Develop an Automatic Road Network Extraction System from Remote Sensing Images

Abstract

In recent years, both the development of high resolution satellite images and the amount of newlineavailable aerial images have expanded and accessible easily. Unfortunately, the technologies newlineused to analyze all of these images have not kept pace, so a large portion of the job is still newlinedone manually by humans, which is costly, time-consuming, and error-prone. Due to these newlinefactors, there is a strong need for efficient and dependable techniques that can automatically newlineanalyze remote sensing images. However, even in high-resolution remote sensing images, newlinebackground and roadways might be difficult to distinguish from complex background images newlinebecause of the occlusion of trees and buildings. Now a day, Deep Learning, which has the newlinecomputing power for massive data, has emerged as the most popular and effective newlineclassification approach. newlineTwo different novel methods are developed for automatic road surface detection from high- newlineresolution remote sensing images based on deep learning that detect effectively, efficiently newlineand fast in manner. First, a modified U-Net is used to construct a semantic segmentation newlinealgorithm for road surface extraction. The modified U-Net has fewer convolution layers than newlinethe normal U-Net. The intersection over union (IOU) yielded a model performance of newline93.71%, while the average segmentation time for a single image was 0.28 seconds. newlineThe second proposed approach, Gradient Descent Sea Lion Optimization (GDSLO) fusion of newlineSea Lion Optimization (SLnO) and Stochastic Gradient Descent (SGD) algorithms. The newlinemodel performance for road surface is measured by evaluation metrics, such as precision, newlinerecall, and F1-measure with the highest values of 0.888, 0.930, and 0.810, respectively. The newlineroad edge performance detected with precision, recall and F1-score are 0.801, 0.76, and 0.786 newlinerespectively. In the same way precision, recall, and F1-score for centerline detection is 0.800, newline0.762, and 0.7999 respectively.