Awesome

HybridSN

This is a personal pytorch-based implement of HybridSN by Jupyter notebook.

The HybridSN is spectral-spatial 3D-CNN followed by spatial 2D-CNN.The 3D-CNN facilitates the joint spatial-spectral feature representation from a stack of spectral bands. The 2D-CNN on top of the 3D-CNN further learns more abstract level spatial representation.

The paper link: "HybridSN: Exploring 3-D–2-D CNN Feature Hierarchy for Hyperspectral Image Classification"

Official keras code: gokriznastic/HybridSN

Requirements

pytorch 1.9.0

scikit-learn 1.0.2

spectral 0.22.2

torchinfo 1.6.1

Results

<img src="results/IP_RGB_origin.jpg" width = "200" height = "200" align=center /> <img src="results/IP_gt.jpg" width = "200" height = "200" align=center /> <img src="results/IP/prediction_masked.jpg" width = "200" height = "200" align=center />

Fig1. IndianPines (a) RGB (b) groundtruth (c) predition

<img src="results/PU_RGB_origin.jpg" width = "100" height = "200" align=center /> <img src="results/PU_gt.jpg" width = "100" height = "200" align=center /> <img src="results/PU/prediction_masked.jpg" width = "100" height = "200" align=center />

Fig2. Pavia University (a) RGB (b) groundtruth (c) predition

<img src="results/SA_RGB_origin.jpg" width = "100" height = "200" align=center /> <img src="results/SA_gt.jpg" width = "100" height = "200" align=center /> <img src="results/SA/prediction_masked.jpg" width = "100" height = "200" align=center />

Fig3. Salinas Scene (a) RGB (b) groundtruth (c) predition