Road Extraction from Unmanned Aerial Vehicle Remote Sensing Images Based on Improved Neural Networks.

Area of Science:

• Remote Sensing
• Computer Vision
• Deep Learning

Background:

• Road extraction is crucial for infrastructure development and is a key area in remote sensing.
• Deep learning methods, particularly neural networks, show promise for accurate road extraction from aerial imagery.
• Existing models like D-LinkNet face challenges with computational efficiency due to large network scales.

Purpose of the Study:

• To improve the computational efficiency and precision of road extraction from Unmanned Aerial Vehicle (UAV) remote sensing images.
• To address the limitations of existing deep learning models, specifically D-LinkNet, in terms of computational performance.
• To develop novel neural network architectures for enhanced road detection.

Main Methods:

• Proposed modifications to the D-LinkNet architecture, including replacing the initial block with a stem block.
• Rebuilt the network using ResNet units to create an improved D-Linknetplus model.
• Introduced 1x1 convolution layers to reduce feature map dimensions and parameters, leading to B-D-LinknetPlus, and validated performance on the Massachusetts Roads Dataset.

Main Results:

• The developed improved neural networks, D-Linknetplus and B-D-LinknetPlus, demonstrated reduced network sizes.
• The enhanced models achieved improved precision in road extraction tasks compared to the original D-LinkNet.
• The modifications successfully addressed the computational inefficiency issues of the base model.

Conclusions:

• The proposed neural network improvements are effective for efficient and precise road extraction from UAV imagery.
• Optimized deep learning architectures can significantly enhance the performance of remote sensing applications.
• Further research into network optimization can lead to more practical and scalable solutions for infrastructure mapping.