Document Type : Research Paper

Author

• Hussein Ali Ali Al-Iedane

Basrah University For Oil and Gas, Iraq

Abstract

Remote sensing’s biggest challenge is identifying roads in VHR satellite images. This project aimed to develop a model for classifying map roads as roads or non-roads. Hand-sort satellite images. Because human error is so common, separation should be done automatically. Transfer learning and a convolutional neural network built the model. The authors classified pictures using three transfer-learning models. AlexNet, VGG-16, Resnet50, and DenseNet-121. CNN’s module has a 95.3% accuracy rate, VGG-16’s 97.6%, Resnet50’s 96.5%, and Alex’s Net’s 84.6%. Because each spatial
feature point can now refer to additional contextual data, road segmentation is more accurate. In the official Deep Learning Challenge, our unprocessed models outperformed every published state-of-the-art ensemble model. The authors also present empirical evaluations of how to use non-local blocks in the base model. The methodology used transfer learning and a large dataset to build an unrevealed best alternative identification procedure. The architecture uses data augmentation, therefore custom data extraction is unnecessary.

Keywords

• Transfer learning
• Image Satellite
• Convolutional Neural Networks (CNNs)
• Road Extraction