Integrity Screening Report: A Wavelet-Augmented Dual-Branch Position-Embedding Mamba Network for Hyperspectral Image Change Detection (DOI: 10.1109/TGRS.2025.3602225)

Academic fraud report · Geng Detector

Summary

This report evaluates a 2025 IEEE TGRS paper proposing a Wavelet-augmented Dual-branch Position-embedding Mamba network for hyperspectral image change detection. Overall verdict: highly suspicious. The most serious finding is a major internal inconsistency between the main comparison results (Section IV-C) and the threshold-sensitivity analysis (Section IV-E3). For three datasets, the OA reported as 'best' in the main results is several percentage points higher than the OA at the paper's own stated optimal threshold. For River, main-text OA is 96.65% but the optimal threshold (0.8) yields 93.1%; Hermiston shows 98.94% vs 94.5% at threshold 0.9; Farmland shows 94.95% vs 92.8% at threshold 0.75. Secondary issues include: a runtime stack inconsistent with PyTorch version compatibility (PyTorch 1.10.1 does not support Python 3.10), and selection of baselines that omits concurrent 2024–2025 Mamba-based HSI-CD methods despite citing them. Confidence is high for the threshold inconsistency (values are explicitly stated in the paper); medium for the PyTorch/Python claim (derivable from official release notes); lower for the baseline-selection critique (interpretive).

Verdict

🟠 Highly suspicious. The paper contains a hard quantitative inconsistency between its main results and its own hyperparameter (threshold) analysis, which is unlikely to be a benign typo and points to possible data fabrication or careless copy-paste of experimental records. Auxiliary issues (incompatible software versions, selective baseline selection) reinforce but do not by themselves prove misconduct.

Key findings

Internal performance contradiction across sections. The main comparison (Sec. IV-C) and the threshold-value study (Sec. IV-E3) report mutually incompatible 'best' Overall Accuracy (OA) values for the same model on the same datasets.
Magnitude of discrepancy is large (≈2.2–3.6 percentage points). This is well outside the range expected from re-tuning or noise.
Software stack inconsistency. Paper claims PyTorch 1.10.1 with Python 3.10.0; PyTorch 1.10.1 predates Python 3.10 support, which began in PyTorch 1.11 (March 2022). Suggests templated boilerplate carried over from an earlier project.
Selective baseline comparison. Among cited references [48]–[54] (2024 Mamba-related works), none appear to be quantitatively compared in the main experiments, despite being directly relevant to the proposed Mamba-based HSI-CD method.
Verification gaps. Figures 1–15 and the data rows of Tables II–VII were not available for pixel-level image reuse checks or statistical tests (Benford, variance homogeneity).

Evidence highlights

Reported numeric discrepancy (main-text OA vs. OA at the paper's stated optimal threshold):

| Dataset | Main-text OA | Optimal threshold | OA at optimal threshold | Δ | |---|---|---|---|---| | River | 96.65% | 0.80 | 93.1% | −3.55 pp | | Hermiston | 98.94% | 0.90 | 94.5% | −4.44 pp | | Farmland | 94.95% | 0.75 | 92.8% | −2.15 pp |

Source locations cited in the report: Section IV-C (Comparison Results) vs. Section IV-E3 (Threshold Value), and Section IV-B3 (Implementation Details).

Notes

What is well-supported: The numerical contradictions in the table above are taken directly from the paper's own text as quoted in the source report; they are therefore reliable points of inquiry regardless of how the discrepancy arose.
What is interpretive: The 'copy-paste after Ctrl+H' hypothesis is one plausible explanation; alternative explanations (e.g., the main results used a different train/val split, a different metric definition, or a post-hoc tuned threshold not disclosed in Sec. IV-E3) need to be ruled out by examining code, logs, or raw predictions. The author response to a data-and-code request is the decisive next step.
What could not be verified: Pixel-level image reuse in Figures 1–15; Benford-law or variance checks on Tables II–VII (only headers were extractable). These are flagged as 'cannot verify' rather than cleared.
Recommended actions from the original report: request raw data, code, and threshold-selection logs; raise the issue on PubPeer; consider editorial/institutional escalation depending on the author's response.
Uncertainty: This is a preprint/online screening report, not an institutional investigation. Final determination of misconduct requires verification by the publisher or the authors' institution.