PaperlayerView exact revision on arXiv

Source-linked AI summary

Graph-Based Semi-Supervised Hyperspectral Image Classification with Distance-Aware Spatial Measure

Sérgio J. M. Almeida, José C. M. Bermudez

arXiv:2609.29367v1eess.IVeess.SP

TL;DR

Hyperspectral image classification faces challenges including limited labeled training data, while graph-based semi-supervised methods have an open question about incorporating pixels’ relative spatial positions. The paper adds a spatial-position kernel to a composite kernel, and reports improved results against those previously reported in the literature.

  • Problem

    Hyperspectral image classification is challenging, and supervised methods need a good-quality training set that may be difficult to obtain.

    Source: §1 · pp. 2–3

  • Method

    The method adds a spatial-2D kernel based on pixel positions to a composite kernel that also includes spectral and spatial-spectral information.

    Source: §1 · p. 8, §2.1 · p. 9, §4.1 · p. 15

  • Results

    The proposed composite kernel adds significant information to learning and yields better classification results than those previously reported in the literature.

    Source: §5 · pp. 26–27

  • Takeaways & Limitations

    The conclusion identifies the spatial-2D term as a distinct source of information for semi-supervised hyperspectral image classification.

    Source: §5 · pp. 26–27

    The introduction notes that implementation of the supervised methods under discussion can be complex and require significant computational effort.

    Source: §1 · p. 5

Abstract

from arXiv · show

The classification of hyperspectral images (HIs) still presents several challenges. One of them is the difficulty to obtain a large set of labeled samples to train the classifier. Semi-supervised learning methods have received much attention recently, as they require the initial labeling of a reduced number of image pixels and lead to very good results for practical application. One of the open problems in graph-based semi-supervised HI classification is the need to consider relative spatial relationship between pixels in the image to improve the smoothness of the solution. Kernel-based approaches using composite kernels have led to very good results, in which one kernel addresses the spectral properties of the pixels while a second kernel addresses some spatial properties. Most available solutions employ a spectral-spatial kernel which considers the spectral properties of a spatial region about each pixel. This work proposes a composite kernel approach that includes a third kernel dealing exclusively with the relative spatial position of the pixels. Experiments with real HI images show that the use of the new spatial kernel has led to improved classification results when compared to those previously reported in the literature. These results also shed some light on the relative contributions of the spectral and spectral-spatial kernels in semi-supervised HI classification.

1. Introduction

The introduction motivates semi-supervised hyperspectral classification in light of scarce labels and numerous spectral channels, emphasizing the importance of graph construction. It presents a coordinate-based spatial term added to a composite kernel to complement spectral-spatial approaches that may miss physical separation between pixels.

1. Introduction

  • Few labeled pixels and many spectral channels make HI classification difficult, increasing computational cost and the risk of dimensionality-driven overfitting.The paper identifies the high cost of obtaining labeled samples as a key obstacle to supervised learning.Source: §1 · pp. 2–3
  • Atmospheric, illumination, and environmental variation can give pixels of the same material distinct spectral representations.This spectral variability makes classification harder even within small regions of a single material.Source: §1 · pp. 2–3
  • Pixels of the same class may also lie in distant image regions, creating spatial variability that challenges classification algorithms.The paper notes that this arrangement is not uncommon.Source: §1 · p. 3
  • With labeled data scarce and data volumes growing, semi-supervised learning is presented as a natural route toward better classifier performance.The introduction also emphasizes that graph construction is important to performance in graph-based transductive SSL.Source: §1 · pp. 3–4
  • In hyperspectral supervised segmentation, conditional random fields encourage local neighboring pixels to share labels, yielding an improved classification map.The introduction describes this as a common post-processing approach after an initial pixel-level classification.Source: §1 · p. 5

