REVIEW 4 major objections 6 minor 3 references
Examination of PCA Utilisation for Multilabel Classifier of Multispectral Images
T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash
Pith's one-line read PCA preprocessing helps fine-tuned DINOv2 but hurts ResNet50 in multispectral image classification.
desk verdict Useful practical question, but the PCA effect is confounded with first-layer reinitialization and the paper's own conclusion contradicts Table 5; needs reruns before the claim can be trusted. read the letter →
The pith
A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.
The reading
What carries the argument
The carrying mechanism is a paired experimental pipeline: PCA reduces each 13-band image to 3 principal components, the encoder's first layer is replaced with a randomly initialised 3-channel layer, and a fixed three-layer classifier maps the encoder output to nine sigmoid outputs. Four design choices—encoder (ResNet50, a convolutional residual network, vs DINOv2, a vision transformer), pretraining source (a soft-contrast learning run on Earth-observation data vs a natural-image pretraining), and fine-tuned vs frozen weights—give the paired comparisons that pin down where PCA helps. The PCA itself is fit on a 40% sample of training images so that the comparison reflects the effect of compres
What would settle it
Re-run the DINOv2 fine-tuned configuration with and without PCA over at least five random seeds and several PCA fits on different training subsets. If the F1 gap does not remain substantially positive and the sign does not stay consistent across seeds, the architecture-dependent conclusion is not established. A useful second check: repeat the ResNet50 frozen runs, where PCA cost about 6 F1 points, to see if that loss reproduces.
Extended reading notes
Core claim
On the SSL4EO-S12-ML benchmark—13-band satellite patches labelled into nine land-cover classes that can co-occur—the paper finds that inserting PCA between the raw image and a deep encoder is not neutral. With ResNet50 as encoder, every configuration (soft-contrast pretraining or natural-image pretraining, frozen or fine-tuned) scored lower with PCA than without, with F1 dropping by up to about 6 points. With DINOv2 as encoder, the direction flips when the model is fine-tuned: PCA raised fine-tuned DINOv2 F1 from 31.16 to 62.97 with soft-contrast weights and from 39.56 to 62.95 with natural-image pretraining. With frozen DINOv2, PCA again hurt. The paper attributes the difference to convolut
Load-bearing premise
The load-bearing premise is that the with-PCA and without-PCA runs differ only in PCA preprocessing; since the first encoder layer is randomly reinitialised, PCA is fit on only 40% of the training images, and each configuration runs once, random initialisation or training instability could produce the observed gap instead of PCA itself.
Editorial extensions
If this is right
- For convolutional encoders, PCA preprocessing should not be used as a default: it lowered F1 in every ResNet50 configuration tested, so the spectral-spatial information it discards is apparently load-bearing for CNN features.
- For fine-tuned transformer encoders, PCA can more than double F1 on this task, making it a candidate cheap preprocessing win when compute or input bandwidth is limited.
- For frozen encoders, PCA should be avoided: pretrained weights cannot adapt to the reduced input, and even a trainable first layer does not recover the lost distribution.
- PCA's size and speed benefits are real but modest: parameter counts and model sizes drop slightly, and ResNet50's average inference time decreases, so adopting PCA for efficiency is reasonable where it does not cost accuracy.
- Pretraining the contrastive encoder on PCA-reduced data, as the paper proposes, is the direct next experiment: if it closes the gap, dimensionality reduction can be folded into representation learning rather than bolted on.
Reading between the lines
- A plausible mechanism, which the paper reports but does not demonstrate, is that PCA's channel decorrelation removes the spatial-spectral co-occurrence structure that convolutional kernels are built to exploit, while self-attention in DINOv2 can recombine three high-variance components adaptively. Testing this would require probing intermediate feature statistics with and without PCA.
- The paper fixes the PCA output at three components. Whether DINOv2's gain comes from compression per se or from this particular projection is untested; replacing PCA with a random 3-dimensional linear projection, or varying the number of components, would separate the two.
- Because each configuration was run once, the reported DINOv2 gap (31.16 vs 62.97) could be inflated by initialization luck; a multi-seed replication would likely retain the direction but produce different magnitudes.
- For frozen encoders, PCA may act less like denoising and more like a domain shift—the pretrained weights were fitted to 13-channel inputs, so even the trainable first layer cannot fully compensate. This predicts that PCA will hurt any frozen pretrained encoder fed out-of-distribution channels, not just these two.
Editorial analysis
A structured set of objections, weighed in public.
