acceptodds
Under review as a conference paper at ICLR 2027

MARA-NIR: Masked Autoencoder Representation Analysis of Near-Infrared Spectra

Abstract

Near-infrared (NIR) spectroscopy predicts the chemical composition of a sample from a single one-dimensional spectrum. Before calibration, chemometricians preprocess this spectrum with fixed formulas, such as scatter corrections and derivative filters. These formulas tell us in advance what a useful representation of NIR spectra computes, which is rare in representation learning. We use them to ask what a masked autoencoder learns without labels. We pretrain one on 7.2 million NIR spectra of food and feed and compare its encoder with the formulas, with untrained encoders and with supervised models. Its first layer rediscovers Savitzky-Golay derivative filters, which neither untrained encoders nor supervised models trained on soil carbon learn. Deeper in the encoder, soil organic carbon becomes more predictable, while untrained encoders stay flat. We then test the frozen encoder on the public LUCAS soil library, a new domain measured on a new instrument. With a linear model, it matches tuned chemometric models for soil organic carbon when all labels are used, and with 50 to 400 labels it stays within 0.05 in of the best of them.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
near-infrared spectroscopy, self-supervised learning, masked autoencoders, representation analysis, chemometrics

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_pqiSOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper investigates representations learned by a masked autoencoder for near-infrared spectroscopy. The authors pretrain on approximately 7.2 million food and feed spectral scans from 1.3 million measurements. The model divides spectra into contiguous patches, uses positional encodings based on absolute wavenumber, and trains a Transformer through reconstruction of independently normalized masked patches.

Critical strengths
  1. Practical relevance and valuable transfer evaluation with limited labels. Chemical-property labels require costly laboratory measurements, making unlabeled pretraining valuable for reducing annotation requirements. Transfer from food and feed to soil measured with a different instrument provides meaningful evidence under both material and instrument shifts, extending beyond evaluation within the pretraining domain.

  2. A concrete, testable approach to representation analysis. Savitzky–Golay (SG) filters provide a mathematically defined reference for interpreting learned operations. Beyond visualizing weights, the authors examine their findings using untrained encoders, phase-randomized controls with matched smoothness, polynomial subspace analysis, and layerwise probes.

Critical weaknesses

The link from filter similarity to chemometric functionality and transfer gains remains insufficiently established.

The analyses support the emergence of first-layer weights resembling SG templates and a low-rank polynomial structure. However, these observations do not establish that the filters perform the corresponding derivative or baseline-removal operations within the network. High template correlation does not guarantee the expected response to constant or linear baselines, depending on derivative order. If Pearson correlation is used, adding a constant to every filter coefficient leaves the correlation unchanged while altering the response to the local spectral mean. Structural similarity therefore needs to be distinguished from functional equivalence.

The connection to downstream gains also remains observational. The 90% masking model has a substantially higher template-match rate than the 75% model, yet performs no better in few-label prediction; the distributions of matched derivative orders also differ. This does not rule out the value of derivative features, but indicates that match rate alone cannot explain representation quality. Targeted interventions or controlled filter replacements are needed to establish whether the identified structures contribute to the reported transfer gains.

Because the architecture and training objective largely follow the standard MAE framework, the strength of this mechanistic explanation is central to my assessment of the paper’s substantive contribution.

Questions for authors
  1. Does the correlation measure center the learned filter weights? Beyond high correlation, do matched filters exhibit the expected responses to constant, linear, and other low-degree polynomial inputs for their identified derivative orders? Direct comparisons of learned-filter and SG outputs would help assess functional correspondence beyond weight similarity.

  2. Does “rediscovering chemometric preprocessing” refer to similarity between local weights and SG templates, derivative-like behavior of local mappings, or suppression of physical interference while preserving chemical information in the overall representation? Given that Appendix E explicitly states that the encoder does not remove scatter by itself, which interpretation does the current evidence support? This distinction materially affects my assessment of the main analytical contribution.

