PaperlayerView exact revision on arXiv

Source-linked AI summary

Revealing Imaging Characteristics of Thermal Wakes from Underwater Vehicles Without Cooling-Water Discharge

Guoyi Zhang, Yanjin Du, Guangsheng Xu, Zhengyao Zhao, Siyang Chen, Jingjing Gao, Yuezheng Zhou, Yiyang Wu, Bin Wu, Tongsu Zhang, Lejia Ye, Donghe Wang, Jie Wang, Kaiye Hu, Zhipeng Ren, Han Wang, Boo Cheong Khoo, Xiangpeng Xu, Xiaohu Zhang

arXiv:2610.35038v1cs.CV

TL;DR

Small uncrewed underwater vehicles can evade conventional sonar and magnetic sensing, and realistic evidence about distinguishing their thermal wakes from natural clutter is limited. The paper establishes a real-world benchmark and analyzes wake structure, finding local similarity to background vortices but global organization along vehicle trajectories. It concludes that detection should account for this global physical structure.

  • Problem

    Realistic evidence is limited on whether displacement-induced thermal wakes can be distinguished from complex natural backgrounds, despite the need to detect small underwater vehicles that evade conventional sensors.

    Source: Introduction · pp. 1–2

  • Method

    The authors establish AVP-IR-Wake, a real-world benchmark of 161,485 infrared frames, and analyze wake and background structure alongside representative detectors.

    Source: Introduction · p. 2, Results · p. 3

  • Results

    Individual wake vortices are locally indistinguishable from background vortices, but the overall wake follows the vehicle trajectory while background clutter follows wind-driven thermal fronts.

    Source: Introduction · p. 2, Imaging Characteristics of Infrared… · p. 8

  • Takeaways & Limitations

    Detection approaches suited to thermal wakes must represent global trajectory-constrained morphology and wind-organized background structure rather than rely only on local features.

    Source: Discussion · p. 10

    The study does not establish environment-specific quantitative parameterizations of wake intensity or morphology, which require precise coupled oceanic, atmospheric and hydrodynamic measurements.

    Source: Discussion · p. 11

Abstract

from arXiv · show

Small uncrewed underwater vehicles are proliferating and threaten maritime security, yet evade sonar and magnetic sensors built for larger submarines. Infrared thermal wakes, generated by hydrodynamic displacement that modulates the cold-skin layer, offer a non-acoustic, non-magnetic alternative even without cooling-water discharge. Existing studies, however, rely on idealized simulations that fail to capture real-world backgrounds, offering limited guidance for target-background discrimination. We establish AVP-IR-Wake, the first large-scale, real-world, physically diverse benchmark of 161,485 frames collected over three years. We show that thermal wakes are extended structures of discrete vortices, violating the compactness assumption of conventional targets. We reveal local indistinguishability: individual wake vortices are statistically identical to wind-induced background vortices in temporal, frequency, and sparse domains, sharing low-frequency information and long-range correlations. This challenges the assumptions of compact, transient, high-frequency targets in conventional detection. Yet we establish global morphological separability: although locally inseparable, the wake as a whole is organized by vehicle trajectory, whereas background clutter is organized by wind-driven thermal fronts. This explains why representative detectors fail and shows that reliable detection must shift from local feature extraction to global trajectory-constrained physical modeling. Because the cold-skin layer varies globally and seasonally, forecasting wake observability will require integrating physical oceanography and atmospheric science, linking underwater surveillance to the changing ocean surface.

Introduction

Small underwater vehicles can evade sonar and magnetic sensors, while realistic displacement-induced thermal wakes remain poorly characterized against complex natural backgrounds. AVP-IR-Wake and its analyses show that conventional detectors struggle with locally indistinguishable vortices, while the wake’s global trajectory-constrained morphology distinguishes it from wind-organized clutter.

