acceptodds
Under review as a conference paper at ICLR 2027

One or More? Physics-Guided Super-Resolution Decomposition of Underwater Multispectral LiDAR Echoes

Abstract

Underwater multispectral LiDAR jointly captures range and spectral information, providing multidimensional observations for detailed underwater sensing. However, echoes from closely spaced targets can overlap within a single waveform, while water-induced temporal broadening further reduces peak separability. These effects make accurate target counting and physical spacing estimation challenging. To study this problem under realistic conditions, we develop an underwater multispectral LiDAR system and use it to acquire two real-world full-waveform datasets: a single-target reference library spanning target ranges, reflectances, and laser intensities, and a multiband multi-target library covering different degrees of echo overlap with physical spacing annotations. We propose Aquatic Super-Resolution (AquaSR), a physics-guided framework for underwater echo super-resolution and decomposition. AquaSR incorporates a backbone-agnostic prior-constrained joint multiband decomposition (PC-JMD) module. Its central principle is to share target count and physical spacing across bands while retaining band-specific echo shapes, amplitudes, first-target delays, and backgrounds. A frozen decoder pretrained on the single-target library provides a continuous single-echo shape prior that constrains component reconstruction. The module determines target count by comparing reliability-weighted reconstruction errors under single- and dual-target hypotheses and estimates shared spacing by fusing cross-band geometric evidence. Physical-attribute experts then refine the spacing using echo morphology and energy features, followed by component reconstruction. Experiments on measured data show that AquaSR outperforms representative classical waveform decomposition methods and deep learning baselines and recovers target range structure at sub-pulse-width scales. Experiments with multiple backbone configurations further demonstrate the cross-architecture compatibility of PC-JMD. These results support precise ranging and structural analysis of closely spaced underwater targets.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
Underwater detection, Multispectral LiDAR, Full-waveform decomposition

Reviews

Among:
Shape:
1234this paper, Top 32%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.50 over 2 reviews, above 68% of the others.
Reviewer_ZWRqOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper studies target counting and physical spacing estimation for underwater multispectral LiDAR when closely spaced targets produce overlapping, effectively single-peaked echoes. Water-induced temporal broadening further hides the second peak, so counting visible peaks does not determine the number of targets.

The authors build an underwater multispectral LiDAR system at four wavelengths and collect two measured full-waveform datasets in a 60 m water tank. One is a single-target reference library. The other is a multiband multi-target library with physical spacing labels.

They propose AquaSR. Its core PC-JMD module shares target count and physical spacing across bands while keeping echo shape, amplitude, first-target delay, and background band-specific. A frozen MMD-WAE decoder pretrained on single-target echoes supplies a continuous single-echo shape prior. Target count is chosen by comparing reliability-weighted reconstruction residuals under one- and two-target hypotheses, and six physical-attribute experts then refine the spacing.

Validation on a range-disjoint test set compares AquaSR with five classical decomposition methods and one learning baseline (AFD-Net), together with multi-seed module ablations, six backbones, and runtime analysis. AquaSR reports single-target and dual-target counting with the lowest conditional spacing MAE (0.48 cm) and bias (0.25 cm).

Critical strengths
  1. A real measurement system with a physically grounded and internally consistent method. The authors build a four-wavelength underwater multispectral LiDAR and collect two real full-waveform datasets. The PC-JMD factorization, which shares target count and physical spacing while keeping shapes band-specific, is supported by a curvature analysis and an independent frozen-template BIC study.

  2. Counting from competing generative explanations rather than a standalone classifier, with cross-architecture evidence. Deciding the target count from one- versus two-target reconstruction residuals under a frozen single-echo prior is more principled than a feature classifier. The module ablations and six backbones substantiate both the design choices and the claimed backbone agnosticism.

Critical weaknesses

1.Limited external validity. All results use static targets in one tank under a single water condition on a small held-out set whose single/dual counts are not reported. The headline challenge of water-induced broadening is never varied along the turbidity or broadening axis, even though a new water condition would push the frozen single-echo shape prior out of distribution, and that prior directly drives the one-versus-two decision.