  3. What invariances and information losses do per-spectrum input normalization and per-patch target normalization introduce, respectively? In particular, how does the emphasis on local shape induced by target normalization relate to the observed polynomial and derivative-like filters? How do the authors distinguish representational preferences induced by the task design from preprocessing mechanisms learned from spectral data?

3 comments
AuthorsOpenReview ↗

Response to Reviewer pqiS (Q1, Q2)

We appreciate that pqiS finds the transfer evaluation with limited labels practically relevant and valuable, and the approach to representation analysis concrete and testable.

Q1: However, these observations do not establish that the filters perform the corresponding derivative or baseline-removal operations within the network. High template correlation does not guarantee the expected response to constant or linear baselines, depending on derivative order. If Pearson correlation is used, adding a constant to every filter coefficient leaves the correlation unchanged while altering the response to the local spectral mean. Structural similarity therefore needs to be distinguished from functional equivalence.

A1: We ran the requested tests to distinguish structural similarity from functional equivalence, and added them to Section 3 and Appendix C. Light scattering from particle size and packing adds offsets and slopes to a spectrum, and derivatives remove a constant offset and a linear baseline locally (Rinnan et al., 2009). A derivative operation should therefore respond little to the level and slope of a patch, and strongly to its curvature and cubic shape, which follow the overtone and combination bands (L031-042). We added each of these 4 shapes to one patch at a time and measured how much its token changes, at every layer. From the first block on, level and slope cause less than 6% of the change, and curvature and cubic shape 81 to 84%, over 6 seeds with 75% masking and closely with 90%. An untrained encoder gives each of the four about 2%, the chance share among the 48 directions we test (L232-239, Table 10). The pretrained encoder thus performs the derivative and baseline-removal operations, as the Savitzky-Golay (SG) derivatives of chemometrics do. Because the Pearson correlation ignores a constant added to a filter, we also report the level alone. The first layer gives it 3 to 5%, slightly above chance, and from the first block on the encoder gives it less than 1% (Table 10): the network removes the level, as a derivative does.

Q2: The connection to downstream gains also remains observational. The 90% masking model has a substantially higher template-match rate than the 75% model, yet performs no better in few-label prediction; the distributions of matched derivative orders also differ. This does not rule out the value of derivative features, but indicates that match rate alone cannot explain representation quality. Targeted interventions or controlled filter replacements are needed to establish whether the identified structures contribute to the reported transfer gains.

A2: We followed the request for targeted interventions and controlled filter replacements in a study of 32 pretrainings, added to Section 3 and Appendix C (L240-247, L969-977). We froze the first layer to SG filters on 5 seeds and to random filters on 5 seeds, pretrained the rest, and compared each with the learned encoder of the same seed (Figure 4a). With SG filters, transfer stays at the level of the learned encoder on every seed, losing at most 0.013 in R2R^2 (median over all settings). With random filters, transfer drops on 3 of 5 seeds. The response test separates them: all 13 encoders that transfer, the 6 learned, the 5 SG and the 2 remaining random ones, give level and slope at most 4% of their output change, and the 3 that drop give them 13 to 20% (Figure 10). With the derivative operations built in, transfer is reliable. Without them, it depends on whether the trainable layers rebuild them.

Following Yosinski et al. (2014), we then also kept the blocks after the first layer at their random weights, up to 12 random layers in total, on 2 seeds (L270-278, Figure 4b). Up to 8 random layers, transfer stays within 0.10 of the learned encoder. Beyond 8, it falls on both seeds, by 0.25 with 12 random layers. Yet these encoders still solve the pretraining task: with 12 random layers, the validation loss of the reconstruction is 0.029, against 0.028 with 6. Solving the pretraining task therefore does not guarantee transfer.

This also explains why the 2 mask ratios transfer alike despite different match rates and derivative orders: both respond little to level and slope and mostly to curvature and cubic shape (Tables 8 and 10). These results suggest that the derivative operations, not the match rate, contribute to the transfer gains.

AuthorsOpenReview ↗

Response to Reviewer pqiS (Q3, Q4)

Q3: Because the architecture and training objective largely follow the standard MAE framework, the strength of this mechanistic explanation is central to my assessment of the paper’s substantive contribution.

