A Cross Entropy Based Deep Neural Network Model for Road Extraction from Satellite Images.

Area of Science:

• Remote Sensing
• Computer Vision
• Deep Learning

Background:

• Accurate road network extraction from satellite imagery is crucial for various applications.
• Existing deep learning models often face challenges in balancing extraction precision and computational efficiency.

Purpose of the Study:

• To propose an efficient and accurate deep convolutional neural network model for road network extraction from satellite images.
• To improve road boundary smoothness and clarity.

Main Methods:

• Developed an encoder-decoder deep convolutional neural network model incorporating ResNet-18 and Atrous Spatial Pyramid Pooling.
• Utilized a modified cross-entropy loss function and the PointRend algorithm for training and boundary refinement.
• Employed the augmented DeepGlobe dataset and asynchronous training for model development.

Main Results:

• The proposed E-Road model demonstrates a reduced parameter count and shorter training time.
• Achieved significant performance improvements, ranging from 5.84% to 59.09%, over state-of-the-art deep models.
• The model accurately predicts road networks in complex environments.

Conclusions:

• The E-Road model offers an effective solution for precise and efficient road network extraction from satellite images.
• The model's performance indicates its potential for real-world geospatial applications.