Satellite radar images of the ocean are almost entirely empty. In the xView3-SAR benchmark, ship pixels make up about 0.005% of each Sentinel-1 scene, and each scene is roughly 29,400 × 24,400 pixels (xView3-SAR). Practical systems handle this with a cascade. A cheap "gate" model looks at each tile and throws away the ones that seem empty. A heavy detector then processes the rest. This saves a lot of compute, but it creates two problems that current work leaves open. First, any ship the gate discards is lost for good, and nobody controls how often that happens. Second, the detector is trained on a different set of tiles than it sees at deployment, because at test time it only ever sees the tiles the gate lets through.
We propose to build the cascade so that the two stages cover for each other, and to choose the gate's discard threshold so the fraction of vessels it loses is bounded ahead of time on held-out scenes. The gate looks at a coarse, wide-context view that includes bathymetry, and it is trained to reject the kind of clutter that fools the detector. The detector is trained only on the tiles that survive the gate. Its training negatives are therefore the hard backgrounds, such as harbours, rough seas and rocky coasts, that it will actually face, not uniformly sampled empty ocean. The claim is that this combination beats both a single detector run over every tile and an independently trained gate. It should win on detection F1, on nearshore false alarms and on runtime. It should also keep its promised small-vessel recall across different sea states and coastlines.
What the prior work does and where this departs
Three SAR systems already use a gate in front of a detector, and all three motivate this proposal. Van Meerbeeck et al. train a MobileNetV2 tile classifier separately from a Faster R-CNN detector (Optimizing ship detection efficiency). Their gate keeps 99.5% of baseline AP in 44% of the time. Pushing it to a 4× speedup drops relative AP to 92.7%. The detector never adapts to the gate, and the threshold is a hand-picked operating point with no statement of how many ships it loses. Xu et al. use a K-distribution CFAR (constant false alarm rate, a classical detector that flags pixels much brighter than a locally estimated clutter statistic) in front of YOLOv4-tiny (HISEA-1 on-board detection). Their system still has 14.12% false alarms and 11.42% missed detections, concentrated around ports. LS-SSDD-v1.0 shows that mixing pure-background tiles into detector training cuts Faster R-CNN false-alarm probability from 76.27% to 26.26% (LS-SSDD-v1.0). The file also reports that this mixing lowers detection probability. Because those background tiles are sampled uniformly, most of them are easy open water. Our gate-matched training replaces uniform sampling with the specific backgrounds that survive the gate.
In medical imaging, GF-Screen trains a "glance" model with reinforcement learning (Group Relative Policy Optimization), rewarding it by how well a downstream segmenter performs on the patches it keeps (GF-Screen). The discarding improved accuracy, not just speed: +25.6% DSC with 83% of background discarded. That result supports our main hypothesis that dropping distracting background can raise accuracy. We deliberately do not copy its reinforcement-learning machinery. In ship search the quantity the reward approximates, "does this tile contain a vessel the detector would find", has a direct label. A supervised gate that is weighted by the detector's own errors is simpler, and we test it head-to-head against a GRPO gate.
The gradient-aware losses of Wu et al. constrain part of the design (GA Loss). They show that patches containing no foreground give standard losses near-zero or erratic gradients, which lets small classes collapse. Our detector will see many vessel-free tiles, so we include their Top-K hard mining as a competing detector-side remedy, not as part of our method.
Finally, the sea–land modules of Ke et al. (Scene-aware SAR detection) and the context-maximization result of Ma and Nie (Large context for aneurysm segmentation) explain why the gate should be the wide-context stage. A wide view resolves "is this coastline or open sea". The detector can then keep full resolution for single-pixel vessels. This addresses the tension noted after xView3: large context helps with scene ambiguity but blurs out small targets.
Core Claim
The claim has two parts. First, calibrating the gate threshold on held-out scenes bounds vessel loss per scene at a chosen level. Second, training the detector on gate-surviving tiles makes the cascade more accurate, not only faster, than the same detector run over every tile. The accuracy gain should show up mainly as fewer nearshore false alarms, with no loss in small-vessel recall beyond the promised bound.
