Soils probably hold more plastic than the oceans, yet we can barely measure it. The best available method, pyrolysis-gas chromatography-mass spectrometry, burns a chemically digested soil sample and reads polymer fragments. It is destructive and slow, and it is still confused by natural soil organic matter for exactly the polymers farmers spread most, polyethylene mulch film among them. Cheap handheld near-infrared (NIR) spectrometers could scan fields in seconds. The difficulty is that NIR light only sees blurred "echoes" of molecular vibrations: overtones (roughly double or triple the base vibration frequency) and combination bands (sums of two base frequencies). The sharp, diagnostic base vibrations live in the mid-infrared (MIR), which needs laboratory instruments.
This note proposes a spectral model whose internal wiring encodes that physics. Each NIR channel is connected only to the MIR channels that can physically produce it, at overtone and combination positions, with a small learnable offset for the fact that real overtones fall slightly short of exact integer multiples. The model is trained to translate field NIR into lab-quality MIR and to estimate polymer content. The bet is that physically placed sparse connections let a model learn from the few hundred soil samples that can realistically be spiked with plastic, and still generalize to soils it has never seen. Dense or purely local models need more data than this setting will ever provide.
Core Claim
Audio research has shown that treating each frequency bin as independent fails on structured signals, and that the cure is to model dependencies across bins. Three designs do this in different ways: grouped "banded" filters for neighbouring leakage, learned global frequency-transformation matrices, and attention across frequency. Optimal transport with a harmonic cost matrix is a hand-designed version of the same idea. Soil spectra are a harsher setting than audio. They have more than 1,700 channels, the labelled sets are tiny (the inter-laboratory test set has 206 soils), and the cross-frequency relation is known in advance from vibrational physics but is anharmonic rather than strictly integer. The claim is that a coupling matrix restricted to physically allowed overtone and combination positions, each widened into a narrow band, fits this setting better than any of the audio designs alone. Concretely, it should give lower detection limits for polyethylene, polypropylene, polyethylene terephthalate and polystyrene in soils not seen during training, and it should reduce the systematic NIR-to-MIR reconstruction error that current models show in the 3750–3000 cm⁻¹ region.
What Exists and Where This Sits
The closest work is Self-supervised and Multi-fidelity Learning for Extended Predictive Soil Spectroscopy. It pretrains a variational autoencoder on more than 334,000 MIR scans from the USDA KSSL library and maps NIR spectra into the same 32-dimensional latent space. On the proficiency-test benchmark it raises NIR inorganic-carbon concordance from 0.77 to 0.85. It supplies our pretraining data and our strongest baseline, and it motivates the proposal because the authors report persistent offsets in the hydroxyl region. It does not model polymers. Its link from NIR to MIR runs through a generic bottleneck, with no explicit cross-frequency structure. We keep its MIR pretraining and replace the bottleneck route with an overtone-wired translator.
Three audio papers supply the mechanism, and they also constrain it.
- Meta-Learning for Adaptive Filters with Higher-Order Frequency Dependencies shows that overlapping banded groups beat both independent bins and non-overlapping blocks. Its Banded-3 model is 3.21 dB better in signal-to-distortion ratio than a Kalman filter, at a tenth of the computation. We reuse banding to absorb local shifts.
- CFTNet learns a full frequency-transformation matrix. That works at 257 bins but would mean millions of free parameters at 1,700 channels with a few hundred labels, so we treat it as a competitor, not a component.
- A Convolutional Neural Network with Non-Local Module for Speech Enhancement finds that global attention helps, but performance drops once more than two non-local blocks are stacked. That is a warning that unconstrained global coupling overfits.
Optimal spectral transportation is the conceptual ancestor: a cost matrix that makes moving energy between harmonics cheap. It improved piano-transcription F-measure from 0.624 to 0.684. Its integer-harmonic template, however, is exactly what fails for anharmonic molecular overtones. We learn the offset instead of fixing it.
On the soil side, Quantification of microplastic targets in environmental matrices using pyrolysis-gas chromatography-mass spectrometry provides the reference measurement and documents matrix interference for polyethylene, polystyrene and polyvinyl chloride. Global Soil Salinity Prediction by Open Soil Vis-NIR Spectral Library supplies the tabular chemometric baselines. The ecological stakes are set by work showing that plastics carry agrochemicals into organisms. In Enhanced bioaccumulation and toxicity of Fenpropathrin by polystyrene nano(micro)plastics in the model insect, silkworm (Bombyx mori), nanoplastic co-exposure raised fat-body pesticide levels 43-fold. Other studies show belowground fauna respond sensitively to habitat change (Euedaphic Rather than Hemiedaphic or Epedaphic Collembola…), yet these organisms are rarely monitored. Mapping plastic load cheaply is the precondition for studying those effects at field scale.
