STIRUnet: SwinTransformer and inverted residual convolution embedding in unet for Sea-Land segmentation

Area of Science:

• Remote Sensing
• Geospatial Analysis
• Deep Learning

Background:

• Accurate coastline extraction is crucial for coastal management, erosion monitoring, and ocean construction.
• Complex nearshore environments with tidal flats, suspended sediments, and biological communities challenge traditional methods and reduce segmentation accuracy.
• Existing methods using Convolutional Neural Networks (CNNs) struggle with pixel-level refinement and comprehensive image understanding.

Purpose of the Study:

• To develop an advanced deep learning model for precise coastline extraction from optical remote sensing images.
• To overcome limitations of existing methods in handling complex nearshore environments and small-scale features.
• To improve sea-land segmentation accuracy in medium-high-resolution imagery.

Main Methods:

• Proposed a novel U-shaped deep learning model named STIRUnet.
• Integrated SwinTransformer for global modeling capabilities with an improved CNN utilizing an inverted residual module.
• Employed global supervised feature learning and layer-by-layer feature extraction.

Main Results:

• Identified suspended sediments and coastal biological communities as key factors causing coastline blurring.
• Demonstrated that recovering minute features like narrow watercourses and microscale structures significantly enhances edge details.
• Achieved more realistic sea-land segmentation outcomes on GF-HNCD and BSD remote sensing datasets.

Conclusions:

• The STIRUnet model offers significant improvements for accurate sea-land information extraction in complex marine environments.
• Findings provide novel insights into mixed-pixel identification and enhance the understanding of factors affecting coastline delineation.
• The study highlights the importance of detailed feature recovery for realistic segmentation in remote sensing applications.