| Strategic noise maps are essential for environmental exposure assessment under the European Environmental Noise Directive, yet producing them with physics-based propagation models can be data- and compute-intensive at continental scale. This paper presents a transferable surrogate noise-mapping pipeline that predicts the average level Lden on a dense 10 m grid using multi-class semantic segmentation. We train on a collection of 5-band Lden rasters derived from CNOSSOS-EU simulations for 11 labelled European cities, and deploy the same model without fine-tuning to six additional cities4. Inputs are a four-channel stack of open or satellite-derived predictors: (i) ground absorption coefficients from Copernicus Urban Atlas 2018 land-use and infrastructure products, (ii) Euclidean distance-to-road, (iii) smoothed OpenStreetMap maximum speed, and (iv) Local Climate Zones (LCZ). Our proposed architecture, BigResSegNet-ASPP, is a deep residual encoder–decoder augmented with convolutional block attention (CBAM) and atrous spatial pyramid pooling (ASPP) to capture multi-scale road-to-neighbourhood context. To address severe class imbalance, we optimize a composite objective combining class-weighted cross-entropy with label smoothing, weighted Dice overlap, and focal modulation. We provide a fully tiled training/inference implementation, a reproducible preprocessing protocol, and an evaluation framework including mIOU and band-wise noise error. |
*** Title, author list and abstract as submitted during Camera-Ready version delivery. Small changes that may have occurred during processing by Springer may not appear in this window.