Introduction

  • Earlier supervised methods can be computationally complex, while PDE and wavelet preprocessing produced accuracy comparable to simpler despite added complexity.Source: §1 · pp. 5–6
  • Existing composite-kernel methods [35] encode neighborhood information in spectral feature space, but do not address same-class pixels in spatially distinct regions.Physical distance can help compensate for spectral variability between same-class pixels and distinguish spectrally similar pixels from different, distant regions.Source: §1 · p. 7
  • The proposed composite kernel adds a spatial term based on pixel coordinates directly to spectral-spatial similarity for hyperspectral image classification.The paper describes this as a new spatial term in composite kernels and distinguishes it from supervised approaches that use position information.Source: §1 · p. 8
  • The proposed spatial information is incorporated into a simple-to-apply semi-supervised algorithm that the authors report leads to very good classification results.Source: §1 · p. 8

2. Graph-Based Method Proposed

The method builds pixel similarity from spectral information, spectra averaged over a surrounding region, and image-coordinate distance. It combines the corresponding kernels through a weighted sum.

2.1. Construction of the spatial-spectral kernel

  • Each pixel is represented by its spectrum, the average spectra in a surrounding window, and its 2-D image coordinates.The windowed representation averages spectral values band by band.Source: §2.1 · pp. 8–9
  • The three representations yield spectral, spatial-spectral, and spatial-2D similarity matrices, with RBF kernels used for all three.Their bandwidths are chosen to equalize the dynamic ranges of the component kernels.Source: §2.1 · pp. 9–10
  • The proposed spatial-2D adjacent matrix adds pixel-coordinate distance as a distinct contribution to similarity.It complements spectral and spatial-spectral information rather than replacing them.Source: §2.1 · pp. 9–10

2.2. The weighted-summation composite kernel

  • The weighted-summation composite kernel forms W by linearly combining the three component matrices.The design adjusts the component bandwidths to align their value ranges and requires choosing weights β, λ, and γ.Source: §2.2 · p. 10

3. Problem Solution

The solution constructs a graph over labeled and unlabeled pixels and uses its affinity matrix in a semi-supervised label-propagation procedure. The iteration combines graph-based propagation with the initial labels until convergence.

3. Problem Solution

  • The algorithm relies on the consistency assumption that nearby points and points on the same cluster or manifold tend to share a class.This motivates a classification function that is smooth with respect to the graph structure.Source: §3 · p. 11
  • The graph-based solution uses the composite weight matrix as the affinity matrix over labeled and unlabeled pixels.Each pixel maps to a class-score vector, while the label matrix encodes the initial labels of labeled pixels.Source: §3 · p. 11
  • The method iterates F(k + 1) = αSF(k) + (1 − α)Y until convergence, with α in (0, 1).S is the normalized affinity matrix, and the text states that the iteration converges for α in this interval.Source: §3 · p. 12

4. Experimental Results and Discussions

The experiments apply the proposed method to two real hyperspectral image datasets and compare with previously published results. The supplied passages describe the evaluation setup but do not report its classification outcomes.

4. Experimental Results and Discussions

  • The evaluation applies the proposed composite-kernel approach to the AVIRIS Indian Pines and ROSIS Pavia University datasets.The datasets were selected to permit comparison with previously published results, although the software was unavailable.Source: §4 · p. 12
  • The evaluation uses Algorithm 1 for semi-supervised hyperspectral image classification.Source: §4 · pp. 12–13
  • The AVIRIS Indian Pines image has 224 spectral bands and spatial dimensions of 145 × 145 pixels.Source: §4 · p. 12

14 end

The experiments assess an auxiliary graph based on pixels’ two-dimensional distances alongside spectral and spectral-spatial kernels on Indian Pines and Pavia University data. They report stronger accuracy than prior methods in tested Indian Pines comparisons, with the relative value of similarities and spatial-kernel omission varying across the images.

