Forensic Report: SRF: SpectrumRecombineFormer for Hyperspectral Image Classification (DOI: 10.1145/3715698)

Academic fraud report · Geng Detector

Summary

This report assesses the ACM TOMM 2025 paper 'SRF: SpectrumRecombineFormer for Hyperspectral Image Classification' by Jing et al. The overall verdict is 'highly suspicious.' Based solely on the provided text (image-based forensic analysis was not possible), the manuscript exhibits strong hallmarks of 'assembly-line' academic production. Key findings include: (1) a citation error in which reference [20]—a hyperspectral denoising paper—is mischaracterized in the introduction as a classification framework (SSRT-UNet); (2) extensive verbatim duplication between Section 1 and Section 2.2, suggesting copy-paste composition; (3) contradictory training-epoch reporting (300 epochs in Section 4.2 vs. 100 epochs in Section 4.3.4), undermining reproducibility; and (4) suspiciously poor baseline results, notably PyFormer achieving 1.85% on Houston2013 Class 6 (Water) versus >90% for other methods, indicating potential misconfiguration of competing models. No image-level manipulation could be assessed. Confidence is moderate given textual evidence alone; final determination of misconduct requires institutional investigation.

Verdict

🟠 Highly suspicious. Text-only analysis reveals citation misattribution, large-scale verbatim duplication, methodological inconsistencies, and unreliable baseline comparison data, all consistent with low-rigor 'paper-mill' style production. No image-based forensics could be performed.

Key findings

Citation misattribution (Reference [20]): The introduction states Fu et al. [20] proposed a classification framework named SSRT-UNet combining RNN and Transformer. However, the bibliography entry describes [20] as a 2024 paper on hyperspectral image denoising via spatial-spectral recurrent transformer. The cited work appears mislabeled as a classification method.
Verbatim duplication between Section 1 and Section 2.2: The transformer-based-HSI-classification paragraph is essentially identical across the introduction and the related-work section, including the same citations and phrasing (He et al. [26] HSI-BERT; Hong et al. [29] SpectralFormer; Zhong et al. [71] SSTN; Mei et al. [43]; Ahmad et al. [3]).
Contradictory epoch reporting: Section 4.2 sets training to 300 epochs for all datasets, while Section 4.3.4 records time per 100 epochs, and Table 9 lists 'Train (s)' without clarifying whether it reflects 100 or 300 epochs.
Implausible baseline accuracy: In Table 11 (Houston2013), PyFormer reports 1.85% on Class 6 (Water) while competing methods exceed 90%; on Class 15 (Running Track) it reports 71.21% versus ~97% for others, raising concerns about baseline tuning integrity.
Image forensics not feasible: Original figures were unavailable, so pixel-level checks (duplication, splicing, post-hoc accuracy edits) could not be executed.

Evidence highlights

Finding 1 — Reference [20] mismatch: Introduction describes SSRT-UNet (RNN+Transformer, classification). Reference list: 'Guanyiman Fu... 2024. Hyperspectral image denoising via spatial-spectral recurrent transformer.'
Finding 2 — Duplicated paragraph: Section 1 and Section 2.2 share the sentence 'Recent studies have investigated the application of Transformer networks for HSI classification... He et al. [26]... Hong et al. [29]... Zhong et al. [71]... Mei et al. [43]... Ahmad et al. [3]...' with minimal variation.
Finding 3 — Epoch inconsistency: Section 4.2: '...the number of epochs set to 300 for all datasets.' Section 4.3.4: 'The time required to train for 100 epochs for each method was recorded...'
Finding 4 — Houston2013 Table 11 outliers: PyFormer Class 6 (Water) = 1.85%; Class 15 (Running Track) = 71.21%, in contrast to peer methods in the 90%+ range.
DOI preserved: https://doi.org/10.1145/3715698

Notes

All conclusions derive from text provided by the user; high-resolution figures were not supplied, so image-based fraud checks (Western blot reuse is inapplicable here, but classification-maps splicing/splicing and figure reuse could not be assessed) were skipped.
The severity ratings and tagging follow the original Chinese report (🟠 moderate-high; 🟡 moderate).
Mislabeling a denoising paper as a classification framework is a factual error that, combined with copy-paste related-work prose, lowers confidence in the manuscript's editorial rigor.
PyFormer's collapse on Houston2013 (1.85% on a class where peers exceed 90%) is not, by itself, proof of data fabrication, but combined with the other findings it constitutes reasonable cause for editorial follow-up (e.g., requesting code, logs, and clarified epoch accounting).
Any definitive finding of academic misconduct requires an official investigation by the journal or institution; this report is informational only.