---
title: 'DarkVesselNet: Multi-Modal Dark Vessel Detection'
url: https://www.emergentmind.com/topics/darkvesselnet
type: topic
---

# DarkVesselNet: Multi-Modal Dark Vessel Detection

Searching arXiv for the cited DarkVesselNet paper and closely related dark-vessel detection work.
DarkVesselNet is a multi-modal remote sensing and trajectory-reasoning stack for dark-vessel detection that fuses Sentinel-1 SAR, Sentinel-2 optical imagery, AIS trajectories, geospatial foundation-model backbones, TGARD-style gap detection, and a Pi-DPM-inspired anomaly head [2606.00445]. It is presented as an auditable alerting system rather than a single detector: SAR supplies day-and-night physical observations, optical imagery contributes contextual interpretation when available, AIS provides cooperative motion reports, and trajectory reasoning evaluates whether AIS gaps or rendezvous patterns are suspicious [2606.00445]. In this framing, dark-vessel detection is an evidence-fusion problem under missingness, and the output is intended to be a triage alert with a trace rather than an automatic legal conclusion [2606.00445].

## 1. Conceptual scope and problem definition

The paper distinguishes several operational categories that structure the entire stack: a **SAR object** is a radar-bright candidate; a **vessel candidate** is a SAR object consistent with a vessel in size or context; an **AIS-matched vessel** is one plausibly associated with a track; an **AIS-unmatched vessel** is one with no plausible AIS explanation under the matching policy; and a **dark-vessel alert** is a review-worthy suspicious candidate [2606.00445]. It also states explicitly that **confirmed illegal activity** is outside the model’s authority [2606.00445].

This taxonomy matters because the system does not equate AIS absence with wrongdoing. The paper states that missing broadcasts may arise from receiver coverage gaps, device failure, timing mismatch, deliberate disabling, or other benign causes [2606.00445]. DarkVesselNet is therefore framed as a mechanism for surfacing evidence patterns rather than inferring intent or legality [2606.00445]. A plausible implication is that the system’s design priorities—traceability, calibration, and explicit missingness handling—follow directly from this narrow evidentiary framing.

The broader literature in the supplied record situates this framing within a wider class of dark-vessel systems. Sea-Scan, for example, defines “dark vessel” detection as identifying vessel transit events from unlabeled or AIS-silent DAS data, meaning events where a vessel is acoustically present but not accompanied by a corresponding AIS report [2606.21326]. Likewise, the xView3-SAR benchmark is motivated by the fact that vessels may not carry AIS, may turn it off, or may be operating illicitly, making image-based or radar-based detection necessary [2206.00897]. DarkVesselNet differs in emphasizing cross-modal evidence fusion and trajectory reasoning rather than a single sensing modality [2606.00445].

## 2. End-to-end pipeline and sensor stack

The architecture is presented as a pipeline from area of interest to calibrated alert [2606.00445]. The sequence begins with acquisition of Sentinel-1 SAR and Sentinel-2 optical scenes for an AOI and collection of AIS trajectories over a time window around the SAR acquisition [2606.00445]. Preprocessing then applies speckle reduction via a **Lee filter** to SAR imagery, performs cloud masking and band-ratio feature extraction such as NDVI and NDWI on optical imagery, coregisters SAR, optical, and contextual layers, and builds availability masks for missing modalities [2606.00445].

Sentinel-1 SAR is the primary all-weather sensor in the stack [2606.00445]. The paper emphasizes its utility at night and through clouds, while also noting speckle and coastal clutter [2606.00445]. The Lee-filter implementation is validated by tests showing that constant imagery remains effectively unchanged and that synthetic speckle variance is reduced [2606.00445]. This suggests that the SAR branch is designed around stability and software verifiability rather than around a claimed benchmark detector.

Sentinel-2 optical imagery is treated as contextual support rather than the primary signal [2606.00445]. The reported band-ratio formulas are NDVI and NDWI, and the text presents them as standard discriminants for contextual reasoning around coastlines, water, and land [2606.00445]. The tests check band-ratio behavior and output ranges [2606.00445]. The paper also repeatedly references coastline, port, and infrastructure context as additional signals because near-shore false positives are a major failure mode [2606.00445].

