physics.ao-phOct 30, 2026
SaveSATFRONT: Detecting Atmospheric Fronts with Satellite Data using Machine Learning
Organizations: University of Amsterdam, Amsterdam, The Netherlands · EUMETSAT, Darmstadt, Germany
Abstract
Atmospheric fronts are associated with much of the high-impact weather in the mid-latitudes, yet their identification and classification remain largely manual and subjective. Existing deep learning approaches rely primarily on gridded reanalysis or forecast data, limiting their applicability to satellite observations available for time-critical applications. We introduce SATFRONT, a machine learning framework for detecting cold, warm, and occluded fronts over Europe from geostationary satellite observations, with reanalysis data as an optional additional input. The model is trained on frontal analyses from five European national meteorological services and evaluated against UK Met Office analyses for 2024. Using satellite observations alone, SATFRONT achieves a Critical Success Index (CSI) of 0.428 at a 150 km tolerance, within the range of agreement between analyses from different meteorological services (0.277-0.501) and higher than that of an operational rule-based front detector that uses satellite data (0.133). Adding ERA5 reanalysis increases the CSI to 0.591. ERA5 substantially improves the placement of individual fronts, although its effect on their annual spatial distribution varies by front type. These results demonstrate that cold, warm, and occluded fronts can be detected directly from geostationary satellite observations, providing a basis for automated frontal analyses at the high temporal frequency required for time-critical applications.
Figures & tables
Figure 1: Schematic of a low-pressure system (L) with a cold front (blue, triangles), a warm front (red, semicircles), and an occluded front (purple, alternating symbols). The symbols point in the direction in which each front moves. The shaded area between the cold and the warm front is the warm sector, and arrows indicate the motion of the air masses.
Figure 2: Satellite observations (top) and ERA5 variables (bottom) for 25 February 2024, 12:00 UTC, with the fronts of the UK Met Office analysis overlaid. Top left: the 11.0 m brightness temperature is close to the temperature of cloud tops, so high, cold clouds appear bright. Top right: the 6.2 m brightness temperature mainly reflects water vapour in the upper troposphere. Bottom left: ERA5 temperature and wind at 850 hPa, about 1.5 km above sea level. Bottom right: ERA5 specific humidity at 850 hPa with mean sea level pressure contours (hPa). The satellite channels show clouds and moisture, whereas ERA5 provides temperature, humidity, wind, and pressure directly.
Figure 3: Surface analysis charts for 8 November 2021, 00:00 UTC, from (a) the UK Met Office, (b) KNMI, (c) DWD, and (d) Met Éireann. Blue lines with triangles, red lines with semicircles, and purple lines with alternating symbols mark cold, warm, and occluded fronts. The services agree on the main frontal systems but differ in the position, extent, and type of several fronts. Chart sources: (a) © Crown Copyright 2021. Information provided by the National Meteorological Library and Archive – Met Office, UK. (b) KNMI (CC0). (c) Deutscher Wetterdienst (DWD), obtained via Wetterzentrale.de. (d) © Met Éireann, source: met.ie.
Figure 4: The SATFRONT architecture with service-specific output heads. Each stack of input planes represents a group of input channels. The front planes show the 11.0 m satellite channel and the ERA5 temperature at 850 hPa for 25 February 2024, 12:00 UTC. The ERA5 variables (dashed outline) are used only in the satellite+ERA5 configuration. The encoder reduces the spatial resolution in four stages of Swin Transformer blocks, and the decoder restores it and combines its features with those of the encoder through skip connections (dashed arrows). The values below the blocks give, for each stage, the number of Swin Transformer blocks, the number of channels, and the spatial resolution relative to the input. Each output head is a convolution that maps the shared features to class scores. The output plane shows the fronts predicted by the UK Met Office head of SATFRONT with satellite+ERA5 inputs for the same time. The single-service models have one output head.
