Dense Air Pollution Estimation from Sparse in-situ Measurements and Satellite Data

TL;DR

The paper tackles the challenge of estimating ground-level NO$_2$ concentrations globally when ground truth data are sparse and unevenly distributed. It proposes a dense estimation framework that uses uniformly random offset sampling to generate a grid of NO$_2$ predictions from multispectral satellite data, implemented with UNet and Autoencoder backbones and dual heads for NO$_2$ regression and land-cover segmentation. The approach achieves a notable improvement over point-wise methods, achieving a Mean Absolute Error of $4.98\,mu g/m^3$ and surpassing the point-wise baseline by $9.45\%$, while reducing computational load and enabling scalable global assessment; it also generalizes to new regions such as the US West Coast. The combination of random sampling, dual-task learning, and patch-based prediction demonstrates practical utility for continual, large-scale environmental monitoring and policy support, with robust performance across varying prediction spaces and geographies. Overall, the dense estimation method provides a scalable, accurate, and region-robust tool for global NO$_2$ monitoring using satellite data.

Abstract

This paper addresses the critical environmental challenge of estimating ambient Nitrogen Dioxide (NO$_2$) concentrations, a key issue in public health and environmental policy. Existing methods for satellite-based air pollution estimation model the relationship between satellite and in-situ measurements at select point locations. While these approaches have advanced our ability to provide air quality estimations on a global scale, they come with inherent limitations. The most notable limitation is the computational intensity required for generating comprehensive estimates over extensive areas. Motivated by these limitations, this study introduces a novel dense estimation technique. Our approach seeks to balance the accuracy of high-resolution estimates with the practicality of computational constraints, thereby enabling efficient and scalable global environmental assessment. By utilizing a uniformly random offset sampling strategy, our method disperses the ground truth data pixel location evenly across a larger patch. At inference, the dense estimation method can then generate a grid of estimates in a single step, significantly reducing the computational resources required to provide estimates for larger areas. Notably, our approach also surpasses the results of existing point-wise methods by a significant margin of $9.45\%$, achieving a Mean Absolute Error (MAE) of $4.98\ μ\text{g}/\text{m}^3$. This demonstrates both high accuracy and computational efficiency, highlighting the applicability of our method for global environmental assessment. Furthermore, we showcase the method's adaptability and robustness by applying it to diverse geographic regions. Our method offers a viable solution to the computational challenges of large-scale environmental monitoring.

Figures (8)

• Figure 1: Comparison of point-wise and dense estimation approaches for NO2 concentration estimation from satellite imagery. The point-wise approach (left) generates predictions at a single location, denoted by the red dot. The dense estimation approach (right) estimates NO2 concentrations over a designated area, indicated by the red square, enhancing computational efficiency for regional-scale estimations.
• Figure 2: High-level overview showcasing the process of estimating NO2 concentrations using a deep learning model. The model inputs are randomly sampled windows consisting of satellite data from Sentinel-2 and Sentinel-5P, which then pass through a convolutional neural network to extract relevant features. The NO2 Regression Head produces a concentration estimate, which is subsequently aligned with ground truth data for validation. Concurrently, the Land Cover Classification Head processes the same features to classify land cover, aiding in the interpretation of NO2 distribution. Losses from both heads are combined to optimize the model, with the ultimate goal of providing accurate, high-resolution estimations of ground-level NO2.
• Figure 3: Illustration of the uniformly random offset sampling technique. The base image (left) shows the full 200 x 200 pixels area from which samples are taken. Individual samples (right) are 128 x 128 pixels, each randomly offset within the larger image. This sampling strategy diversifies the pixel location of the measurement point while retaining the original geospatial station coordinates.
• Figure 4: Distribution of ground-truth pixel coordinates of 2871 samples within the sampling window for varying prediction space sizes.
• Figure 5: Illustration showcasing the dual-task approach where one head is dedicated to the NO2 estimation, and the other head focuses on land cover classification.