A3: The mechanistic explanation now rests on 3 levels, each tested against a control. The filters resemble SG templates, 2 to 39 times more often than filters of the same smoothness, and per derivative order far more than random filters in the same subspace (Tables 1 and 11). The encoder performs the derivative operation: from the first block on, it removes level and slope, while an untrained encoder responds to them at chance (A1, Table 10). These operations contribute to transfer: frozen SG filters transfer as well as learned ones on every seed, while random filters lose transfer on 3 of 5 seeds. Every one of these encoders that transfers removes level and slope, including the 2 random ones, and the 3 that drop do not (A2, Figures 4 and 10).

Q4: Does the correlation measure center the learned filter weights? Beyond high correlation, do matched filters exhibit the expected responses to constant, linear, and other low-degree polynomial inputs for their identified derivative orders? Direct comparisons of learned-filter and SG outputs would help assess functional correspondence beyond weight similarity.

A4: Yes. The correlation is computed after removing the mean of both the filter and the template (L843-844). We therefore tested each matched filter directly: an SG derivative of order k gives no response to polynomials below degree k, and a linear filter's response to each polynomial is its weight on it. The matched filters put 80 to 99% of their response on their own degree and at most 13% on the lower ones, against 17 to 84% for random filters in the same subspace (L930-940, Table 11). Per seed, only 1 to 9 of 29 to 73 first-derivative filters still respond more to a constant, which the encoder removes from the first block on (Table 10). The functional correspondence thus holds for the matched filters, beyond weight similarity.

AuthorsOpenReview ↗

Response to Reviewer pqiS (Q5, Q6)

Q5: Does “rediscovering chemometric preprocessing” refer to similarity between local weights and SG templates, derivative-like behavior of local mappings, or suppression of physical interference while preserving chemical information in the overall representation? Given that Appendix E explicitly states that the encoder does not remove scatter by itself, which interpretation does the current evidence support? This distinction materially affects my assessment of the main analytical contribution.

A5: We apologise for the ambiguity. Chemometric preprocessing comprises 2 families of methods, scatter corrections and spectral derivatives (Rinnan et al., 2009). Scatter corrections such as SNV standardise each spectrum, to reduce the multiplicative and additive effects of particle size and packing. Spectral derivatives such as SG act locally: they remove the offset and slope of the baseline within a window and sharpen overlapping bands. Pipelines usually apply both, SNV followed by a derivative, and part of the scatter remains, which also carries information such as particle size (L1101-1115). Our tuned pipelines apply SNV beforehand, and cross-validation then adds an SG derivative in 87% of cases (L1115-1125). The encoder is set up comparably: it uses the patching and input normalisation that PatchTST (Nie et al., 2022) popularised for transformers on time series, which subtracts the mean of each spectrum and divides by its standard deviation (L139-140, L152-154). This coincides with SNV. It thus applies the standard pipeline, SNV followed by a derivative, and learns the derivative itself, as derivative-like behaviour of local mappings (A1 and A4, Tables 10 and 11). In this sense, it rediscovers chemometric preprocessing. We added these points to Section 2 and Appendix E.

Q6: What invariances and information losses do per-spectrum input normalization and per-patch target normalization introduce, respectively? In particular, how does the emphasis on local shape induced by target normalization relate to the observed polynomial and derivative-like filters? How do the authors distinguish representational preferences induced by the task design from preprocessing mechanisms learned from spectral data?

A6: Per-spectrum input normalisation: the encoder is invariant to the offset and scale of each spectrum, as with SNV. No embedding changes by more than 10−610^{-6} under either (L1083-1089). It loses only these 2 numbers per spectrum. Extended multiplicative scatter correction (EMSC) still moves the embedding by about half the distance between samples. The scatter it removes therefore stays in the embedding (Table 15).

