Source-linked AI summary
Estimating regional ground-level PM2.5 directly from satellite top-of-atmosphere reflectance using deep learning
Huanfeng Shen, Tongwen Li, Qiangqiang Yuan, Liangpei Zhang
TL;DR
Conventional satellite PM2.5 estimation typically retrieves AOD from TOA reflectance before estimating ground-level PM2.5, raising the question of whether the intermediate retrieval can be bypassed. This study uses a deep learning model to estimate PM2.5 directly from TOA reflectance and reports competitive performance against AOD-PM modeling in the Wuhan Urban Agglomeration.
Problem
The study asks whether ground-level PM2.5 can be estimated directly from satellite TOA reflectance, avoiding the intermediate AOD retrieval used by conventional methods.
Source: §5 · p. 19Method
A deep belief network relates ground-level PM2.5 to satellite TOA reflectance, observation angles, and meteorological factors.
Source: §1 · p. 5Results
Sample-based cross-validation yielded R2 of 0.87 and RMSE of 9.89, with competitive performance versus AOD-PM modeling and finer resolution and larger spatial coverage than AOD-derived PM2.5.
Source: §5 · p. 19Takeaways & Limitations
In the Wuhan Urban Agglomeration, TOA-reflectance-derived PM2.5 had finer resolution and larger spatial coverage than AOD-derived PM2.5.
Source: §5 · p. 19The model used only R1, R3, R7, and observation angles, and its validation was limited to the small Wuhan Urban Agglomeration region.
Source: §5 · p. 19
Abstract
from arXiv · showhide
Almost all remote sensing atmospheric PM2.5 estimation methods need satellite aerosol optical depth (AOD) products, which are often retrieved from top-of-atmosphere (TOA) reflectance via an atmospheric radiative transfer model. Then, is it possible to estimate ground-level PM2.5 directly from satellite TOA reflectance without a physical model? In this study, this challenging work are achieved based on a machine learning model. Specifically, we establish the relationship between PM2.5, satellite TOA reflectance, observation angles, and meteorological factors in a deep learning architecture (denoted as Ref-PM modeling). Taking the Wuhan Urban Agglomeration (WUA) as a case study, the results demonstrate that compared with the AOD-PM modeling, the Ref-PM modeling obtains a competitive performance, with out-of-sample cross-validated R2 and RMSE values of 0.87 and 9.89 ug/m3 respectively. Also, the TOA-reflectance-derived PM2.5 have a finer resolution and larger spatial coverage than the AOD-derived PM2.5. This work updates the traditional cognition of remote sensing PM2.5 estimation and has the potential to promote the application in atmospheric environmental monitoring.
1. Introduction
The usual satellite-based approach estimates ground-level PM2.5 from AOD retrieved from TOA reflectance, prompting the question of whether PM2.5 can instead be estimated directly from TOA reflectance. This study proposes deep learning for that direct relationship and evaluates it in the Wuhan Urban Agglomeration.
1. Introduction
- The study asks whether ground-level PM2.5 can be estimated directly from TOA reflectance, bypassing the intermediate AOD retrieval used in conventional satellite-based workflows.AOD products are often retrieved from TOA reflectance using an atmospheric radiative transfer model.Source: §1 · p. 4
- Deep learning is proposed to fit the complex, nonlinear relationship between TOA reflectance, auxiliary factors, and ground-level PM2.5.The introduction notes that both AOD retrieval and the relationship between AOD and PM2.5 are nonlinear.Source: §1 · pp. 4–5
- The study develops a deep belief network linking ground-level PM2.5 with TOA reflectance, observation angles, and meteorological factors, and validates it in the Wuhan Urban Agglomeration using 2016 data.The stated objective is to estimate PM2.5 using TOA reflectance rather than AOD products.Source: §1 · p. 5
2. Study region and data
The study uses the Wuhan Urban Agglomeration and 2016 ground-monitoring data, alongside satellite observations and meteorological inputs, to develop matched datasets for Ref-PM and AOD-PM modeling.
2.1. Study region
- The Wuhan Urban Agglomeration, centered on Wuhan and including eight neighboring cities, is the study region, and the study period is 2016.The region is in Hubei province, central China.Source: §2.1 · pp. 5–6
2.2. Ground-level PM2.5 measurements
- Hourly PM2.5 measurements from 104 stations were averaged to daily means, excluding station-days with fewer than 18 hourly observations.The measurements came from CNEMC and HPEMCS stations.Source: §2.2 · p. 7
2.4. Meteorological data
- GEOS 5-FP meteorological inputs included relative humidity, air temperature, wind speed, surface pressure, and planetary boundary layer height.Meteorological parameters were extracted around the Aqua overpass time.Source: §2.4 · pp. 8–9
2.5. Data pre-processing and matching
- The researchers created 0.01-degree grids for Ref-PM and 0.03-degree grids for AOD-PM, averaging station measurements within each grid and matching satellite and resampled meteorological data.Data were re-projected to a common coordinate system before extracting observations at PM2.5 monitoring locations.Source: §2.5 · p. 9
3. Deep learning-based Ref-PM modeling for the estimation of PM2.5
The Ref-PM model uses a deep belief network to estimate ground-level PM2.5 directly from TOA reflectance and related inputs, without first retrieving AOD. It is trained against station measurements and evaluated with sample- and site-based cross-validation.
