ConvNeXtV2 CBAM Framework for Urban Drone Noise Footprint Prediction from Flight Paths and Building Morphology
Authors
Files
Abstract
Urban drone route noise prediction requires spatial assessment of both flight geometry and the surrounding built environment. This paper specifies a ConvNeXtV2 network with convolutional block attention for predicting event sound exposure maps from synchronized flight paths and building morphology. Route information is encoded as dwell time, altitude, speed, heading, and source energy rasters, while a parallel morphology stream represents building height, occupancy, facade orientation, and visibility. Hierarchical fusion and a multiscale decoder support detailed footprint reconstruction. The numerical evaluation uses six comparison cases across 120 route events on 96 by 96 receptor grids, covering five urban forms and five operating heights. Within this comparison collection, the event-averaged root mean squared error (RMSE) is 2.40 dB for the complete attention case and 3.10 dB for the backbone case, giving a relative difference of 22.6%. Mean absolute error is approximately 1.91 dB, and the fraction of receptors within 3 dB is approximately 79.0%. Morphology stratification, altitude conditioning, footprint thresholds, and adverse input analyses examine the behavior of these comparison cases. The architecture specification links source timing and urban geometry to event exposure prediction, while the evaluation protocol combines receptor errors, spatial footprint metrics, and event- level statistics. The contribution is an encoding and assessment framework whose predictive accuracy requires subsequent network training and independent acoustic validation.