No paper in the file applies inter-frequency dependency models to microplastic detection. The novelty is not "deep learning on soil spectra", which the multi-fidelity paper already does. It is the physics-placed sparse coupling, and its test on the hardest version of the problem: trace polymers inside organic-rich soils.
The Model
The input is an NIR spectrum $x \in \mathbb{R}^{F_N}$; the reconstruction target is an MIR spectrum $y \in \mathbb{R}^{F_M}$, with $F_M$ above 1,700. Each MIR channel at wavenumber $\nu_m$ is predicted from NIR channels lying in a set of allowed positions:
$$\mathcal{A}(m) = \{\, j : |w_j - (n\,\nu_m - \chi_n(\nu_m))| \le \delta \ \text{for } n \in \{2,3\} \,\} \ \cup\ \{\, j : |w_j - (\nu_m + \nu_k - \chi_c)| \le \delta,\ k \in \mathcal{K}(m) \,\}$$
The symbols are defined as follows:
- $w_j$ is the wavenumber of NIR channel $j$.
- $\chi_n(\nu)$ is a small learnable anharmonic shift, parameterised as a smooth function of $\nu$ with a handful of spline coefficients.
- $\delta$ is the band half-width, which plays the role of the overlapping bands in Wu et al.
- $\mathcal{K}(m)$ is a short list of partner fundamentals for combination bands, chosen from known bending and stretching regions. Examples are the C–H stretch near 2900 cm⁻¹ paired with C–H bends, which yields the combination band near 4300 cm⁻¹.
Gated attention runs only over $\mathcal{A}(m)$. Its output feeds the frozen MIR decoder of the multi-fidelity paper's pretrained autoencoder through a residual path, so the model inherits that decoder's soil prior and adds explicit cross-frequency structure on top of it. A polymer head reads the reconstructed fundamental regions, where polymer bands are sharp, and outputs a mass fraction for each polymer.
The connection count grows roughly as $F_M$ times a few times $\delta$. That is linear in channel count, not quadratic, which answers the scaling problem that dense attention runs into. The learned shift curves $\chi_n$ are themselves checkable: they should recover the known anharmonicity of C–H and O–H stretches, which gives an interpretability test that black-box baselines cannot offer.
Research Program
The work has three stages.
Stage one stays on existing benchmarks. We train NIR-to-MIR translation on the 2,106 paired NIR–MIR soils, starting from the KSSL MIR pretraining. We then evaluate soil-property prediction on the 206-sample North American Proficiency Testing set, using the same nine properties and Lin's concordance correlation coefficient as the multi-fidelity paper. We also measure reconstruction error per spectral region, with particular attention to 3750–3000 cm⁻¹. This stage asks whether physical wiring helps at all before any plastic is involved.
Stage two builds the missing benchmark, since no labelled spectral data on soil microplastics exists in the file. We select archived KSSL-type soils spanning texture and organic-carbon content. Into these we spike pristine and UV-aged fragments of polyethylene, polypropylene, polyethylene terephthalate and polystyrene, singly and in mixtures, at a ladder of mass fractions down to environmentally relevant levels. Every specimen is scanned on both a portable NIR instrument and a lab MIR instrument. The weighed spike mass is the primary ground truth. This choice is deliberate: Py-GC-MS itself suffers matrix interference for polyethylene and polystyrene, so it cannot serve as the reference for those polymers. A smaller set of field soils from mulched fields is measured by Py-GC-MS following the cited protocol as an external check.
Stage three analyses where gains come from. We inspect the learned shift curves and the attention weights for each polymer, and we test whether the model separates the polymer's C–H overtone from that of soil organic matter. That separation is the specific confusion the chemistry literature reports.
Baselines
The naive reference has two parts. For polymer content it predicts the training-set mean, and for reconstruction it outputs the training-set mean MIR spectrum. Any model that fails to beat these on soils it has not seen has learned nothing transferable.
The chemometric standard is partial least squares regression (PLSR) on preprocessed NIR, with the number of components chosen by grouped cross-validation. We also include XGBoost on the top 30 principal components, following the salinity study (Global Soil Salinity Prediction by Open Soil Vis-NIR Spectral Library). That study reports R² 0.59 and RMSE 1.96 dS m⁻¹ for this baseline on electrical conductivity, and R² 0.39 for full-spectrum PLSR. These numbers come from a different target, so they calibrate expectations rather than set the bar. PLSR is the method practitioners would actually use, which makes it the bar a chemist cares about.
The strongest existing method is the multi-fidelity approach of Self-supervised and Multi-fidelity Learning for Extended Predictive Soil Spectroscopy. Its NIR encoder maps into the frozen 32-dimensional MIR latent, and a shallow regressor is fine-tuned on the spiked data. Its reported numbers are an average concordance of 0.91 from MIR, and an improvement in NIR inorganic-carbon concordance from 0.77 to 0.85 on the proficiency set. For microplastic quantification from soil NIR, no reported result was found on Lacuna. Our first task is therefore to reproduce the 0.85 inorganic-carbon figure before comparing anything.
