NSLSR: Nearshore Ship SWIR-LWIR Registration

• The paper introduces NSLSR as a registered paired SWIR–LWIR ship dataset, addressing data and task gaps in coastal multimodal perception.
• NSLSR leverages a two-stage registration process—manual rigid correction followed by heterogeneous matching—to align 1,205 image pairs with 2,818 annotated objects.
• NSLSR underpins LSFDNet, a single-stage fusion-detection network that jointly optimizes cross-modal features, delivering state-of-the-art performance in nearshore ship detection.

Nearshore Ship Long-Short Wave Registration (NSLSR) is a registered nearshore ship dataset of paired short-wave infrared (SWIR) and long-wave infrared (LWIR) images introduced as the data foundation for the single-stage fusion-and-detection framework LSFDNet. In current arXiv usage, NSLSR denotes a benchmark for joint SWIR–LWIR image fusion and ship detection in complex coastal scenes rather than a ship hydrodynamics algorithm. The dataset contains 1,205 registered SWIR–LWIR image pairs at 640 × 512 resolution and 2,818 annotated objects, and was created to address the scarcity of practical SWIR–LWIR registered ship datasets and the lack of benchmarks for multimodal nearshore ship perception under variable illumination, fog, shoreline clutter, and partial occlusion (Guo et al., 28 Jul 2025).

1. Definition, nomenclature, and problem setting

NSLSR is introduced in the paper “LSFDNet: A Single-Stage Fusion and Detection Network for Ships Using SWIR and LWIR” as “Nearshore Ship Long-Short Wave Registration”, and the text also refers to it as the “Nearshore Ship Long-Short Wave Registered Dataset” (Guo et al., 28 Jul 2025). The operational content of the acronym is the pairing and registration of LWIR and SWIR image modalities for nearshore ship detection. This suggests that “long-short wave” in this context refers to infrared wavelength bands rather than a coastal hydrodynamic long-wave/short-wave decomposition.

The dataset addresses two gaps identified by the authors. The first is a data gap: publicly available registered SWIR–LWIR maritime datasets are described as scarce, with the Infrared Ship Dataset (ISD) from Shandong University identified as the only public dataset available before NSLSR for this direction, but limited in scale and diversity. The second is a task gap: prior multimodal fusion work is described as concentrating on visible–infrared fusion rather than SWIR–LWIR fusion, while maritime ship detection work is described as relying mainly on single-modal imagery (Guo et al., 28 Jul 2025).

The target environment is explicitly nearshore marine scenes collected over different time periods and under diverse lighting conditions. The dataset is intended for realistic coastal detection problems where ships must be detected against complex shorelines, sea-surface clutter, weak visibility, and partial occlusion. In the paper’s formulation, NSLSR is not merely an archive of paired imagery; it is the principal benchmark used to train and evaluate a coupled fusion-and-detection network, so dataset design and task design are inseparable (Guo et al., 28 Jul 2025).

2. Data acquisition, registration, and annotation

NSLSR was collected with a binocular synchronous system comprising one SWIR camera and one LWIR camera, both producing images at 640 × 512 resolution (Guo et al., 28 Jul 2025). The SWIR sensor is specified as an uncooled InGaAs infrared FPA detector with 15 μm pixel pitch and 0.9–1.7 μm spectral response. The LWIR sensor is specified as an uncooled VOx infrared FPA detector with 12 μm pixel pitch and 8–14 μm spectral response (Guo et al., 28 Jul 2025). The paper does not specify whether this binocular platform was shore-based, vessel-mounted, tripod-mounted, or airborne.

Registration is central to the dataset’s definition. Each SWIR image and LWIR image in a pair is spatially aligned so that corresponding scene content, especially ship targets, occupies matching positions across modalities. The alignment procedure has two stages. First, rigid transformation correction between SWIR and LWIR image pairs was manually corrected. Second, soft deformations were aligned using a heterogeneous image registration algorithm, specifically MINIMA: Modality Invariant Image Matching. After this two-stage process, the authors discarded poorly registered pairs (Guo et al., 28 Jul 2025). The paper does not report a numerical registration-accuracy metric or a separate registration-quality benchmark.

