Research on SAR Image Marine Oil Spill Detection under Frequency-Domain Decoupling and Boundary Constraints

Abstract

To address the theoretical challenges in Synthetic Aperture Radar (SAR) marine oil spill monitoring, where coherent speckle
noise is highly coupled with weak oil film textures in the spatial domain and geometric boundaries are blurred due to continuous down
sampling, traditional deep learning models based on spatial convolution often struggle to effectively suppress multiplicative noise while
enhancing features. Consequently, this paper proposes a theoretical deep learning framework integrating frequency domain decoupling
mechanisms with explicit boundary constraints (FDB-Net). Transcending the limitations of feature aggregation solely in pixel space, this
method innovatively constructs a Frequency Domain Decoupling (FDD) module. By leveraging the global perception and orthogonal
sparsity of the Fourier Transform, a learnable spectral filter is designed to dynamically separate high-frequency speckle noise from non
stationary oil film signals, theoretically circumventing the "noise amplification" effect inherent in spatial convolution. Furthermore, target
ing the inherent defect of topological structure loss in encoder-decoder architectures, an explicit geometric supervision branch based on
high-frequency residual information is introduced. This branch feeds the predicted edge probability map back to the main segmentation
network as a gating signal, achieving refined shaping of oil spill contours. Coupled with a joint optimization strategy of spectral consisten
cy and spatial semantic accuracy, FDB-Net establishes a feature representation system characterized by "Spatial-Frequency Synergy" and
"Region-Boundary Complementarity, " providing a highly interpretable structured solution for the refined interpretation of weak marine
ecological elements under complex sea states.