Interpretable Patch Screening and Guided Attention U-Net For Floating Marine Debris Detection

Abstract

Floating marine debris is spectrally ambiguous and represents only a small fraction of Sentinel-2 imagery. In this work, we present two complementary debris-aware implementations, which are evaluated with the official partitions of the Marine Debris Archive (MARIDA). First, we propose an interpretable patch-screening baseline that translates eleven-band patches to 54 spectral, index and texture descriptors and compares logistic regression, radial-basis support vector machines and Random Forests. On 359 held-out patches, the Random Forest achieves 0.774 accuracy, 0.585 precision, 0.683 recall, 0.630 F1, 0.821 ROC-AUC and 0.640 average precision. Second, the proposed approach introduces a 350,414-parameter Physics-Guided Spectral-Spatial Attention U-Net for sparse pixel segmentation. It augments the eleven spectral bands with differentiable NDVI, NDWI, Floating Debris Index and Plastic Index channels, learns patch-conditioned spectral gates, applies multiscale convolution and bottleneck self-attention, and is optimised with hard-negative binary cross-entropy, recall-oriented Tversky loss and an edge term. Only on validation data the threshold of 0.75, two-pixel component filter and morphological closing are selected On the locked test set, the Proposed method achieves 0.983 ROC-AUC, 0.713 average precision, 0.201 precision, 0.530 recall, 0.291 F1 and 0.170 IoU.