Introduction

  • Displacement-induced wakes can form without cooling-water discharge, but their imaging characteristics under realistic environmental variability remain insufficiently understood.The vehicle transports warmer subsurface water upward, perturbing the cold-skin layer; natural cold-skin conditions are difficult to reproduce in indoor wave tanks.Source: Introduction · pp. 1–2

Results

  • AVP-IR-Wake contains 161,485 real-world infrared frames collected over three years across multiple lakes and vehicle types.Most targets are below 0 dB SNR, while target sizes, background complexity, wind, season and collection time vary substantially.Source: Results · pp. 3–4, Figure 2 · p. 4

Imaging Characteristics of Infrared Thermal Wakes

  • Individual wake vortices share long-range temporal correlations with background vortices, and low-rank/sparse decomposition absorbs the wake into its low-rank component.The sparse component remains nearly empty, so the wake does not satisfy the low-rank and sparse assumptions.Source: Imaging Characteristics of Infrared… · pp. 5–6, Figure 3 · p. 5
  • Wake and background clutter share low-frequency information in spatial and spatiotemporal domains, leaving no frequency band that separates them.The retina-inspired filter produces no distinguishable response to wake sequences.Source: Imaging Characteristics of Infrared… · pp. 6–7, Figure 4 · p. 6
  • The wake is globally organized along the vehicle trajectory, whereas background vortices follow wind-driven thermal fronts.Gusts reduce both target and background vortex intensities; at some submergence depths, the wake can remain observable despite a weak cold-skin gradient.Source: Figure 5 · p. 7, Imaging Characteristics of Infrared… · p. 8
  • Across detector categories, average precision remains at or near zero, and several methods fail to converge on AVP-IR-Wake.The tested priors include compactness, high-frequency content, sparsity, temporal profiles and Gaussian dispersion.Source: Imaging Characteristics of Infrared… · pp. 8, 10, Table 1 · p. 9

Discussion

The paper argues that thermal-wake detection depends on matching detector assumptions to wake physics and the coupled ocean–atmosphere state. It identifies trajectory-constrained morphology as a global cue and calls for predictive work across broader environmental regimes.

Toward Detectors Matched to Thermal Wake Physics

  • Detectors should test for trajectory-constrained global structures while allowing vortices to occur randomly, rather than relying on local compactness, high-frequency content, or sparsity.The background should be represented as wind-organized thermal fronts, not random noise; environmental metadata may help future detectors.Source: Discussion · p. 10

Implications for Underwater Vehicle Stealth

  • Thermal-wake detectability depends on the coupled ocean–atmosphere boundary layer, making stealth a state-dependent property rather than a fixed platform capability.The discussion says nocturnal cooling makes identical operating conditions more detectable at night, while the depth–stealth tradeoff changes with wind speed.Source: Discussion · p. 10

Underwater Vehicle Observability as An Emergent Property of The Coupled Ocean–Atmosphere System

  • Wake observability varies with the ocean–atmosphere system: environmental coupling obscures local and spectral cues while preserving trajectory-constrained global morphology.The paper says future prediction requires interdisciplinary work spanning hydrodynamics, oceanography, atmospheric forcing, remote sensing, and computer vision.Source: Discussion · p. 10

Limitations and Future Work

  • The study characterizes general wake structure under realistic variability but does not establish environment-specific quantitative parameterizations of wake intensity or morphology.Predictive relationships require future observations across broader oceanic regimes and precise coupled-variable measurements.Source: Discussion · p. 11

Method

The methods combine physically diverse field data, trajectory- and physics-informed annotation, acquisition-day splitting, and frequency-domain analyses to characterize wake observability. The operating-envelope estimate and analyses are tied to specified vehicle, camera, and filtering conditions.

AVP-IR-Wake Dataset Collection

  • The estimated maximum detectable dive depth was 4.5D, where D is vehicle diameter, using the hydrodynamic model, vehicle CAD and motion profile, and camera sensitivity.The estimate assumes the model’s original parameters and settings and a camera sensitivity of 20 mK.Source: AVP-IR-Wake Dataset Collection · p. 11
  • Field experiments ran from 2023 to 2025 across multiple lakes and inland seas, using ROVs and AUVs under varied operating and environmental conditions.Tianjin Port was the primary site because it more closely approximates port surveillance; AUV safeguards were reconfigured to permit operation below their original depth limit.Source: AVP-IR-Wake Dataset Collection · p. 11

Data Annotation and Splitting

  • Annotations combined trajectory-shaped morphology, depth- and speed-dependent temporal lag, CFD predictions, camera geometry, and expert review to localize wakes.Gust-induced fronts aligned with vehicle motion required multi-expert cross-validation; IoU thresholds were 0.9 for strip-like wakes and 0.8 for vortical targets.Source: Data Annotation and Splitting · pp. 11–12
  • Videos were split by acquisition day into training and test sets at about a 3:2 ratio to assess generalization across unseen operating conditions.Of 252 videos, 52 were excluded; redundant frames under the same conditions were filtered, and the two principal wake morphologies had consistent proportions across splits.Source: Data Annotation and Splitting · p. 12

