Spatio-Temporal Regression of LTE/5G Signal Metrics for ITU/3GPP Quality Classification on Drive-Test Data from Manaus

This paper investigates the feasibility of predicting LTE/5G signal-quality metrics (RSRP, RSRQ, SINR, and RSSI) from purely spatio-temporal inputs (latitude, longitude, hour, weekday) and using the predicted values to recover the ITU/3GPP composite quality label. A drive-test dataset of 17,646 measurements collected in Manaus, Brazil (April 13-16, 2026) is processed through a four-stage pipeline: (i) anomaly detection and removal, (ii) per-metric ITU/3GPP threshold-based labeling, (iii) multi-output regression of the four signal metrics, and (iv) re classification of the predicted metrics for label validation. Seven regression models are compared on the same stratified 80/20 split. The best performer, a Decision Tree, achieves R2 = 0.884 averaged across the four metrics and recovers the original ITU/3GPP label with 86.51% accuracy. The results indicate that spatio-temporal features alone capture a substantial portion of the signal-quality structure in the studied area, supporting the feasibility of prospective coverage estimation without on-device measurement.