Research Program
The gate. The gate is a small CNN applied to a 4× downsampled view of each tile. Its input channels are the VH and VV radar polarizations, bathymetry and wind speed, the ancillary rasters that xView3 shows help separate rocks from ships. Its receptive field extends into neighbouring tiles so it can see the coastline. It is trained as a binary "contains a vessel" classifier. Each background tile's loss is weighted by the false-alarm score a reference detector, trained on all tiles, produces in that tile. As a result, the gate spends its capacity rejecting exactly the clutter the detector confuses with ships, and its errors become complementary to the detector's.
The threshold. Let $s(t)$ be the gate score of tile $t$, and let $L(\tau)$ be the fraction of annotated vessels in a scene that lie in tiles with $s(t) < \tau$, meaning tiles that would be discarded. On a set of calibration scenes separate from both the training and test scenes, we choose the most aggressive threshold that satisfies
$$\hat\tau = \max\Big\{\tau : \tfrac{1}{n+1}\Big(\sum_{i=1}^{n} L_i(\tau) + 1\Big) \le \alpha\Big\},$$
where $n$ is the number of calibration scenes, $L_i$ is the loss on scene $i$, and $\alpha$ is the target vessel-loss rate. This is the standard finite-sample correction for choosing a threshold on a monotone loss that is bounded by one. It yields an average-case guarantee if test scenes are exchangeable with calibration scenes.
Scenes are not exchangeable across sea states, sensors and coastlines, so we also calibrate separately within strata: offshore versus nearshore (from distance-to-shore), and length bins (under 20 m versus larger, using xView3 length labels). We report whether the per-stratum bound holds on test scenes.
The detector. The detector is a full-resolution anchor-free detector with a fine P3 pyramid level, following the small-ship motivation of Quad-FPN (Quad-FPN). It is trained with a sampler that draws negatives only from tiles the gate passes at threshold $\hat\tau$. This is iterated once: gate, then detector, then refit the gate weights using the new detector's false alarms, then recalibrate.
Baselines
The naive reference is the same detector run over every tile, trained with uniform pure-background mixing as in LS-SSDD's PBHT. It isolates what the cascade adds. On LS-SSDD, the reported points of comparison are Faster R-CNN's false-alarm probability falling from 76.27% to 26.26% with PBHT, and Double-Head R-CNN's best entire-scene mAP of 75.74% (LS-SSDD-v1.0).
The strongest cascade competitor is the decoupled MobileNetV2 gatekeeper of Van Meerbeeck et al. It is the closest published system and the one that exposes the speed–recall trade-off. Their reported numbers on the 900-tile LS-SSDD subset are a Faster R-CNN baseline AP of 0.711 in 811 seconds, 99.5% of that AP at 44% of the time, and 92.7% at 25% of the time. We reimplement it on our splits with the same detector as ours, so the only difference is gate training and threshold choice. We compare full AP-versus-time curves. Beating it means a higher AP at matched time across the curve. It also means meeting the stated vessel-loss bound where their threshold choice gives none.
The strongest full-benchmark reference is the xView3 winning system: aggregate score 0.6177 and detection F1 0.7702, versus 0.1904 for the Faster R-CNN baseline (xView3-SAR). We cannot claim to reproduce that system. Instead we report our detector-only and cascade numbers with an identical backbone, and judge success by the gain of the cascade over its own detector-only version. Matching or exceeding 0.7702 detection F1 would be a bonus, not the criterion. All methods get the same training data, the same number of GPU-hours, the same hyperparameter-search budget, and inference within xView3's 15-minute single-GPU limit.
Three ablations isolate the ingredients. First, gate weighting is turned off, so the gate is a plain classifier. Second, gate-matched negative sampling is turned off, so the detector is trained on uniform PBHT. Third, the calibrated threshold is replaced by a threshold tuned on the test-like validation set, which separates the bound from the operating point.
Why not just…?