2.An application-driven integration of established techniques with a new dataset, whose two sources of value are not yet secured. The cross-band geometric-consistency idea is already the basis of MSWD and Rclonte, and the WAE prior, competing-hypothesis model selection, and physics-guided experts are also established. The novelty is their non-trivial combination for underwater multispectral LiDAR rather than a new mechanism.

Questions for authors

1.Robustness to water-induced broadening. What are the optical properties of the tank water (attenuation/scattering coefficients), and are training and test data taken from the same water body? Because the frozen single-echo prior drives the one-versus-two decision, can you report counting and spacing results over two or three turbidity levels, or a synthetic sweep of the broadening-kernel width, separating zero-shot transfer from recalibration?

2.Isolating the cross-band sharing gain. The ablations only remove components within AquaSR. Can you add a controlled comparison that aligns hard sharing of count and spacing with the soft cross-channel consistency of MSWD and the cross-wavelength matching of Rclonte under the same data and backbone, to isolate the gain attributable to the proposed factorization?

3.Test-set composition. Please state the single/dual event counts and distance distribution of the main test set in Table 1 (the percentages imply roughly 44 single and 57 dual events) and list the ranges used for training, validation, and test.

1 comment

Thank you for your valuable time and constructive feedback. Due to time constraints, we have temporarily placed the new experiments and analyses in the Appendix (highlighted in blue, starting from page 23), and will integrate the key results into the main text in the final revision. We hope our responses address your concerns, and we appreciate your understanding.

Q1 Water conditions and zero-shot transfer

The tank contained tap water, with a recorded attenuation coefficient of 0.1 m⁻¹ at 500 nm; the scattering coefficient was not measured separately. Training and testing largely used this water, although different ranges were sometimes acquired on different days (Appendix F.1).

Archived single-target data acquired after three days without water replacement provide a transfer check. The water appeared more turbid, but turbidity was not quantified. At 532 nm, pulse width narrowed from 2.312 to 2.173 ns (about 6%). Without retraining or recalibration, AquaSR correctly identified the cases as single-target; dual-target spacing was not evaluated under this condition (Appendix F.2–F.3).

Spatial variation along the long water path prevented a reproducible quantitative turbidity sweep. The apparatus has been dismantled for miniaturization; marine measurements at different depths are planned for early next year (Appendix F.4).

Q2 Controlled cross-band comparison

We compared independent decomposition (I), soft consistency (S), MSWD-style (M), Rclonte-style (R), and native AquaSR using identical splits, backbone, and frozen single-echo decoder. AquaSR includes reliability fusion, consensus, and correction.

S penalizes count-probability (Jensen–Shannon) and spacing (Smooth-L1) disagreement, with validation-tuned weights and a spacing penalty during refinement. M retains energy ranking, completion, and shared-center fitting (zero offsets for aligned inputs). R retains center ranking, correspondence screening, completion, and per-band fitting with adjustable spacings.

M and R use a shared decoder in place of Gaussian components, matched initialization, one-/two-target hypotheses, residual count selection, and fitting budgets, while retaining distinct cross-band mechanisms. Event-level MAE counts each synchronized multiband event once; the main comparison aggregates valid bands. Across three seeds, AquaSR has the lowest mean MAE (0.505 ± 0.026 cm) and smallest standard deviation. All five methods reach 100% counting accuracy, isolating spacing differences (Appendix G.2; Table G.2).

Q3 Independent events and held-out spacings

The 10 m test set comprises 26 independent events (11 single-target, 15 dual-target). The 44 and 57 entries in band-weighted metrics are valid single- and dual-target band waveforms (101 total), not independent events. Each measured split contains one dual-target event at each of 15 nominal spacings (Appendix H.1; Table H.1).

