Automatic extraction of surface water based on lightweight convolutional neural network

Area of Science:

• Remote Sensing
• Hydrology
• Artificial Intelligence

Background:

• Accurate surface water extraction is crucial for global water cycle studies and water resource management.
• Existing methods using multi-spectral remote sensing struggle with shadows, leading to inaccuracies and manual adjustments.
• Shadows in images have spectral similarities to water, complicating traditional water extraction techniques.

Purpose of the Study:

• To develop a fast, large-area, and automatic method for surface water extraction.
• To improve water extraction accuracy in complex scenarios affected by shadows.
• To overcome limitations of traditional index-based water extraction methods.

Main Methods:

• Introduced thermal infrared band data for pre-treatment.
• Proposed a lightweight neural network (EDCM) combining image classification and semantic segmentation models.
• Employed multi-scale training of samples using lightweight convolutional networks to capture multi-scale context information.

Main Results:

• The EDCM model achieved the highest accuracy in three highly heterogeneous test scenarios.
• Achieved an accuracy exceeding 95.28% in all selected test areas.
• Demonstrated superior performance compared to existing methods in complex environments.

Conclusions:

• The EDCM model offers a robust solution for high-precision surface water extraction in challenging areas.
• The integration of thermal infrared data and a lightweight neural network enhances automatic water mapping capabilities.
• The proposed method addresses the limitations of traditional approaches, enabling efficient and accurate remote sensing monitoring of surface water.