Beyond here and now: estimating pollution levels across space and time from street view images with deep learning

Keywords: Deep learning; computer vision; air pollution; noise pollution; street-view images; environmental modelling
Abstract Type: Paper Abstract

Abstract

Deep learning augmented computer vision techniques have led to an emerging field of image-based inference of environmental pollution and its sources. The real-world application of image-based pollution models crucially relies on their spatial and temporal generalisability, but is currently understudied, particularly in low-income countries where there is limited infrastructure for measuring complex patterns of noise and air pollution. We used two complementary classification models, based on convolutional neural networks, in both an end-to-end approach and as a feature extractor (object detection), to estimate spatially and temporally resolved noise levels and fine particulate matter (PM2.5) and in Accra, Ghana. Models were trained on data from a unique dataset of 2.1 million images paired with air and noise measurements, collected at 145 representative locations over 15 months. Both end-to-end and feature-based approaches outperformed null model benchmarks for predicting PM2.5 and noise at single locations, but accuracy diminished when applied to unseen locations. End-to-end model accuracy for predicting PM2.5 was associated with characteristics of images based on atmospheric visibility, while specific objects such as vehicles and people were relatively more important for noise models. Model accuracy was reduced when tested on images from locations unseen during training but improved by including data from a greater number of locations during training, even if the total quantity was reduced. The results demonstrate challenges and potential of spatiotemporal, image-based noise and air pollution estimation, and that reliable, environmental modelling with image data continues to need integration with sensor networks.