Document Type : Original Research Paper
Authors
Department of Surveying and Geomatics Engineering, Faculty of Civi Engineering, Shahid Rajaee Teacher Training University, Tehran, Iran
Abstract
Background and Objectives: Roads are known as vital and essential elements in the development of cities, because they play a very important role in communication and transportation and represent the extent of urban development and growth. In order to increase accuracy and efficiency in road detection and classification, researchers have designed and used automatic methods based on deep learning algorithms. These approaches, due to their superior capabilities in recognizing patterns and complex features of images, have effectively replaced traditional methods and have significantly improved the accuracy and speed of road detection.
Methods: In this paper, an improved UNet3+ encoder-decoder model has been used for road detection from remote sensing images. In this proposed model, pyramid pooling and spatial and channel attention modules are used to improve road detection results. The spatial attention module is used in the proposed network architecture to improve the network's focus on important locations in feature maps. The channel attention module also allows the network to more focus on important information and perform better at tasks such as feature detection and classification. The pyramid pooling module is designed to receive multi-scale information. This module helps the network to understand different spatial scales by applying averaging at different levels and then resizing the averaged features to the size of the original feature map.
Findings: The evaluation of the capabilities of the proposed network in detecting secondary roads in areas with less residential density and with soil and vegetation cover shows the superiority of this network over the original version of UNet3+. The improved network proposed in this paper was able to detect roads more accurately. This shows the power of the network in detecting roads in conditions where there is less environmental interference. Quantitative results obtained from this network show the fact that the use of spatial and channel attention modules and pyramid pooling module has been able to increase the accuracy, recall, F1 score and IOU measures by 6, 15.6, 8.3 and 17.4, respectively, compared to the original version of the UNet3+ network.
Conclusion: The challenges raised in the automatic roads detection from remote sensing images, including the effect of shadows and obstruction of the road with buildings and vegetation cover, and the similarity of the secondary roads with the soil background can lead to a decrease in the accuracy of recognizing roads from remote sensing images. The use of improved UNet3+ encoder-decoder architecture capabilities in this research was able to reduce some of these challenges and increase the accuracy of the detection results of secondary roads in areas with soil and vegetation.
Keywords
Main Subjects
COPYRIGHTS
© 2024 The Author(s). This is an open-access article distributed under the terms and conditions of the Creative Attribution-NonCommercial 4.0 International (CC BY-NC 4.0) (https://creativecommons.org/licenses/by-nc/4.0/)
)