Per-patch target normalisation: as in He et al. (2022), the mean and scale of each hidden patch are removed from its reconstruction target (L752-757). The loss is therefore invariant to the level and scale of each hidden patch, while the encoder still sees each visible patch in full. This emphasis on local shape could at most explain why the encoder ignores the level. Suppose it did. Then every model with the same normalisations would ignore the level. In our study (Tables 10 and 12), models with the same normalisations, trained on less data, still give the level 1.5 to 8%, against less than 1% for our encoder. Ignoring the level therefore cannot be explained by the target normalisation alone. For the slope, no assumption is needed: the target normalisation leaves it in the target, yet our encoder removes it, and responds mostly to curvature and cubic shape, 81 to 84% from the first block on (Table 10).

Task design or learned mechanism: to distinguish representational preferences induced by the task design from preprocessing mechanisms learned from spectral data, we pretrained the same model, with the same objective and both normalisations, on less data: LUCAS and food and feed spectra of the same size. With the same task design, these models respond mostly to slope and curvature, 87 to 95% from the first block on, and are far from untrained, with 3 to 8% on other directions against 92% (Table 12). The task design thus induces a preference for local shape, but not the derivative operators: the higher-order operators form with masked pretraining on the full food and feed corpus, as a preprocessing mechanism learned from spectral data. We added these comparisons to Section 6 and Appendices C and I (L949-957, Tables 12 and 23, L459-467).

Reviewer_K3muOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The paper trains MAE on food and feed NIR spectra, analyzes its representations against known chemometric operators, and evaluates frozen embeddings for predicting soil properties on LUCAS.

Critical strengths

Connecting learned representations to known spectral operators is useful, and the phase randomized controls, random encoders, and chemometric baselines strengthen the analysis.

Critical weaknesses

There are a few problems: the filter resemblance does not provide a sufficient explanation. No removal, replacement, or freezing experiment tests the contribution of weight shapes to transfer performance. I noticed that the two models have matching-filter proportions of 21% and 65%, yet similar downstream performance, and soil prediction stabilizes before the filters fully develop. In my opinion this only supports a descriptive observation more than a mechanism explaining transfer.

Also I found the comparison between masked and supervised learning fundamentally connfusing. The MAE is trained on repeated scans from approximately 1 million~ measurements whereas the supervised models learn from 20k~ soil samples. Domain, dataset size, training duration, and task all change simultaneously. The attribution of observed differences to training objective is questionable.

My biggest concern is that the only external dataset also participates in method selection, LUCAS is not an untouched validation set and pretraining variability is unmeasured(each mask ratio with one pretraining run). I cannot consider this result as generalizable or sound.

Questions for authors

I do not have any specific questions, if answered, could significantly change my ratings.

2 comments
AuthorsOpenReview ↗

Response to Reviewer K3mu (Q1)

We appreciate that K3mu finds connecting learned representations to known spectral operators useful, and that the phase-randomised controls, random encoders and chemometric baselines strengthen the analysis.

Q1: There are a few problems: the filter resemblance does not provide a sufficient explanation. No removal, replacement, or freezing experiment tests the contribution of weight shapes to transfer performance. I noticed that the two models have matching-filter proportions of 21% and 65%, yet similar downstream performance, and soil prediction stabilizes before the filters fully develop. In my opinion this only supports a descriptive observation more than a mechanism explaining transfer.

A1: To go beyond the resemblance of the filters, we probed the encoder with the perturbations that derivative filters are expected to remove. We added a constant level or a linear slope to one patch at a time, and for comparison a curvature or a cubic shape, which a derivative keeps, and measured how much the encoding of the patch changes. From the first block on, level and slope cause less than 6% of the change, and curvature and cubic shape 81 to 84%, over 6 seeds with 75% masking and closely with 90%. An untrained encoder gives each of the four about 2%, the chance share among the 48 directions we test (Table 10). The encoder thus removes what the filters are expected to remove. We added this test to Section 3 and Appendix C.

