Deep learning maps fog, haze and sea fog across eastern China and tests pollution cuts

A convolutional neural network that fuses geostationary satellite imagery, weather reanalysis and ground-based particulate measurements can now identify and quantify the full range of low-visibility events over eastern China, from clean land fog and polluted fog-haze to dry haze and sea fog. The framework, developed by Yuting Liang, Shitong Zhao, Dantong Liu and colleagues at Zhejiang University and partner institutions, is described in a research article published in Atmospheric Chemistry and Physics. Beyond simply detecting where visibility drops below 2 kilometres, the model quantifies the intensity of each event and links it to aerosol loading, allowing the team to run controlled sensitivity experiments on air pollution. Their headline result is striking: cutting overall PM2.5 concentrations by 20 percent in winter would shrink the fog-haze area by about 43 percent, yet the population exposed to those events would fall by only 15 percent, because the most polluted air sits directly over the most densely populated land.

Low-visibility events are far more than an inconvenience for travellers. They degrade atmospheric transparency through light extinction by suspended droplets and aerosol particles, threatening transport safety and, when the culprit is polluted fog-haze, human health. Analysis of weather records from 2412 stations in February showed that fog accounted for 73 percent of the relevant weather types at inland sites and 59 percent at coastal sites, with rain contributing a further 18 and 17 percent respectively and all other types below 5 percent. Fog itself forms when near-surface air becomes sufficiently saturated for aerosol particles to activate into cloud droplets, and the microphysical properties of those droplets, including their size spectrum, are influenced by aerosol concentration. A fog carrying high aerosol loadings can exert adverse health impacts, which is why distinguishing clean fog from polluted fog-haze matters for policy as well as for science.

Traditional satellite-based fog detection has long relied on fixed thresholds applied to one or a few parameters. Nighttime fog is typically flagged using the brightness temperature difference between infrared channels, while daytime algorithms exploit visible and 3.9-micrometre shortwave infrared reflectance to separate water clouds from snow and ice, with textural thresholds distinguishing smooth fog from rougher cloud tops and cloud-top-to-surface temperature differences confirming low-level fog. These approaches struggle to generalize across changing surface and cloud conditions, are hampered at night by the absence of visible channels, and are contaminated by mid- and high-level clouds. Machine-learning methods can instead learn nonlinear relationships among many variables, and the new study shows that integrating large-scale multi-source data improves generalization further.

The team built their dataset for 2019 and 2020 around China’s FY-4A geostationary satellite, drawing on brightness temperatures at 3.72 and 10.8 micrometres together with the solar zenith angle. The 10.8-micrometre thermal infrared channel reveals cloud-top temperature, separating low-level fog from colder elevated cloud layers, while the 3.72-micrometre channel distinguishes fine haze from fog because only the larger fog droplets trigger strong Mie scattering. The brightness temperature difference between the two channels enhances the spectral contrast among fog, haze and the background. Meteorological variables came from the ERA5 reanalysis, with the spread between 2-metre temperature and dew point serving as a proxy for near-surface saturation that separates ground-contacting fog from unsaturated low stratus or aerosol-dominated haze. Hourly PM2.5 concentrations from the China National Environmental Monitoring Centre represented aerosol loading over land, hourly visibility observations from the NOAA/NCDC Integrated Surface Database Lite provided the ground truth for labelling, and 1-kilometre WorldPop population data supported the exposure analysis.

The labelling scheme defined low visibility as station observations at or below 2 kilometres, mapped onto the nearest 4-kilometre FY-4A grid cell to create a binary mask for training. All datasets were matched hourly and resampled to that grid, then cropped into 64-by-64-pixel windows for the network. The researchers quantified the station-to-grid representativeness uncertainty using nearest-station distance statistics and leave-one-out cross-validation. The final dataset comprised 1440 low-visibility events, split 70 percent for training, 15 percent for validation and 15 percent for testing. Two model variants were trained: a land-based Sat-PM model incorporating PM2.5, and a maritime Sat model without it, since anthropogenic influence over open ocean is minimal.

Architecturally, the network stacks three convolutional blocks, each pairing a 3-by-3 convolutional layer with batch normalization and rectified linear unit activation, followed by max pooling that shrinks the spatial resolution from 64 by 64 to 8 by 8 while deepening the features from 32 to 128 channels. Fully connected layers then compress 8192 features down to 512 before a Softmax output yields low-visibility and clear probabilities. A dual-branch attention mechanism addresses the redundancy of multi-source inputs: a variable-wise module weights the importance of each input channel through global average pooling, while a spatial attention module uses average and max pooling with a 7-by-7 convolution kernel to focus on informative regions such as fog boundaries. Training used a hybrid loss combining balanced focal loss, with a focusing parameter of 3.0, and dice loss in equal measure, a combination designed to overcome the severe class imbalance between rare low-visibility samples and frequent clear conditions.

