Deep convolutional neural network (DCNN) for road network extraction from fusion of airborne laser scanning (ALS) data and highest resolution image in the urban environment

Abstract

Brazil is a big country that has a challenging extension area and millions of kilometers of roads distributed irregularly. The road information is crucial for government planners and the economy (e.g., logistics). Since the 70s, several authors have created methods for automatic road extraction. All have advantages and disadvantages. However, the flexibility requirements with big data and handcraft processes affect the accuracy and quickness of extraction. After 2000 years, new machine learning approaches arise in many fields of knowledge. Among them is Remote Sensing. These tools permit to extract information of extensive data from multiple sources. Since the year 2014, deep neural convolutional networks have obtained exciting results in the Remote Sensing field, mainly due to high accuracy, automatic feature extraction, and flexibility. LASER Detection and Ranging (LiDAR) is a data source that could supply more information than only visible spectral range. The research aims to create a methodology of automatic road extraction in the urban environment using data fusion LiDAR-Orthoimages through the neural convolutional network (U-Net architecture). The methodology was split into two steps to analyze the data source contribution on road extraction, mainly in occlusion and obstruction cases. Lastly, this research uses deep learning as a tool for automatic information extraction in Remote Sensing that, nowadays, is state of the art. So, the methodology purposes road extraction quickly with quality control. (AU)