We also ran the replacement and freezing experiments, in a study of 32 pretrainings added to Section 3 and Appendix C. We replaced the learned first layer by fixed Savitzky-Golay filters on 5 seeds and by fixed random filters on 5 seeds, and pretrained the rest. With Savitzky-Golay filters, transfer stays at the level of the learned encoder on every seed, losing at most 0.013 in R2R^2 (median over all settings). With random filters, the outcome depends on the run: transfer drops on 3 of 5 seeds (Figure 4a). All 13 encoders that transfer, the 6 learned, the 5 Savitzky-Golay and the 2 remaining random ones, give level and slope at most 4% of their output change, and the 3 that drop give them 13 to 20% (Figure 10). The derivative operators thus give a reliable route to transfer, while a random first layer leaves the trainable layers to find one. Following Yosinski et al. (2014), we kept more layers at their random weights. Up to 8 random layers, transfer stays within 0.10 of the learned encoder. Beyond 8, it falls on both seeds, although the network still solves its pretraining task (Figure 4b). Solving the pretraining task therefore does not guarantee transfer.

The response also explains why the models with 21% and 65% matching filters transfer alike: they respond in the same way (Tables 8 and 10). It explains why soil prediction stabilises early as well: by epoch 50 of 800, the learned encoders already give level and slope at most 4% of their output change, as the encoders that transfer well do, while the filters keep sharpening (Figures 3 and 10). These results suggest that the derivative operators contribute to transfer: the filters resemble them, the encoder performs them, and with them transfer is reliable.

AuthorsOpenReview ↗

Response to Reviewer K3mu (Q2, Q3)

Q2: Also I found the comparison between masked and supervised learning fundamentally connfusing. The MAE is trained on repeated scans from approximately 1 million~ measurements whereas the supervised models learn from 20k~ soil samples. Domain, dataset size, training duration, and task all change simultaneously. The attribution of observed differences to training objective is questionable.

A2: We apologise for the confusion. To find what makes the encoder learn the derivative operators, we now change one factor at a time and test both the first-layer filters and, with the response test of Section 3, whether the encoder removes level and slope (Section 6, Appendices C and I, Tables 12 and 23).

Domain: we pretrained the masked autoencoder on the unlabelled LUCAS spectra and on food and feed spectra of the same size, with the same compute budget. Both domains behave alike: the first layer stays mostly noise, with 67 to 86% of the weight outside level, slope, curvature and cubic shape (Table 23), and the response goes mostly to slope and curvature, 87 to 95% from the first block on. They keep the slope instead of removing it.

Objective: on the same LUCAS spectra and with the same architecture, the supervised transformer learns derivative filters once its first layer gets a higher learning rate (Appendix I), with 20 to 21% matching a Savitzky-Golay template (Table 4), but gives level and slope 71 to 83% of its response: its filters resemble derivatives, yet it does not remove level and slope.

Size: only the masked autoencoder on the full food and feed corpus does both. 15 to 68% of its filters match, and it removes level and slope, with less than 6% of its response. The higher-order operators form with masked pretraining on the full food and feed corpus. Whether this comes from the size or the domain at that scale remains open, since no soil library approaches the size of the pretraining corpus.

Q3: My biggest concern is that the only external dataset also participates in method selection, LUCAS is not an untouched validation set and pretraining variability is unmeasured(each mask ratio with one pretraining run). I cannot consider this result as generalizable or sound.

A3: We added 2 untouched external soil libraries, KSSL and NIRsoil (Section 5, Appendix G, L349-350, L1175-1183). We also added ranges over 6 pretraining seeds with 75% masking and 2 with 90% (L187-190, Table 1, Figures 3 and 6), and a paired bootstrap and a randomisation test (Section 5, Table 3, L413-421). On both new libraries, the embedding stays within 0.06 in R2R^2 of the best tuned chemometric model for organic carbon with 50 to 400 labels, as on LUCAS, with a standard deviation over seeds below 0.035 (Figure 6). With all labels, it is not significantly behind the best chemometric model for organic carbon, in a paired bootstrap and a randomisation test, on KSSL, on NIRsoil (0.83 against 0.87) and on 4 of 5 LUCAS subsets (Tables 3 and 18).

Reviewer_CtcCOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper pretrains a masked autoencoder on 7.2M unlabelled NIR spectra of food and feed, compares its first-layer filters with Savitzky-Golay derivative templates against untrained and same-smoothness controls, probes soil organic carbon layer by layer, and evaluates frozen embeddings on the LUCAS soil library under domain and instrument shift.

