Road Features Extraction Using Convolutional Neural Network

Global Positioning System (GPS) provides road connectivity around the globe. It is still challenging to maintain proper road connectivity due to laking of updated, improper, and incomplete road datasets. Extraction of road features from High-Resolution Remote Sensing Images (HRRSI) is challenging due to the complex structure of road, shadows, and occlusion caused by trees, cars, and buildings. Traditional techniques (i.e., SVM, ANN, etc.) provide fragmented and poor road connectivity. A CNN (Convolutional Neural Network) based model is proposed to overcome this issue. In this work, Massachusetts Road Dataset (i.e., publically available) is used to extract the road features from HRRSI. The collected dataset consists of 1178 images, where 1108 images are used for training, 49 for testing, and 14 for validation. The result analysis shows that the proposed model achieved 91.81% accuracy in extracting the complete road network.

I. Introduction

The study of natural resources and the environment increasingly depends on Remote Sensing (RS) data. RS data has numerous applications, including urban planning, crop monitoring, land use monitoring and management, traffic management, natural disaster recovery, defense and intelligence, climate change analysis, deforestation, etc. Different earth observation satellites (like Sentinel, QuickBird, Landsat, Aqua, SWOT, EarthCare, etc.) are used to collect the HRRSI. With the advancement of technology and data storage capacity, a large amount of spatial data is generated daily. However, only a small number of areas are effectively using RS data, as it is quite challenging to process thousands of images and extract relevant information from them. However, Remote Sensing Images (RSI) are extremely useful in many areas, such as urban planning, GPS navigation, automatic vehicle navigation, traffic management, etc. [1].