Revealing Imaging Characteristics of Thermal Wakes from Underwater Vehicles Without Cooling-Water Discharge
Authors: 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
Organizations: School of Aeronautics and Astronautics, Sun Yat-sen University, Shenzhen, 518107, Guangdong, China · Changchun Institute of Optics, Fine Mechanics and Physics, Chinese Academy of Sciences, Changchun 130033, Jilin, China · Institute of Machinery Manufacturing Technology, China Academy of Engineering Physics, Mianyang, 621900, China · College of Shipbuilding Engineering, Harbin Engineering University, Harbin, 150006, Heilongjiang, China · School of Energy Science and Engineering, Harbin Institute of Technology, Harbin, 150001, Heilongjiang, China · Mechanical Engineering Department, National University of Singapore, 117575, Singapore
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.
Figures & tables
Figure 1: Thermal wakes as infrared visual targets. a , Schematic of wake formation. The vehicle’s hydrodynamic displacement transports warmer subsurface water upward, modulating the cold-skin layer and producing a surface thermal signature even without cooling-water discharge. b and c , Data were mainly collected at Tianjin Port using both AUVs and ROVs. d , Representative images of thermal wakes under different conditions. Deep submergence produces discrete vortex clusters, while shallow submergence produces continuous stripe-like wakes. These images show that the same target can appear in fundamentally different forms depending on operating conditions. e , The wake extends approximately 42 times the vehicle length and persists for several minutes, far exceeding the scale of the vehicle itself.
Figure 2: Statistical characterization of AVP-IR-Wake. a , SNR distribution. b , aspect-ratio distribution. c , object-area distribution. d , object position bias. e , background complexity. f , wind speed distribution. g , vehicles distribution. h , season distribution. i , collection time distribution.
Figure 3: Indistinguishability of individual wake vortices and background vortices. a , Sampling points. P1 and P2 are located on two constituent vortices of the wake; P3 to P6 are located on background false alarms. b , Temporal profiles of pixel intensity at the six sampling points. c , Autocorrelation functions of the six time series. d , Negative time-lagged cross-correlation matrix. e , Positive time-lagged cross-correlation matrix. The correlation between target and background points is comparable to or even higher than that between the two target vortices. f , A representative wake image at t1+3 s. g , Low-rank component obtained by low-rank and sparse decomposition. h , Sparse component. The wake is absorbed into the low-rank component, whereas the sparse component contains no discernible wake structure. These results show that individual wake vortices are not statistically distinguishable from background vortices, with additional results provided in Supplementary Figure S2 and S3 , and that the wake does not satisfy the low-rank and sparse assumptions (Further results are provided in Supplementary Figure S4 ).
Figure 4: Frequency-domain inseparability of thermal wakes and background clutter. a , Multiband decomposition of a wake image. The wake energy is concentrated in the low-frequency subband (L2_LL), whereas all high-frequency subbands (L1_HH, L1_HL, L1_LH, L2_HH, L2_HL, L2_LH) contain only background texture and noise. The same low-frequency dominance is observed in background regions, so no spatial-frequency band separates the wake from the clutter (Further results are provided in Supplementary Figure S5 and Figure S6 .). b , Motion sensitivity analysis using a retina-inspired spatiotemporal filter. The filter produces no distinguishable response in wake sequences (Seq.1 and Seq.2), but responds strongly to satellite small targets [ 39 ] (Seq.3) and gaseous objects [ 40 ] (Seq.4). This indicates that the wake also lacks a distinctive temporal-frequency signature in the spatiotemporal domain, whereas other targets possess one. Together, these results show that the wake and background clutter share the same low-frequency information in both the spatial and spatiotemporal domains, making frequency-domain separability unattainable.
Figure 5: Global morphological separability between thermal wakes and background vortices. a , Wake morphology is trajectory-constrained: individual vortices resemble background vortices, but the overall wake follows the vehicle trajectory (red arrows), forming a distinctive global structure. b , Local vortices appear stochastically in time and space with non-fixed spacing, due to non-uniform vehicle motion. c , Wind generates thermal fronts that organize background vortices into a characteristic global structure. Thus, the wake is organized by the trajectory, whereas background vortices are organized by wind-driven thermal fronts. d , Gusts weaken the cold-skin temperature gradient, reducing both target and background vortex intensities. Because the global trajectory structure is the key to target-background separability, the entire target region was delineated as the bounding box following expert recommendations. This inevitably results in background clutter occupying the dominant spatial proportion within these target boxes. To quantify this, we extracted UMAP [ 41 ] embeddings of 130,000 randomly sampled target boxes and corresponding background regions from the same images. The two classes are completely inseparable in the embedding space (Supplementary Fig. S7 ), quantitatively confirming that the visual features within the target boxes are overwhelmingly governed by the surrounding background clutter.
