Source-linked AI summary
Precipitation Nowcasting with Satellite Imagery
Vadim Lebedev, Vladimir Ivashkin, Irina Rudenko, Alexander Ganshin, Alexander Molchanov, Sergey Ovcharenko, Ruslan Grokhovetskiy, Ivan Bushmarinov, Dmitry Solomentsev
TL;DR
Precipitation nowcasting is limited by the geographic reach of ground-based radar, restricting short-range rain forecasts and alerts. This paper combines geostationary satellite imagery with radar observations, NWP-derived features, and machine learning to extend nowcasting beyond radar coverage. The resulting system supports Yandex precipitation maps and related alerts, with coverage expansion toward broader and potentially global services.
Problem
Radar-dependent nowcasting is geographically constrained, while users need precise precipitation timing and location for short-range decisions.
Method
The system combines radar observations, geostationary satellite imagery, NWP-derived atmospheric features, and machine learning for precipitation detection and nowcasting.
Results
The system provides nowcasting beyond ground-based radar coverage with quality similar to traditional radar-based nowcasting and supports Yandex precipitation maps and alerts.
Takeaways & Limitations
Satellite data extends precipitation nowcasting to territories without radar coverage and supports broader weather-service delivery.
Takeaways & Limitations
The deployed system is currently limited to the Meteosat-8 field of view, and expansion to other satellites is expected to require regional retraining and feature adjustments.
Abstract
from arXiv · showhide
Precipitation nowcasting is a short-range forecast of rain/snow (up to 2 hours), often displayed on top of the geographical map by the weather service. Modern precipitation nowcasting algorithms rely on the extrapolation of observations by ground-based radars via optical flow techniques or neural network models. Dependent on these radars, typical nowcasting is limited to the regions around their locations. We have developed a method for precipitation nowcasting based on geostationary satellite imagery and incorporated the resulting data into the Yandex.Weather precipitation map (including an alerting service with push notifications for products in the Yandex ecosystem), thus expanding its coverage and paving the way to a truly global nowcasting service.
1 INTRODUCTION
Weather services support decisions from immediate conditions through two-hour nowcasts, but radar-dependent precipitation maps remain geographically constrained. This study extends nowcasting with satellite imagery and NWP products, targeting regions beyond radar coverage.
- 1 INTRODUCTION: A radar-based precipitation map provides house-level forecasts two hours ahead at 10-minute intervals and supports human-readable rain alerts.The service can express when precipitation will begin, where it will occur, and how long it will last.
- 1 INTRODUCTION: Searches for “karta osadkov” were only 10 times less frequent than global English “weather map” searches in summer 2018, indicating strong interest in the precipitation-map feature.Interest peaks corresponded to severe thunderstorms in the Moscow region.
- 1 INTRODUCTION: The precipitation-map data also adjusts current-weather displays across Yandex services and reaches partners and offline users, at least doubling the product’s effective audience.Yandex.Weather had approximately 5 million daily active users and more than 24 million monthly unique cross-device users in December 2018.
- 1 INTRODUCTION: Traditional numerical forecasts lack sufficiently precise precipitation locations and timing, while radar extrapolation is accurate initially but cannot model later physical precipitation development.These limitations motivate combining high-resolution observations with other forecasting information.
- 1 INTRODUCTION: Radar-based precipitation products are poorly scalable because coverage depends on costly infrastructure, leaving many Russian and developing-country regions without radar networks.Russia’s Siberian and Ural federal districts are identified as target regions, together representing about 30 million people.
- 1 INTRODUCTION: The system aims to recreate radar-derived precipitation fields from satellite imagery and NWP products, then verify predictions against ground-based weather stations.The target is practical nowcasting over territories with little or no radar coverage.
2 DATA SOURCES
Precipitation nowcasting requires high-resolution, frequent, direct, and broadly distributed observations that no single source supplies. The paper therefore combines radar, satellite, and NWP information, using geostationary imagery to extend coverage beyond radar networks.
