Satellite Imagery-Based Urban Noise Prediction Model - Case study over Vancouver, Canada
Abstract
Understanding the spatial variability of noise in an urban environment is a critical step towards protecting the long-term health of its citizens. Currently, urban noise modeling methods require heavily refined input data layers and large pools of calibration data. These requirements make urban noise modeling a financially draining and labor-intensive process for municipalities. By comparison, a satellite image-based model utilizes widely available imagery datasets, making it a more practical noise-modeling alternative. This encourages responsible authorities to characterize environmental noise, which is an important initial step in mitigating this health hazard.
The goal of this study is to develop a convolutional neural network-based model that predicts annual average noise levels in the urban environment. The model is given about 160,000 100-meter satellite image patches extracted from meter-scale WorldView2 imagery over Vancouver, BC. These patches are fed into a pre-trained VGG-16 network and the model is fine tuned to (1) extract features relevant to the propagation of noise through urban environments (such as roadways, building shapes, and surface types) and (2) estimate noise levels using the extracted features. Our model was trained on mean annual noise values (LDEN dBA) from an industry standard noise model ( based on traffic and road data, topography, and land use) over Metro Vancouver, BC. Preliminary results have shown that the model reliably picks up the broad spatial variation of noise in the study area, with areas closer to the mean noise levels (e.g. residential and urban) performing better than areas with more extreme noise values (high values in the case of large highways and low values in mountainous areas). This study demonstrates the feasibility of using high resolution satellite imagery in a deep learning framework for predicting urban noise, as well as different model architectures that have proven most effective for this goal.- Publication:
-
AGU Fall Meeting Abstracts
- Pub Date:
- December 2020
- Bibcode:
- 2020AGUFMA043.0014L
- Keywords:
-
- 3333 Model calibration;
- ATMOSPHERIC PROCESSES;
- 3337 Global climate models;
- ATMOSPHERIC PROCESSES;
- 3339 Ocean/atmosphere interactions;
- ATMOSPHERIC PROCESSES;
- 1942 Machine learning;
- INFORMATICS