Training, validation, and testing use 8, 12, and 10 m data, respectively. For each configuration, 1,000 technical repetitions are screened and averaged; excluded channels are listed in Table H.2 (rationale in Appendix H.2). Training and validation also use superpositions of measured single-target echoes. The test set is measured-only, and no 10 m data enter prior learning, synthetic-data construction, or fine-tuning.

In ten leave-one-measured-spacing-out evaluations, each selected spacing's measured training/validation examples were removed while its 10 m test event was retained. Absolute errors improved at 13, 23, 28, and 35 cm (0.21→0.16, 0.85→0.40, 1.48→1.02, and 0.29→0.03 cm). Nine of ten full-test MAEs were below 1 cm; their mean was 0.681 cm versus the 0.48 cm reference, with the 28 cm exclusion improving to 0.46 cm. Other folds worsened (Appendix I.1–I.3; Table I.1). Appendix J.3 sets out our plan to release the dataset and core code upon acceptance and extend the dataset through future marine trials.

Others: Clarification on Critical Weaknesses 2

We thank the reviewer for recognizing the application-driven contribution and new dataset. Aligned with our Primary Area (Applications to Physical Sciences). The methodological contribution is a coupled, physics-guided inference formulation for multiband underwater LiDAR: shared physical-coordinate consensus before wavelength-specific synthesis, a frozen measured single-echo prior coupled to reconstruction-based count selection, and count-preserving geometric correction. Appendix G.4 explains these mechanisms relative to MSWD/Rclonte, with controlled comparisons (G.2–G.3) and the main-text ablations (Table 2).

Reviewer_b9dLOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

The problem that the author addresses is whether, when two underwater objects are close together, their reflected laser pulses can merge into one apparent echo. Can we still determine that there are two objects and estimate the distance between them? Objects can have the same physical separation at every wavelength, while their echo shapes and strengths can differ.

Its central assumption is that the objects have the same physical separation at every wavelength, while their echo shapes and strengths can differ.

They propose AquaSuperresolution (AquaSR), a physics-guided method for underwater echo-guided superresolution and decomposition. The method is:

  1. Learns single-object echo shapes from measured data and freezes that learned decoder.
  2. Combines evidence across wavelengths to estimate shared target spacing.
  3. Constructs one-target and two-target explanations and compares their reconstruction errors to choose the count.
  4. Refines the spacing using six branches that examine slopes, curvature, intensity, and energy distribution.

Reports 100% on target counting (table 1)

Critical strengths

Unique method that presumably has application areas well beyond the tested setup.

Innovative use of sharing geometry while allowing different waveform shapes across wavelengths .

The authors do more than report neural-network accuracy. In Appendix A, they use a frozen measured template to show that single-target and overlapping dual-target observations can produce distinguishable fitting scores.

The authors build a complete Acquisition platform and collect both single-target reference echoes and overlapping multi-target observation.

Critical weaknesses

The main evaluation holds out one distance, 10 m, using one apparatus and static manufactured targets in a dark water tank. Is this universal enough?

That leaves important questions unanswered:

  • Does the method transfer to different water conditions?
  • Does it handle unseen materials ?

Table 2: MEA goes from 0.55 cm for no correction to 0.50 cm for full model. The The additional branches provide a modest 0.05 cm average improvement.

In the supplied text, several important details remain insufficient: • Exact training, validation, and test dataset sizes and acquisition grouping. • Learning rates, batch sizes, training durations, loss function, …

There is no code and dataset available . Figures are very small

Questions for authors
  1. Can you demonstrate the improvement under a matched training protocol?
How does AquaSR compare with AFD-Net adapted to the same measured training data and setup, were competing methods given the same one-/two-target constraint?

  2. What exactly are the independent training, validation, and test units?


  3. Does the method generalize to genuinely unseen configurations?

1 comment

Thank you for your valuable time and constructive feedback. Due to time constraints, all new experiments and analyses are temporarily placed in the Appendix and will be selected and integrated into the main text in the camera-ready revision. Each response below points to the corresponding revisions, which are highlighted in blue starting from page 23 in the Appendix. We hope our responses address your concerns and appreciate your understanding.