A concise summary of the operational stack is given below.

| Stage | Reported function | Reported components |
|---|---|---|
| Acquisition | Collect imagery and AIS windows | Sentinel-1, Sentinel-2, AIS trajectories |
| Preprocessing | Prepare sensor inputs and masks | Lee filter, cloud masking, NDVI, NDWI, coregistration |
| Encoding | Convert imagery to tokens | GeoBackbone adapter, patch tokens \((B,N,D)\) |
| Trajectory reasoning | Analyze AIS history | TGARD-style gap detection, Haversine checks |
| Fusion and alerting | Combine evidence and score anomalies | Missingness-aware fusion, Pi-DPM-inspired anomaly head, temperature scaling |

In comparison with other dark-vessel detection paradigms in the supplied record, this sensor stack is explicitly heterogeneous. Sea-Scan uses distributed acoustic sensing over a 120 km subsea link and detects vessel passages from DAS time–distance representations independently of AIS [2606.21326]. The satellite transshipment system of [2404.07607] instead combines PlanetScope optical imagery with AIS cross-referencing to identify dark ship-to-ship transfers. DarkVesselNet occupies a different design point by combining SAR, optical, AIS, and static context into a single missingness-aware architecture [2606.00445].

## 3. Backbone abstraction and fusion under missingness

A central architectural component is the **GeoBackbone adapter**, which normalizes different geospatial foundation models behind one API and returns patch tokens of shape \((B, N, D)\) regardless of the underlying model [2606.00445]. Supported backbones mentioned in the paper include **Prithvi-2, Clay, SatMAE++, DOFA, SatlasNet, and RemoteCLIP-style models** [2606.00445]. Each backbone is described by metadata such as patch size, embedding dimension, expected bands, Hugging Face model identifier, and license constraints [2606.00445]. The paper also notes that in CPU-only tests, a lightweight fallback projection mimics token output so that the rest of the system can be validated without downloading large weights [2606.00445].

This adapter design keeps downstream fusion and anomaly heads fixed while allowing encoder substitution [2606.00445]. That design choice aligns with the paper’s emphasis on a software-grounded, auditable stack. The validation evidence includes tests for backbone token shapes, supported backbone list, and shape consistency under the adapter [2606.00445]. A plausible implication is that DarkVesselNet treats backbone variability as an interface problem rather than a monolithic model-selection problem.

The fusion mechanism is explicitly **missingness-aware** [2606.00445]. One representation given in the paper is
\[
h = \operatorname{MLP}([e_s,e_o,e_a,m_s,m_o,m_a,c]),
\]
where \(e_s\), \(e_o\), and \(e_a\) are SAR, optical, and AIS embeddings, \(m_s\), \(m_o\), and \(m_a\) are modality-availability masks, and \(c\) denotes context features [2606.00445]. A second variant multiplies embeddings by availability masks before fusion and still includes the masks explicitly [2606.00445]. The stated purpose is to prevent the model from confusing absent evidence with zero-valued evidence [2606.00445].

The architecture figure is described as implementing fusion with LayerNorm, evidence \(Q,K,V\), missingness attention, fusion MLP plus residual, and residual add/norm, with the attention matrix shown as a \(QK^\top/\sqrt{d}\) heatmap [2606.00445]. The paper also gives a conceptual Bayesian view in which the alert is a posterior over latent event \(Y\) conditioned on SAR, optical, AIS, and context observations [2606.00445]. Because the notation is reported as garbled in the source, the safe factual interpretation is limited to the intended idea of posterior-style evidence fusion [2606.00445].

This architecture can be contrasted with the weakly supervised spatio-temporal fusion in Sea-Scan, where a UniFormer-based encoder, Feature Pyramid Network, temporal branch, and spatial branch are combined multiplicatively as \(M_{t,c} = P_t \cdot P_{t,c}\) [2606.21326]. Both systems use factorization and gating to stabilize noisy evidence, but DarkVesselNet applies this principle across sensing modalities and metadata rather than across temporal and channel axes within DAS [2606.00445].

## 4. AIS trajectory reasoning and matching logic

AIS is represented as sequences of the form \((t,\phi,\lambda,\text{sog},\text{cog})\), where \(\phi\) and \(\lambda\) are latitude and longitude, and \(\text{sog}\) and \(\text{cog}\) are speed over ground and course over ground [2606.00445]. The paper uses Haversine distance for geodesic separation,
\[
d = 2R\arcsin\sqrt{\sin^2(\Delta\phi/2)+\cos\phi_1\cos\phi_2\sin^2(\Delta\lambda/2)},
\]
and reports tests including a zero-distance sanity check, a one-degree latitude sanity check, and gap-related checks [2606.00445]. These geometry routines are used to assess whether motion between AIS observations is plausible [2606.00445].

The **TGARD-style gap detection** component reasons about time duration of missing AIS, distance that would need to be covered, and feasibility under vessel motion constraints [2606.00445]. The current implementation is said to test short-gap skipping, synthetic gap emission, and infeasible-gap detection [2606.00445]. Importantly, TGARD is not treated as a standalone anomaly detector but as one evidence source in the multimodal alert stack [2606.00445].