Why not just sweep the gate threshold and pick a good operating point? This is a real threat. A plain sweep traces the same curve that calibration selects from, so on a single in-distribution test set it may match our accuracy exactly. The calibration step adds a different thing: a stated vessel-loss level that holds on new scenes and, after stratifying, on nearshore and small-vessel subsets. The decisive check is whether the vessel-loss level chosen by sweeping on validation scenes is violated on held-out scenes more often than the calibrated one. We also check whether the per-stratum bound holds when the naive global bound fails for small vessels. If neither happens, we will report the calibration as no better than a sweep and present the method's value as gate-matched training alone.
Why not just fix the detector's loss with hard negative mining or gradient-aware losses, without a gate? Top-K hard mining and the gradient-aware cross-entropy of Wu et al. (GA Loss) address the dilution of training signal by empty tiles. They could recover much of the detection-probability loss that LS-SSDD attributes to PBHT. They do nothing for inference cost, and they do not give the detector wide coastline context. We run this alternative as a baseline at equal compute. If it matches our nearshore false-alarm rate, the accuracy claim falls to the loss function and the cascade's contribution is speed alone. We would report it that way.
Why not just use CFAR as the gate? CFAR is cheap and needs no training. However, sliding-window versions mis-set the threshold when nearby ships enter the clutter-estimation window and are slow on large scenes: 1,669 s versus 47.56 s for the superpixel variant on TerraSAR-X (SP-NW-CFAR). As a front end it still left 14.12% false alarms in harbours (HISEA-1). We include superpixel CFAR as a gate inside our same calibration procedure. This tests whether the learned, context-aware gate is what matters or whether any calibrated filter would do.
Why not just train the gate with reinforcement learning, as GF-Screen does? Reinforcement learning optimizes a downstream reward directly, but it adds variance and many tuning choices. Our supervised gate with detector-weighted negatives targets the same signal. We train a GRPO gate with the detector's per-tile F1 as reward, using the same compute. If it clearly wins, the supervised shortcut is insufficient and we adopt it.
Proposed Experiments
The primary benchmark is xView3-SAR, split by scene into training, calibration and test sets. We report aggregate score, detection F1, nearshore and offshore F1 separately, recall for vessels under 20 m, the fraction of tiles discarded, and full-scene wall time. The secondary benchmark is LS-SSDD-v1.0 entire-scene evaluation with AP, detection probability and false-alarm probability, which gives direct comparison to PBHT and to Van Meerbeeck et al.
The first decisive check is whether the cascade's detection F1 exceeds the detector-only F1 while discarding most tiles. GF-Screen's medical result predicts a gain, mainly from nearshore false alarms. A null result would mean that discarding background does not help SAR accuracy. The second check is whether realized vessel loss on test scenes stays at or below $\alpha$ for α between 0.5% and 5%, both globally and per stratum. The third check is a cross-dataset transfer: a gate calibrated on xView3 is tested on LS-SSDD scenes to measure how much the bound degrades under sensor and region shift, and whether recalibrating on a few target scenes restores it. The fourth check is the error-overlap analysis: the fraction of detector false alarms that the gate removes, before and after gate weighting. This directly tests whether the two stages' errors have become complementary.
References
- xView3-SAR: Detecting Dark Fishing Activity Using Synthetic Aperture Radar Imagery
- Optimizing ship detection efficiency in SAR images
- On-Board Real-Time Ship Detection in HISEA-1 SAR Images Based on CFAR and Lightweight Deep Learning
- LS-SSDD-v1.0: A Deep Learning Dataset Dedicated to Small Ship Detection from Large-Scale Sentinel-1 SAR Images
- Glance and Focus Reinforcement for Pan-cancer Screening
- Gradient-Aware for Class-Imbalanced Semi-supervised Medical Image Segmentation
- Scene-aware SAR ship detection guided by unsupervised sea-land segmentation
- Exploring Large Context for Cerebral Aneurysm Segmentation
- Quad-FPN: A Novel Quad Feature Pyramid Network for SAR Ship Detection
- Fast Superpixel-Based Non-Window CFAR Ship Detector for SAR Imagery