Discretized Urban Path Loss Modeling at 3.5 GHz using Geospatial Features and Machine Learning Techniques: From Regression to Classification

Accurately modeling and predicting urban path loss is a challenging task due to the fact that conventional regression models yield continuous estimates that can hide important performance nuances across different attenuation regimes. Therefore, this study aims to transform the continuous path loss prediction problem into a classification task by discretizing measurements into finely tuned 10 dB bins (implying a nominal discretization error of ±5 dB). The motivation behind this approach lies in enhancing model interpretability and robustness, particularly in complex urban environments where signal attenuation is influenced not only by distance but also by structural obstructions. This way, we conducted a measurement campaign at 3.5 GHz in Cartagena, Spain, collecting 2381 valid samples and deriving geospatial predictors from cadastral and digital surface models; class imbalance was addressed with SMOTE. We trained machinelearning models Gradient Boosting (GB), k-Nearest Neighbors, Support Vector Machines, and Neural Networks and evaluated performance with cross-validation (CV) and a 10% hold-out test set. On the full dataset, GB —which stood out from the other algorithms considered— achieved an Area Under the Receiver Operating Characteristic curve (AUC) of 0.993 (CV) and 0.991 (test), with test accuracy = 0.879 and F1-score = 0.878. Cluster-specific analysis showed strong separability across attenuation regimes, with test accuracies ranging from 0.869 (Cluster 3) to 0.945 (Cluster 4) and AUCs from 0.954 to 0.989. Permutation-based importance and information-criteria analyses consistently identified total transmitter–receiver distance as the dominant predictor, while initial and final free-space segments and the number/height of obstructing elements provided additional discriminative power.