Critical strengths
  • Using chemometric operators as a known reference for representation analysis is an original and checkable idea.
  • The controls (phase-randomised filters, untrained encoders, extensively tuned chemometric baselines) are careful, and limitations such as the 25-label regime and pH/carbonate results are reported honestly.
Critical weaknesses
  • The supervised control does not really support the claim about the training objective. The first-layer filters of the supervised transformer remain very close to their initialisation (median cosine similarity of 0.97). This suggests that the first layer barely changed during training, rather than showing that supervision cannot produce derivative filters. Also, the supervised models are trained on a different domain and with much less data than the Masked AutoEncoder. A fairer comparison would be to pretrain the same model on unlabeled LUCAS spectra and compare it against supervised training on the same data.

  • Limited practical benefit. When all labels are available, the untrained encoder already achieves an R² of 0.89, compared with 0.90 for the pretrained one. Also, after removing the contribution of 15 PLS components, the embedding residual provides very little additional predictive power (R² = 0.05–0.14). The gap is even more noticeable in the 25-label setting, which the paper describes as the starting point for new calibrations: the embedding achieves only 0.25, compared with 0.54 for Cubist.

  • Limited experimental scope. Pretraining is done using a single instrument family, and evaluation relies on just one external dataset, with a reconstructed rather than the original split from Stevens et al. (2013). This makes it difficult to know how well the results would generalise to other instruments and datasets.

  • Mismatch issue. The abstract claims that the model matches chemometric approaches for SOC prediction when all labels are available. However, it still falls behind on mineral soils (0.68 vs. 0.73), which is also the largest subset. There is also an inconsistency in the reported MAE band-shift values: Table 4 reports 1.1%, while Table 11 reports 0.67% on LUCAS.

Questions for authors
  • Can you pretrain the MAE on unlabelled LUCAS spectra and compare it with supervised training on the same data?

  • Why is absorbance used as the main setting when reflectance improves results in Appendix F?

2 comments
AuthorsOpenReview ↗

Response to Reviewer CtcC (Q1, Q2)

We are glad that CtcC finds chemometric operators as a known reference an original and checkable idea, and the controls careful. The revision adds the experiments CtcC suggested.

Q1: The supervised control does not really support the claim about the training objective. The first-layer filters of the supervised transformer remain very close to their initialisation (median cosine similarity of 0.97). This suggests that the first layer barely changed during training, rather than showing that supervision cannot produce derivative filters. Also, the supervised models are trained on a different domain and with much less data than the Masked AutoEncoder. A fairer comparison would be to pretrain the same model on unlabeled LUCAS spectra and compare it against supervised training on the same data.

A1: We raised the learning rate of the first layer, and the supervised transformer now learns derivative filters (Figure 8, Table 4). This agrees with earlier work on supervised CNNs for spectra (Cui and Fearn, 2018, Zhang et al., 2020). We removed the opposite claim from the abstract and updated Section 6. As suggested, we also pretrained the same model on the unlabelled LUCAS spectra, and on food and feed spectra of the same size with the same compute budget. Domain, dataset size and task then no longer change together. On this data, the masked autoencoder's first layer stays noisy, while the supervised transformer learns lower-order derivative filters. The higher-order operators form with masked pretraining on the full food and feed corpus (Table 23). Whether more soil data would change this remains open, since no soil library approaches the size of the pretraining corpus.

Q2: Limited practical benefit. When all labels are available, the untrained encoder already achieves an R² of 0.89, compared with 0.90 for the pretrained one. Also, after removing the contribution of 15 PLS components, the embedding residual provides very little additional predictive power (R² = 0.05–0.14). The gap is even more noticeable in the 25-label setting, which the paper describes as the starting point for new calibrations: the embedding achieves only 0.25, compared with 0.54 for Cubist.

