[Paper Review] A Joint Morphological Profiles and Patch Tensor Change Detection for Hyperspectral Imagery

Zengfu Hou, Wei Li|arXiv (Cornell University)|Jan 20, 2022

Remote-Sensing Image Classification4 citations

TL;DR

This paper proposes a novel joint morphological profiles and patch-tensor change detection (JMPT) method for hyperspectral imagery that enhances detection accuracy by jointly exploiting spatial-spectral features through patch-based tensor reconstruction and multi-scale morphological profiles. The method achieves state-of-the-art performance on two real datasets, with AUC scores of 83.769% (Hermiston) and 96.119% (Yancheng), outperforming existing methods despite higher computational cost.

ABSTRACT

Multi-temporal hyperspectral images can be used to detect changed information, which has gradually attracted researchers' attention. However, traditional change detection algorithms have not deeply explored the relevance of spatial and spectral changed features, which leads to low detection accuracy. To better excavate both spectral and spatial information of changed features, a joint morphology and patch-tensor change detection (JMPT) method is proposed. Initially, a patch-based tensor strategy is adopted to exploit similar property of spatial structure, where the non-overlapping local patch image is reshaped into a new tensor cube, and then three-order Tucker decompositon and image reconstruction strategies are adopted to obtain more robust multi-temporal hyperspectral datasets. Meanwhile, multiple morphological profiles including max-tree and min-tree are applied to extract different attributes of multi-temporal images. Finally, these results are fused to general a final change detection map. Experiments conducted on two real hyperspectral datasets demonstrate that the proposed detector achieves better detection performance.

Motivation & Objective

• To address the limitation of traditional change detection methods in fully exploiting spatial and spectral correlations in multi-temporal hyperspectral images.
• To overcome the loss of spectral information caused by dimensionality reduction in existing single-band or multi-band algorithms.
• To develop a method that preserves the 3D structure of hyperspectral data while enhancing feature representation for improved change detection.
• To integrate morphological profiles and tensor-based reconstruction to better capture local spatial structures and spectral variations.
• To achieve higher detection accuracy and better separability between changed and unchanged pixels.

Proposed method

• A patch-based tensor strategy is used to reshape non-overlapping local image patches into 3D tensor cubes, preserving spatial structure.
• Three-order Tucker decomposition is applied to the tensor cubes to extract low-rank representations and reconstruct robust multi-temporal hyperspectral data.
• Multiple morphological profiles—including max-tree and min-tree—are computed to extract topological and intensity-based attributes from the reconstructed data.
• The morphological profiles and tensor-reconstructed features are fused to generate a final change detection map.
• The method leverages both spectral variation and spatial context by combining morphological analysis with tensor-based dimensionality reduction and reconstruction.
• The framework is designed as a dual-pipeline system to process spatial and spectral features in parallel before fusion.

Experimental results

Research questions

• RQ1 Can joint processing of morphological profiles and patch-tensor representations improve change detection accuracy in hyperspectral imagery?
• RQ2 How does preserving the 3D structure of hyperspectral data through tensor decomposition enhance feature representation compared to 2D flattening?
• RQ3 To what extent do max-tree and min-tree morphological profiles improve separability between changed and unchanged pixels?
• RQ4 How does the proposed JMPT method compare to state-of-the-art methods in terms of AUC and computational efficiency?
• RQ5 Can the fusion of tensor-based reconstruction and morphological features better suppress background noise and highlight true changes?

Key findings

• The JMPT method achieved an AUC of 83.769% on the Hermiston dataset, outperforming all compared methods, including TDRD (80.215%) and SALA (80.445%).
• On the Yancheng dataset, JMPT achieved the highest AUC of 96.119%, significantly exceeding TDRD (95.275%) and SALA (95.004%).
• Statistical separability analysis showed that the interval between changed (red) and unchanged (blue) pixel distributions was largest for JMPT, indicating superior class separability.
• ROC curves demonstrated that JMPT's curve was consistently closest to the upper-left corner, confirming its superior detection performance across both datasets.
• The method exhibited higher computational cost, with execution times of 190.62 seconds on Hermiston and 49.40 seconds on Yancheng, primarily due to morphological feature extraction and Tucker decomposition.