A Convolutional Neural Network with Spatial Location Integration for Nearshore Water Depth Inversion

Area of Science:

• Remote Sensing and Geospatial Analysis
• Coastal Engineering and Management
• Machine Learning and Deep Learning Applications

Background:

• Accurate nearshore water depth is critical for navigation, coastal protection, and disaster mitigation.
• Existing remote sensing methods struggle with insufficient feature extraction for precise nearshore water depth inversion.
• Integrating spatial information can enhance the performance of deep learning models in geospatial tasks.

Purpose of the Study:

• To develop an improved method for nearshore water depth inversion using remote sensing data.
• To address the limitations of feature extraction in current nearshore bathymetry techniques.
• To evaluate the effectiveness of a novel convolutional neural network incorporating spatial location information.

Main Methods:

• Proposed a convolutional neural network with spatial location integration (CNN-SLI).
• Integrated pixel spatial location as additional channels into the input data for the CNN.
• Utilized GF-6 remote sensing imagery and electronic nautical chart data for experiments near Nanshan Port.

Main Results:

• CNN-SLI achieved superior nearshore water depth inversion accuracy with RMSE of 1.34 m, MAE of 0.94 m, and R² of 0.97.
• The model demonstrated consistent performance across shallow and deep waters, outperforming Lyzenga, MLP, and CNN models.
• CNN-SLI showed better generalization ability on independent datasets compared to traditional and other deep learning models.

Conclusions:

• The CNN-SLI model significantly enhances nearshore water depth inversion accuracy by effectively utilizing spatial information.
• Integrating spatial location data into deep learning architectures is crucial for improving remote sensing-based bathymetry.
• The proposed CNN-SLI method offers a robust and accurate solution for coastal zone management and scientific research.