A Machine-Learning-Based Connectivity Model for Complex Terrain Large-Scale Low-Power Wireless Deployments

Abstract

We evaluate the accuracy of a machine-learning-based path loss model trained on 42 157 324 RSSI samples collected over one year from an environmental wireless-sensor network using 2.4 GHz radios. The 2218 links in the network span a 2000 km2 basin and are deployed in a complex environment, with large variations of terrain attributes and vegetation coverage. Four candidate machine-learning algorithms were evaluated in order to find the one with lowest error: 1) Random Forest; 2) AdaBoost; 3) Neural Networks; and 4) ${K}$ -Nearest-Neighbors. Of the candidate models, Random Forest showed the lowest error. The independent variables used in the model include path distance, canopy coverage, terrain variability, and path angle. We compare the accuracy of this model to several well-known canonical (free space, plane earth) and empirical propagation models (Weissberger, ITU-R, COST235). Unlike canonical models, machine-learning algorithms are not problem-specific: they rely on an extensive dataset and a flexible model architecture to make predictions. We show how this model achieves a 37% reduction in the average prediction error compared to the canonical/empirical model with the best performance. This paper presents an in-depth discussion on the strengths and limitations of the proposed approach as well as opportunities for further research.