Geospatial Contextual Learning in AI-Driven Ensemble Framework for Wireless Systems

Accurate channel estimation is necessary for reliable link-level communication and spectral efficiency in 6G communications. In this paper, we present a channel estimation and received power prediction algorithm using a Random Forest (RF) along with a Deep Neural Network (DNN) trained on Demodulation Reference Signal (DMRS)-based Minimum Mean Square Error (MMSE)/Least Squares (LS) estimates. The simulated testbed involves geospatial metadata collected from OpenStreetMap (OSM), including latitude and longitude coordinates, and Clustered Delay Line (CDL) channels modeled according to Third Generation Partnership Project (3GPP) standards in urban, semi-urban, and rural environments using MATLAB. Physical Layer (PHY), equalizer, Channel Quality Indicator (CQI), etc. are included in the input parameters. Our framework is flexible and can be used for different types of deployment configurations. Results demonstrate remarkable performance enhancements in terms of Mean Squared Error (MSE) and Mean Absolute Error (MAE) compared to conventional channel estimation approaches, particularly under low-SNR and mobility conditions.