PaperlayerView exact revision on arXiv

Source-linked AI summary

Hyperspectral Image Classification With Context-Aware Dynamic Graph Convolutional Network

Sheng Wan, Chen Gong, Ping Zhong, Shirui Pan, Guangyu Li, Jian Yang

arXiv:1909.11953v1cs.LGcs.CVeess.IVstat.ML

TL;DR

HSI classification needs spatial context, but conventional methods can misrepresent complex regions by assuming nearby pixels share labels. CAD-GCN learns a region-induced graph, captures long-range relations, and dynamically refines graph connections; experiments on three benchmark datasets report superior classification performance over competing methods.

  • Problem

    Conventional spatial-context methods assume neighboring pixels share land-cover labels, limiting contextual modeling in irregular or inhomogeneous regions.

  • Method

    CAD-GCN projects pixels into a learned region graph, uses successive graph convolutions for long-range context, and dynamically refines edge weights and connections.

  • Results

    CAD-GCN achieves the best OA and Kappa coefficient among competitors on the University of Pavia dataset and exceeds R-2D-CNN by 9.94% OA.

  • Takeaways & Limitations

    Learning and refining contextual relations improves CAD-GCN’s HSI representation and classification performance across the reported benchmark datasets.

Abstract

from arXiv · show

In hyperspectral image (HSI) classification, spatial context has demonstrated its significance in achieving promising performance. However, conventional spatial context-based methods simply assume that spatially neighboring pixels should correspond to the same land-cover class, so they often fail to correctly discover the contextual relations among pixels in complex situations, and thus leading to imperfect classification results on some irregular or inhomogeneous regions such as class boundaries. To address this deficiency, we develop a new HSI classification method based on the recently proposed Graph Convolutional Network (GCN), as it can flexibly encode the relations among arbitrarily structured non-Euclidean data. Different from traditional GCN, there are two novel strategies adopted by our method to further exploit the contextual relations for accurate HSI classification. First, since the receptive field of traditional GCN is often limited to fairly small neighborhood, we proposed to capture long range contextual relations in HSI by performing successive graph convolutions on a learned region-induced graph which is transformed from the original 2D image grids. Second, we refine the graph edge weight and the connective relationships among image regions by learning the improved adjacency matrix and the 'edge filter', so that the graph can be gradually refined to adapt to the representations generated by each graph convolutional layer. Such updated graph will in turn result in accurate region representations, and vice versa. The experiments carried out on three real-world benchmark datasets demonstrate that the proposed method yields significant improvement in the classification performance when compared with some state-of-the-art approaches.

I. INTRODUCTION

HSI classification benefits from spatial context, but fixed neighborhood assumptions and limited GCN receptive fields can misrepresent complex regions. CAD-GCN addresses this by learning region graphs, long-range relations, and dynamically refined contextual connections.

  • I. INTRODUCTION: Pixelwise spectral methods ignore spatial correlation, making them vulnerable to spectral variability and imperfect classification.
  • I. INTRODUCTION: Fixed spatial-context methods assume nearby pixels share labels, causing errors in object interiors, irregular regions, and class boundaries.
  • I. INTRODUCTION: CAD-GCN combines GCNs, graph projection and re-projection, and dynamic graph refinement to model HSI contextual relations.The method learns a region graph from image pixels, captures long-range relations through graph convolutions, and refines similarities and connections during processing.
  • I. INTRODUCTION: GCN aggregates information over graph neighborhoods, allowing contextual modeling on non-Euclidean and irregular HSI regions.This flexibility is intended to preserve complex areas such as target boundaries.
  • I. INTRODUCTION: Traditional graph convolution has a small receptive field, whereas CAD-GCN uses a learned region graph to capture long-range contextual relations without a dense fully connected graph.
  • I. INTRODUCTION: Dynamic graph refinement updates region similarities and filters improper inter-class edges as representations evolve.

A. Hyperspectral Image Classification

HSI classification methods increasingly incorporate spatial structure, while graph convolution provides a framework for aggregating information over irregular neighborhoods. CAD-GCN applies this framework to learned region graphs and uses graph-based filtering for contextual representation.

  • A. Hyperspectral Image Classification: HSI classification methods include spectral, structural filtering, morphology, and graph-based approaches, reflecting the importance of spatial information alongside spectra.
  • A. Hyperspectral Image Classification: Graph convolution aggregates neighboring node features, but constructing an appropriate neighborhood remains an open challenge for spatial graph methods.
  • A. Hyperspectral Image Classification: GCN was adopted because it directly processes graph-structured non-Euclidean data and can aggregate information across contextual node neighborhoods.
  • A. Hyperspectral Image Classification: CAD-GCN projects pixels into a learned region graph, applies dynamic graph convolutions, and re-projects region representations into image grids.The pipeline converts 2D pixel data into graph data and produces classification outputs through region-to-pixel interpolation.
  • A. Hyperspectral Image Classification: The graph is formally represented as G = (V, E), with an adjacency matrix encoding edges between graph nodes.
  • A. Hyperspectral Image Classification: Spectral graph convolution uses graph Fourier representations and polynomial filtering, while CAD-GCN uses first-order localized filtering with K = 1.

