linlin-yu / uncertainty-aware-hsic

Uncertainty-aware Graph-based Hyperspectral Image Classification

This repository presents the experiments of the paper:

Uncertainty-aware Graph-based Hyperspectral Image Classification
Linlin Yu, Yifei Lou, Feng Chen
International Conference on Learning Representations (ICLR), 2024.

[paper]

Requirements

To install requirements:

conda env create -f environment.yaml
conda activate uhsic

Data Preprocessing & Running Experiments

The experiments include three datasets: PaviaU, KSC and Houston2013.

1. Download the raw files including the feature and classification ground-truth matrix and save them under the folder raw_data/{dataset}/；
2. Run an unsupervised unmixing model and get the endmember matrix as the prior knowledge for the proposed architecture; In our paper, we use the '[Blind Hyperspectral Unmixing Based on Graph Total Variation Regularization (https://ieeexplore.ieee.org/document/9200736)]' to generate the predicted abundance matrix and endmember matrix, and note that we need to run a permutation algorithm to match the endmember matrix with the material label (we provide sample code in the data/find_perm). The generated matrix should be saved under the folder data/{dataset}/unmxing/; Then we need to
3. Run data/data_preprocess.py, which will generate a folder under data/{dataset}/raw for required matrices;
4. For 'GKDE' based models, first run alpha_prior_generation.py and probability_teacher_generation.py, which will generate and save GKDE teacher and probability teacher tensors under folder teacher;
5. For experiments related to misclassification detection, please execute the Python files that end with clearngraph. For out-of-distribution (OOD) detection experiments, run the Python files ending with oodgraph. For experiments involving softmax graph convolutional networks (GCN), execute the Python files that begin with classification. For experiments on enhanced GCN (EGCN) models based on Gaussian Kernel Density Estimation (GKDE), run Python files starting with GKDE. Lastly, for experiments related to 'GPN' based models, please run Python files beginning with GPN.

Cite

Please cite our paper if you use the model or this code in your own work:

@inproceedings{
yu2024uncertaintyaware,
title={Uncertainty-aware Graph-based Hyperspectral Image Classification},
author={Linlin Yu and Yifei Lou and Feng Chen},
booktitle={The Twelfth International Conference on Learning Representations},
year={2024},
url={https://openreview.net/forum?id=8dN7gApKm3}
}