The final dataset contains 1,205 registered SWIR–LWIR image pairs and 2,818 annotated objects, with the paper stating that all ship objects in the images are annotated (Guo et al., 28 Jul 2025). The annotation regime is effectively single-class ship detection; no multi-class taxonomy is reported. The paper does not specify the annotation file format.

A notable practical issue is the split specification. In the dataset section, NSLSR is said to be divided into training and testing subsets with a 9:1 ratio. In the experimental section, however, the reported split is 844 images for training and 361 images for testing, which sums to 1,205 but does not match a 9:1 ratio (Guo et al., 28 Jul 2025). The experimentally used split is therefore clearly 844/361, while the textual ratio statement remains inconsistent. Within the 361-image test set, 118 images are used for fusion evaluation, whereas the full 361-image test set is used for detection evaluation. No validation split is described.

3. Modal complementarity and nearshore scene content

The rationale for NSLSR rests on the complementary sensing properties of SWIR and LWIR. The paper describes SWIR in general as approximately 0.9–2.5 μm, while the specific dataset camera operates over 0.9–1.7 μm; LWIR is given as 8–14 μm (Guo et al., 28 Jul 2025). SWIR is said to preserve more texture and structural detail and to provide strong target-to-water contrast because seawater strongly absorbs SWIR radiation, making the background darker and ships relatively bright. SWIR is also described as having better fog penetration than visible imagery and maintaining clarity in thin fog, smoke, and aerosol conditions (Guo et al., 28 Jul 2025).

LWIR supplies a different advantage profile. Because it is based on thermal emission, it is much less sensitive to illumination changes and remains useful in evening and nighttime conditions. Its weakness is reduced fine detail and edge sharpness, attributed in the paper to thermal diffusion and noise (Guo et al., 28 Jul 2025). The fusion problem therefore has a physically grounded objective: preserve thermal saliency from LWIR while preserving detail and texture from SWIR.

NSLSR’s scene content is defined by the practical difficulties of nearshore detection. The paper repeatedly emphasizes complex coastal scenarios and diverse lighting conditions, with qualitative examples including normal lighting, evening, nighttime, fog / low visibility, sea-surface clutter/noise, complex coastal backgrounds, small or weak targets, and partial occlusion (Guo et al., 28 Jul 2025). One highlighted qualitative case involves a small distant boat partly blocked by the mast of a larger nearby vessel. The paper does not provide a finer-grained metadata taxonomy such as exact locations, harbor names, weather labels, or sea-state labels.

The comparison with ISD further clarifies NSLSR’s intended role. ISD is reported to contain 1,044 image pairs at 300 × 300 resolution, only 28 unique ship instances, and long-range imagery at 10–12 km with monotonous backgrounds and limited diversity. NSLSR is positioned instead as a nearshore dataset with more object instances, more diverse backgrounds, and more practical utility for learning both fusion and detection in cluttered coastal settings (Guo et al., 28 Jul 2025).

4. Use of NSLSR in LSFDNet

NSLSR is the principal benchmark for LSFDNet, a single-stage multimodal fusion-detection network designed to exploit registered SWIR/LWIR pairs (Guo et al., 28 Jul 2025). The architecture has three named components: Multi-Task Feature Extraction (MTFE), the Multi-Level Cross-Fusion (MLCF) branch, and the Object Detection branch. The central design decision is to avoid a sequential “fuse first, detect later” pipeline. Instead, fusion and detection are jointly optimized with bidirectional cross-task interaction.

Given registered inputs $I_{SW}$ and $I_{LW}$, a base extractor produces shallow features $F_{SW}$ and $F_{LW}$. A fusion extractor produces modality-specific fusion features $\tilde{F}_{SW}$ and $\tilde{F}_{LW}$, while a YOLO-based detection backbone, cited in the paper as YOLOv12, processes the detection stream (Guo et al., 28 Jul 2025). A preliminary fused feature $F_f$ is passed into the detector, and an attention-weighted detection feature $F_{det\_attn}$ is returned to the fusion branch. Because NSLSR pairs are spatially registered, these cross-modal and cross-task feature transfers are meaningful at the pixel and feature levels.

