STMNet: Single-Temporal Mask-Based Network for Self-Supervised Hyperspectral Change Detection - Integrity Review (DOI: 10.1109/TGRS.2024.3523541)

Academic fraud report · Geng Detector

Summary

This report assesses allegations of academic misconduct against Zhou et al.'s STMNet paper published in IEEE TGRS (2025, online Dec 2024). Verdict: Highly suspicious. Three substantive concerns are raised: (1) a core methodological flaw in which horizontal spectral flipping is used to simulate physical land-cover change in a single-temporal hyperspectral image, contradicting the physics of real spectral transitions; (2) a single-seed (seed=1) deep-learning experimental setup that violates publication norms for stochastic methods, raising strong cherry-picking concerns; and (3) an implausible result claiming an unsupervised method reaches 99.08% OA on the Farmland dataset, outperforming supervised ReCNN (98.97%) with 0.2% labels, while other deep models collapse (BCNN: 94.30%). Pixel-level image forensics was not possible from the text-only source. Timeline and references appear consistent. Overall confidence is moderate-high; final determination requires the editorial investigation and additional experiments requested.

Verdict

🟠 Highly suspicious. Multiple serious methodological and experimental issues cast strong doubt on the validity of the reported results. Not confirmed misconduct, but warrants editorial scrutiny.

Key findings

Spectra-flipping simulation is physically unsound (Finding 1). The authors propose to simulate the "second-temporal" change by laterally flipping the spectrum of a center pixel and using it to fill masked regions. Real land-cover change is a non-linear spectral transition, not a mirror reflection. This undermines the entire self-supervised signal.
Single-seed experiments violate reporting norms (Finding 2). The paper explicitly states training-sample selection uses seed = "1" only. For a deep-learning method whose results depend strongly on data partitioning, reporting only one run is a textbook cherry-picking risk and prevents reproducibility.
Implausible unsupervised > supervised result (Finding 3). On the Farmland dataset, the unsupervised STMNet allegedly reaches 99.08% OA, exceeding supervised ReCNN (98.97%) trained with 0.2% labels, while the classical CVA already achieves 95.25%. If the dataset is this easy, the supervised ceiling should be much higher; instead, BCNN drops to 94.30%. The inconsistency suggests an unreliable comparison setup.
Image forensics not performed (Finding 4). Text-extracted PDF content prevented pixel-level duplication/manipulation checks on Figures 9–11 and architecture diagrams.
No timeline anomalies (Finding 5). Latest citations are from May 2024; submission in September 2024 is temporally consistent. Hardware (RTX 3090) is plausible.

Evidence highlights

Direct quote on spectral flip: *"we extract the center position of a spectral band... and perform a lateral flip operation on it as the spectrum of the changing pixel point... filling the masked region with this spectrum can simulate the real change sample."*
Direct quote on random seed: *"we selected the training samples by using one as a random seed."*
Quantitative claims from Table I (Farmland): STMNet unsupervised OA = 99.08%; ReCNN supervised (0.2% labels) OA = 98.97%; BCNN OA = 94.30%; classical CVA OA = 95.25%.
DOI: 10.1109/TGRS.2024.3523541.
Year: 2025 (online December 2024), journal: IEEE Transactions on Geoscience and Remote Sensing (TGRS).

Notes

The reviewer explicitly recommends requesting multi-seed (≥3) variance results and raw training logs from the authors, raising the spectra-flipping physics question on PubPeer, and notifying IEEE TGRS of the non-standard single-seed experimental protocol.
Limitations of this assessment: only text-extracted content was available; image-based duplication and statistical re-analyses could not be performed. Confidence is moderate-high but not definitive; institutional investigation is required for a formal ruling. Uncertainty is explicitly noted in the original disclaimer.
No evidence was found of citation anomalies, future references, or implausible equipment claims.