GeoStatNet: Geometric-Statistical Feature Fusion for Robust Ship Detection in Degraded Maritime Environments

Abstract

Arbitrary-oriented ship detection in optical remote sensing images is a critical task for maritime defense, yet perfor-mance often degrades under adverse environmental conditions. Although modern deep learning methods perform well in clear conditions, they struggle in the presence of sea fog, heavy haze, or low illumination, where the signal-to-noise ratio decreases, and the ship signature becomes indistinguishable from background clutter. To address this, we propose GeoStatNet (Geometric-Statistical Network), a novel architecture that integrates differen-tial geometry and second-order statistical modeling directly into the feature pyramidal hierarchy. Unlike conventional methods that simply fuse multi-scale features, GeoStatNet explicitly mod-els the physical properties of the target. We introduce a unified Geometric-Statistical FPN comprised of three mathematically grounded components: the Local Coherence Excitation (LCE), which leverages the principal gradient directions of the structure tensor to isolate directional ship features from directionless atmospheric noise; the Differential Shape Prior (DSP), which utilizes the Determinant of the Hessian to inject illumination-tolerant geometric-statistical convexity cues; and the Global Multivariate Attention (GMA), which employs second-order au-tocorrelation pooling to model and suppress repetitive sea surface textures. Extensive experiments on the SCCOS and HRSC2016 datasets demonstrate that GeoStatNet achieves state-of-the-art performance and exhibits superior stability in degraded defense environments.