Frequency-Domain Analysis

  • Wavelet decomposition measured energy fractions across seven subbands and found wake energy concentrated in the lowest-frequency subband, with no band separating wakes from background clutter.The qualitative result persisted with db2 wavelets and three decomposition levels.Source: Frequency-Domain Analysis · p. 12
  • A five-layer retina-inspired filter processed consecutive grayscale frames to characterize motion and transient intensity changes in wake and comparison sequences.Its final response combined ganglion-cell output and accumulated temporal signal, then applied gamma correction, bilateral filtering, and normalization.Source: Frequency-Domain Analysis · p. 12

Data availability

The comparative analysis used two external infrared datasets whose access is available upon request.

Data availability

  • IRSatVideo-LEO comprises 200 image sequences and 91,021 frames, and access is available upon request via its dataset form.It is a semi-simulated benchmark for multi-frame infrared small-target detection in satellite videos, developed from Landsat 7–8 imagery.Source: Data availability · p. 13
  • GOD-Video comprises 600 videos and 141,017 frames covering multiple gas types, with access available upon request via its dataset website.The dataset is licensed under CC BY-NC 4.0.Source: Data availability · p. 13
  • The two external datasets were used in Fig. 4b to contrast thermal wakes with other classes of infrared targets.The comparison included satellite small-target images and gaseous-object videos.Source: Data availability · p. 13

Competing interests

The authors declare no known competing financial interests or personal relationships that could have influenced the reported work.

Competing interests

  • The authors declare no known competing financial interests or personal relationships that could have appeared to influence the work.Source: Competing interests · p. 14

S1 Comparison with Simulated Imaging

Synthetic wake imaging either omits realistic environmental backgrounds or combines wakes with backgrounds without preserving their physical coupling. Neither strategy provides a valid proxy for benchmarking detection algorithms against real-world clutter.

S1 Comparison with Simulated Imaging

  • CFD temperature fields can reproduce idealized wake morphology but omit environmental coupling from wind, waves, solar radiation, and resulting clutter.Source: S1 Comparison with Simulated Imaging · p. 15
  • Embedding synthetic wakes in real scenes ignores physical links between wake characteristics and background conditions, and high-sensitivity, long-term ocean imagery is scarce.Source: S1 Comparison with Simulated Imaging · p. 15
  • Neither computational simulation nor target embedding provides a valid proxy for benchmarking wake detectors.Simulations use idealized targets, while embedded wakes can be physically inconsistent with their backgrounds.Source: S1 Comparison with Simulated Imaging · p. 15

S2 Metrics

The dataset characterizes difficulty and physical diversity with SNR and VIE, while detection performance is reported using COCO-style metrics. VIE also defines four levels of background complexity.

S2.1 Dataset Metrics

  • SNR and VIE are the dataset's two metrics for characterizing difficulty and physical diversity.Source: S2.1 Dataset Metrics · p. 15
  • VIE quantifies grayscale-distribution fluctuation, and sequence-average VIE classifies background complexity as low, medium, high, or very high.The respective ranges are 0–1,000, 1,000–5,000, 5,000–10,000, and >10,000.Source: S2.1 Dataset Metrics · p. 15

S2.2 Evaluation Metrics

  • Evaluation reports COCO-style AP, AP50, AP75, APs, APm, APl, and AR detection metrics.AP averages precision–recall area across IoU thresholds from 0.50 to 0.95; the size-specific AP metrics follow COCO area ranges.Source: S2.2 Evaluation Metrics · p. 16

S3 Temporal correlation analysis

Temporal correlation analysis extracts intensity time series from infrared-image points and characterizes autocorrelation and pairwise cross-correlation. The accompanying figure reports that correlations between wake and background points can match or exceed correlations between wake vortices.

S3 Temporal correlation analysis

  • At each point of interest, the method extracts an intensity time series and computes autocorrelation across lags.Cross-correlation is defined for pairs of time series across positive and negative lags.Source: S3 Temporal correlation analysis · p. 16
  • The analysis constructs positive- and negative-correlation matrices from pairwise time-series cross-correlations.The maximum lag is 500, and time series are mean-subtracted before correlation analysis.Source: S3 Temporal correlation analysis · p. 16
  • Correlations between wake and background points are comparable to or higher than correlations between the two wake vortices.Source: Figure S2 · p. 17

S4 Training Process Visualization

Many detectors converge but achieve near-zero test performance under substantial shifts in external conditions. The train–test split probes whether they learn target representations that generalize across those conditions.

