A hybrid deep learning approach for accurate water body segmentation in satellite imagery

Abstract

Abstract Precise water body segmentation in satellite imagery plays a vital role in environmental monitoring, water resource management, and disaster prevention. This study introduces a high-performance segmentation framework leveraging Sentinel-2 imagery, integrating advanced methodologies to enhance data integrity and segmentation accuracy. To increase dataset diversity and model generalization, StyleGAN3-based augmentation was implemented, yielding a 5% accuracy improvement over conventional methods. An Attention-Guided Denoising Autoencoder with Skip Connection (AG-DAES) was utilized for noise reduction, effectively preserving spatial details and strengthening segmentation robustness. To address missing and corrupted pixels, Bi-ConvRNN was employed for pixel restoration, significantly boosting performance. Additionally, the Particle Swarm Dandelion Optimization (PSDO) algorithm was used for hyperparameter tuning, contributing to an additional 3–5% accuracy gain. Feature extraction was refined through the Multiscale Strip Convolution Module (MSSCM), enhancing spatial-spectral representation and leading to a 6–8% accuracy increase. The segmentation process was executed using the Map U-Net model, which, after integrating all proposed improvements, achieved state-of-the-art accuracy exceeding 99%. A comparative study demonstrated that the proposed framework outperforms existing methods, particularly in complex scenarios involving vegetation interference, occlusions, and mixed land–water transitions. This adaptable and scalable approach sets a new standard for water body segmentation in satellite image analysis, offering a powerful tool for future research in the field.