Figure 5: Ground truth (left) and predictions of SATFRONT with satellite+ERA5 inputs (middle) and satellite-only inputs (right) for 25 February 2024, 12:00 UTC, using the UK Met Office head. The grey background shows the 11.0 m brightness temperature where it is below 250 K, corresponding to cold cloud tops. The dashed line marks the UK Met Office region (59 ∘ W–28 ∘ E, 38 ∘ –76 ∘ N). Only pixels inside this region are scored, and only UK Met Office labels inside it contribute to the training loss. Ground-truth fronts that extend beyond the region are therefore not used. For visualisation, the satellite-only predictions are shown after a post-processing step that fills gaps of up to approximately 12 pixels using a morphological closing and removes regions smaller than 120 pixels. The satellite+ERA5 predictions are shown without post-processing. All quantitative results in this paper use predictions without post-processing. Appendix F compares satellite-only predictions with and without post-processing.
Figure 6: Overall CSI on the 2024 test set as a function of the tolerance (50–250 km). Solid lines show satellite+ERA5 inputs, and dashed lines show satellite-only inputs. The Baseline U-Net and single-service SATFRONT are trained only on UK Met Office labels, whereas SATFRONT is trained on labels from all five services.
| Model | Input | Class | CSI | POD | SR | F1 | FB |
|---|---|---|---|---|---|---|---|
| SATFRONT (ours) | Satellite | Overall | .428 | .561 | .691 | .618 | 1.005 |
| Cold | .462 | .600 | .734 | .660 | 1.122 | ||
| Warm | .378 | .546 | .612 | .577 | 1.149 | ||
| Occluded | .445 | .535 | .726 | .616 | 0.743 | ||
| Baseline U-Net | Satellite | Overall | .473 | .660 | .610 | .631 | 1.025 |
| ASII-SAT | Satellite | Overall | .133 | .355 | .237 | .265 | 2.332 |
Table 1: Test scores at the 150 km tolerance on the 2024 test set, with UK Met Office labels as ground truth. SATFRONT is trained on labels from all five services and evaluated with its UK Met Office head. The Baseline U-Net is trained only on UK Met Office labels. ASII-SAT and ASII-SAT+NWP are the two modes of the operational ASII product. The models are grouped by input, and bold marks the best overall value of each metric within each group (for FB, the value closest to 1).
Figure 7: Ground truth (left) and satellite-only predictions of SATFRONT (middle) and the Baseline U-Net (right) for 25 February 2024, 12:00 UTC (top) and 22 November 2024, 00:00 UTC (bottom). Predictions from both models are shown after the same post-processing used in Figure 5 , which fills small gaps and removes regions smaller than 120 pixels. The dashed line marks the UK Met Office region, as in Figure 5 . Fronts predicted outside it, for example over North Africa, are not included in the scores.
Figure 8: Frequency of cold, warm, and occluded fronts (rows) over the 2024 test year within the UK Met Office region, expressed as a percentage of the 1 462 test samples. Columns 1–3 show the ground truth and SATFRONT with satellite+ERA5 and satellite-only inputs. Column 4 (“pattern residual”) shows pixels where the satellite+ERA5 map is closer to the ground truth than the satellite-only map (red) or farther from it (blue). For this column only, each map is first divided by its mean so that the comparison reflects spatial pattern rather than overall frontal frequency. It can therefore differ from the pixel percentages reported in Table 3 , which are computed from the unnormalised maps. All maps are smoothed for visualisation using a Gaussian filter with a standard deviation of 150 km.
| Predictor | Overall | Cold | Warm | Occl. |
|---|---|---|---|---|
| DWD MO | .376 [.363, .392] | .481 | .399 | .247 |
| FMI MO | .277 [.263, .291] | .322 | .266 | .244 |
| KNMI MO | .468 [.452, .480] | .522 | .436 | .446 |
| Met Éireann MO | .501 [.486, .519] | .543 | .454 | .508 |
| SATFRONT, satellite-only MO | .428 | .462 | .378 | .445 |
| SATFRONT, satellite+ERA5 MO | .591 | .616 | .568 | .588 |
Table 2: Agreement between services on the 2024 test set, with the UK Met Office (MO) analyses as ground truth. Values are CSI at 150 km, with 95% block-bootstrap confidence intervals for the overall service scores. Each service row uses only the samples and geographic region shared with the UK Met Office. The SATFRONT rows use all test samples and the full UK Met Office region.