On independent test data, the terrestrial Sat-PM model achieved an overall accuracy of 0.93, an F1-score of 0.78 and a recall of 0.81, while the maritime Sat model reached an accuracy of 0.91 with a fog-specific F1-score of 0.71. After the network flags low-visibility pixels, a physically interpretable attribution step assigns event types: over land, pixels with dew point depression below 3 degrees Celsius and PM2.5 below 75 micrograms per cubic metre are labelled fog, those with the same humidity but PM2.5 at or above 75 micrograms per cubic metre become fog-haze, and drier polluted pixels become haze; over ocean, low-visibility pixels are attributed to sea fog. The 75-microgram threshold follows the Grade-II 24-hour limit of China’s ambient air quality standard. Validation against 821 stations showed that 62 percent had correlations above 0.6 between modelled fog intensity and observed visibility reduction, confirming that the model captures event intensity rather than merely binary occurrence.

The seasonal and diurnal behaviour of the four event types diverged sharply. Land fog-haze showed the strongest contrast, with its accumulated winter area roughly 21 times the summer value, while accumulated winter haze was about 1.7 times summer. Sea fog covered a larger area in winter too, with a peak median area about 1.75 times the summer value, whereas land fog showed a weaker contrast, its maximum winter extent of about 140,000 square kilometres in February against 95,000 square kilometres in July, roughly 1.5 times summer. Land fog and fog-haze peaked before dawn and dissipated after sunrise, consistent with classic radiation fog driven by nocturnal cooling; winter fog peaked later, at 07:00 to 08:00, than summer fog at 05:00 to 06:00, reflecting delayed sunrise. Winter haze peaked overnight at 05:00 and fell to a minimum by mid-afternoon as boundary-layer mixing diluted pollutants, but summer haze instead built after sunrise and peaked from late morning to midday under photochemical production. Summer sea fog persisted into midday, unlike winter sea fog, because the weak diurnal variation in sea surface temperature delays marine boundary-layer warming and fog dissipation.

The sensitivity experiments are where the framework becomes a policy tool. Holding all satellite and meteorological inputs fixed at winter conditions and scaling only the PM2.5 input, the team found that reducing PM2.5 to 80, 60 and 40 percent of original levels cut the mean February fog-haze area by approximately 43, 75 and 95 percent respectively, from a baseline with peak areas exceeding 70,000 square kilometres. Reducing winter PM2.5 to 40 percent shifted the simulated cumulative distribution of fog-haze area onto the observed summer distribution, and conversely raising summer PM2.5 to 250 percent reproduced the winter distribution, indicating that aerosol loading contributes substantially to the seasonal contrast. Notably, lower PM2.5 increased clean-fog coverage, meaning some polluted fog-haze shifted category into cleaner fog rather than disappearing, consistent with aerosol effects on droplet numbers and light extinction. The exposure metric, a population-weighted intensity, responded more sluggishly: the first 20 percent PM2.5 cut removed fog-haze mainly from less populated regions, and only after reductions beyond 40 percent did exposure fall faster than area, with each further 10 percent cut yielding an 18 percent exposure decrease. The authors stress that the population-weighted intensity is an indicator of potential exposure rather than a direct epidemiological estimate of health risk, and that the 40 percent figure marks an effective range under tested winter conditions rather than a universal threshold. Their broader message is that visibility improvement alone is a poor proxy for reduced exposure, and that air-quality assessment should combine event attribution with population weighting.

Subject of Research: Deep learning classification and quantification of low-visibility events over eastern China

Article Title: Classification and quantification of low-visibility events using deep learning over eastern China

Article References: Liang, Y., Zhao, S., Liu, D., Wu, C., Yi, L., Jia, X., Xu, H., & Wang, J. (2026). Classification and quantification of low-visibility events using deep learning over eastern China. Atmospheric Chemistry and Physics, 26(20), 15003-15015. https://doi.org/10.5194/acp-26-15003-2026

Image Credits: AI Generated

DOI: 10.5194/acp-26-15003-2026

Keywords: deep learning, low-visibility events, fog-haze, sea fog, PM2.5, eastern China, FY-4A satellite, convolutional neural network, population exposure, air quality policy, visibility, Atmospheric Chemistry and Physics