Organizations: Machine Intelligence Unit, Indian Statistical Institute, Kolkata, West Bengal 700108, India · Siksha ’O’ Anusandhan, Bhubaneswar, Odisha 751030, India · Computer Vision and Pattern Recognition Unit, Indian Statistical Institute, Kolkata, West Bengal 700108, India · International Institute of Information Technology, Bhubaneswar, Odisha 751003, India
Hyperspectral images (HSI) possess discriminating spectral characteristics with respect to land utilization due to high-dimensional information. Recently, unsupervised representation learning using graph neural networks has unveiled these utilization-specific complex spectral relations. This article has three main contributions. First, it proposes an index, the Mean Absolute Local Neighborhood Inverse Spectral Index (MALNISI), which measures the similarity between two pixels based on their neighborhood information. The index is used to construct a weighted adjacency matrix, giving the HSI a graph structure. Second, it proposes an upgraded Graph Attention Network that learns compact latent representations of hyperspectral data from the constructed graph. In this network, an enhanced multi-head attention mechanism incorporates the MALNISI weights as an auxiliary feature to emphasize learned attention between neighboring pixels. To alleviate over-smoothing and preserve fine-grained spectral details, the architecture integrates residual connections, skip projections, and feature normalization, which together enhance representation capability, stabilize learning, and accelerate convergence. Third, it proposes a weighted contrastive loss function that incorporates neighborhood-aware MALNISI between connected pixels, encouraging strongly connected pixels in the high-dimensional space to remain close in the latent space, and weakly connected pixels to stay apart. The resulting discriminative low-dimensional embeddings are then clustered using spectral clustering. Experimental analysis on six benchmark HSI datasets demonstrates the efficiency and necessity of the three aforementioned contributions, and the comparative study highlights superior performance over fourteen state-of-the-art approaches.
Figures & tables
Fig. 1: Architectural overview of the proposed MALNISI-GAT framework, illustrating the three-stage semantic segmentation procedure: (1) MALNISI-based weighted graph construction, (2) unsupervised representation learning via multi-head Graph Attention Network with weighted contrastive loss, and (3) spectral clustering for segmentation.
Category
Methods
Mechanism
Key Advantages
Limitations
Linear Reduction
PCA, LDA
Linear projection to reduce spectral redundancy.
Computationally simple; fast execution.
Fails to capture non-linear structures.
Centroid-Based
K -means, FCM
Iterative refinement of cluster centers.
Intuitive; incoorporates explainability.
Assumes spherical cluster shapes.
Density-Based
DBSCAN, CFSFDP
Identification of clusters as high-density regions.
Finds arbitrary shapes; robust to noise.
Struggles with the curse of dimensionality.
Graph-Based
FSCAG, SGCNR, BGPC, SAPC
Pixel-to-anchor adjacency and bipartite graph projection.
Near-linear scalability; captures connectivity.
Quality depends on anchor selection/kernels.
Subspace-Based
EnSC, SSC-BP, HESSC, SGLSC, EGFSC
Representation as a union of low-dimensional subspaces.
Handles spectral variability and spatial edges.
Sensitive to manual graph design and noise.
Deep Learning
DCN, DSCRLE, DSCFS
End-to-end non-linear feature extraction via AEs/CNNs.
High discriminative power; non-linear modeling.
Often suffers from fragmented feature integration.
TABLE I: Taxonomy of SOTA HSI unsupervised segmentation methods.
Challenges in SOTA
Proposed Solution (MALNISI-GAT Network)
Spectral Noise Interference
Integrates noise-robust MALNISI in initial graph construction.
Spatially Distributed Spectrally Similar Regions
Graph networks have transitive feature aggregation across multiple layers that strongly pull same-category pixels.
Complex & Non-Compact Data
Weighted contrastive loss forces to maintain spectrally consistency while embedding.
TABLE II: Summary of technical challenges in SOTA and proposed solutions in the MALNISI-GAT framework.
Fig. 2: Illustration and analysis of the MALNISI-based graph construction, where different colors represent different pixel category.
Fig. 3: Detailed illustration of the proposed MALNISI-GAT architecture.
Fig. 4: Geometric visualization of the proposed weighted contrastive loss.
Dataset
Sensor
Location
Image Size
Bands
Categories
Data Type
Indian Pines (IndianP) [ 38 ]
AVIRIS
Indiana, USA
145×145
200
16
Agricultural
Pavia University (PaviaU) [ 38 ]
ROSIS
Pavia, Italy
610×340
103
9
Urban
Salinas [ 38 ]
AVIRIS
Salinas Valley, CA, USA
512×217
204
16
Agricultural
HongHu [ 39 ]
Nano-Hyperspec
Honghu, Hubei, China
940×475
270
22
Agricultural
Botswana [ 38 ]
Hyperion
Okavango Delta, Botswana
1476×256
145
14
Wetland / Vegetation
Kennedy Space Center (KSC) [ 38 ]
AVIRIS
Florida, USA
512×614
176
13
Wetland / Vegetation
TABLE III: Description of the benchmark HSI datasets used for the performance evaluation.