Referee Report
Summary. The paper studies whether PCA preprocessing (reducing 13-band multispectral images to 3 principal components) helps or hurts multi-label land-cover classification with ResNet50 and DINOv2 encoders, using encoder weights pretrained on ImageNet or SoftCon, in both fine-tuned and frozen regimes, followed by a three-layer classifier. The central claim is that PCA's effect is strongly architecture- and training-strategy-dependent: it helps fine-tuned DINOv2, hurts ResNet50, and generally hurts frozen encoders. The paper also reports model size and inference-time comparisons. The experimental matrix is clearly described, but the PCA/no-PCA comparisons are confounded by unequal first-layer initialization, results come from single runs without seeds, and the concluding explanation is contradicted by one row of Table 5.
Significance. If the claimed architecture dependence were established, the paper would be a useful practical guide for dimensionality-reduction choices in multispectral image classification, and its negative result for ResNet would caution against treating PCA as a neutral preprocessing step. The paper does provide reproducible-looking measurements of model size and inference time, and it benchmarks a reasonable set of configurations. However, the central comparison does not isolate PCA: the first encoder layer is randomly reinitialized in some arms and pretrained in others, and no seed variation is reported. As presented, the evidence is not sufficient to support the architecture-dependence conclusion. The core experimental design can be repaired with additional runs and controlled initialization, so the work is a candidate for major revision rather than rejection.
major comments (4)
- [§3.3 and §4.1, Table 5] The PCA versus no-PCA comparison conflates input dimensionality with first-layer weight initialization. For ImageNet-pretrained encoders (3-channel input), the no-PCA arm (13-channel input) cannot use the pretrained first layer, so it is randomly initialized and trainable, while the PCA arm uses the original pretrained first layer. For SoftCon-pretrained encoders (13-channel input), the situation is reversed: the PCA arm loses the pretrained first layer, while the no-PCA arm keeps it. Thus the flagship DINOv2/SoftCon/fine-tuned gap (F1 62.97 vs 31.16) and the frozen comparisons (e.g., DINOv2/ImageNet/frozen F1 63.17 vs 56.99) compare different initialization states as well as different input dimensionalities. The conclusion that PCA itself is architecture-dependent is not supported unless first-layer initialization is equalized across arms (e.g., by reinitializing in both arms, or by ada
- [§6 vs. Table 5, DINOv2/ImageNet/frozen row] The conclusion states that 'in frozen encoder configurations, PCA degraded performance' and 'applying PCA to frozen encoders appears counterproductive.' This is contradicted by Table 5: for DINOv2/ImageNet/frozen, PCA improves F1 from 56.99 to 63.17, accuracy from 75.04 to 79.96, and precision from 63.29 to 71.20. The same table shows that the effect is not uniform across frozen configurations. The textual conclusion needs to be revised to describe the actual pattern, and any interpretation must account for the initialization confound.
- [§4.1, Table 5] Every configuration is run once, with no seeds, no error bars, and no statistical comparison. Given the extreme variance suggested by the no-PCA fine-tuned DINOv2 results (F1 31.16, precision 24.01), a single run cannot distinguish a true PCA effect from random initialization or training instability. The paper should report means and standard deviations over at least three seeds, or otherwise justify why a single run is adequate. This is load-bearing for the central claim.
- [§3.1] The claim that the PCA components 'collectively explaining 99.9% of the total variance' is ambiguous. Since the pipeline reduces to 3 dimensions, the reader infers that the first 3 principal components explain 99.9%, but this is not stated explicitly. In addition, PCA was fit on only 40% of randomly selected training samples, and no details are given about whether data were standardized before applying PCA (the Related Work section describes standardization as part of PCA). Please clarify the exact number of retained components, the variance they explain, and the preprocessing steps, as these directly affect the interpretation of the results.
minor comments (6)
- [§2.2, Eq. (3)] The notation for the SoftCon loss is hard to parse: Yi j = yi ⊗ yTj and Xi j = zi ⊗ zTj are not defined in the text; using '⊙' or explicitly stating that ⊗ denotes an outer product would improve clarity.
- [Table 2] The entries 'resnet50' and 'dinov2' use inconsistent capitalization, and the footnote '*' explains that encoder weights are not updated except the first layer, but the textual description of which configurations this applies to is not explicit until §4.1. Consider merging Table 2 with the explanation.
- [Table 4] The heading 'Size Comprehension' appears to be a typo; presumably 'Size Comparison' is intended.
- [§4.1 and Table 3] The methodology states inference times are computed as average per sample over 2,510 samples passed five times, but Table 3 reports times 'per batch of 64 samples.' Please clarify the batch size and the conversion.