Methods
Venue
AP(%)
AP50(%)
AP75(%)
APs(%)
APm(%)
APl(%)
AR(%)
Single-Frame Detector
DNANet [ 43 ]
TIP’23
0.4
1.5
0.1
0.0
0.2
0.5
14.7
UIUNet [ 44 ]
TIP’23
0.7
2.7
0.1
0.0
0.0
1.4
3.4
YOLOH [ 45 ]
TIP’24
0.7
2.8
0.0
0.0
0.0
0.8
4.9
MSHNet [ 46 ]
CVPR’24
0.0
0.0
0.0
0.0
0.0
0.0
0.0
RT-DETR [ 47 ]
CVPR’24
4.7
15.9
1.4
0.0
0.3
5.9
30.6
Hyper-YOLO [ 48 ]
TPAMI’24
0.5
2.3
0.0
0.0
0.3
0.6
2.7
Table 1: Quantifying the Breakdown of Conventional Detection Assumptions on AVP-IR-Wake.
Figure S1: Comparison with Simulated Imaging. Simulated images can be computed from temperature fields and a rough sea surface emission model by reverse Monte Carlo (RMC); however, these data fail to capture the complex environmental information necessary to establish a realistic background. We thank Associate Professor Cheng’an Wang for providing the simulated data and computational results.
Figure S2: Indistinguishability of individual wake vortices and background vortices. a , Sampling points. P1 and P2 are located on two constituent vortices of the wake; P3 to P7 are located on background false alarms. b , Temporal profiles of pixel intensity at the seven sampling points. c , Autocorrelation functions of the seven time series. d , Negative time-lagged cross-correlation matrix. e , Positive time-lagged cross-correlation matrix. The correlation between target and background points is comparable to or even higher than that between the two target vortices.
Figure S3: Indistinguishability of individual wake vortices and background vortices. a , Sampling points. P1 to P3 are located on two constituent vortices of the wake; P4 to P7 are located on background false alarms. b , Temporal profiles of pixel intensity at the seven sampling points. c , Autocorrelation functions of the seven time series. d , Negative time-lagged cross-correlation matrix. e , Positive time-lagged cross-correlation matrix. The correlation between target and background points is comparable to or even higher than that between the two target vortices.
Figure S4: The wake does not satisfy the low-rank and sparse assumptions.
Figure S5: Multiband decomposition of a wake image. The wake energy is concentrated in the low-frequency subband (L2_LL), whereas all high-frequency subbands (L1_HH, L1_HL, L1_LH, L2_HH, L2_HL, L2_LH) contain only background texture and noise. The same low-frequency dominance is observed in background regions, so no spatial-frequency band separates the wake from the clutter.
Figure S6: Relative frequency-domain energy distribution across the entire dataset. Boxplots showing the Mean Squared Error (MSE) between the original images and their respective wavelet subbands across the entire 161,485-frame dataset. To facilitate comparison, the MSE values are normalized by the global maximum MSE across all frames and subbands. The y-axis is presented on a logarithmic scale. Even after this normalization, the lowest-frequency subband (L2_LL) exhibits an MSE nearly two orders of magnitude lower than the high-frequency detail subbands. This confirms that the image energy is overwhelmingly concentrated in the lowest frequency band, while high-frequency bands contain only noise. Boxes represent the interquartile range (IQR, 25th–75th percentiles), center lines indicate the median, and whiskers extend to 1.5× IQR.
Figure S7: Visualization of target and background feature distributions. a , UMAP embedding of 130,000 randomly sampled pure background regions. b , UMAP embedding of 130,000 randomly sampled target bounding boxes. c , Combined overlay of both distributions. Because the target annotation is designed to capture the entire trajectory-constrained global morphology, background clutter inevitably occupies the dominant spatial proportion within these target boxes. Consequently, the feature distributions of the target boxes ( b ) and the pure background regions ( a ) are completely inseparable in the embedding space, as demonstrated by the perfect overlap in c . This quantitatively confirms that the visual features extracted from the target boxes are overwhelmingly governed by the surrounding environmental clutter rather than the target itself. Furthermore, the target samples are scattered across multiple distinct clusters rather than forming a compact group, verifying the extremely high intra-class variance of thermal wakes under diverse environmental conditions. d–f , Parallel analysis using t-SNE embeddings for background regions ( d ), target boxes ( e ), and their combined overlay ( f ). The consistent inseparability and high intra-class variance observed in both UMAP and t-SNE confirm that these phenomena are intrinsic to the data rather than artifacts of a specific dimensionality reduction technique.
Figure S8: Training process. a , Training loss curve of Hyper-YOLO, a representative single-frame detector. b , Training loss curve of SSTNet, a representative multi-frame detector. Both detectors achieve near-zero AP on the test set, yet they converge successfully. However, their validation loss barely decreases, because changes in environmental conditions cause large simultaneous variations in both target and background characteristics. This highlights the stringent generalization requirement for detectors applied to unseen scenarios in real-world applications.