- 2 DATA SOURCES: No single observation source provides the spatial resolution, temporal resolution, direct rainfall measurement, and global coverage required for precipitation nowcasting.The paper frames source combination as necessary because these requirements differ from those of NWP.
- 2 DATA SOURCES: Ground stations directly observe precipitation, but their usual reporting intervals provide insufficient spatial and temporal information for high-resolution precipitation fields.The SYNOP protocol specifies accumulated precipitation reporting once every 12 hours, while many stations report conditions every three hours.
- 2 DATA SOURCES: Ground-based radar reconstructs 2 × 2 km precipitation fields every 10 minutes within ranges up to 250 km, but coverage is limited outside densely developed regions.Most radars in the Russian network are concentrated in the western part of the country.
- 2 DATA SOURCES: Low-Earth-orbit satellite constellations provide global precipitation coverage but revisit locations infrequently; GPM covers 65°S–65°N at three-hour resolution.Their narrow orbital swaths create long intervals between consecutive passes of one satellite.
- 2 DATA SOURCES: Geostationary satellites continuously observe clouds across the visible Earth disk, but precipitation detection from visible and infrared imagery is difficult and prior methods lacked user-facing accuracy at 10-minute alert horizons.This paper uses Meteosat-8 data together with radar data and NWP models to design and evaluate a new detection approach.
3 RELATED WORK
Related work frames satellite precipitation detection as a coverage-oriented alternative to radar-based nowcasting, using heuristics or machine learning to infer precipitation from imagery. Prior methods include physics-based retrievals, conventional machine learning, semantic-segmentation CNNs, and Conv-LSTM or optical-flow approaches for forecasting.
- Geostationary satellite imagery offers global, uninterrupted coverage, but precipitation must be inferred indirectly through heuristic estimation or binary detection.This paper focuses on precipitation detection rather than regression.
- Physics-based methods combine atmospheric assumptions, cloud properties, satellite measurements, and sometimes NWP or radar calibration, but can be limited to convective rain or daylight operation.The MPE method misses cyclonic and frontal precipitation, while PP retrievals use visible wavelengths and therefore operate only during daylight.
- Conventional machine-learning studies compare trees, neural networks, and SVMs, but pixel-wise splitting may overfit because atmospheric fields are smooth across time and space.Reported results were best under daytime conditions, while stacked denoising autoencoders were proposed to reduce overfitting.
- Precipitation detection resembles semantic segmentation because multichannel image inputs receive pixel-wise labels, motivating convolutional neural networks developed for image-processing tasks.CNNs have been applied to satellite imagery, including road extraction and building detection, although aerial-image architectures are less diverse than those used for large semantic-segmentation datasets.
- Radar nowcasting commonly estimates motion with optical flow and advects precipitation fields, while Conv-LSTM models add complexity to represent artifacts and precipitation emergence or disappearance.Conv-LSTM was proposed by [27] and improved by [28].
4 PRECIPITATION DETECTION
The precipitation detection pipeline aligns radar and satellite data, augments satellite imagery with NWP and other features, and trains a UNet-based classifier. The resulting model outperforms physics-based approaches across time periods, while GFS features further improve F1 score.
- Data preparation: The pipeline remaps radar and satellite data to a common grid, converts satellite imagery to radar’s 10-minute timestep, and adjusts scan timestamps.Radar observations beyond 200 km are discarded before aggregation, while optical-flow interpolation supplies intermediate satellite frames.
- Features: GFS atmospheric fields, topography, and solar altitude augment satellite imagery to provide signals not easily inferred from imagery alone.The GFS inputs include convective precipitation rate, cloud work function, cloud water, precipitable water, and convective potential energy.
- Model and training: The main detector is a five-block UNet trained for three precipitation levels, with Dice loss improving F1 scores over binary cross-entropy alone.The network produces separate output maps for light, medium, and heavy precipitation, and is trained with Adam for 250000 iterations.
