Performance Analysis in the Segmentation of urban asphalted roads in RGB satellite images using K-Means++ and SegNet

Case study in São Luís-MA

Abstract

The design and manual insertion of new terrestrial roads into geographic databases is a frequent activity in geoprocessing and their demand usually occurs as the most up-to-date satellite imagery of the territory is acquired. Continually, new urban and rural occupations emerge, for which specific vector geometries need to be designed to characterize the cartographic inputs and accommodate the relevant associated data. Therefore, it is convenient to develop a computational tool that, with the help of artificial intelligence, automates what is possible in this respect, since manual editing depends on the limits of user agility, and does it in images that are usually easy and free to access.

To test the feasibility of this proposal, a database of RGB images containing asphalted urban roads is presented to the K-Means++ algorithm and the SegNet Convolutional Neural Network, and the performance of each was evaluated and compared for accuracy and IoU of road identification.

Under the conditions of the experiment, K-Means++ achieved poor and unviable results for use in a real-life application involving tarmac detection in RGB satellite images, with average accuracy ranging from 41.67% to 64.19% and average IoU of 12.30% to 16.16%, depending on the preprocessing strategy used. On the other hand, the SegNet Convolutional Neural Network proved to be appropriate for precision applications not sensitive to discontinuities, achieving an average accuracy of 87.12% and an average IoU of 71.93%.