Mapping and Discriminating Rural Settlements Using Gaofen-2 Images and a Fully Convolutional Network

Area of Science:

• Remote Sensing
• Geographic Information Systems
• Artificial Intelligence

Background:

• Rapid rural construction in China creates complex settlement patterns, mixing old and new developments.
• Accurate mapping of rural settlements is vital for effective land management and decision-making.
• Existing methods struggle with the irregular morphology and distribution of rural settlements, despite advances in High Spatial Resolution (HSR) imagery and deep learning.

Purpose of the Study:

• To develop and evaluate a novel framework for accurate rural settlement mapping using deep learning and HSR satellite imagery.
• To leverage representation learning from deep learning models to overcome challenges in mapping irregular rural settlements.
• To provide a robust method for understanding the spatial characteristics of rural settlements for land management.

Main Methods:

• A novel end-to-end deep learning architecture combining a dilated residual convolutional network (Dilated-ResNet) and a multi-scale context subnetwork.
• Utilizing Gaofen-2 High Spatial Resolution (HSR) satellite images as input data.
• Learning high-resolution feature representations and refining multi-scale features for settlement extraction.

Main Results:

• The proposed framework achieved an overall accuracy of 98% and a Kappa coefficient of 85% in mapping rural settlements in Tongxiang city.
• Demonstrated comparable and improved performance against existing rural settlement mapping methods.
• Successfully mapped and discriminated between new and old rural settlements with high precision.

Conclusions:

• The developed framework offers an effective solution for accurate and timely rural settlement extraction from HSR images.
• The method provides significant benefits for other convolutional neural network (CNN)-based approaches, especially when current ground truth data is unavailable.
• Opens new avenues for obtaining spatial-explicit understanding of rural settlements, supporting informed land management and policy decisions.