SATELLITE IMAGE PROCESSING TECHNIQUES FOR MONITORING COASTAL EROSION AND CLIMATE CHANGE IMPACT

Abstract

Coastal erosion and climate change are increasingly altering shoreline environments, creating risks for ecosystems, infrastructure, and coastal communities. Satellite remote sensing and deep learning offer promising approaches for improving large-scale and repeatable coastal monitoring. This study aimed to develop a deep learning–based semantic segmentation framework for coastal satellite imagery to classify major coastal features and support more efficient shoreline monitoring. A supervised semantic segmentation approach was implemented using a U-Net model. The dataset consisted of labeled coastal satellite images with four classes: water, surf, sediment, and other. RGB inputs were resized, normalized, and divided into training, validation, and test sets. The model was trained using categorical cross-entropy with class weights and evaluated using pixel accuracy, Intersection over Union (IoU), Dice coefficient, confusion matrix analysis, and qualitative visual assessment. The model achieved a pixel accuracy of 0.8007, a mean IoU of 0.6006, and a mean Dice coefficient of 0.7449. The highest class-wise performance was observed for water and other, with IoU values of 0.7059 and 0.7062 and Dice coefficients exceeding 0.82. Lower performance for surf and sediment reflects the complexity of transitional shoreline zones and spectral overlap between adjacent classes. Qualitative assessment indicated that the model preserved overall shoreline morphology, although some boundary-level errors remained. Deep learning based semantic segmentation demonstrates reliable performance for coastal feature classification and shows potential for scalable shoreline monitoring. Further improvements may be achieved through the incorporation of multi-spectral inputs, expansion of datasets, and the development of boundary-sensitive model designs.