Road Extraction From Remote Sensing Images Using U-Net

The enhancement of the transportation database through updated road networks is a critical aspect with a wide range of applications. Traditional methods rely on onsite measurements or human interpretation of remote sensing images for mapping purposes. However, these methods have significant drawbacks, including high expenses, lengthy processing times, and a substantial workforce requirement. To overcome this challenge, we propose a method leveraging the U-Net architecture for road extraction from multispectral satellite images. The satellite imagery used is sampled from the DigitalGlobe Images dataset, covering images captured over Thailand, Indonesia, and India, with a ground resolution of 50 cm/pixel. The U-Net model is trained on both original and augmented data, employing convolutional layers, max-pooling, and up-sampling to extract and reconstruct road masks. Performance evaluation is conducted using custom validation sets, measuring metrics such as IOU and loss. Results demonstrate the effectiveness of the proposed system, achieving high accuracy and validation scores. The extracted road networks contribute to enhancing transportation databases, offering valuable insights for urban planning and management.