| Correlation | Error | Satellite+ERA5 | ||
|---|---|---|---|---|
| Class | Satellite-only | Satellite+ERA5 | reduction (MSS) | closer ( ) |
| All fronts | .402 | .559 | 20.5% | 54.1% |
| Cold | .649 | .764 | 3.5% | 43.0% |
| Warm | .345 | .521 | 48.5% | 53.6% |
| Occluded | .420 | .550 | 2.0% | 50.1% |
Table 3: Comparison of the annual frontal frequency maps for the 2024 test year with the UK Met Office ground truth, using the measures defined in Section 3.5 . A negative error reduction indicates that the satellite+ERA5 map has a larger mean squared error than the satellite-only map.
| Group | Channel or variable | Description |
| Satellite (GEO-Ring, 8) | 3.8 m | Infrared window; low clouds and fog at night; includes reflected sunlight during the day |
| 6.2 m | Water vapour in the upper troposphere | |
| 7.3 m | Water vapour in the middle troposphere | |
| 8.6 m | Infrared window; cloud phase | |
| 9.7 m | Ozone absorption | |
| 11.0 m | Infrared window; temperature of cloud tops and the surface |
Table 4: Input channels. The satellite-only configuration uses the 15 channels of the first two groups, and the satellite+ERA5 configuration uses all 43 channels. The descriptions of the satellite channels give their main sensitivity.
Figure 9: Ground truth (left) and predictions of SATFRONT with satellite+ERA5 inputs (middle) and satellite-only inputs (right) for 22 July 2024, 12:00 UTC (top) and 22 November 2024, 00:00 UTC (bottom), using the UK Met Office head. As in Figure 5 , the dashed line marks the UK Met Office region, the satellite-only predictions are shown after post-processing, and the satellite+ERA5 predictions are shown without post-processing.
Appendix figures & tables4 assets
Supplementary material from the paper’s appendix.
Appendix
| Tolerance (km) | ||||||
|---|---|---|---|---|---|---|
| Input | Class | 50 | 100 | 150 | 200 | 250 |
| Satellite | Overall | .253 | .348 | .428 | .494 | .548 |
| Cold | .270 | .377 | .462 | .529 | .581 | |
| Warm | .220 | .302 | .378 | .444 | .501 | |
| Occluded | .267 | .366 | .445 | .508 | .561 | |
| Satellite+ERA5 | Overall | .374 | .505 | .591 | .649 | .692 |
Table 5: Test CSI of SATFRONT for each front class at every tolerance, with UK Met Office labels as ground truth.
Figure 10: Extraction of the fronts from a UK Met Office chart for 1 January 2018, 00:00 UTC. (a) Original chart. (b) The chart after georeferencing to latitude and longitude with a thin plate spline estimated from manually selected control points. (c) Pixels selected by colour for each front class. (d) The selected pixels after removing isolated pixels and closing small gaps. (e) Skeletons with a width of one pixel, widened for display. (f) Paths traced along the skeletons, each in a different colour. Grey pixels are the short branches formed by the frontal symbols, which are removed. (g) Paths after spline smoothing. (h) The extracted fronts in geographic coordinates. Original chart © Crown Copyright 2018. Information provided by the National Meteorological Library and Archive – Met Office, UK. Panels (b)–(h) show our processing of this chart to extract the frontal lines.
Figure 11: Ground truth (left) and satellite-only predictions of the Baseline U-Net without (middle) and with (right) the auxiliary features for three test samples, shown without post-processing. Without the features, the model predicts large spurious warm-front regions over North Africa. With the features, these regions become much smaller but do not disappear. The dashed line marks the UK Met Office region, outside which predictions are not scored and labels do not contribute to the training loss.
Figure 12: Ground truth (left) and satellite-only predictions of SATFRONT without (middle) and with (right) post-processing for two test samples. The dashed line marks the UK Met Office region.
More in this topic
Climate Science
Climate science studies Earth's climate system, its past and future changes, and the physical and human drivers of climate risks.
See all 63 papers