Highly Accurate Multi-Modal LTE Channel Prediction via Semantic Segmentation of Satellite Images

Predicting the coverage of the base stations in mobile networks is a critical task for mobile operators to identify the geographical area covered by the cellular base stations. It also enables the network operators to discover the coverage gaps and optimally choose the locations for new base stations. Existing prediction models use ray tracing techniques that are computationally expensive and depend on three-dimensional maps, which are costly and need to be regularly updated. This paper proposes an efficient and highly accurate multi-modal channel model prediction algorithm using numerical features and satellite images with semantic segmentation (SS) to extract the environmental characteristics. Experimental measurements were gathered and combined with two-dimensional satellite maps from a real LTE network in the Cairo region for an accurate evaluation. Using the proposed architecture with SS and introducing new numerical features, we achieved a mean absolute error (MAE) of 1.57 dB and 2.21 root-mean-square error (RMSE) with a 23.7% enhancement over the state-of-theart techniques and a 61.04% reduction in system complexity in terms of the number of trainable parameters.