithms based on signal processing:

Method	Description
Brovey	Component substitution with band ratios
IHS	Intensity-Hue-Saturation transformation
SFIM	Smoothing Filter-based Intensity Modulation
Gram-Schmidt	Gram-Schmidt spectral sharpening
HPF	High-Pass Filter injection

Deep Learning Models¶

Neural network-based approaches:

Model	Architecture	Parameters	Description
`pnn`	3-layer CNN	~50K	Basic baseline
`pannet`	ResNet + High-pass	~340K	Residual learning
`drpnn`	Deep ResNet	~300K	Deeper network
`pannet_cbam`	PanNet + CBAM	~340K	Attention-enhanced
`mspannet`	Multi-scale FPN	~500K	Feature pyramid
`panformer`	Transformer	~1M	Cross-attention
`panformer_lite`	Window Transformer	~370K	Efficient transformer

Model Selection Guide¶

Recommendations

Quick experiments: Use pnn or pannet
Best quality: Use panformer_lite with 100+ epochs
Spectral preservation: Use pannet_cbam with spectral_focus loss
Limited compute: Use classic methods (no training needed)

Factory Function¶

All models can be created using the factory function:

from models import create_model, AVAILABLE_MODELS

# List available models
print(AVAILABLE_MODELS)
# ['pnn', 'pannet', 'drpnn', 'pannet_cbam', 'mspannet', 'panformer', 'panformer_lite']

# Create a model
model = create_model('panformer_lite', ms_bands=4)