The AIS matching policy centers on a SAR acquisition time \(t_s\). Candidate AIS messages are drawn from a window \([t_s-\Delta_t, t_s+\Delta_t]\), and interpolation may be used when messages bracket \(t_s\) and the implied motion is plausible [2606.00445]. The paper gives a score-like formulation involving spatial distance, heading consistency, and vessel-length compatibility, and also provides an uncertainty-aware likelihood form,
\[
\ell(a\rightarrow s)= -\frac{1}{2}(p_a(t_s)-p_s)^\top\Sigma^{-1}(p_a(t_s)-p_s) -\eta|\ell_a-\ell_s|,
\]
where \(\Sigma\) captures uncertainty from AIS interpolation, SAR geolocation, and time offset [2606.00445].

This trajectory-reasoning layer differentiates DarkVesselNet from image-only pipelines. The optical STS system of [2404.07607] cross-references image detections with AIS using a surrounding 500 m area and a \(\pm 12\) hour temporal window, classifying a transfer as dark if fewer than two distinct vessel identities are found [2404.07607]. DarkVesselNet instead integrates AIS reasoning directly into its anomaly stack and treats gaps and matching plausibility as features for alert generation [2606.00445].

## 5. Anomaly head, calibration, and evidence trace

The anomaly head is described as **Pi-DPM-inspired**, though explicitly not as a full diffusion sampler [2606.00445]. It pools scene tokens, projects AIS points through an MLP, fuses both, and outputs both an alert logit and a reconstructed AIS or trajectory segment [2606.00445]. The paper gives the structure
\[
h = f_{\theta}\left([\text{pool}(E_{\text{scene}}), \text{pool}(E_{\text{AIS}})]\right), \quad y = W_s h,\quad \hat{\tau}=W_r h,
\]
while noting formatting corruption in the source [2606.00445]. Later it provides a more general additive alert decomposition,
\[
z = z_{\text{sar}} + z_{\text{ais}} + z_{\text{traj}} + z_{\text{opt}} + z_{\text{ctx}},
\]
followed by calibrated probability
\[
\hat{p}=\sigma(z/T)
\]
using temperature scaling [2606.00445].

The paper emphasizes that raw logits are not probabilities and proposes **temperature scaling** as the calibration method [2606.00445]. It also specifies an expected calibration error:
\[
\mathrm{ECE} = \sum_{b=1}^{M} \frac{|S_b|}{N} \left| \operatorname{acc}(S_b)-\operatorname{conf}(S_b) \right|,
\]
which is meant to be reported after fusion rather than only on a SAR detector [2606.00445]. This indicates that calibration is treated as an end-to-end system property.

A distinct trace-completeness metric is also defined:
\[
\begin{aligned}
T=\frac{1}{N}\sum_{i=1}^{N} &\mathbf{1}\{\text{sensor evidence}_i\}\mathbf{1}\{\text{AIS decision}_i\}\\
&\times\mathbf{1}\{\text{context decision}_i\}\mathbf{1}\{\text{uncertainty reported}_i\}.
\end{aligned}
\]
The paper states that the trace is part of the system output, not optional metadata [2606.00445]. The evidence trace includes modality availability, AIS match status, gap or rendezvous evidence, context, and calibration bucket [2606.00445]. This is one of the paper’s clearest departures from purely score-centric detector design.

A plausible implication is that DarkVesselNet’s alert semantics are closer to decision-support systems in safety-critical settings than to benchmark-oriented object detectors. That interpretation is consistent with the paper’s repeated warning that a black-box probability is insufficient and that ambiguity, missingness, and near-shore clutter must be reported explicitly [2606.00445].

## 6. Validation status, recommended evaluation, and limitations

The reported evaluation is primarily software-grounded rather than benchmark-grounded [2606.00445]. The repository passes **15 tests** in the current run, with checks covering SAR speckle filtering behavior, optical band-ratio utilities, Haversine checks, gap logic, sensor fusion and coregistration, backbone token shapes, anomaly-head output shapes, and Space smoke behavior [2606.00445]. The validation matrix lists 6 checks for SAR and optical behavior, 4 for trajectory reasoning, 3 for fusion and backbone, and 2 for the anomaly head [2606.00445]. The paper also states that a local run of `uv run --extra dev pytest -q` yields 15 passing tests [2606.00445].

Specific software validations are enumerated with unusual precision: SAR speckle filtering is tested via constant-image idempotence and reduced variance on synthetic speckle; optical utilities via NDVI and NDWI range and shape checks; Haversine distance via zero-distance and one-degree latitude checks; TGARD gap logic via short-gap skipping and synthetic gap detection; coregistration via identity coregistration shape; the backbone adapter via output shape \((B,N,D)\); and the anomaly head via output shapes and backward-pass support [2606.00445]. The evidence currently available is therefore implementation-centric [2606.00445].