Model inputs
- The Ref-PM formulation predicts PM2.5 from reflectance terms R1, R3, and R7, observation angles, RH, WS, TMP, PBL, PS, NDVI, and geographic PM2.5 correlation inputs.The geographic inputs include nearby PM2.5 from neighboring grids and PM2.5 observations from prior t days.Source: §3 · p. 10
DBN model
- A deep belief network represents the complex input-output relationship, using stacked restricted Boltzmann machine layers and a back-propagation prediction layer.The described network uses two RBM layers, each with 15 neurons.Source: §3 · p. 10
Model training
- The DBN is pre-trained without supervision on input data, excluding station PM2.5 measurements, to initialize coefficients; it is then fine-tuned by back-propagation using MSE between estimated and ground-observed PM2.5.Source: §3 · p. 11
Prediction and mapping
- The trained model is validated and used to predict PM2.5 where ground monitoring stations are absent.The stated workflow uses a third step for validation and prediction at unmonitored locations.Source: §3 · p. 11
Validation
- Ten-fold sample-based and site-based cross-validation assess predictive power and potential overfitting using R2, RMSE, MPE, and RPE.Sample-based folds divide observations; site-based folds divide monitoring stations.Source: §3 · p. 12
4. Results and discussion
Across the Wuhan study, Ref-PM achieved competitive cross-validation performance and stronger spatial results than AOD-PM. Reflectance-derived estimates also provided finer resolution and more valid estimates on bright surfaces.
4.1. Descriptive statistics
- 5 to 256.96 µg/m3 was the PM2.5 range, with a mean of 49.43 µg/m3; R1 and R3 correlations with PM2.5 were 0.10 and 0.12.The authors interpret the low correlations as indicating a nonlinear relationship between TOA reflectance and PM2.5.Source: §4.1 · pp. 13–14
4.2. Performance of the Ref-PM modeling
- Sample-based (site-based) cross-validation R2 and RMSE were 0.87 (0.79) and 9.89 (12.97) for Ref-PM, versus 0.86 (0.72) and 10.42 (15.30) for AOD-PM.Fig. 5 shows the sample-based and site-based cross-validation scatter plots for both models.Source: §4.2 · p. 14, Figure 5 · p. 15
- 70% of grids had Ref-PM R2 values above 0.80, and 81% had RMSE below 15; the corresponding AOD-PM shares were 59% and 69%.Fig. 6 maps R2 and RMSE for both models under site-based cross-validation.Source: §4.2 · pp. 15–16, Figure 6 · p. 16
4.3. Discussion
- Reflectance-derived PM2.5 had finer resolution and more valid estimates on bright surfaces than AOD-derived PM2.5.The authors attribute the larger spatial coverage largely to estimates over bright surfaces, where AOD-based estimation may be missing.Source: Figure 7 · p. 17, §4.3 · pp. 17–18
- 0.80 was the sample-based cross-validated R2 on bright surfaces, compared with 0.87 on dark target regions.The authors describe the bright-surface estimation accuracy as acceptable.Source: §4.3 · p. 18
5. Conclusions and future work
In the Wuhan Urban Agglomeration, Ref-PM estimates ground-level PM2.5 directly from satellite TOA reflectance, avoiding intermediate AOD retrieval. It achieves competitive cross-validation performance and finer-resolution, broader-coverage maps than AOD-derived PM2.5, while its input selection and regional validation remain limited.
5. Conclusions and future work
- Ref-PM avoids intermediate AOD retrieval to estimate ground-level PM2.5 directly from satellite TOA reflectance, without a physical model.The authors identify potential to extend the approach to satellite monitoring of other environmental features.Source: §5 · p. 19
- In WUA, Ref-PM modeling achieved competitive performance compared with AOD-PM modeling, with cross-validation R2 and RMSE values of 0.87 and 9.89 ug/m3, respectively.Source: §5 · p. 19
- In WUA, reflectance-derived PM2.5 maps have finer resolution and larger spatial coverage than AOD-derived PM2.5.The study maps daily PM2.5 distributions in the agglomeration.Source: §5 · p. 19
- The model uses only R1, R3, R7, and observation angles; whether surface reflectance or other satellite parameters improve accuracy remains open.The authors note that the parameter selection was inspired by the dark target algorithm.Source: §5 · p. 19
- Validation is confined to the small WUA region, and application in larger geographical regions may bring new challenges.The authors identify larger-region application and validation as future work.Source: §5 · pp. 19–20