Multiscale superpixelwise prophet model for noise-robust feature extraction in hyperspectral images. [Dataset]

Abstract

In this paper, we have proposed Multiscale Superpixelwise Prophet model (MSPM), a novel spectral-spatial feature mining framework for noise-robust feature extraction and effective data classification of the HSI. First, we demonstrate that the Prophet model is able to enhance the HSI features in terms of reduced intraclass variance and enlarged interclass diversity. Second, superpixelwise image segmentation has found particularly useful for grouping local spectrally similar pixels and reducing the high intra-class heterogeneity and inter-class homogeneity of different land cover classes in the HSI. Third, our MSPM model has successfully exploited spectral data at different noise levels. The Prophet model has also contributed noticeably to the classification, especially to the uncorrected datasets. The superpixelwise segmentation and the Prophet model can supplement to each other in enhancing the features in the spatial and spectral domains, respectively. As a result, the joint spectral-spatial features can more effectively characterize both the corrected and uncorrected HSI datasets, especially with very limited training samples. Our MSPM model has significantly outperformed a few state-of-the-art approaches, including several deep learning models along with much more training samples. The improved classification results from the uncorrected datasets have enabled potentially a new and fully automatic roadmap for interpreting the HSI where conventional wisdom of pre-filtering of unwanted bands can be skipped. The accompanying file contains supplementary figures and tables.

Citation

MA, P., REN, J., SUN, G., ZHAO, H., JIA, X., YAN, Y. and ZABALZA, J. 2023. Multiscale superpixelwise prophet model for noise-robust feature extraction in hyperspectral images. [Dataset]. IEEE transactions on geoscience and remote sensing [online], 61, article 5508912. Available from: https://doi.org/10.1109/tgrs.2023.3260634/mm1

• Acceptance Date: Mar 17, 2023
• Online Publication Date: Mar 27, 2023
• Publication Date: Dec 31, 2023
• Deposit Date: Apr 4, 2023
• Publicly Available Date: Apr 4, 2023
• Publisher: Institute of Electrical and Electronics Engineers (IEEE)
• DOI: https://doi.org/10.1109/tgrs.2023.3260634/mm1
• Keywords: Hyperspectral image (HSI); Multiscale prophet model; Spectral-spatial feature mining; Superpixel segmentation
• Public URL: https://rgu-repository.worktribe.com/output/1930754
• Related Public URLs: https://rgu-repository.worktribe.com/output/1925440 (Journal article)
• Type of Data: Supplementary figures and tables.
• Collection Date: Jan 5, 2023
• Collection Method: For performance assessment, three publicly available HSI datasets are used in our paper. The descriptions of the datasets and the experimental settings as well as some ablation studies are presented in detail as follows. The first dataset is the Indian Pines, which was collected using the Airborne Visible/Infrared Imaging Spectrometer (AVIRIS) sensor over the Indian Pines study site in Northwest Indiana, USA. This dataset contains 145×145 pixels with a low spatial resolution of 20 m/pixel and 220 spectral bands in the wavelength range of 0.4 − 2.5μm after removing 4 invalid bands. In addition, twenty water absorption bands (104-108, 150-163 and 220) were often discarded [40] before data classification, leaving 200 spectral reflectance bands for analysis and testing. There are 16 manually defined land-cover classes, which are mostly different kinds of crops. The second dataset is Salinas, which was also acquired via AVIRIS in the Salinas Valley in California, USA. It has 512×217 pixels and 224 spectral bands with a spatial resolution of 3.7m. Similar to the first dataset, twenty water absorption bands are usually discarded. In addition, the corresponding ground-truth map also has 16 classes. The third is the Pavia University dataset, namely PaviaU, which was acquired by the Reflective Optics System Imaging Spectrometer sensor over Pavia, Northern Italy. It contains 114 spectral bands in a spectral range of 0.43 − 0.86μm and 610×340 pixels with a spatial resolution of 1.3m. Similarly, for avoiding the effect of water absorption, the available number of bands was reduced from 114 to 103. The ground truth has nine classes of land covers.