Double-Branch Multi-Scale Contextual Network: A Model for Multi-Scale Street Tree Segmentation in High-Resolution...

Area of Science:

• Remote Sensing
• Computer Vision
• Urban Ecology

Background:

• Street trees are crucial for urban green spaces and require efficient management.
• Accurate segmentation of street trees from remote sensing data is challenging due to varying tree sizes.
• Traditional methods often fail to capture all street trees, impacting urban green space management.

Purpose of the Study:

• To develop an advanced deep learning model for precise street tree segmentation from high-resolution remote sensing images.
• To address the limitations of existing methods in handling diverse street tree sizes and complexities.
• To introduce a new benchmark dataset for evaluating street tree segmentation models.

Main Methods:

• Proposed the Double-Branch Multi-Scale Contextual Network (DB-MSC Net) featuring a dual-branch architecture.
• Incorporated a Multi-Scale Contextual (MSC) block with parallel dilated convolutions and transformer blocks for enhanced feature extraction.
• Integrated a Channel Attention Mechanism (CAM) in the decoder to effectively fuse RGB and NDVI features.

Main Results:

• The DB-MSC Net achieved superior performance compared to Unet, HRnet, and SETR.
• Demonstrated significant improvements in segmentation accuracy, with Overall Accuracy (OA) increasing by at least 0.16% and mean Intersection over Union (mIoU) by at least 1.13%.
• The proposed model's segmentation accuracy meets the practical requirements for effective urban green space management.

Conclusions:

• The DB-MSC Net offers a robust solution for accurate street tree segmentation in urban environments.
• The model's ability to handle multi-scale features and fuse complementary data sources (RGB, NDVI) is key to its success.
• This research contributes to advancing remote sensing applications in urban planning and green space management.