An ML-Based Location Tracking System for LoRa Mesh Networks in GPS-Denied Environments

LoRaWAN technology is a cornerstone in developing low-power wireless communication solutions for smart cities, agriculture, and remote environmental monitoring applications. LoRa enables long-range communication with minimal power usage, making it a suitable technology for battery-operated IoT devices. However, LoRa mesh networks rely on GPS to identify the location of its LoRa End Devices (EDs), which may be unreliable or unavailable in GPS-denied environments. This paper presents a location tracking and prediction system for LoRa mesh networks using Machine Learning (ML). In particular, we evaluate the performance of various ML models for predicting the geographic coordinates of LoRa EDs across the network. We trained four ML models, namely Linear Regression, Random Forest, KNeighbors Regressor, and Decision Tree Regressor, with a dataset collected by conducting real-world testbed and physical LoRaWAN hardware. Comprehensive hyperparameter tuning was conducted to enhance model performance. Our experimental results demonstrate that Random Forest, with its ability to capture complex and non-linear relationships, significantly outperforms other models in terms of accuracy and robustness. While lightweight ML models like Linear Regression offer competitive performance, they fall short in explanatory power.