The MLCF module is the core fusion mechanism. It is composed of three Multi-Feature Attention (MFA) blocks and is intended to aggregate complementarity across three axes: multimodal, multiscale, and multitask (Guo et al., 28 Jul 2025). The first MFA fuses $\tilde{F}_{SW}$ and $\tilde{F}_{LW}$ into $I_{LW}$0; a second path produces $I_{LW}$1; a third MFA combines them into the preliminary fused feature $I_{LW}$2. The MFA itself operates at the patch level and uses self-attention within modality and cross-attention across modalities. The paper gives, for example, the SWIR self-attention update as

$I_{LW}$3

with an analogous expression for LWIR, followed by cross-attention between modalities (Guo et al., 28 Jul 2025).

NSLSR annotations are used not only for detection supervision but also to shape the fusion objective through the Object Enhancement (OE) loss. The total loss is

$I_{LW}$4

where $I_{LW}$5 is the fusion loss and $I_{LW}$6 is the YOLO detection loss (Guo et al., 28 Jul 2025). The fusion loss is decomposed as

$I_{LW}$7

with reported hyperparameters

$I_{LW}$8

A gamma-enhanced LWIR image is defined as

$I_{LW}$9

and the object-focused loss is built over annotated ship regions. This design is motivated by the claim that generic fusion losses may preserve excessive sea-surface background structure such as undulations, highlights, and glare, whereas NSLSR requires object-sensitive fusion (Guo et al., 28 Jul 2025).

5. Benchmark protocol, reported performance, and ablation findings

On NSLSR, LSFDNet is trained end-to-end with Adam, learning rate $F_{SW}$0, linear decay, 500 warm-up iterations, 30,000 total iterations, and batch size 8, on NVIDIA GeForce RTX 4090 hardware (Guo et al., 28 Jul 2025). Fusion is evaluated on 118 test images using EN, SF, SD, SCD, VIF/VI, and Qabf. Detection is evaluated on the full 361-image test set using Precision (P), Recall (R), mAP$F_{SW}$1, and mAP$F_{SW}$2 (Guo et al., 28 Jul 2025).

For fusion on NSLSR, LSFDNet reports:

• EN: 7.181
• SF: 21.022
• SD: 64.723
• SCD: 1.427
• VIF/VI: 0.611
• Qabf: 0.520

The paper states that LSFDNet is either best or second-best on all six metrics on NSLSR, and best on SF, SD, SCD, VIF, and Qabf, while second-best on EN, behind DifFusion’s 7.216 (Guo et al., 28 Jul 2025). Qualitatively, the reported interpretation is that LSFDNet suppresses sea-surface noise while preserving ship contours, detail, and brightness.

For detection on NSLSR, the reported results are:

• SWIR only: $F_{SW}$3, $F_{SW}$4, $F_{SW}$5, $F_{SW}$6
• LWIR only: $F_{SW}$7, $F_{SW}$8, $F_{SW}$9, $F_{LW}$0
• SeA + YOLOv12s: $F_{LW}$1, $F_{LW}$2, $F_{LW}$3, $F_{LW}$4
• IGNet + YOLOv12s: $F_{LW}$5, $F_{LW}$6, $F_{LW}$7, $F_{LW}$8
• Fusiondif + YOLOv12s: $F_{LW}$9, $\tilde{F}_{SW}$0, $\tilde{F}_{SW}$1, $\tilde{F}_{SW}$2
• Swin + YOLOv12s: $\tilde{F}_{SW}$3, $\tilde{F}_{SW}$4, $\tilde{F}_{SW}$5, $\tilde{F}_{SW}$6
• CAFF-DINO: $\tilde{F}_{SW}$7, $\tilde{F}_{SW}$8, $\tilde{F}_{SW}$9, $\tilde{F}_{LW}$0
• DEYOLO: $\tilde{F}_{LW}$1, $\tilde{F}_{LW}$2, $\tilde{F}_{LW}$3, $\tilde{F}_{LW}$4
• LSFDNet: $\tilde{F}_{LW}$5, $\tilde{F}_{LW}$6, $\tilde{F}_{LW}$7, $\tilde{F}_{LW}$8