- [§5] The discussion attributes the ResNet decline to PCA 'removing useful spectral-spatial information' and the DINOv2 improvement to 'minimising redundant spectral information.' These are post hoc interpretations; the study does not directly measure information content or redundancy. Flagging them as hypotheses would be more appropriate.
- [§3.1] The random selection of 100,430 samples for PCA fitting is not seeded, so the PCA transform is not reproducible. A fixed random seed should be reported.
Circularity Check
No circularity: results are direct experimental measurements, not derived from fitted predictions or self-citation chains.
full rationale
This paper makes no derivation-based claim that reduces to its own inputs. The central result—that PCA's effect depends on architecture and training strategy—is an empirical observation from training runs reported in Table 5. The PCA transform is fit unsupervised on 40% of training images (Section 3.1) and is not optimized against the classification labels or the final F1 scores; no parameter is fitted to the outcome it is then said to predict. The model-size and inference-time comparisons are direct measurements, and the smaller size with PCA is a mechanical consequence of reducing input channels, not a disguised prediction. There are no load-bearing self-citations: the cited SoftCon and SSL4EO works are by other authors and serve as dataset/pretraining sources, not as justification for the paper's conclusion. The paper itself flags a genuine methodological limitation in Section 4.1: when the pretrained encoder's input dimension mismatches the number of PCA or raw channels, the first layer is randomly initialized and left trainable, so PCA/no-PCA comparisons are partially confounded by initialization and trainability. That is a validity threat to the architecture-dependence claim, not a circularity. No equation in the paper is equivalent by construction to any other, and nothing is renamed as a result. Accordingly, the appropriate circularity score is 0.
Assumptions & free parameters
free parameters (2)
- Number of PCA components =
3
- PCA training subset fraction =
40% (~100,430 samples)
assumptions (3)
- domain assumption Top 3 principal components retain enough spectral information for multi-label classification
- domain assumption SSL4EO-S12-ML labels from Google Dynamic World are correct and complete
- domain assumption Reinitializing and training only the first encoder layer restores compatibility with the new input dimensionality
Cite this review
Pith. "Pith review of Examination of PCA Utilisation for Multilabel Classifier of Multispectral Images." pith.science (2026). https://pith.science/paper/4OWQI7VC
@misc{pith2026250901691,
author = {Pith},
title = {Pith review of: Examination of PCA Utilisation for Multilabel Classifier of Multispectral Images},
year = {2026},
howpublished = {\url{https://pith.science/paper/4OWQI7VC}},
note = {Machine review of arXiv:2509.01691}
}
read the original abstract
This paper investigates the utility of Principal Component Analysis (PCA) for multi-label classification of multispectral images using ResNet50 and DINOv2, acknowledging the high dimensionality of such data and the associated processing challenges. Multi-label classification, where each image may belong to multiple classes, adds further complexity to feature extraction. Our pipeline includes an optional PCA step that reduces the data to three dimensions before feeding it into a three-layer classifier. The findings demonstrate that the effectiveness of PCA for multi-label multispectral image classification depends strongly on the chosen deep learning architecture and training strategy, opening avenues for future research into self-supervised pre-training and alternative dimensionality reduction approaches.
Figures
Reference graph
Works this paper leans on
-
[2020]
ImageNet: A large-scale hi- erarchical image database
arXiv: 2011 . 10566 [cs.CV] . URL: https : // arxiv . org / abs / 2011.10566. [Den09] Jia Deng et al. “ImageNet: A large-scale hi- erarchical image database”. In: 2009 IEEE Conference on Computer Vision and Pattern Recognition. 2009, pp. 248–255. DOI: 10. 1109/CVPR.2009.5206848. [He15] Kaiming He et al. Deep Residual Learn- ing for Image Recognition . 201...
arXiv 2011
-
[2024]
Scikit-learn: Ma- chine Learning in Python
arXiv: 2304 . 07193 [cs.CV] . URL: https : // arxiv . org / abs / 2304.07193. [Ped11] F. Pedregosa et al. “Scikit-learn: Ma- chine Learning in Python”. In: Journal of Machine Learning Research 12 (2011), pp. 2825–2830. [Ram24] Billy G. Ram et al. “A systematic review of hyperspectral imaging in precision agriculture: Analysis of its current state and futu...
arXiv 2011
-
[6251]
DOI: 10 . 1109 / TGRS . 2016 . 2584107. [Che20] Xinlei Chen and Kaiming He. Exploring Simple Siamese Representation Learning
work page 2016
Reviewed August 5, 2026 · model on record in the stance chip above.
Discussion (0). Sign in to comment.