4.1. First Experiment

  • The first experiment used a 68 × 86 Indian Pines subset with four labeled classes; early-growth corn and soybeans made land-cover classification challenging at 20 m spatial resolution.Source: §4.1 · pp. 14–15
  • Removing spectral-spatial Ws while retaining Wω and Wd improved accuracy, indicating Wd provided better spatial information than Ws.Case 2 removed Ws while maintaining the relative weights of Wω and Wd.Source: §4.1 · p. 16
  • Accuracy comparisons with results reported in indicate that the proposed spatial-2D component provided important spatial information, and that Wω should be preferred over Ws when only one is employed.Source: §4.1 · p. 16
  • With two spectral terms combined, accuracies for all four classes were better than the best results reported in for all kernel compositions; Case 2 led to overall best results for all classes.Table 1 reports classification accuracy (%) for different numbers of labeled samples per class for the Indian Pines Image.Source: §4.1 · p. 16, Table 1 · p. 17
  • For Indian Pines, Ws could be disregarded with reduced computational complexity because accuracy did not strongly depend on combining Wω and Ws, even when λ was small.The passage attributes this to averaging that leaves mean spectra largely unchanged across extensive same-class areas.Source: §4.1 · p. 17

4.2. Second Experiment

  • The second experiment selected five Pavia University classes with moderate amounts of labeled data for a balanced method assessment.The selected classes were Gravel, Trees, Painted metal sheets, Bitumen, and Shadows.Source: §4.2 · p. 18
  • For the PU image, classification performance varied little across kernel weights, and the case omitting Wω yielded the best accuracies in most tested cases.The results differed from Indian Pines, where performance was more sensitive to weight choice.Source: §4.2 · p. 18
  • On the PU image, the spatial-spectral representation xs contributed more relevant spectral information than xω, while the proposed xd could improve accuracy.These findings come from comparisons of cases 2 and 3, and cases 1 and 4, respectively.Source: §4.2 · pp. 19–20, Table 3 · p. 20

4.3. Third Experiment

  • For the Indian-Pines image, the proposed method provided more accurate results than in all cases; Table 4 reports classification accuracy results for the proposed method and.Source: §4.3 · p. 21, Table 4 · p. 21

4.4. Fourth Experiment

  • The proposed method achieved superior classification accuracy to the method in for every class except class 13, and its overall average was significantly superior.Source: §4.4 · p. 22

4.5. Fifth Experiment

  • In Table 7, the proposed method achieved superior accuracies to the best results reported in for the Indian-Pines image in every class except Class 3, where the difference was small.Source: §4.5 · p. 24, Table 7 · p. 25

4.6. Discussion

  • The relative effectiveness of spectral and spectral-spatial similarities depends on the image: neighborhood windowing with 2D spatial distance may help Pavia but offers limited gains for Indian Pines.The discussion relates this difference to greater spectral separation among Pavia classes and tightly clustered Indian Pines signatures.Source: §4.6 · p. 24

5. Conclusions

The paper adds a kernel based exclusively on pixels’ spatial properties to graph-based semi-supervised transductive hyperspectral image classification. Experiments with real images show improved classification results and indicate that the relative contributions of spectral-based components depend on the image.

5. Conclusions

  • The method adds a term to the typical composite kernel that considers pixels’ spatial properties exclusively.It is proposed for graph-based semi-supervised transductive hyperspectral image classification.Source: §5 · p. 26
  • Unlike common approaches that derive spatial properties from spectra in a region around each pixel, the added term represents spatial properties directly.The paper notes that such region-based information remains spectral in nature.Source: §5 · p. 26
  • The approach does not require spectral operations around each image pixel and can be applied to a simple graph-based learning method.Source: §5 · p. 26
  • The proposed composite kernel yielded better classification results than those previously reported in the literature.The experiments used real hyperspectral images.Source: §5 · pp. 26–27
  • The relative contributions of the spectral-based components depend on the properties of the hyperspectral image being classified.Source: §5 · p. 27
  • The authors conjecture that the approach may also support directed-graph classification methods emphasizing labeled samples’ influence on unlabeled samples.They identify this as a possible direction for future work.Source: §5 · p. 27
Loading 2609.29367v1…