Thus LSFDNet achieves the best reported mAP$\tilde{F}_{LW}$9 and mAP$F_f$0 on NSLSR (Guo et al., 28 Jul 2025). The strongest listed competitor in $F_f$1 is CAFF-DINO at 0.706, so LSFDNet’s gain is 0.064 absolute, which the paper describes as about 10%.

The ablation study is also entirely on NSLSR. For fusion, the full model M7 reaches the same fusion metrics listed above. The paper states that removing OE loss reduces performance, especially SD, while removing the entire MLCF causes much larger degradation; the authors further state that multimodal aggregation is especially crucial and that feeding detection features back to fusion improves fusion quality (Guo et al., 28 Jul 2025). For detection, removing fusion feature $F_f$2 yields

$F_f$3

whereas full LSFDNet gives

$F_f$4

This corresponds to gains of +0.009 in $F_f$5 and +0.045 in $F_f$6, supporting the paper’s claim that fusion features materially improve downstream detection on NSLSR (Guo et al., 28 Jul 2025).

6. Interpretive boundaries, limitations, and relation to adjacent nearshore research

A recurrent source of ambiguity is the acronym itself. In the current literature, NSLSR denotes a registered SWIR–LWIR nearshore ship dataset and benchmark (Guo et al., 28 Jul 2025). This can be confused with physically grounded “long-wave/short-wave” ship–wave analysis. The distinction matters because several nearby research directions concern actual nearshore hydrodynamics rather than infrared image registration.

For example, “Down-scale marine hydrodynamic analysis at the Norwegian coast — the NORA-SARAH open framework” presents a multiscale coastal hydrodynamic downscaling chain,

$F_f$7

for reconstructing nearshore wave conditions and loads around coastal assets, and its relevance to ships is interpretive rather than a definition of NSLSR (Wang et al., 6 Sep 2025). Likewise, single-point breaking-wave classification from nearshore pressure-gauge records (Holand et al., 2023), spectrogram-based identification of linear and nonlinear ship-wake components at fixed sensors (Pethiyagoda et al., 2016), and non-hydrostatic modeling of low-frequency response of moored floating structures in coastal waters (Rijnsdorp et al., 2022) all concern physical wave registration in the hydrodynamic sense, not the SWIR–LWIR image-registration benchmark designated by NSLSR.

Within its own domain, NSLSR has several explicit limitations. The paper does not provide:

• geographic coordinates or named collection sites,
• exact platform mounting details,
• a per-condition breakdown such as day/fog/night counts,
• a class taxonomy beyond ship,
• an annotation file format,
• registration error statistics,
• a validation split (Guo et al., 28 Jul 2025).

The split discrepancy between the stated 9:1 partition and the experimentally used 844/361 split is another practical limitation. More broadly, NSLSR is presently tied to a single application family: registered SWIR–LWIR fusion and detection in nearshore scenes. It does not itself encode ship dynamics, sea-state metadata, hydrodynamic long-wave/short-wave decomposition, or multimodal synchronization with radar, LiDAR, AIS, GNSS, or wave gauges. A plausible implication is that NSLSR should be viewed as a perception benchmark rather than a complete maritime situational-awareness corpus.

The paper states that the source code and dataset are available at https://github.com/Yanyin-Guo/LSFDNet (Guo et al., 28 Jul 2025). That public-release claim is important because NSLSR’s primary significance lies in enabling reproducible work on a modality pairing that had been largely absent from public maritime benchmarks: registered SWIR + LWIR under realistic nearshore conditions.