S4 Training Process Visualization

S5 Physical Basis of Trajectory–Morphology Consistency

An advection–diffusion model links the surface-temperature perturbation to a vehicle-localized source and its past trajectory. The resulting wake’s overall support follows the trajectory, while its internal vortex texture is set by local hydrodynamic structure.

S5 Physical Basis of Trajectory–Morphology Consistency

  • When thermal relaxation is long relative to vehicle motion, the surface-temperature anomaly is a superposition of contributions from the vehicle’s past positions along its trajectory.The model represents the surface perturbation as evolving through advection and diffusion, with a localized source induced by volumetric displacement.Source: S5 Physical Basis of… · pp. 17–18
  • Correlations between target and background sampling points are comparable to or higher than correlations between the two target vortices.The figure compares temporal profiles and autocorrelation and cross-correlation functions at points on wake and background vortices.Source: Figure S3 · p. 18
  • The wake’s spatial support follows the vehicle path, whether straight or curved, while its internal structure depends on the lateral source structure rather than trajectory.The source structure reflects local instabilities such as vortex shedding and turbulent mixing, and its scale and evolution depend on vehicle diameter, depth, speed, and stratification.Source: S5 Physical Basis of… · p. 18

S6 Geometric Basis of Camera-to-Wake Relative Positioning

Camera pose and vehicle trajectory provide a first-order estimate of where to search for a wake in an image. The estimate is approximate because wake drift can occur, and the geometric cue is used only to restrict annotation search.

S6 Geometric Basis of Camera-to-Wake Relative Positioning

  • Projecting the vertical line through the vehicle’s position under a pinhole camera model provides a first-order estimate of the wake’s image region.The projection uses vehicle and camera positions, camera altitude, and camera intrinsic and rotation matrices.Source: S6 Geometric Basis of Camera-to-Wake… · p. 18
  • Wake drift from the vehicle’s vertical projection can result from surface currents, wind stress, and residual cold-skin motion after the vehicle passes.The displacement is described in terms of effective surface drift velocity and elapsed time, and is typically modest while the wake remains visible.Source: S6 Geometric Basis of Camera-to-Wake… · p. 19
  • The wake and background share low-frequency dominance, so no spatial-frequency band separates them; the wake also does not satisfy low-rank and sparse assumptions.Across the dataset, the lowest-frequency subband’s normalized MSE is nearly two orders of magnitude lower than that of high-frequency detail subbands.Source: Figure S4 · p. 19, Figure S5 · p. 20, Figure S6 · p. 21
  • The geometric cue restricts the manual-annotation search region; it is not treated as an exact target location.The workflow combines it with trajectory consistency, hysteresis, CFD-assisted morphology prediction, and expert cross-validation.Source: S6 Geometric Basis of Camera-to-Wake… · p. 21

S7 Fréchet Inception Distance (FID) Analysis

The analysis uses Fréchet Inception Distance to quantify train–test feature-distribution shift across acquisition days with different environmental conditions. The reported distance is large, alongside evidence of inseparable target and background features and poor detector generalization.

S7 Fréchet Inception Distance (FID) Analysis

  • Target-box and pure-background feature distributions overlap completely in both UMAP and t-SNE embeddings, with target samples scattered across multiple clusters.The target boxes include surrounding clutter that occupies the dominant spatial proportion, and their features are consequently governed mostly by environmental clutter.Source: Figure S7 · p. 22
  • 252.467 is the dataset’s FID, indicating substantial train–test distribution discrepancy associated with simultaneous shifts in target and background characteristics across environmental conditions.The acquisition-day split places different conditions in the training and test sets; examples include calm nighttime wakes and fragmented wakes under strong daytime winds.Source: S7 Fréchet Inception Distance (FID)… · pp. 21–22
  • Hyper-YOLO and SSTNet converge in training but achieve near-zero test AP, while validation loss barely decreases as environmental conditions change.The figure identifies these as representative single-frame and multi-frame detectors, respectively.Source: Figure S8 · p. 23

Related papers

Loading 2610.35038v1…