DANet-SMIW: An Improved Model for Island Waterline Segmentation Based on DANet

Abstract

Segmentation of island waterline contributes to shoreline movement analyzing, environmental monitoring, and integrated coastal zone management. To achieve high efficiency and high accuracy of island waterline segmentation in remote sensing images, we proposed a model for island waterline segmentation based on DANet (DANet-SMIW). In DANet-SMIW model, different indexes (Normalized Difference Water Index and OTSU) were taken as new channels added to input dataset, which enhanced waterline's spectral information. DANet backbone network was improved by dense connection of DenseNet. Loss function, consisting of binary cross entropy loss and Dice loss, was used to resolve the sample imbalance problem, and then, rough results of island waterline segmentation were refined by boundary refinement module (BRM). In total, 2042 island images were taken as experiment dataset, which were cropped from Landsat-8 images and divided into 1634 images for training, 100 images for testing, and 308 images for validation, and DANet-SMIW model was compared against other models, including FCN-32s, DeepLabv3+, PSPNet, Dense-ASPP, PSANet, ICNet, DuNet, and PIDNet. Results demonstrated that DANet-SMIW model achieved the highest values with pixel accuracy and Mean Intersection over Union and possessed higher segmentation efficiency than most other models. Collectively, DANet-SMIW model was an integrated accurate and efficient model for island waterline segmentation in remote sensing images.

Keywords

• DANet
• deep learning
• island waterline
• semantic segmentation