Two learned-coupling competitors complete the set: a CFTNet-style dense frequency-transformation block (CFTNet) and a non-local attention block across frequency (A Convolutional Neural Network with Non-Local Module for Speech Enhancement). Both are placed at the same position in the network as our sparse layer.
Three ablations isolate the key ingredient:
- Random sparse wiring keeps exactly the same number of connections per MIR channel but places them at random wavenumbers.
- Band-only wiring keeps local neighbours but drops the overtone and combination edges.
- Integer-only wiring fixes $\chi_n = 0$, which corresponds to the transport-paper assumption.
Fairness rules apply to every comparison. All neural models share the KSSL pretraining, the same paired and spiked training splits, and the same hyperparameter search budget, counted in trials. Tabular baselines get an equal search budget.
Beating a baseline means the following. On leave-one-soil-group-out splits, our model must achieve lower root-mean-square error in polymer mass fraction and a lower estimated detection limit (blank mean plus three blank standard deviations, propagated through the calibration) than the multi-fidelity baseline. The gap must hold across the resampled splits, not just one. The model must also beat the random-sparse ablation, which shows that the physical placement matters and not merely the sparsity.
Why not just…?
Why not just use PLSR on the NIR spectrum? It is linear, data-efficient and trusted, and for a single strong polymer band in a single soil type it may well match us. That is a real threat. It is weakest in the case that matters, where a polyethylene C–H overtone sits on top of a soil-organic-matter C–H overtone whose shape changes from soil to soil. The salinity study shows PLSR trailing nonlinear models even on bulk properties. The leave-one-soil-group-out split decides the question. If PLSR ties us on unseen soil groups, the physics wiring is not earning its complexity, and we would report that outcome.
Why not just fine-tune the multi-fidelity autoencoder on spiked samples? This is the most plausible alternative. A 32-dimensional latent learned from bulk soils may simply have no room for a trace polymer signal: the authors note that compression filters out "noise and redundancy", and a 0.1% plastic signal looks like exactly that. The region-wise reconstruction error and the recovery of spiked polymers will show whether the latent keeps polymer information. We also run a variant that widens the latent, to separate a capacity problem from a structural one.
Why not just let dense attention or a learned frequency-transformation matrix discover the couplings? With a few hundred labelled spiked soils and more than 1,700 channels, a dense coupling has far more freedom than the data constrain. The non-local paper already saw performance fall when global blocks were stacked. If dense coupling matches ours at full data but loses at a quarter of the data, the physics prior is buying sample efficiency. If it matches at every data size, the prior is unnecessary. The learning-curve experiment distinguishes the two cases.
Why not just measure MIR in the lab, or run Py-GC-MS? For accuracy on a few samples, that is the right choice, and lab MIR serves as our upper-bound model. The proposal targets the regime where neither scales: field-wide mapping with cheap instruments, and soils where Py-GC-MS is itself compromised by organic interference for polyethylene and polystyrene. If NIR detection limits stay far above the field concentrations measured by Py-GC-MS, the conclusion is that portable NIR is unsuited to this task. That would be a useful negative result in its own right.
Decisive Checks
Four results would settle the question:
- Learning curves. Polymer RMSE plotted against the number of spiked training soils, for our model, the multi-fidelity baseline, dense coupling and PLSR, on soil groups held out from training. The claim predicts the largest advantage at small sample sizes.
- Wiring ablation. Overtone wiring compared against random sparse wiring with equal connection count. Losing this comparison would falsify the central mechanism.
- Reconstruction error. Error in the 3750–3000 cm⁻¹ region compared with the multi-fidelity model on the 2,106 paired soils. This tests the hydroxyl-region offset the original authors could not explain.
- Weathering transfer. Train on pristine fragments, test on UV-aged ones. Aging shifts band positions slightly, which the learnable shift and band width are designed to absorb, and which integer-only wiring should not.
If all four go our way, the outcome is a measurement tool that makes field-scale studies of the belowground plastic problem feasible.
References
- Self-supervised and Multi-fidelity Learning for Extended Predictive Soil Spectroscopy
- Meta-Learning for Adaptive Filters with higher-order Frequency Dependencies
- CFTNet: Complex-valued Frequency Transformation Network for Speech Enhancement
- A Convolutional Neural Network with Non-Local Module for Speech Enhancement
- Optimal spectral transportation with application to music transcription
- Quantification of microplastic targets in environmental matrices using pyrolysis-gas chromatography-mass spectrometry
- Global Soil Salinity Prediction by Open Soil Vis-NIR Spectral Library
- Enhanced bioaccumulation and toxicity of Fenpropathrin by polystyrene nano(micro)plastics in the model insect, silkworm (Bombyx mori)
- Euedaphic Rather than Hemiedaphic or Epedaphic Collembola Are More Sensitive to Different Climate Conditions in the Black Soil Region of Northeast China