An Optimized Algorithm For Road Extraction With Improved Connectivity And Completeness In Remote Sensing Imagery

Extracting road networks from satellite images is a promising method for effective and timely updating of dynamic changes in road networks.However,pixel-based methods often generate fragmented roads and cannot predict topological correctness due to occlusion by other objects and complex traffic environments,making it a challenge.This paper aims to propose a new road network extraction method to improve connectivity and integrity of the road network.We propose an encoder-decoder structure that combines multi-scale contextual information and introduces a strip convolution kernel to capture road shape and features.We also combine segmentation and graph-based methods and use a multi-start point tracking algorithm to address road discontinuity issues.Firstly,we use the encoder-decoder structure combined with ResNet and ASPP modules to extract roads.By utilizing ResNet and ASPP modules at different scales to capture more contextual information,our method can more accurately identify road areas.Secondly,we propose a technique called "strip convolution kernel"for road feature extraction.This technique considers the shape and features of the road,and can better capture road boundaries and details.Compared to traditional square convolution kernels,strip convolution kernels can better adapt to the long and narrow characteristics of roads and prevent irrelevant areas from interfering with feature learning.Our results demonstrate that this method can effectively extract road features and provide strong support for road extraction tasks.Finally,we combine segmentation-based and graph-based methods and use a multi-start point tracking algorithm to address road discontinuity issues.The advantages of segmentation-based and post-processing methods are good integrity and alignment,while the disadvantages are poor connectivity.On the other hand,graph-based methods have good connectivity but poor integrity and alignment.One of the key points of this research is to highlight the advantages of combining segmentation-based and graph-based methods to improve the accuracy and integrity of road extraction.Experimental results on the SpaceNet and DeepGlobe datasets demonstrate that our proposed method performs well in road extraction,surpassing some SOTA methods.Specifically,our method achieves good performance on F1-score,IoU,and APLS metrics.These results demonstrate the superiority of our proposed method.