An end-to-end deep learning pipeline for urban noise monitoring with spatiotemporal analysis

Urban noise pollution is a critical public health problem affecting millions of people around the world. Current acoustic monitoring systems largely focus on optimizing classification accuracy but fail to meet operational requirements such as calibrated probability outputs, deployment formats, and spatiotemporal integration. This study presents a comprehensive urban noise monitoring pipeline extending from raw sensor data to city-scale policy support. The proposed framework uses a CNN14-based ensemble architecture on the SONYC-UST dataset collected from 56 acoustic sensors deployed across New York City. The system performs simultaneous multi-source detection across 35 hierarchical acoustic presence labels and provides reliable probability calibration validated by a Brier score of 0.1269. In terms of technical infrastructure, the pipeline offers export in TorchScript format that can run on edge devices without Python dependency, processing capacity of more than 2,000 audio segments per second on an NVIDIA GeForce RTX 3060 and training time of less than six hours per fold. Integrated 250-meter spatial grid mapping and hourly-daily temporal aggregation modules enable direct integration of predictions into urban planning workflows. Independent external corroboration was performed using NYC 311 complaint records across 3,460 matched grid-day observations. Permutation testing revealed a statistically significant relationship between predicted noise scores and community-reported disturbance. This study demonstrates that model accuracy alone is not sufficient for deployable urban noise monitoring systems, and that a holistic approach encompassing calibration, deployment, and spatial integration is required.