From Data Uncertainty to Spectral Consistency: Curriculum-Derived Semantic Understanding in GPS-Denied Maritime Environments

Maritime vision systems—whether deployed on unmanned surface vessels, aerial drones,or coastal surveillance cameras, rely on accurate extraction of three sparse semanticboundaries—the Visible Horizon Line (VHL), True Horizon Line (THL), and Shoreline(SL)—to stabilise cameras, geo-register imagery to digital elevation models, and providelast-resort localisation when GPS is jammed or denied. Yet these contours are challengingto detect in scenes dominated by vast, low-texture sky-and-water regions, specularreflections, wave clutter, haze, and rapidly changing weather; fewer than 1% of pixelsbelong to the true boundaries, and horizon cues often blur into clouds or disappearbehind vessel wakes. Classical Canny–Hough or super-pixel heuristics collapse undersuch conditions, while generic deep networks demand dense, costly annotation. To overcomethese obstacles, this thesis introduces a unified, data-efficient framework—builtfrom four complementary advances—that achieves state-of-the-art semantic-boundarydetection and multi-boundary estimation for robust, GPS-denied maritime navigation.The first work builds a framework on top of a streamlined, low-latency UNet-like architecturethat ingests RGB frames alongside inertial measurement unit (IMU) streams,enabling cross-modal feature alignment. Evaluated on multiple benchmarks, our networkachieves highly accurate predictions for the key semantic boundaries while sustainingreal-time performance. The work shows a first step towards extracting multiplesemantic boundaries from a single image using a simple framework. To the best of ourknowledge, this is the first work to attempt to extract all three semantic boundaries simultaneouslyand within real-time constraints. The extraction of multiple boundaries,or boundaries in general, applies to many autonomous tasks, such as those involvingUAVs, USVs, and land-based vehicles. These domains require the detection of surfaceedges and horizons. Enhancing contour detection will improve downstream tasks,including safer navigation in complex environments, obstacle detection, and collisionavoidance.Second, recognising that even modest label noise severely degrades thin-boundary prediction,we introduce a novel lightweight Edge Prior Module (EPM) that guides earlylayers to suppress water-reflection artefacts and sky texture. The proposed frameworkpredicts multiple high-quality semantic boundaries in complex scenarios. The maritimeenvironment presents numerous obstacles and photometric distortions, making the taskparticularly challenging when using classical methods. We train a robust and deployablemodel that overcomes numerous challenges, such as varying lighting conditions,obfuscations, and reflections from water surfaces.Third, to address the semantically pixel class imbalance and accelerate convergence, wedevise curriculum learning with Fourier Spectral Alignment (FSA). The FSA loss alignsthe predictions and ground truth, forcing the model to capture both low-frequencyglobal shape and high-frequency edge detail. Our method and mechanisms demonstratea significant improvement over the current state of the art in semantic segmentationon the LaRS dataset. The results showcase the detection capabilities of the approach,setting a new state-of-the-art.Fourth, to curb data and annotation costs, we introduce policy-driven curriculum distillation,where image selection is framed as a Markov decision process solved by aDeep Q-Network (DQN). At each epoch, the agent evaluates image embeddings, predicteddifficulty, and past validation gains to assign a utility score, admitting onlythe most informative samples. The saliency-informed curation compresses the trainingset while preserving key performance metric scores compared to full-data training,reducing annotation effort and GPU compute. Because the policy re-scores incomingdata in real-time, it can continuously adapt the distilled subset to evolving operationalscenarios while retaining the model’s predictive strength.Collectively, these four contributions—multi-contour semantic segmentation, edge-priorconditioning, spectral curriculum learning, and reinforcement learning-guided datasetdistillation—compose a principled pipeline for maritime boundary detection. The resultingsystem establishes new benchmarks across ReMaSTrED300, MaSTr1325, andLaRS, achieves robust, real-time performance in GPS-denied conditions, and reducesdata requirements by more than half. Beyond maritime navigation, the proposedtechniques generalise to any sparse-label, edge-centric vision task, offering a scalableblueprint for boundary-aware perception in autonomous robotic navigation.