Mapping the terraces on the Loess Plateau based on a deep learning-based model at 1.89 m resolution

Area of Science:

• Earth and Environmental Sciences
• Remote Sensing
• Artificial Intelligence

Background:

• Terraces are crucial for soil conservation and agriculture on the Loess Plateau.
• Limited high-resolution (<10m) terrace maps hinder regional research.
• Previous studies focused on specific areas due to data limitations.

Purpose of the Study:

• To develop a high-resolution terrace distribution map for the Loess Plateau.
• To apply a novel deep learning approach for regional terrace extraction.
• To provide a foundational dataset for sustainable land management.

Main Methods:

• Developed a deep learning-based terrace extraction model (DLTEM) using the UNet++ framework.
• Integrated high-resolution satellite imagery, digital elevation models, and GlobeLand30 data.
• Incorporated texture features, topography, vegetation correction, and manual adjustments for a 1.89m resolution map.

Main Results:

• Generated the Terrace Distribution Map of the Loess Plateau (TDMLP) at 1.89m spatial resolution.
• Achieved high classification accuracy: 98.39% with test samples and 96.93% with field validation points.
• Demonstrated the effectiveness of the DLTEM for regional-scale terrace mapping.

Conclusions:

• The TDMLP offers an unprecedented, accurate resource for Loess Plateau terrace research.
• This high-resolution mapping facilitates studies on the economic and ecological value of terraces.
• The findings support sustainable development initiatives on the Loess Plateau.