Abstract:

To address the challenges of weak feature extraction and sample imbalance in maritime small sea surface targets detection, a cost-sensitive learning-based integrated detection model combining deep belief network （DBN） and extreme gradient boosting （XGBoost） is proposed. The DBN automatically extracts deep-level features from radar echo signals, overcoming the subjectivity of manual feature selection. The XGBoost classifier incorporates a cost-sensitive matrix to alleviate data imbalance between sea clutter dataset and target echoes. Tests on the Ice Multiparameter Imaging X-Band Radar 1993 radar dataset demonstrate that under 0.512 s observation time with 0.001 false alarm rate, the model improves detection rate by 56% compared to the three-feature detector; achieves 0.875 detection rate at observation time 1.024 s observation time, outperforming Hurst index detector, three-feature detector, and fast adaptive clustering gradient boosted decision trees method. Theoretical derivation enables adaptive false alarm rate control mechanism that overcomes the limitations of fixed thresholds. This study verifies the synergistic advantages of deep feature learning and cost-sensitive algorithms, providing an enhanced solution for marine monitoring systems.

Key words: sea surface small target detection, cost-sensitive learning, imbalanced datasets