2 AD−1 2 )x (6)

The normalized GCN operator is stabilized before repeated graph convolutions are applied.

  • 2 AD−1 2 )x (6): GCN replaces the unnormalized propagation operator with a renormalized adjacency matrix to avoid numerical instabilities and exploding or vanishing gradients.
  • 2 AD−1 2 )x (6): Each graph-convolution layer applies the propagation operator, trainable weights, and an activation function such as softplus.

B. Pixel-to-Region Assignment

CAD-GCN learns a soft pixel-to-region assignment to transform HSI pixel grids into a region graph. This learned representation reduces reliance on inaccurate pre-computed regions while supporting longer-range contextual modeling.

  • B. Pixel-to-Region Assignment: CAD-GCN learns a soft assignment matrix that maps each pixel to neighboring regions, transforming the original HSI into a region graph.The matrix is parameterized by region anchor points, and each assignment reflects a pixel’s relation to adjacent regions.
  • B. Pixel-to-Region Assignment: Learned region features reduce the negative impact of inaccurate pre-computed region features.
  • B. Pixel-to-Region Assignment: SLIC-based initialization prevents severely imbalanced assignments that could concentrate most pixels in one region.Average spectral signatures within SLIC regions initialize the region anchor points before further gradient-based updates.
  • B. Pixel-to-Region Assignment: Figure 2 depicts initialized regions with yellow boundaries and blue arrows showing a pixel’s assignments to neighboring regions.
  • B. Pixel-to-Region Assignment: Successive graph convolutions on the learned region graph enable contextual relations among regions beyond pixel-level receptive fields.The region graph is acquired after learning pixel-to-region assignments and encoding region features.

C. Dynamic Graph Refinement

CAD-GCN dynamically refines graph similarities and connectivity as graph representations evolve. It combines learned layer-specific distance metrics with an edge filter that suppresses weak, potentially inter-class connections.

  • C. Dynamic Graph Refinement: The learned metric replaces Euclidean distance in the adjacency construction used to encode region similarities.The resulting adjacency remains restricted by neighborhood connectivity while using the learned distance measure.
  • C. Dynamic Graph Refinement: CAD-GCN learns a layer-specific positive semi-definite metric to measure node similarities as graph representations change across convolutional layers.A single similarity matrix is considered insufficient because node representations are updated at each layer.
  • C. Dynamic Graph Refinement: An edge filter removes low-weight connections that are more likely to represent inter-class relations and harmful feature aggregation.The filter applies a threshold β^(l) at each graph convolutional layer; the implementation can share one threshold across layers.
  • C. Dynamic Graph Refinement: The graph convolution uses the refined adjacency together with the evolving layer representations, with H(0) = X as the initial state.

D. Region-to-Pixel Assignment

After region-level dynamic graph convolution, CAD-GCN re-projects learned region features onto the original pixel grid. Soft assignment and interpolation preserve diverse pixel representations and contextual details.

  • D. Region-to-Pixel Assignment: Graph re-projection maps the learned region representation H(L) back to the 2D pixel grid using the soft assignment matrix P.The operation is expressed as O = PH(L), where L is the number of graph convolutional layers.
  • D. Region-to-Pixel Assignment: Linear interpolation gives pixels assigned to the same region diverse feature representations, preserving HSI contextual details.
  • D. Region-to-Pixel Assignment: The complete training procedure initializes regions with SLIC, refines the graph through convolution, interpolates features back to 2D grids, and updates weights by full-batch gradient descent.Label prediction is then performed through the graph convolution and re-projection operations.
  • D. Region-to-Pixel Assignment: CAD-GCN trains end-to-end with cross-entropy loss comparing network outputs against labels of labeled pixels.The labels are represented as a class-indicator matrix over the labeled-pixel indices.

IV. EXPERIMENTAL RESULTS

The experiments evaluate CAD-GCN on three real-world HSI benchmarks against state-of-the-art methods. Performance is assessed using per-class accuracy, OA, AA, and the kappa coefficient.

  • IV. EXPERIMENTAL RESULTS: CAD-GCN is evaluated on Indian Pines, University of Pavia, and Salinas using per-class accuracy, OA, AA, and kappa coefficient.

