Adaptive Coastline Detection based on Deep Q-Network-Guided Contour Refinement

Accurate coastline detection from satellite imagery is crucial for coastal management and navigation safety, yet existing methods suffer from maritime false positives and inconsistent outputs.We propose a novel framework combining Constrained Deep Q-Learning with Precise Sea Area Cleanup that introduces three key innovations: (1) HSV-based dualstream feature extraction for robust water-land discrimination, and (2) context-aware action constraints with curiosity-driven exploration to prevent artifacts.Unlike traditional approaches requiring extensive training data, our single-image reinforcement learning method adapts to new environments without retraining.Experiments on St Andrews (Scotland) imagery achieve F1-score of 0.984, IoU of 0.969, and perfect recall.Validation on temporal and geographical variations confirms strong generalization.The framework's interpretability and operational consistency make it practical for diverse coastal monitoring applications without site-specific calibration.