- Model and training: Training on small 96x96-pixel crops mitigates overfitting caused by the limited geographical training area and radar-labeled coverage.Without mitigation, the network can memorize relief and geographical location, undermining predictions outside radar coverage.
- Metrics and results: The UNet consistently outperforms physics-based methods across reported periods and metrics, while its nighttime performance drops only slightly.The evaluation uses F1 averaged across temporal and spatial dimensions because ordinary accuracy is uninformative under class imbalance; Figure 4 reports performance over local time.
- Metrics and results: Adding GFS features further increases the UNet model’s F1 score, and the UNet performs better than the pointwise model.The reported UNet advantage is attributed to its larger receptive field, which captures relationships between adjacent atmospheric locations.
5 NOWCASTING
The nowcasting stage predicts future precipitation fields from consecutive reconstructed fields using either optical-flow extrapolation or a convolutional neural network. Preliminary experiments found that the simpler optical-flow approach slightly outperformed the neural network on composited radar and satellite data.
- Nowcasting methods: Future precipitation fields are predicted from consecutive reconstructed fields using optical-flow extrapolation or a convolutional neural network.The neural network uses blocks that model optical-flow extrapolation through spatial transformer layers.
- Nowcasting results: In preliminary experiments, optical flow provided slightly better predictions than the convolutional neural network on composited radar and satellite data.The neural network had previously been superior in a single-radar setting, but that advantage did not transfer to the combined data source.
6 POST-LAUNCH PERFORMANCE
Post-launch evaluation found that satellite-based precipitation nowcasting performs below radar accuracy but above traditional forecasts, while extending useful coverage to radar-sparse regions and increasing user engagement.
- 6 POST-LAUNCH PERFORMANCE: The satellite-based map expanded nowcasting coverage, particularly across the Ural and Siberian districts that were previously underserved by radar.The coverage comparison contrasts availability before and after satellite imagery was introduced.
- 6 POST-LAUNCH PERFORMANCE: Satellite-based accuracy was lower than radar accuracy but better than the traditional forecast, with precipitation imbalance and F1 scores similar to radar results.The Siberian radar used for verification was not included in training, enabling evaluation in a region without radar observation.
- 6 POST-LAUNCH PERFORMANCE: A/B testing showed a statistically significant increase in daily active users in Siberian and Ural areas where the rain map had previously been mostly unavailable.This result supported the product’s rollout in late September.
- 6 POST-LAUNCH PERFORMANCE: Satellite and Meteum detection metrics were evaluated across the full satellite-map territory, whereas radar metrics covered only radar-observed areas.A satellite spacecraft malfunction caused a gap in the metrics at the beginning of November.
7 CONCLUSIONS AND FUTURE WORK
The launched system combines radar observations, geostationary satellite imagery, machine learning, and NWP-based atmospheric and surface information to extend nowcasting beyond radar coverage. Its current deployment is geographically limited and requires further data fusion and regional adaptation.
- 7 CONCLUSIONS AND FUTURE WORK: The deployed system combines ground-based radar observations, geostationary satellite imagery, machine learning, and NWP-derived physical information for precipitation nowcasting.Satellite data enables coverage of territories without ground-based radars while maintaining quality similar to a traditional radar-based nowcast.
- 7 CONCLUSIONS AND FUTURE WORK: The current system covers a region centered on European Russia within Meteosat-8’s field of view, although the approach could extend to other geostationary satellites.Scaling to Himawari and GOES could support global services, but regional weather differences are expected to require retraining and adjusted input features.
- 7 CONCLUSIONS AND FUTURE WORK: A sharp boundary between radar and satellite data can confuse users, motivating more sophisticated fusion such as blending and inpainting along the border.The authors experimented with erasing conflicting border observations and reconstructing the missing region [11].