THE ROLE OF 3D TREES MODEL IN 5G RADIO PREDICTION
Accurate and detailed mapping models are used as fundamental data to solve the radioplanning tasks. Such models have consisted of natural and artificial barriers causing difficulties for radio signal propagation. Under the artificial objects, one understands various buildings and engineering constructions, under the natural — the surface of the Earth as a relief and the vegetation on it.
It is known, that the higher the frequencies of radio signals, the more their propagation is affected by obstacles influence. It means any kind of objects between transmitter and receiver works as a barrier deteriorating the radio signal propagation. Therefore, in the case of 5G networks, which frequencies are: frequency range 1 (FR1) is from 450 MHz to 6 GHz, which includes the LTE frequency range and frequency range 2 (FR2) is from 24.25 GHz to 52.6 GHz, such impact should be taken into account. The vegetation presented by array crowns of trees and even every single vegetation unit and bushes has to be considered as obstacles and embedded into the mapping models.
A contemporary city, as a rule, involves lots of parks, squares, recreation areas, tree plantations, etc. In some cities, the buildings are completely shrouded in foliage.
Thus, there is a task to properly identify the vegetation and map it in the obstacle models with the required accuracy and details. The 5G network deployed without bearing in mind the 3D trees does not correspond to the 5G planning demands and leads to lower connectivity. On the other hand, 3D trees models ensure an accurate and reliable 5G network planning process allowing placing a required number of base stations for the stable signal and full coverage.
In the event of 4G networks, it is usually enough to present the vegetation as generalized polygons showing the uninterrupted common foliage areas as a single object with an average height. For the 5G models, the crowns are shown separately, each one with unique height. This task is more technologically complex since it is required to cure the problem of more accurate matching between the mapping model, which is developed, and the geometry of the crown in reality. Often, in practice for such models, one uses aerial images enabling higher resolution for more precision recognition.
Comparative predictions without-with vegetation heights in clutter height
In order to estimate how the involving of the vegetation model may affect the prediction of the radio-signal propagation, the following test was performed. Based on the 3D vector data of the city of Maintal (Germany), a pure clutter model of vegetation was generated (image).
Within the explored area, a simulated position of 3-sectors LTE site (18 height, 0, 120, 240 degrees, tilt 6 deg) was placed. As a result, there are two comparative prediction models — before and after adding the vegetation clutter height. Models do not include any signal losses (indoor losses, losses in clutter classes, losses in the propagation model).
The vegetation is diverse in nature and can be presented differently in the image depending on various types, color, height, and season. To recognize the foliage contours VISICOM develops and applies the Machine Learning methods based on Convolutional Neural Networks and the Deep Learning Techniques. The experts of VISICOM accomplished the training of neural network models using the training set comprised of 30 thousand objects of various trees worldwide. It allowed reaching a high training accuracy — 95–98% for the test data. Obviously, such accuracy cannot be reached for the real data but due to the obligatory quality check process and data verification, which are made in manual mode, the final models are very close to reality. Each new task and every new sample let expand the training set with new species and permanently improve the recognition standards.
PRAGUE CASE STUDY
· Test zone for 5G trial — 27.sq.km
· Total amount of buildings with roof detail (LOD quality) — 30 633
· Total area covered by vegetation — 11.2 sq.km
· Total quantity of vegetation canopies polygons — 243 000
Recently Visicom has finalized a project for Prague, Czech Republic — 3D models of 1m resolution for 5G test network.
Machine learning methods based on Convolutional Neural Networks and the Deep Learning Techniques here were applied for foliage object recognition. Buildings and vegetation outlines and heights were derived from Pleiades stereo images of 0.5m resolution and air photos. In order to get the heights (AGL — above ground level) of the recognized vegetation, it is required to generate a height grid model. It can be produced based on the DSM (Digital Surface Model) or LIDAR data, the DSM in turn can be created based on aerial or high-resolution satellite images.
As you see, the vegetation covers the biggest part of the territory and with no doubts should be taken into account while radioplanning tasks solving.