A. Datasets

The experiments use three real-world hyperspectral datasets, including Indian Pines and University of Pavia, with dataset-specific spatial, spectral, and class distributions.

  • The evaluation includes Indian Pines, University of Pavia, and Salinas benchmark datasets.
  • 1) Indian Pines:: Indian Pines contains 145×145 pixels, 200 retained spectral bands, and 16 land-cover classes after removing noisy and water-absorption bands.
  • 2) University of Pavia:: University of Pavia contains 610 × 340 pixels, 103 spectral channels, and 9 land-cover classes.

3) Salinas:

Salinas is evaluated with randomly selected labeled examples for training and validation, while multiple GCN, CNN, and traditional methods provide comparison baselines.

  • 3) Salinas:: Salinas contains 512 × 217 pixels, 204 retained spectral bands, and 16 land-cover classes.
  • B. Experimental Settings: The experiments typically select 30 labeled pixels per class for training, or 15 when a class has fewer than 30 pixels.
  • B. Experimental Settings: Ninety percent of labeled examples train the network, 10% form the validation set, and all unlabeled examples are used for testing.
  • B. Experimental Settings: CAD-GCN is compared with GCN, S2GCN, R-2D-CNN, CNN-PPF, MFL, and JSDF across ten independent labeled-pixel splits.

C. Classification Results

Across three datasets, CAD-GCN achieves the strongest reported overall comparison results and produces more compact, less noisy classification maps, especially near difficult boundaries and heterogeneous regions.

  • 1) Results on the Indian Pines Dataset:: On Indian Pines, CAD-GCN reaches top-level OA, AA, and Kappa performance with small standard deviations.
  • 1) Results on the Indian Pines Dataset:: Indian Pines maps from CAD-GCN are more compact and show fewer errors near class boundaries than the compared methods.
  • CAD-GCN achieves the best OA and Kappa coefficient among competitors across the three datasets, improving OA over R-2D-CNN by 9.94% and CNN-PPF by 3.60%.
  • 2) Results on the University of Pavia Dataset:: On University of Pavia, CAD-GCN maps are closer to ground truth than other methods, while fixed coarse graph convolutions produce more errors.
  • 3) Results on the Salinas Dataset:: On Salinas, CAD-GCN yields over 10% higher OA than R-2D-CNN and approximately 8% higher OA than CNN-PPF.
  • 3) Results on the Salinas Dataset:: Salinas classification maps from CAD-GCN are less noisy in the Grapes untrained and Vineyard untrained regions than those of other methods.

D. Impact of the Number of Labeled Examples

CAD-GCN consistently outperforms the compared methods across three datasets as the number of labeled examples per class varies from 5 to 30. Its relatively high OA with only 5 or 10 examples per class indicates stable performance under limited labeling.

  • Impact of labeled examples: CAD-GCN consistently outperforms GCN, S2GCN, and the other competitors across Indian Pines, University of Pavia, and Salinas datasets.The comparison varies labeled examples per class from 5 to 30 in increments of 5 and reports OA.
  • Impact of labeled examples: With only 5 or 10 labeled examples per class, CAD-GCN still achieves relatively high OA, suggesting stability under limited training data.
  • Parameter sensitivity: The four hyperparameters T, η, u, and β are critical for satisfactory CAD-GCN performance, with dataset-specific settings reported for all three datasets.The selected settings are T = 1500, η = 0.001, u = 60, β = 0.01 for Indian Pines; T = 500, η = 0.001, u = 210, β = 0.05 for University of Pavia; and T = 2000, η = 0.0001, u = 110, β = 0.02 for Salinas.

F. Ablation Study

The ablation study shows that graph projection, dynamic node-similarity refinement, and the edge filter jointly support CAD-GCN’s classification performance. Runtime comparisons further indicate that graph projection improves efficiency, especially on larger datasets.

  • Ablation study: Removing any one of graph projection, dynamic node-similarity refinement, or the edge filter inevitably reduces OA across the three datasets.The reduced models are CAD-GCN-v1, CAD-GCN-v2, and CAD-GCN-v3, respectively, and the corresponding results are reported in Tables VII–IX.
  • Running time: Graph projection significantly reduces graph size, giving CAD-GCN higher efficiency than GCN on the University of Pavia and Salinas datasets.Running times are compared for GCN, S2GCN, R-2D-CNN, CNN-PPF, and CAD-GCN under matched labeled-pixel settings.
  • Conclusion: Across three real-world HSI datasets, CAD-GCN yields better performance than state-of-the-art HSI classification methods.The method learns pixel-to-region assignments for long-range context and dynamically updates node similarities and connective relationships during graph convolution.
Loading 1909.11953v1…