Q1 Matched AFD-Net comparison

We adapted AFD-Net on the same measured training and validation data as AquaSR and evaluated it on the same test set, in both its original four-class form and a binary one-/two-target version matching AquaSR's count constraint. The original manuscript baseline followed AFD-Net's published simulation/training procedure; the new comparison isolates the effect of matched measured-data adaptation (Appendix G.1; Table G.1).

Four-class and binary AFD-Net achieve 85.15% and 89.11% overall counting accuracy, with valid dual-target spacing outputs of 44/57 and 48/57, respectively. Their conditional spacing MAEs are 3.172 and 3.324 cm. AquaSR achieves 100% counting accuracy, 57/57 valid outputs, 0.478 cm MAE, and 0.683 cm RMSE. Conditional errors must be interpreted alongside output coverage.

Q2 Independent units and consistency screening

One independent measured unit is a synchronized multiband acquisition event at a fixed range and target configuration; its bands and repeated acquisitions are correlated observations. The 8/12/10 m train/validation/test split has one measured dual-target event at each of 15 nominal spacings per range, with four intended bands and 1,000 technical repetitions per configuration. Training and validation also use superposed measured single-target echoes; testing is measured-only. No 10 m records enter prior learning, synthetic-data construction, or fine-tuning.

At 10 m, 11 single-target and 15 dual-target events yield 44 and 57 valid band waveforms, respectively; each spacing contributes one independent dual-target event (Appendix H.1; Table H.1).

The input waveforms are means after consistency screening. Appendix H.2, Table H.2, and Figures H.1–H.4 document retained/excluded channels and the observed acquisition corruption, transient extra returns, visible suspended particles, and normal records that informed screening.

We also removed anomalous raw repetitions from affected groups and recomputed their means, restoring four-band inputs. Target counts remained correct for the 11, 23, and 28 cm test configurations; Table H.3 reports their before/after spacing estimates.

Q3 Generalization to held-out spacings

The original split tests transfer across propagation ranges. We added ten leave-one-measured-spacing-out evaluations across 7–58 cm: a selected spacing's measured training/validation examples are withheld while its 10 m test event remains. Appendix I.1–I.3 and Table I.1 provide all paired predictions and full-test MAEs.

Withholding the 13, 23, 28, and 35 cm spacings reduces the selected event's absolute error from 0.21 to 0.16, 0.85 to 0.40, 1.48 to 1.02, and 0.29 to 0.03 cm, respectively. Nine of ten full-test MAEs remain below 1 cm; the mean is 0.681 cm against the original 0.48 cm. The 28 cm fold improves full-test MAE to 0.46 cm, while the 9 cm fold reaches 1.30 cm. The results therefore include both improvements and degradations.

We additionally performed a limited zero-shot single-target evaluation using archived measurements acquired from the same tank after three days without water replacement. The results and their limitations are reported in Appendix F.

To our knowledge, prior multi-target waveform decomposition studies have not released their measured datasets. If our paper is accepted, we will release our complete dataset and core code, and add real marine data collected in subsequent sea trials at different depths and in different regions early the next year. We have also reviewed your concern regarding the small text in some figures. The fonts in certain sub-figures of Figures 1 and 2 are indeed somewhat small, and we will reformat these figures in the final version.

Comments on the paper

1 comment

Thank you for your time and constructive feedback. Given the tight schedule, new experiments and analyses are temporarily placed in the Appendix starting from page 23 (highlighted in blue and indexed in each response) and will be integrated into the main text in the final revision. We hope our responses address your concerns and appreciate your understanding.

Q1 Frozen learned decoder

With identical training and evaluation settings, freezing the learned decoder improves spacing MAE from 0.54 to 0.48 cm, RMSE from 0.74 to 0.68 cm, and signed bias from 0.42 to 0.25 cm. Both versions retain 100% single- and dual-target counting accuracy (Appendix G.3; Table G.3).

