HyperspecAE

This repository contains the pytorch implementation for the paper: Hyperspectral Unmixing Using a Neural Network Autoencoder (Palsson et al. 2018). As a POC, the dataloaders and parameters corresponding to experiments on the Samson dataset are presented.

Dependencies

• PyTorch 1.8.0
• Python 3.7.10

Quick Start

Data

The datasets used in the paper are publicly available and can be found here.<br> Download the Samson dataset from the above-mentioned source. Follow the directory tree given below:<br>

|-- [root] HyperspecAE\
    |-- [DIR] data\
        |-- [DIR] Samson\
             |-- [DIR] Data_Matlab\
                 |-- samson_1.mat
             |-- [DIR] GroundTruth
                 |-- end3.mat
                 |-- end3_Abundances.fig
                 |-- end3_Materials.fig

Training

The shell script that trains the model (samson_train.sh) can be found in the run folder. You can simply alter the hyperparameters and other related model options in this script and run it on the terminal.<br> You can refer to opts.py to explore other command line arguments to customize model training.

Abundance Map and End-Member Extraction

The shell script that extracts the abundance maps and end-members (extract.sh) can be found in the run folder. Ensure that the charateristics of the model match exactly with the pre-trained version to be used for extraction.<br>

Results

The following are the results of a deep autoencoder, Configuration Name: LReLU (see paper). You can experiment with other configurations by altering the command line arguments during model training.

Pre-trained Model

The pre-trained model for this configuration can be found here.

Abundance Maps

Left: Tree, Middle: Water and Right: Rock.

Extracted Spectral Signature (End-Members)

Left: Tree, Middle: Water and Right: Rock.

References

Original work by the authors

@article{palsson2018hyperspectral,
  title={Hyperspectral unmixing using a neural network autoencoder},
  author={Palsson, Burkni and Sigurdsson, Jakob and Sveinsson, Johannes R and Ulfarsson, Magnus O},
  journal={IEEE Access},
  volume={6},
  pages={25646--25656},
  year={2018},
  publisher={IEEE}
}