A2: We appreciate CtcC's attention to the practical value. That an untrained encoder is close on LUCAS organic carbon surprised us too. Saxe et al. (2011) attribute it to the architecture. Across libraries and properties, the pretrained encoder is ahead of the untrained one with all labels, 0.80 against 0.76 on average and best or tied in 9 of 11 settings (Table 3), and with 50 to 400 labels in every setting (Table 22). To test what pretraining adds, we combined the embedding with a chemometric model, with both the model and the weight chosen by cross-validation on the calibration data. With all labels, the combination beats the best chemometric pipeline by a small but significant margin, in a paired bootstrap and a randomisation test (Table 3, L413-421). With an untrained encoder instead, cross-validation gives it less weight, and its combination rises less above the best chemometric pipeline. The residual R2R^2 of 0.05 to 0.15 (Table 19) therefore reflects information that the chemometric models do not use. The 25-label setting remains hard and we report it (L446-447, Table 20). From 50 labels on, one linear probe on the frozen embedding stays close to the best tuned pipeline for organic carbon on every library and beats PLS with cross-validated preprocessing, the usual reference in NIR calibration (Wold et al., 2001) (Figure 6). We added these tests to Section 5 and Appendix G.

Beyond these results, we also see value in how a calibration is built. Today each calibration belongs to one instrument, material and property, and starts from scratch with its own search over preprocessings and models (L043-047, Rinnan et al., 2009). This holds for PLS, Cubist, memory-based learning (MBL) and 1D CNNs alike, and MBL also needs the whole training library at prediction time (Ramirez-Lopez et al., 2013). Our representation is learned once, without labels, from spectra of a low-cost handheld instrument family, which are cheap to collect at scale, and each new task needs only one linear model on it. Without any search over preprocessings or models, it already matches the best tuned pipelines in several settings, on a new material measured with higher-resolution benchtop instruments (Table 3). A calibration then no longer depends on an exhaustive search to reach the best performance. We hope this finding encourages the community to explore this direction, with more instruments, materials, prediction heads and tasks.

AuthorsOpenReview ↗

Response to Reviewer CtcC (Q3, Q4, Q5, Q6)

Q3: Limited experimental scope. Pretraining is done using a single instrument family, and evaluation relies on just one external dataset, with a reconstructed rather than the original split from Stevens et al. (2013). This makes it difficult to know how well the results would generalise to other instruments and datasets.

A3: We extended the experimental scope to 3 external libraries, adding KSSL and NIRsoil to LUCAS, and to 6 pretraining seeds. We added these results to Section 5 and Appendix G (L118-120, L198-199, L349-350, L1175-1183). Each library was measured on a different benchtop instrument, with a wider range and a higher resolution than the single instrument family used for pretraining. The encoder reads every instrument on one fixed wavenumber grid, with each patch encoded by its absolute wavenumber. On all 3 libraries, the embedding stays within 0.06 in R2R^2 of the best chemometric model for organic carbon with 50 to 400 labels, and with all labels, combined with a chemometric model, it gives the best organic-carbon model (Figure 6, Table 3). This transfer from 1 instrument to 3 is hard in NIR calibration (Workman, 2018). As future work, we plan to pretrain on multiple instruments (L522-524). The split of Stevens et al. (2013) is not published. We reproduced it as closely as we could, and every model is compared on this same split (L404).

Q4: Mismatch issue. The abstract claims that the model matches chemometric approaches for SOC prediction when all labels are available. However, it still falls behind on mineral soils (0.68 vs. 0.73), which is also the largest subset. There is also an inconsistency in the reported MAE band-shift values: Table 4 reports 1.1%, while Table 11 reports 0.67% on LUCAS.

A4: We revised the abstract, and the Introduction and Conclusion accordingly. It now says that the encoder matches the chemometric models on most LUCAS subsets and on KSSL with all labels. Section 5 reports the mineral soils, now over 6 seeds, at 0.69 against 0.73 (L408-410), and combined with a chemometric model, the embedding recovers the gap, at 0.75 (Table 18). Table 4 compares all models at block 12, where the masked autoencoder moves by 0.7 to 1.3% over 6 seeds. Table 15, formerly Table 11, reports the output of the encoder, the embedding used for every prediction. We state the layer of both band-shift values in the revision.

