ASCEND-UNet: An Improved UNet Configuration Optimized for Rural Settlements Mapping

Area of Science:

• Remote Sensing
• Geographic Information Systems
• Computer Vision

Background:

• Rural revitalization strategies in China have led to complex settlement patterns.
• Accurate differentiation of rural settlement types is crucial for land planning and environmental improvement.
• Existing methods lack automation for rural settlement differentiation from remote sensing data.

Purpose of the Study:

• To develop an automated method for segmenting and classifying dispersed and clustered rural settlement buildings.
• To improve the accuracy of rural settlement differentiation using high-resolution satellite imagery.
• To introduce an improved encoder-decoder network, ASCEND-UNet, for this task.

Main Methods:

• Designed an improved encoder-decoder network, ASCEND-UNet, based on the UNet architecture.
• Incorporated atrous spatial pyramid pooling (ASPP) for multi-scale feature fusion in the encoder.
• Integrated spatial and channel squeeze and excitation (scSE) block at skip connections and hybrid dilated convolution (HDC) block in the decoder.

Main Results:

• ASCEND-UNet demonstrated superior performance compared to UNet, PSPNet, DeepLabV3+, and SegNet.
• Achieved significant improvements in precision (4.67%), recall (2.80%), F1-score (3.73%), and mean intersection over union (MIoU) (6.28%) over the original UNet.
• Ablation experiments confirmed the effectiveness of the ASPP, scSE, and HDC modules.

Conclusions:

• The proposed ASCEND-UNet model provides a more accurate and stable approach for semantic segmentation of rural settlements.
• This novel model enhances automated methods for analyzing diverse rural settlement patterns from remote sensing images.
• The findings support improved rural land planning and living environment assessments.