Higher-order representations for HSI: graphs, hypergraphs and topology

Start with the relationship you need

A hyperspectral image is usually stored as a spatial grid with a spectrum at each pixel. That storage format does not decide which pixels should exchange information during learning. Nearby pixels can belong to different materials, while distant pixels can represent the same land-cover class. Choosing a relational representation is therefore a modelling decision about useful context.

Graphs, hypergraphs and topological representations make different assumptions about that context. A graph records pairwise connections. A hypergraph records membership of groups. A topological construction can distinguish connected components, loops and filled higher-dimensional structures. These tools can complement spectral features, but their names do not establish that they are better than a convolutional network or a carefully tuned spectral baseline.

Before implementing one, write down what a node represents, how relationships are constructed and what information is available at prediction time. That small specification often reveals more than an architecture diagram.

Graphs connect pairs of pixels or regions

In a graph, nodes might represent pixels, patches or superpixels. An edge connects a pair of nodes. Spatial adjacency can connect neighbouring regions; a similarity rule can connect spectra that are close under a chosen distance. The resulting graph need not resemble the original rectangular image grid.

A graph neural network aggregates information along those edges. Its useful context depends on the graph construction: changing the distance metric, neighbourhood size or superpixel segmentation changes who can influence whom. For example, a spectral similarity graph can connect two separated vegetation regions, but it can also connect spectrally similar materials with different semantic labels.

Hong and colleagues' Graph Convolutional Networks for Hyperspectral Image Classification studies CNN and GCN representations and introduces miniGCN for minibatch training and out-of-sample inference. It provides a concrete public example of graph learning for HSI, including CNN–GCN fusion. Its reported comparisons concern the datasets and settings evaluated in that paper.

Graph attention does not create a hypergraph

The original Graph Attention Networks learns coefficients for messages from graph neighbours. Rather than treating each connected neighbour identically, an attention layer can weight its contribution. Multiple attention heads provide several learned aggregations over a neighbourhood.

This changes message weighting, not the underlying definition of an edge. If A attends to B and C through two ordinary edges, the representation is still a graph. Calling that operation higher order because several neighbours are involved would blur the structural distinction. A hypergraph explicitly represents a group such as {A, B, C} as one hyperedge.

Attention weights also need careful interpretation. A large coefficient says that a particular message was weighted strongly inside that computation. It does not by itself prove a physical causal relationship between the pixels or provide a complete explanation of the final prediction.

Hypergraphs make groups explicit

A hyperedge is a set of nodes and may contain more than two members. A binary incidence matrix H records whether node i belongs to hyperedge e. This separates group membership from ordinary pairwise adjacency and allows several overlapping groups to share the same pixel.

The foundational Hypergraph Neural Networks paper develops representation learning through hyperedge convolution. In common formulations, node information is aggregated into hyperedges and redistributed to their members, with normalisation and learned transformations. The precise operator matters: storing groups does not guarantee that every possible joint interaction is retained by the subsequent computation.

For an HSI-specific example, Ma and colleagues' spectral-spatial hypergraph convolution network constructs spectral and spatial hyperedges separately, combines their incidence information and learns hyperedge weights. The paper evaluates Indian Pines and Kennedy Space Center. Those results illustrate a published construction, rather than establishing that all hypergraphs outperform all graphs.

Topology asks about structure across scales

Topology adds another distinction. In a simplicial complex, a filled triangle is a two-dimensional simplex with its edges and vertices included. Three pairwise edges around a triangle form a loop until the face is filled. A three-member hyperedge does not automatically declare either topological interpretation.

Persistent homology follows topological features through a nested sequence of complexes, called a filtration. Components can merge; loops can appear and later disappear. Topological Persistence and Simplification formalises persistence within this growing structure. A persistence diagram or barcode records feature lifetimes.

A public computational example is Ripser, which implements an algorithm for Vietoris–Rips persistence barcodes. Its paper describes implicit representations that avoid constructing and storing the full filtration coboundary matrix. This is software for a topological calculation, not an HSI classifier. A researcher still has to justify how the image is converted into the analysed structure.

For HSI, a researcher could analyse a spectral feature cloud or a scalar image map, but those are different constructions and need different explanations. A loop in a reduced spectral space should not be described as a road loop on the ground. The distance, embedding and filtration define the meaning of the descriptor.

Persistence is also not a certificate of semantic importance. A long-lived feature may reflect acquisition conditions or preprocessing. A short-lived feature may still matter to a small class. Interpreting either requires the scientific question and supporting evidence.

Inspect a small representation before training

The tested Python example below uses four named nodes and two overlapping groups. It builds the incidence matrix, then expands each group into all possible pairwise edges. The expansion is often called a clique expansion. No training data, external package or model download is needed.

Read the two outputs together. The matrix preserves the two named group memberships; the edge list contains only pairs. Once the group identities are discarded, the pairs alone cannot tell you whether they came from those two groups or from several independently specified relationships.

This is a representation exercise, not an implementation of HGNN or a claim about classification accuracy. It is useful when reviewing code that labels an adjacency matrix as a hypergraph without retaining or using group information.

Choose and evaluate the representation step by step

Start with a reproducible baseline and add one relational assumption at a time. Keep the label budget, data split and feature preparation comparable. If a complex model also changes the split or uses additional input data, its score cannot isolate the contribution of the representation.

Document whether the experiment is transductive, where unlabelled target-scene features can participate during training, or inductive, where the model must handle genuinely unseen inputs. Both can be valid. They answer different deployment questions. The weak-supervision article explains the related spatial evaluation risks.

• Define the nodes and specify the spectral and spatial information used to build relationships.
• State the group or edge rule, its parameters and whether that rule changes during training.
• Compare with simpler alternatives under the same split and annotation budget.
• Inspect boundary pixels, minority classes and disconnected regions, rather than reporting only overall accuracy.
• Record construction time, memory use and inference requirements alongside predictive performance.

Common failure modes

Similarity is not the same as class membership. An inaccurate neighbourhood can spread misleading information, and a hyperedge that crosses a class boundary can mix incompatible signals. Increasing the number of connections may make this problem worse rather than provide more useful context.

Superpixels can reduce computation, but a region that merges two materials may already lose a boundary before the network starts. Topological features can depend strongly on scaling and distance choices. Large relational structures can also consume considerable memory even when the neural network itself is modest.

The practical question is whether the chosen relationships encode useful, available information for the task. Answer it with controlled comparisons and clearly stated assumptions. A representation should earn its place through evidence at the intended scale and evaluation setting.

Run the example

Prerequisite: Python 3. Examples use synthetic inputs to explain the calculation. Save the snippet as example.py and run python3 example.py.

from itertools import combinations

nodes = ['A', 'B', 'C', 'D']
hyperedges = [set('ABC'), set('BCD')]
incidence = [[int(node in group) for group in hyperedges]
             for node in nodes]
pairs = sorted({pair for group in hyperedges
                for pair in combinations(sorted(group), 2)})
print('Incidence:', incidence)
print('Pairs:', [''.join(pair) for pair in pairs])

Verified output

Incidence: [[1, 0], [1, 1], [1, 1], [0, 1]]
Pairs: ['AB', 'AC', 'BC', 'BD', 'CD']

Frequently asked questions

References and further reading

1. Graph Convolutional Networks for Hyperspectral Image Classification
2. Graph Attention Networks
3. Hypergraph Neural Networks
4. Hyperspectral image classification using spectral-spatial hypergraph convolution neural network
5. Topological Persistence and Simplification
6. Ripser: efficient computation of Vietoris–Rips persistence barcodes