Q2 Matched AFD-Net comparison

AFD-Net was adapted with AquaSR's measured training and validation sets and tested on the same observations. We evaluated its original four-class count output and a binary one-/two-target output, complementing the main-text baseline trained under AFD-Net's original synthetic-data protocol (Appendix G.1; Table G.1).

The binary version improves overall counting from 85.15% to 89.11% and dual-target counting from 77.19% to 84.21%, while valid dual-target outputs rise from 44/57 to 48/57. Its conditional MAE changes from 3.172 to 3.324 cm, while RMSE decreases from 7.160 to 7.039 cm. AquaSR gives 100% counting accuracy and 57/57 valid outputs with 0.478 cm MAE and 0.683 cm RMSE; output coverage is essential when comparing conditional errors.

Q3(a) Independent units and overlap subsets

An independent unit is a synchronized multiband event at a fixed target configuration. Training/validation/testing use 8/12/10 m data; each measured split contains one dual-target event at each of 15 spacings. Each configuration's 1,000 raw repetitions are screened and averaged, not treated as independent events. Training and validation also use superpositions of measured single-target echoes; testing is measured-only.

The 10 m test set comprises 26 independent events (11 single, 15 dual), with 44 and 57 valid band waveforms and 26,000 technical repetitions (Appendix H.1; Table H.1). Appendix H.2, Table H.2, and Figures H.1–H.4 specify the excluded bands and their screening evidence.

Appendix A.1 shows that visible double peaks begin to disappear in some 8 m bands by 23 cm and are absent across all bands at 16 cm. We therefore evaluated nested 10 m subsets at ≤23, ≤19, ≤18, and ≤16 cm, progressively removing easier large-spacing cases. All subsets retain 100% counting accuracy, with spacing MAE of 0.628, 0.606, 0.508, and 0.496 cm, respectively (Appendix H.3; Table H.4).

Q3(b) Withholding measured spacings

The range-disjoint original split contains the same measured spacings at all three ranges. We therefore ran ten leave-one-measured-spacing-out evaluations across 7–58 cm, withholding each selected spacing's measured training/validation examples while retaining its corresponding 10 m test event.

Absolute error improves in four folds: at 13, 23, 28, and 35 cm, it falls from 0.21 to 0.16, 0.85 to 0.40, 1.48 to 1.02, and 0.29 to 0.03 cm. Nine full-test MAEs are below 1 cm (mean 0.681 cm vs. the original 0.48 cm); withholding 28 cm improves full-test MAE to 0.46 cm. The remaining folds include degradations, but they can also successfully decompose the dual target echo . Appendix I.1–I.3 and Table I.1 give the complete paired results.

Q4 Implementation and release

Appendices J.1 and C.1–C.5 document the calibration relations and network. Inputs comprise four wavelength bands standardized to 512 samples with validity masks. The encoder combines multiscale convolution, local/channel gates, and six Transformer layers. A denoising MMD-WAE learns the single-echo prior; its decoder is frozen during decomposition, and six physical-attribute experts refine spacing.

Appendices J.2 and C.6 define three stages: single-echo prior learning, estimator training, and corrector training. The latter two begin on synthetic data and continue with measured-data fine-tuning; the estimator is frozen for corrector training. Counting, centers, spacing, full-waveform reconstruction, and local windows are supervised; synthetic data additionally supervise components, amplitudes, and latent codes. Candidate cross-entropy, cumulative-distribution, and cross-band consistency terms are used, with measured/synthetic loss mixing and AdamW optimization.

Appendix J.3 describes our plan to release the data. To our knowledge, prior multi-target waveform decomposition studies have not released their precious measured datasets. Upon acceptance of this paper, we will release our complete dataset and core code. Furthermore, once the related experiments are completed early next year, we plan to share real marine data collected from subsequent sea trials across various depths and regions.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.