The paper does not claim completed xView3 leaderboard performance [2606.00445]. Instead, it lays out a recommended evaluation protocol consisting of SAR detection on xView3-style chips, AIS matching under different windows and thresholds, trajectory anomaly detection on matched tracks, fusion ablations, and trace audit for completeness and explanation quality [2606.00445]. Suggested metrics include mAP, recall by vessel length, false positives near shore, match precision and recall, time-offset sensitivity, gap precision, rendezvous precision, end-to-end alert precision, ECE, and trace completeness, as well as
\[
\mathrm{Prec}@B = \frac{1}{B} \sum_{i\in \pi_B} \mathbf{1}\{y_i=\mathrm{dark\ vessel}\},
\]
which the paper motivates as appropriate under a fixed analyst review budget [2606.00445].

The paper includes an ablation-style table with values for SAR only, SAR + AIS, SAR + trajectory, SAR + optical, and full fusion, but it explicitly presents these as recommended or template results rather than definitive benchmark claims [2606.00445]. Because the source characterizes them as illustrative rather than validated, they should not be read as established empirical performance.

The limitations section is unusually explicit. The paper states that the current system is **not live satellite ingest**, that the anomaly head is a surrogate for a fuller physics-informed diffusion model, that backbone models differ in bands, licensing, and modality compatibility, that the system should not be used for enforcement without human review, that it cannot infer legal intent, that AIS spoofing is not fully handled, that optical imagery can be biased by daylight and weather, and that SAR backbones pretrained on optical data may not transfer cleanly [2606.00445]. This suggests that the system’s present contribution is architectural and operational rather than a completed, field-validated surveillance product.

## 7. Position within dark-vessel research

DarkVesselNet belongs to a broader family of systems that detect maritime activity not reliably revealed by AIS, but its defining characteristic is its explicit treatment of detection as multi-modal evidence fusion under uncertainty [2606.00445]. In the provided record, three related lines of work illustrate the surrounding design space.

First, **Sea-Scan** uses distributed acoustic sensing over a subsea fiber link to detect and localize vessels from time–distance DAS measurements, training with AIS-assisted weak labels and then applying the model to AIS-silent segments [2606.21326]. On a balanced test set of 454 vessel-crossing events and 454 noise events, it reports a **97.8% detection rate** with **TP = 444** and **FN = 10**, and a **1.98% false-trigger rate** with **FP = 9** and **TN = 445** [2606.21326]. It also flags 42 of 4,343 noise events as candidate dark-vessel transits after manual review [2606.21326]. Sea-Scan shows that dark-vessel sensing can be achieved with passive subsea acoustics and weak supervision, but it operates in a DAS-specific spatio-temporal setting rather than a general remote-sensing fusion stack [2606.21326].

Second, **xView3-SAR** provides the largest labeled dataset described in the record for dark-vessel detection from SAR imagery, with **991** Sentinel-1 scenes, **243,018** verified maritime object labels, VH and VV bands, and co-located bathymetry and wind rasters [2206.00897]. Its challenge formulation spans maritime object detection, close-to-shore detection, vessel classification, fishing classification, and vessel length estimation [2206.00897]. The benchmark is directly relevant because DarkVesselNet recommends SAR detection on xView3-style chips as part of its evaluation protocol [2606.00445]. xView3-SAR demonstrates the scale, label uncertainty, and tiny-object constraints that any practical SAR-first branch must handle [2206.00897].

Third, the PlanetScope-based dark STS system of [2404.07607] focuses on **dark ship-to-ship transfers** rather than vessel alerts in general. It trains **YOLOv8m** on six classes derived from AIS and imagery and then classifies an STS event as dark when imagery shows a transfer but fewer than two AIS identities are found in the relevant spatial-temporal neighborhood [2404.07607]. It reports **402 dark ship-to-ship transfers** in the Kerch Strait between 2021 and September 2023 [2404.07607]. Relative to that work, DarkVesselNet generalizes from event-specific transfer detection to a broader alerting framework with explicit modality masks, AIS gap reasoning, and audit traces [2606.00445].

Within this landscape, DarkVesselNet’s stated contribution is not a novel single detector but a **system design for auditable dark-vessel alerting** that combines Sentinel-1 SAR and Sentinel-2 optical imagery with AIS trajectory reasoning, wraps multiple geospatial foundation models behind a common token interface, uses TGARD-style gap detection, produces a reviewable alert and reconstruction signal through a Pi-DPM-inspired anomaly head, and preserves availability masks, calibration, and traceability [2606.00445]. Its central scientific claim is therefore architectural: dark-vessel detection should be treated as calibrated, traceable, multi-modal evidence fusion rather than as a standalone image-classification task [2606.00445].

Source: https://www.emergentmind.com/topics/darkvesselnet