Figure S9: CFD-simulated surface temperature fields under different operating conditions.
Figure S10: Statistical characterization of AVP-IR-Wake.
Maritime vessel detectors often face scenes where hulls are small, low-contrast, or blurred, while wakes are longer and easier to detect. This creates a wake-reliance problem: detectors may miss slow or stationary vessels with weak wakes, or produce false positives on wake-like water clutter. We propose HullWake, a hull-first wake-second framework for robust maritime vessel detection. HullWake separates proposal-centered hull evidence from directional wake context, extracts wake cues with bidirectional proposal-anchored corridors, and suppresses wake-dominant predictions through wake response supervision, wake-attenuated consistency, wake-only confidence suppression, and hull--wake decorrelation. We also introduce a wake-oriented evaluation protocol covering weak/no-wake vessels, wake-like hard negatives, worst-group AP, and confidence drop after wake attenuation. Experiments are conducted on Curated-Wake, a wake-oriented maritime dataset of about 10,000 images curated from Ships/Vessels in Aerial Images, the SMD benchmark, and SeaDronesSee, with newly added detection- and segmentation-level wake annotations. Compared with box-only detectors and mask-supervised segmentation baselines, HullWake improves overall AP, weak/no-wake robustness, wake-like false positives, worst-group AP, and confidence stability after wake attenuation.
Yefan Wang, Xingyu Wang, Ruibiao Zhu +1
University of Shanghai for Science and Technology, Shanghai, China · University of Science and Technology Liaoning, Liaoning, China · School of Computing, College of Systems and Society, The Australian National University, ACT, Australia +1
Underwater robots often operate near delicate targets where high-power thrusters resuspend sediments and induce turbulence, degrading image quality at the sensor input. Conventional controllers optimize vehicle-centric objectives, such as tracking and stability, without accounting for the impact of actuation on sensing. We address this actuation-to-perception coupling by exploiting redundancy in over-actuated platforms. For an eight-thruster ROV, multiple thrust allocations can yield the same motion; we search this null space to minimize predicted disturbance in a task-relevant target region while enforcing motion constraints. Our method uses a control-oriented thruster-wake proxy derived from actuator-disk theory with directional attenuation and validated by PIV (R2=0.99 near the wake axis; R2>0.82 in the primary wake region), together with a real-time redundancy-resolving allocator running at 10 Hz (45 ms/solve). Across 440 trials, the approach reduces target-region particle velocity by 67% (p<0.001), improves 3D reconstruction RMSE by 55% versus a disturbance-unaware baseline (1.9±0.4 mm vs. 4.3±1.8 mm), and achieves a 98.5% reconstruction success rate. The framework supports autonomous scanning, which is quantitatively evaluated, and operator-assisted inspection, which is demonstrated in the supplementary materials.
Yuer Gao, Tongqing Xu, Qingyang Liu +1
Smart Manufacturing Thrust, Systems Hub, The Hong Kong University of Science and Technology (Guangzhou), Guangzhou, China
Underwater acoustic target detection and recognition are essential to marine surveillance and intelligent ocean sensing. Despite substantial advances in deep learning, existing systems typically produce binary detection decisions, categorical labels, or scalar confidence scores, offering limited insight into the physical evidence underlying their outputs. Large language models (LLMs) offer an opportunity to move underwater acoustic sensing toward interpretable, evidence-grounded understanding; however, generic audio encoders struggle to capture the distinctive physical characteristics of underwater acoustics, and physically annotated data are scarce. In this paper, the first of a two-part series, we propose a from-signals-to-evidence framework that integrates physics-aware acoustic encoding with a pretrained LLM, organized around a unified feature-evidence-decision chain with a shared reasoning core. An open input layer accommodates heterogeneous information-raw waveforms, time-frequency representations, environmental and geographic data, and textual physical priors-unifying detection and recognition within a common formulation. In Part I, we instantiate this framework for target recognition through an Underwater Acoustic Physics-aware Transformer (UPT), which tokenizes complementary acoustic representations and learns task-disentangled physical representations through four physics-oriented learning tasks. We develop a physics-parameterized data engine that provides traceable supervision at two levels of granularity and introduce an attribute-level evaluation protocol to assess the physical fidelity of generated descriptions. Experiments on simulated recordings of four vessels demonstrate accurate vessel recognition and reliable estimation of shaft frequency, blade-passing frequency, and machinery-related attributes.
Yuge Zhang, Jie Chen, Zhengqiao Zhao +4
School of Artificial Intelligence, Northwestern Polytechnical University, Xi’an, Shaanxi 710129, China · School of Ocean Engineering and Technology, Sun Yat-sen University, Zhuhai, 519000, Guangdong, China · School of Marine Science and Technology, Northwestern Polytechnical University, Xi’an, Shaanxi 710129, China +1