This repository contains the solution for the project of the Pattern Recognition - Machine Learning class of the Department of Informatics and Telecommunications of UoA. The purpose of the project was to classify the pixels of a provided hyperstrectral image into 5 different categories with the help of three different supervised classifiers, which are the following:
- Naive Bayes
- Minimum Eucleidian Distance
- K-Nearest Neighbor (KNN)
The script is split into 2 phases. First, the dataset is loaded and split into training an testing sets. As their names imply, they are used to train and then evaluate the performance of the 3 classifiers. In order to evaluate how accurate each classifier is, we calculate the error rate, a confusion matrix and then use the confusion matrix to find the precision of the trained classifier.
Training the KNN classifier has one more step compared to the other 2 classifiers. Before we train KNN using all points of the training dataset, we initially use 5-fold cross validation to find the best value for K. More specifically, we split the training dataset into 5 subsets and use them to train 5 different KNN classifiers and find their average error rate. This process is done for all values of K from 1 up to 17 by skipping every second value (1->3->5-> ... and so on). We eventually keep the value of K that gave us the lowest average error rate and train the KNN classifier using the whole training dataset.
The second phase of the script uses again all 3 classifiers, but this time we classify the whole dataset (both training and testing datasets). Afterwards, we display the classified pixels which can be seen below:
| Naive Bayes | Eucleidian Distance | KNN | Ground Truth |
|---|---|---|---|
 |
 |
 |
 |
As it can be seen from the above images, KNN has succesfully classified all pixels correctly. More specifically, by running the hsi.m script we get the following error rates:
| Naive Bayes | Eucleidian Distance | KNN |
|---|---|---|
| 0.0259% | 0.0245% | 0.0% |
By looking at the images, it's easy to see that there is a an empty space between the pixels that belong to different classes. This makes the KNN classifier an ideal choice which is also why it provides 100% accuracy in this test. On the other hand, the Naive Bayes and the Eucleidian classifiers rely on calculating the Likelihood parameters which require either a high number of points in our dataset or a low number oh dimensions (or both, preferrably). Unfortunately, this is not the case in this test, which is why those 2 classifiers fail to correctly classify all pixels. As such, the best out of these 2 classifiers for this test is the KNN classifier.
The dataset was provided by the Pattern Recognition - Machine Learning class of the Department of Informatics and Telecommunications of UoA. It represents one area of the Salinas Valley located in California. The provided hyperspectral image has a resolution of 150x150 pixels, contains 204 spectral bands (ranging from 0.2μm to 2.4μm) and a spatial resolution of 3.7 meters (meaning that the hyperstrectral image can be represented by a 150x150x204 cube).
To get the code up and running on your local machine, simply follow the following instructions.
First, you need to have MATLAB installed on your local system. This project was tested on version 2017a, but other versions of the software environment should be compatible.
Get a copy of the project
git clone https://github.com/vsakkas/Hyperspectral-Image-Classification.git
And to enter the directory of the downloaded project, simply type:
cd Hyperspectral-Image-Classification
To run the provided code, open MATLAB, navigate to the scripts directory, run the hsi.m file.
This project is licensed under the MIT License - see the LICENSE file for details