GC-UNet: An Improved UNet Model for Mangrove Segmentation Using Landsat8

Accurate and effective recognition of mangrove on the remote sensing images is necessary in protecting the mangrove ecosystem. Nowadays, many methods have been developed for remote sensing image segmentation. They can be classified as methods based on geographic and methods based on machine learning. However, these methods may be too time-consuming or not accurate enough in some cases. Due to these factors, we propose GC-UNet, a novel method based on UNet, which embeds Global Context Block to get long-range dependency, uses SPConv to focus on intrinsic information and adopts Adaptively Spatial Feature Fusion to process features from different levels. It greatly strengthens the presentation ability of the model. The experimental results demonstrate that our proposed model obviously enhances the performance of the image segmentation. We compare our method with DeepLabV3 and OCNet. Our method achieves a dice coefficient of 0.8604, exceeding the above popular semantic segmentation models.