Hyperspectral Image Classification Using Fusion of Spectral and Spatial Features

Project Description

This project explores the classification of hyperspectral images using a fusion of spectral and spatial features. Hyperspectral imaging (HSI) is a computer vision modality that combines standard 2D imaging with spectroscopy to acquire a three-dimensional hypercube. HSI has been used in various fields, such as remote sensing and medical image diagnosis. The goal of this project is to improve the accuracy and efficiency of hyperspectral image classification by incorporating spatial features along with spectral features.

The project implements and compares two spatial-spectral HSI classification methods and evaluates their performance in terms of accuracy, kappa coefficient, and total time taken.

Implemented Methods

• Semi-supervised learning method for classification based on Spatial Majority Voting [1].
• Edge-Preserving Filters (EPFs) method with PCA-based dimensionality reduction [2].

Installation

1. Install MATLAB (version R2021a or later) from the official website.
2. Clone this repository to your local machine.
3. Install necessary MATLAB toolboxes:

• Image Processing Toolbox
• Statistics and Machine Learning Toolbox

Usage

1. Download the following datasets:

• AVIRIS Indian Pines dataset: Indian Pines Dataset
• Pavia University dataset: Pavia University Dataset
• Pavia Centre dataset: Pavia Centre Dataset

2. Place the downloaded datasets in the data folder within the SpatialMajorityVoting and PCA-EPFs directories.

For Spectral Classifier

1. Open the SpatialMajorityVoting folder in MATLAB.
2. Run the DemoSpectralClassifier MAT file.

For SpatialMajorityVoting Classifier

1. Open the SpatialMajorityVoting folder in MATLAB.
2. Run the DemoSpatialClassifier.m MAT file.

For PCA-EPFs Classifier

1. Open the PCA-EPFs folder in MATLAB.
2. Run the pca_epfssdemo.m MAT file.

Dataset

The dataset used in this project is the AVIRIS Indian Pines dataset, which consists of an HSI image of 145x145 pixels and 224 different spectral bands. The scene includes agricultural, forest, and vegetation objects, as well as structures such as highways, rail lines, and housing. Objects within the scene are distributed into 16 different class labels of crops, such as corn and soybeans.

Results

The performance of the implemented classification methods is evaluated using the overall accuracy, kappa coefficient metrics, and total time taken.

License

This project is released under the MIT License. See LICENSE for details.

Acknowledgements

This project was developed by Nauman Baig under the supervision of Prof. Ling Guan at Ryerson University.

References

[1] Deok Han, Qian Du, and Nicolas H Younan. "Semisupervised classification of hyperspectral remote sensing images with spatial majority voting". In: 2016 9th IAPR Workshop on Pattern Recogniton in Remote Sensing (PRRS). IEEE. 2016, pp. 1-4.

[2] Xudong Kang et al. "PCA-based edge-preserving features for hyperspectral im- age classification". In: IEEE Transactions on Geoscience and Remote Sensing 55.12 (2017), pp. 7140-7151.

[3] Kang Xudong (2020). Hyperspectral Image Classification (https://www.mathworks.com/matlabcentral/fileexchange/69242-hyperspectral-image-classification), MATLAB Central File Exchange. Retrieved April 18, 2020.

[2] Galad_Loth (2020). Hyperspectral Image Classification (https://github.com/galad-loth), Github.