Self-supervised disturbing feature reconstruction network for mangrove biomass estimation with limited data

Area of Science:

• Ecology
• Remote Sensing
• Computer Science

Background:

• Mangrove biomass estimation is vital for ecosystem productivity and carbon cycling.
• Existing deep neural network (DNN) methods struggle with data scarcity in remote sensing applications.
• Accurate mangrove biomass data is essential for global carbon cycle monitoring.

Purpose of the Study:

• To develop a novel DNN for mangrove biomass estimation using limited remote sensing data.
• To address the data scarcity challenge in current biomass estimation techniques.
• To enhance the accuracy of mangrove biomass estimation through advanced feature learning.

Main Methods:

• A self-supervised disturbing feature reconstruction network (SSDFRN) was developed.
• Disturbing feature reconstruction-based self-supervised learning (DFRSSL) was employed, utilizing random feature shuffle and reconstruction.
• A multi-view convolutional neural network (MVCNN) with multi-view cascaded convolution modules (MVCCMs) was integrated.

Main Results:

• The proposed SSDFRN effectively addresses the data scarcity problem in mangrove biomass estimation.
• The MVCNN architecture significantly enhances feature learning capabilities.
• Experimental results on Ximen Island data confirm the outperformance of SSDFRN in mangrove biomass estimation.

Conclusions:

• SSDFRN demonstrates effectiveness in deep feature learning for mangrove biomass estimation.
• The developed method provides a viable solution for biomass estimation with limited data.
• This approach contributes to more accurate ecological assessments and carbon cycle monitoring.