Application for hyper-spectral image preprocessing and classification
Hyper-spectral images classification can solve many problems, such as quantitative monitoring of vegetation changes, flood damage area, urban expansion area, etc. Hyperspectral images are generally noisy, and the data are highly redundant. Therefore, using effective methods to detect and eliminate b...
Elmentve itt :
| Szerző: | |
|---|---|
| További közreműködők: | |
| Dokumentumtípus: | Szakdolgozat |
| Megjelent: |
2018
|
| Tárgyszavak: | |
| Online Access: | http://diploma.bibl.u-szeged.hu/73546 |
| Tartalmi kivonat: | Hyper-spectral images classification can solve many problems, such as quantitative monitoring of vegetation changes, flood damage area, urban expansion area, etc. Hyperspectral images are generally noisy, and the data are highly redundant. Therefore, using effective methods to detect and eliminate bad-bands data is very helpful for image classification. Dimension reduction could greatly improves the utilization of useful information and reduces the amount of computation. Therefore, this paper has focused on image preprocessing, dimension reduction and classification of these parts， through the deep research the author proposed a complete processing flow on Hyper-spectral image and also create an application for the whole processing flow. The application contains five modules(I/O module ,preprocessing module, filtering module, classification module and work flow module) where I/O module can read and save data, preprocessing module can detect bad bands and remove the bad bands, filtering module can reduce noise data, classification module can classify images, work flow module is most important module, this module integrates all the above modules and is an automated process . In this thesis, I used the second derivative of spectral curve of objects to detect the noise pixels, and obtained bad bands according to the percents of noise pixels. After the bad bands detection the user could get a report of the operation then remove these bad bands .Even if the user removed the bad bands, hyper-spectral images still have noise, and these noises will have more or less influence on data dimension reduction and classification. In this thesis, a signal filtering (Savitzky–Golay filter[33]) method was used to filter each pixel of band, and filtered spectrum curve became smooth. Through experimental comparison, it is found that the filtered data classification accuracy is higher than the classification result without filtering operation. In this thesis, PCA(Principal Components Analysis) was used on dimension reduction operation.. Use the main components to represent the original information can reduce the amount of calculations. Experiments have shown that principal component analysis not only reduced the amount of data calculation, but also didn’t had much impact on the results of classification. The confusion matrix of the classification results shown that the classification results obtained by through principal component analysis operation only has 0.05% difference with original classification results. There are supervised classification methods and unsupervised classification methods for image classification. In this paper, ISODATA method and SVM are used to classify images. Then the results of different classification methods are compared, the results shown that the supervised classification method has higher classification accuracy. One of the most important tasks of this paper is to integrate the above-mentioned methods into one application, which makes the image processing and classification work simple. This application is based on the Python language and works properly in the Windows environment. Some functions are referenced from GDAL, Numpy, OpenCV, Spectial site-packages. Most of the methods are written by the author himself. |
|---|