Q5: Can you pretrain the MAE on unlabelled LUCAS spectra and compare it with supervised training on the same data?

A5: Yes, thank you for the suggestion. Beyond the filters reported in A1, we applied the response test of Section 3 to these models. Neither removes level and slope: on the same LUCAS spectra, the masked autoencoder gives 87 to 95% of its response to slope and curvature, and the supervised transformer gives level and slope 71 to 83% (Table 12, Appendix I). Food and feed spectra of the same size behave as LUCAS, and only the masked autoencoder on the full food and feed corpus removes level and slope, with less than 6% of its response (Table 10). The higher-order operators therefore form with masked pretraining on the full corpus, not from the objective alone.

Q6: Why is absorbance used as the main setting when reflectance improves results in Appendix F?

A6: We use absorbance for comparability. LUCAS and NIRsoil are distributed in absorbance and KSSL is converted to it from reflectance (L196-197, Appendix G), and Stevens et al. (2013), whose design we follow, model the LUCAS spectra in absorbance. Absorbance and reflectance are equivalent through A=log⁡10(1/R)A = \log_{10}(1/R). To check the choice, we repeated the few-label evaluation on LUCAS in reflectance: ridge on the embedding gives the same R2R^2 within 0.01 with 50 to 400 labels. We added this check to Appendix F (L1148-1150).

Comments on the paper

1 comment
AuthorsOpenReview ↗

Summary of Responses and Revisions

We thank the reviewers for their time and their careful, constructive reviews, which helped us improve the paper substantially. Their questions shaped the updates below, which take our analysis of the learned representation from descriptive to functional and mechanistic, and test it across more seeds, soil libraries and training data.

We appreciate the following positive feedback:

  1. Using chemometric operators as a known reference is an original and checkable idea, and the controls are careful.
  2. Connecting learned representations to known spectral operators is useful, and the controls strengthen the analysis.
  3. Practical relevance and valuable transfer evaluation with limited labels.
  4. Transfer from food and feed to soil measured with a different instrument provides meaningful evidence under both material and instrument shifts, extending beyond evaluation within the pretraining domain.
  5. A concrete, testable approach to representation analysis, with Savitzky-Golay filters as a mathematically defined reference.

Key updates to the paper:

  1. We tested the response of every layer and of the filters per derivative order, to distinguish structural similarity from functional equivalence (Section 3, Appendix C, Tables 10 and 11): from the first block on, the encoder removes level and slope, while an untrained encoder responds at chance.
  2. We ran targeted interventions in 32 pretrainings with frozen Savitzky-Golay or random layers (Section 3, Appendix C, Figures 4 and 10): frozen Savitzky-Golay filters keep transfer on every seed, frozen random filters lose it on 3 of 5, and every encoder that transfers removes level and slope.
  3. We checked the variability over 52 trained models across seeds, with a paired bootstrap and a randomisation test (Tables 1, 3 and 9, Figures 3 and 6): the filters exceed chance on every seed, and organic carbon R2R^2 varies with a standard deviation below 0.035.
  4. We evaluated the encoder on 2 external soil libraries, KSSL and NIRsoil, measured on other instruments (Section 5, Appendix G, Table 3): on all 3 libraries, the embedding stays within 0.06 in R2R^2 of the best chemometric model for organic carbon with 50 to 400 labels.
  5. We compared the training objectives on the same data, pretraining on unlabelled LUCAS and on food and feed of the same size (Section 6, Appendices C and I, Tables 12 and 23): the higher-order operators form with masked pretraining on the full food and feed corpus.
  6. We combined the embedding with chemometric models to test for complementary information (Section 5, Table 3): with all labels, the combination beats the best chemometric pipeline by a small but significant margin.
  7. We extended the supervised comparison (Section 6, Appendix I, Figure 8, Table 4): with a higher first-layer learning rate, the supervised transformer also learns derivative filters, but does not remove level and slope.

We hope that our replies and revisions address all reviewers' concerns, and we would appreciate any further comments or suggestions.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.