Type
Meric
Purpose
Range
Optimal
Direction
Overall Accuracy (OA)
Overall correct assignment of pixels to corresponding true group
[0,100]
↑
Average Accuracy (AA)
Mean group-wise accuracy with dominant-group bias mitigation
[0,100]
↑
Normalized Mutual Information (NMI)
Similarity between clustering assignments and true group from an information-theoretic perspective with by chance agreement
[0,1]
↑
External Index
Adjusted Rand Index (ARI)
Pairwise similarity between clustering assignments and true group from an information-theoretic perspective
[−1,1]
↑
Adjusted Mutual Information (AMI)
Clustering consistency and structural correspondence between predicted and true group with by chance agreement reduction
[−∞,1]
↑
TABLE IV: Metrics used for the performance evaluation.
TABLE V: Qualitative improvement in discriminative embedding by the proposed framework and comparative spatial segmentation.
Dataset
External Index
Internal Index
Efficiency
OA
AA
NMI
ARI
AMI
FMI
Homogeneity
Completeness
Kappa
Purity
DB
SH
CH
Dunn
Total Inference Time (s)
Inference Time / Sample (s)
Peak Memory (MB)
FLOPS / Sample
IndianP
80.49
57.41
0.7818
0.7143
0.7811
0.7534
0.7559
0.8095
0.7767
0.8049
1.8208
0.206
736.7607
0.0358
5.1013
2.4263e-4
29.54
5.13e+07
PaviaU
90.61
74.76
0.9206
0.9446
0.9205
0.9587
0.8815
0.9633
0.8753
0.9061
1.8759
0.169
3371.7265
0.0515
6.0811
2.932e-5
81.59
3.77e+08
Salinas
86.69
95.74
0.9279
0.7124
0.9279
0.745
0.9201
0.9359
0.8533
0.8669
1.0451
0.4112
7823.4914
0.0182
22.3054
2.0076e-4
155.14
1.43e+09
HongHu
84.46
62.11
0.7964
0.9066
0.7963
0.9259
0.7891
0.8038
0.8049
0.8446
1.865
0.1494
15159.9469
-
7.3074
2.1793e-4
1409.52
1.30e+11
Botswana
80.7
74.02
0.8214
0.7039
0.8198
0.7287
0.8022
0.8414
0.7904
0.807
1.469
0.3249
492.2974
0.0323
2.6438
7.00e-6
9.75
1.33e+05
TABLE VI: Quantitative performance of the proposed framework across all datasets.
Datasets
Metrics
FSCAG [ 22 ]
SGCNR [ 4 ]
HESSC [ 25 ]
NCSC [ 27 ]
SGLSC [ 26 ]
S 3 AGC [ 34 ]
DSCRLE [ 33 ]
BGPC [ 11 ]
SAPC [ 12 ]
EGFSC [ 2 ]
Proposed ± SD
t-statistic
p-value
IndianP
OA
36.13
35.75
34.08
58.65
59.50
49.94
47.80
43.85
55.88
70.08
80.49 ± 0.0728
315.75
3.02e-10
AA
37.82
35.35
35.12
49.46
54.25
42.25
39.88
41.84
52.39
55.74
57.41 ± 0.0683
50.65
4.55e-07
Kappa
0.2955
0.2889
0.2866
0.5346
0.5522
0.4317
0.3996
0.3782
0.4959
0.6580
0.7767 ± 0.0009
303.93
3.52e-10
NMI
0.4513
0.4463
0.4133
0.5911
0.6190
0.5157
0.4779
0.5020
0.5403
0.6910
0.7818 ± 0.0019
105.28
2.44e-08
Purity
0.5187
0.5261
0.5042
0.6404
0.7058
0.5548
0.5270
0.5530
0.5933
0.7517
0.8049 ± 0.0007
159.41
4.64e-09
PaviaU
OA
49.34
57.45
45.03
54.88
62.45
53.38
68.46
74.05
63.94
75.76
90.61 ± 1.5697
17.15
3.39e-05
TABLE VII: Performance comparison with 10 SOTA methods on 4 benchmark HSI datasets with best and second best performance represented in blue and teal . One-sided t-test ( p<0.05 ) is done for statistical significance.
Method
OA
Kappa
NMI
Purity
K -means [ 19 ]
64.56
0.6205
0.7416
0.6456
H-GMM [ 40 ]
66.41
0.6416
0.7493
0.6641
FSEC [ 41 ]
68.20
0.6579
0.8324
0.6820
EnSC [ 24 ]
65.64
0.6331
0.6886
0.6564
SSC-BP [ 23 ]
70.91
0.6875
0.7316
0.7091
DSCFS [ 28 ]
82.39
0.8113
0.8358
0.8239
TABLE VIII: Performance comparison on the Botswana dataset with 6 SOTA methods, with best, second best, and third best performance represented in blue , teal and magenta . One-sided t-test ( p<0.05 ) is done for statistical significance.
Fig. 5: Radar visualization of performance analysis of different similarity indices for segmentation. The angular axes represent different evaluation metrics, the radial axes reflect their normalized range of values, and each polygon corresponds to the effectiveness of the similarity measure.
Fig. 6: Radar visualization of performance analysis of different loss functions.
Fig. 7: Radar visualization of performance analysis under different noise perturbations. The angular axes represent different noise types, the radial axes represent normalized NMI values, and each polygon corresponds to different noise perturbations.
Fig. 8: NMI comparison of traditional and modern feature embedding methods on 6 benchmark HSI datasets. Solid bars denote traditional feature extraction techniques, while hatched bars denote modern embedding methods.
National Engineering Laboratory for Satellite Remote Sensing Applications, Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China · Aerospace Information Research Institute, Chinese Academy of Sciences, Beijing 100094, China · Guangzhou University, Guangzhou 510006, China