Deep Learning-Enabled Road Segmentation and Edge-Centerline Extraction from High-Resolution Remote Sensing Images

Abstract

Nowadays, precise and up-to-date maps of road are of great signi¯cance in an extensive series of applications. However, it automatically extracts the road surfaces from high-resolution remote sensed images which will remain as a demanding issue owing to the occlusion of buildings, trees, and intricate backgrounds. In order to address these issues, a robust Gradient Descent Sea Lion Optimization-based U-Net (GDSLO-based U-Net) is developed in this research work for road outward extraction from High Resolution (HR) sensing images. The developed GDSLO algorithm is newly devised by the incorporation of Stochastic Gradient Descent (SGD) and Sea Lion Optimization Algorithm (SLnO) algorithm. Input image is pre-processed and U-Net is employed in road segmentation phase for extracting the road surfaces. Meanwhile, training data of U-Net has to be done by using the GDSLO optimization algorithm. Once road segmentation is done, road edge detection and road centerline detection is performed using Fully Convolutional Network (FCN). However, the developed GDSLO-based U-Net method achieved superior performance by containing the estimation criteria, including precision, recall, and F1-measure through highest rate of 0.887, 0.930, and 0.809, respectively.