Abstract
Hyperspectral image (HSI) analysis has become an active topic in the field of remote sensing in recent years. HSIs often have hundreds of spectral bands of different wavelengths captured by aircraft or satellites that record land coverage. As one of the most major tasks, identifying detailed classes of pixels has drawn broad attentions and brought about a wide variety of methods due to the enhancement in spectral and spatial resolution. Among them, semi-supervised methods with a small number of training samples and unsupervised methods that do not require any training samples have gained popularity due to the large amount of time and cost required for manual labeling.Therefore, spectral and spatial-based algorithms for HSI classification and clustering are presented in this thesis for their effectiveness in achieving satisfactory results with limited or no labeled data. Specifically, concerning the semi-supervised methods, spatial information of HSI data is utilized in the pre-processing stage and the post-processing stage of the 3-stage method. The pre-processing stage not only improves the coherence of pixels within the same class, mitigates the effects of noise in HSI to improve the representation of training pixels, but also reduces the execution time through the application of Principal Component Analysis. Consequently, the classification results obtained by support vector machines are improved owing to the refined and denoised features. Additionally, the post-processing stage aims to ensure spatial connectivity in the classification map and helps prevent misclassification of isolated pixels in the image. Findings from the experiment show that the 3-stage is capable of significantly decreasing the quantity of training pixels while enhancing the accuracy of classification. As a result, it is of great practical significance since expert annotations are often expensive and difficult to collect. Further to this, the shape adaptive reconstruction is introduced as a variant for the pre-processing stage whose flexible selection of reconstruction window makes it more suitable for those classes manifested as elongated shapes with sharp corners.
While for the unsupervised clustering, the Spatial-Spectral Image Reconstruction and Clustering with Diffusion Geometry (DSIRC) algorithm also incorporates both spatial and spectral information in order to reduce the noise and varied spectral characteristics in the homogeneous region and enhance the accuracy of clustering. Particularly, for each pixel, the DSIRC algorithm estimates both density and purity, providing a quantitative measure of each pixel's significance and representation within the dataset. Moreover, the DSIRC locates spectrally correlated pixels within a data-adaptive spatial neighborhood and reconstructs that pixel's spectral signature using those of its neighbors. Accordingly, the diffusion distance is calculated based on the spatially involved graph which is constructed using reconstruction data to capture intrinsic relationships between pixels. Finally, it locates high-density, high-purity pixels which are far in diffusion distance (a data-dependent distance metric) from other high-density, high-purity pixels and treats these as cluster exemplars, giving each a unique label. Non-modal pixels are assigned the label of their diffusion distance-nearest neighbor of higher density and purity that is already labeled. Strong numerical results indicate that incorporating spatial information through image reconstruction substantially improves the performance of pixel-wise clustering.
In the future directions and perspective, we consider to integrate superpixel segmentation as a strategic data augmentation technique. It aims to mitigate the challenges associated with the scarcity of labeled data in HSI data. Additionally, we advocate for the determination of the optimal number of superpixels by leveraging the intrinsic dataset characteristics, such as size, spatial complexity, and resolution. Besides, combining machine learning algorithms and neural networks is a promising research trajectory, which shows the potential to enhance the efficiency and accuracy of HSI classification tasks. Finally, exploring alternative dimensionality reduction algorithms that can capture the subtle spectral nuances more effectively than existing techniques is also a future research direction.
| Date of Award | 16 Aug 2024 |
|---|---|
| Original language | English |
| Awarding Institution |
|
| Supervisor | Jean-Michel Henri Olivier MOREL (Supervisor) & Hon Fu, Raymond CHAN (External Co-Supervisor) |