Adaptive high-resolution mapping
of air pollution with a novel
implicit 3D representation approach

Ting Zhang1 Bo Zheng1 Ruqi Huang1
1Tsinghua University,
Code
1. We propose HF-SDF, the noval approach to conduct adaptive high-resolution mapping using implicit representation. By fully utilizing the transferable, geometry-aware implicit representation, our method achieves strong transferability by eliminating dependencies on localization-specific predictors, mitigates resolution constraints through continuous modeling, and provides a more accurate and adaptable framework for high-resolution pollution mapping.
2. HF-SDF demonstrates the ability to understand air pollution patterns through continuous reconstructing using noisy, sparse, and incomplete observations.
3. HF-SDF presents a promising way to alleviate data assimulation challenges and enable more scalable multi-modal applications like Earth System Models (ESMs).
Interpolate start reference image.


Model framework. (a)–(d) illustrate the stepwise modeling approach. Specifically, (a.1), (b), (c.1), and (d) represent the high-resolution mapping pipeline for single-day measurements, while (a.2), (b), (c.2), and (d) correspond to the high-resolution mapping pipeline for multi-day measurements.

Abstract

Mapping air pollution at high spatial resolution is essential for understanding, managing, and mitigating the adverse impacts of air pollution. The current approaches for air pollution monitoring (e.g., ground station, satellite) suffer from limited spatial coverage and resolution. Artificial intelligence (AI) has shown great potential to resolve these issues. However, regarding low-quality labeled data and uneven spatial coverage, AI-based methods with transferability for air pollution monitoring are still in their infancy. Here, we introduce an innovative 3D implicit representation, dubbed Height-Field Signed Distance Function (HF-SDF), to reconstruct air pollution concentration maps at spatial resolutions of desire, which can achieve both extensive spatial coverage and fine-scale results with powerful transferability. Our proposed HF-SDF, which integrates a deep network with a geometric constraint, can produce adaptive high-resolution air pollution maps, which are robust to noisy and locally incomplete data.Our framework can learn a continuous, transferable 3D generative model of HF-SDF, by employing an auto-decoder network structure, allowing for the comprehensive reconstruction of a broad spectrum of air pollutants. The evaluation based on two types of air pollution data, including aerosol concentration maps and satellite observations, reaches accuracy rates of 96% and 91%, respectively. HF-SDF reveals immense promise in advancing air pollution monitoring by offering insights into the spatial heterogeneity of pollution distributions.

Concept illustration of HF-SDF

Interpolate start reference image.

Illustration of the HF-SDF design.

Demonstration Video

Adaptive High-Resolution Mapping on TAP

Reconstruction of PM2.5 distribution in central China. In (a), the TAP PM2.5 distribution within China is illustrated at a resolution of 10 km on November 3rd, 2023. For panels (b), (c), and (d), the upper images depict the TAP PM2.5 distribution in Central China, Hubei Province, and Wuhan City, respectively, at a 10 km resolution. The lower images provide detailed patterns of the estimated PM2.5 distribution in the corresponding areas of Central China, Hubei Province, and Wuhan City.
Interpolate start reference image.
Assessment of HF-SDF for ultra-high resolution mapping from sparse observations. In (a), the PM2.5 distribution around the Sichuan Basin from the TAP 1km dataset was downscaled at various scales, where DS 5km, DS 10km, DS 20km, DS 30km, DS 40km illustrate downsampled inputs at resolutions of 5km, 10km, 20km, 30km, and 40km, respectively. Panel (b) illustrates our reconstruction results at a resolution of 1km, inferred from TAP 1km, DS 5km, DS 10km, DS 20km, DS 30km, and DS 40km, respectively. Panel (c) illustrates the difference between our 1km resolution reconstructions and the data from TAP 1km. Chart (d) shows two groups of R and IOA: Ours 1km vs. TAP 1km and Ours vs. Downsampled Input. Chart (e) illustrates two groups of distribution of reconstruction error: Ours 1km vs. TAP 1km and Ours vs. Downsampled Input.
Interpolate start reference image.
High-resolution air pollution mapping with varying unmeasured areas. Panels (a) to (e) show reconstruction outcomes for maps with unmeasured areas of radii 0.05°, 0.15°, 0.25°, 0.35°, and 0.45° respectively. GAP1 to GAP5 denote these unmeasured areas. Each panel displays (left to right): the input map, high-resolution output, complete map (GT), and bias visualization (CDF). See Supporting Information Appendix B.3 for more details.
Interpolate start reference image.
We utilize the coded map network to evaluate the generalization capability of HF-SDF within trained regions, as well as its transferability to entirely unseen locations and different air pollutants. Details regarding the dataset used for model training and testing, as well as model performance, are summarized in Table 1. The training set comprises PM2.5 data (TAP) collected from 2021 to 2022 in six locations: Chengdu (CD), Bazhou (BZ), Taiyuan (TY), Changsha (CS), Xining (XN), and Longnan (LN), each covering a geographic area of 10° × 10°. For additional details, refer to the Supporting Information (Table S7). To address concerns of homogeneity, the training dataset includes 1080 instances of concentration maps, randomly sampled from 180 days between 2021 and 2022 for each location. Subsequently, test sets, which include different regions, times, and air pollutants, were utilized to assess the model's generalization capability and transferability.
Interpolate start reference image.

Satellite Observations Mapping

Satellites provide a global perspective on air quality monitoring; however, limitations such as low spatial resolution and unexpected data gaps reduce data availability. As discussed above, HF-SDF facilitates the continuous reconstruction of air pollution data despite these challenges, enabling effective reconstruction of satellite data. Here, we take the TROPOMI NO2 column as an example to demonstrate the effectiveness of HF-SDF for satellite observations mapping.
Interpolate start reference image.
Experiments on TROPOMI. Panel (a) illustrates the projection of satellite data alongside our continuous reconstruction at various resolutions: (a.1) presents raw NO2 data from TROPOMI, (a.2) showcases our reconstruction projected at a resolution of 10 km, (a.3) displays our reconstruction projected at a resolution of 2 km, (a.4) exhibits our reconstruction projected at a resolution of 1 km, and (a.5) visualizes the reconstructed 3D concentration surface. Panel (b) depicts NO2 concentration maps over China on the 10-day scale. (b.1) features averaged NO2 concentration maps over 10 continuous days from May 1st to 10th, 2018, while (b.2) presents the corresponding reconstruction. Reconstruction on a single day can be found in Supporting Information Figure S6.