Evolving superpixel-level affinity based on contrastive learning and good neighbors for hyperspectral image clustering
Recently, graph clustering has been applied to hyperspectral image (HSI) clustering and proves to be effective on capturing the complex affinity among hyperspectral samples to a certain extent. However, graph clustering based on sample-level affinity usually suffers from a heavy computation overhead...
Saved in:
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
Elsevier
2025-08-01
|
| Series: | International Journal of Applied Earth Observations and Geoinformation |
| Subjects: | |
| Online Access: | http://www.sciencedirect.com/science/article/pii/S1569843225001487 |
| Tags: |
Add Tag
No Tags, Be the first to tag this record!
|
| Summary: | Recently, graph clustering has been applied to hyperspectral image (HSI) clustering and proves to be effective on capturing the complex affinity among hyperspectral samples to a certain extent. However, graph clustering based on sample-level affinity usually suffers from a heavy computation overhead due to the generation of coefficient matrices. Meanwhile, since the spatial information of HSIs may not be sufficiently utilized in these methods, it is difficult for them to obtain robust clustering performance when processing various HSIs. In order to alleviate the negative effect of sample-level affinity, we achieve evolving superpixel-level affinity (ESA) by collaboratively integrating the sample-level contrastive learning (CL) for learning discriminative features with finding the inter-superpixel good neighbors (GN) of each sample. Initially, CL is operated based on the assumption of homogeneous superpixels to pull together samples from the same superpixel and the learned features are fed to the smooth representation model to pursue the representation coefficients that used for generating GN. Then, the superpixel-level affinity is updated by simply converting the similarity between samples in each pair of GN into the similarity of the corresponding superpixels. By combining the homogeneous assumption of superpixels with the updated superpixel-level affinity, further CL is achieved to acquire more GN. In this way, CL and generation of GN works collaboratively to evolutionarily solve the superpixel-level affinity. When the process finishes, density-based spectral clustering is first implemented on several affinity matrices selected by designed growth rate of Eigenvalues of normalized Laplacian matrix to obtain clustering results of superpixels. The superpixel-to-pixels projection is then applied to these results to gain clustering maps. Finally, the majority vote strategy is adopted to realize the final clustering. Comprehensive experiments on four benchmark HSIs demonstrate that the proposed ESA method is superior to the considered baseline counterparts in terms of clustering accuracy. |
|---|---|
| ISSN: | 1569-8432 |