TL;DR
This study pioneers the use of high-resolution satellite imagery (PlanetScope) and deep learning to estimate the oxidative potential (OP) of particulate matter (PM), a crucial but under-monitored indicator of PM's toxicity. The developed models show promising results, capturing over half of the variation in OP with a relatively low error, suggesting a cost-effective way to expand OP monitoring.
What Problem Does This Paper Address?
Particulate matter (PM) air pollution is a major environmental health hazard, but simply measuring PM mass concentration doesn't fully capture its toxicity. Oxidative potential (OP) is a more direct measure of PM's capacity to induce oxidative stress in the lungs, making it a better indicator of health impacts. However, OP is not routinely monitored due to the cost and complexity of measurements, leading to limited exposure data and hindering public health efforts. While satellite imagery is widely used for PM mass estimation, its potential for estimating OP has remained unexplored. This paper aims to bridge this gap by investigating if satellite data, combined with deep learning, can effectively estimate OP.
Methods
The researchers employed a deep convolutional neural network (CNN), specifically a ResNet50 architecture, to extract features from daily 3m/pixel PlanetScope satellite images. These extracted image features were then fed into a multilayer perceptron (MLP) along with meteorological variables (temperature, humidity, wind, etc.) to predict the 24-hour average oxidative potential (OPAA and OPDTT). They explored several modeling approaches:
- Baseline Model: Used only meteorological variables.
- Random Features Model: Used random features extracted from satellite images, combined with meteorology.
- Transfer Learning Model: Leveraged features pre-trained on the large ImageNet dataset, combined with meteorology.
- Fine-tuning Model: Further adapted the ImageNet pre-trained features to the specific air quality task.
- Contrastive Learning (SimSiam) Model: Explored self-supervised learning to extract features from unlabelled satellite images.
The models were trained and evaluated using a standard train/validation/test split. Performance was assessed using R², Root Mean Squared Error (RMSE), and Normalized Mean Absolute Error (NMAE).
Datasets
The study utilized the following datasets:
- Satellite Imagery: Daily 3m/pixel PlanetScope satellite images (RGB and TOAR bands) for the period of January 1, 2017, to December 31, 2021, covering three air quality monitoring stations in Grenoble, France.
- Ground-Truth Air Quality Data: 24-hour average measurements of PM10 mass concentration, and two types of oxidative potential assays: OPAA (ascorbic acid) and OPDTT (dithiothreitol). These were collected at three stations (urban center, urban background, suburban) over several years.
- Meteorological Data: Hourly data for temperature, relative humidity, surface pressure, and wind components from the ERA5-Land reanalysis (approx. 10 km resolution), and planetary boundary layer height from ERA5 reanalysis (approx. 30 km resolution).
Key Insights and Ideas
- Satellite Imagery as a Proxy for OP: The study demonstrates that satellite imagery, particularly when combined with meteorological data, can capture a significant portion of the variability in oxidative potential, a metric previously unaddressed by remote sensing.
- Image Features are Informative, but Complementary: While image features alone can estimate a substantial amount of OP variation (up to ~50% for OPAA), they perform best when combined with meteorological data, suggesting that both sources provide complementary information.
- ImageNet Features Outperform Random/Contrastive Features: Pre-trained features from ImageNet (transfer learning and fine-tuning) generally yielded better results for OP estimation compared to randomly initialized CNN features or contrastive learning features, indicating the transferability of generic image features to this specific task.
- OP is Less Spatially Correlated than PM Mass: The study found that OPAA and OPDTT were less correlated between different monitoring stations compared to PM10 mass concentration, highlighting the localized nature of OP drivers.
- Potential for Low-Cost OP Monitoring: The success of this approach suggests a pathway to developing a cost-effective method for expanding the temporal and spatial coverage of OP estimates, which is crucial for public health research and policy.
Key Results
- Combined Model Performance: The best models, using both image features and meteorological data, captured over half of the variation in OPAA (test set R² = 0.62) and almost half of the variation in OPDTT (test set R² = 0.48).
- Image-Only Performance: Even using satellite images alone, the models achieved respectable results, capturing about half of the variation in OPAA (test set R² = 0.49) and one-third in OPDTT (test set R² = 0.36).
- Error Rates: Relative Mean Absolute Error (MAE) was around 32% when using both image and meteorological data, and about 37% when using only satellite images.
- ImageNet Transfer Learning Effectiveness: Fine-tuning ImageNet-derived features generally led to the best performance, suggesting that adapting generic image features is beneficial.
Limitations and Future Work
- Dataset Size: The primary limitation is the relatively small size of the OP dataset (1354 measures over five years from three locations). This limited the number of satellite image-OP pairs available for training (210), potentially affecting model robustness and the ability to learn complex relationships.
- Spatial Resolution of Meteorological Data: The ERA5 reanalysis data had a lower spatial resolution (10-30 km) than the satellite imagery (3m), which might have limited the model's ability to capture fine-scale meteorological influences.
- Limited Temporal Coverage: The study focused on daily averages, and future work could explore finer temporal resolutions.
- Additional Variables: The authors suggest exploring other variables like solar radiation or ozone concentration, as OPDTT is sensitive to secondary organic aerosols.
- Broader Applicability: Future studies should validate these findings in other geographical areas and with more diverse OP measurements to confirm the approach's broad applicability and ability to capture spatial OP variations.
- Increasing Training Data: Investigating the impact of larger training datasets on model performance is recommended.
Summary updated