The purpose of this project is to classify the pixels of a hyperspectral image to specific categories. These images can be represented as 3D arrays where the first two dimensions represent the coordinates of the pixel and the 3rd dimension represents the different spectral bands, known as the spectral signature of the pixel.
The classification to the 5 categories is done by three different classifiers (supervised):
- Naive Bayes classifier
- minimum Euclidean distance classifier
- k-nearest neighbor classifier
The program uses some points of the dataset for training and some others for testing. In the case of the KNN classifier, before doing the actual training, the training set is first used to calculate the optimal K value (from 1 to 17). This is done with the use of the cross validation method, by splitting the training dataset to 5 equal parts, but with distinct points, and with the same ratio of categories. Specifically for every K value, 5 knn classifiers are trained, one from each part of the training set while using the remaining 4 for testing. Then by calculating the average for every K, we pick the one which provides the lowest error rate.
The only thing required is having a MATLAB enviroment installed.
Get a copy of the project
git clone https://github.com/v-pap/hyperspectral_image_classification.git
In order to use the program you will have to open the MATLAB suite and browse the directory with the contents of the project.
Then in the src directory you will have to run the project_m_file.m file.
The dataset was provided by the pattern recognition / machine learning class of the informatics and telecommunications department of the UoA and it represents one area of the Salinas Valley located in California. The hyperspectral image has a resolution of 150x150 pixels with 204 spectral bands (from 0.2μm to 2.4μm) and a spatial resolution of 3.7 meters.
This project is licensed under the MIT License