Deep Learning Image Segmentation for Satellite Images of National Highways of Bangladesh

Roads are the connection network of a country that upholds the economy. Road quality needs to be measured and repaired from time to time to keep this connection stable. The Roads and Highways Division of Bangladesh spends a huge of money with exorbitant time and labor for this purpose. An automation system using satellite images for road quality measurement of Bangladesh highways can be the best solution that would be much more efficient than the existing manual process and may reduce the time and cost significantly. An efficient satellite image Database is crucial for this intelligent system. No significant research has been done in this sector and no suitable dataset is available in this domain yet. Previous work has been done for a database creation that has many limitations and worked for only straight roads. Curve and covered road detection methods were absent [1]. In this paper, a deep learning-based system that extracts and segments the road from satellite images is proposed for the national highways of Bangladesh. It can be used to classify roads afterward into the automation system. Satellite images come with other objects together rather than only roads. It is needed to segment the roads from the satellite images based on the International Roughness Index (IRI) [2] [3]. Google Earth Pro desktop application is used to collect the satellite images at their highest resolution of 8192x4639 and created our dataset according to the Road Roughness Survey Results of Bangladesh [3]. The total dataset is divided into the training and testing set with a ratio of 1:6. To build the proposed system, nine different deep learning models Deeplabv3, DeeplabV3plus, MAnet, LinkNet, Unet, FPN, Unetplusplus, PAN, and PSPNet are used. First, the training dataset is used for manual segmentation, and the augmentation process is performed on the segmented data before training the model. Then the nine deep learning models were experimented with in different train:validation ratios as 60:40, 70:30, and 80:20. The experimental results show that DeeplabV3plus gained better among all the models maintaining more than 97% accuracy in all ratios. In addition, it provides better extraction than other models in visual inspection.