Nowcasting: forecasting the next three hours

At 3 pm a thunderstorm builds over the next ridge. The daily forecast said "thundershowers likely in the district", which was true and not much help. What the officer on duty needs to know is whether it will rain over a particular town, a bridge or a catchment in the next hour or two. That is the question nowcasting tries to answer. This article explains what nowcasting means in weather forecasting, how it is done, how it differs from the forecasts you see each morning, and where its limits lie.

What is nowcasting in weather?

Nowcasting is forecasting with local detail, by any method, from the present out to six hours ahead, including a detailed description of the present weather. That is the working definition in the World Meteorological Organization's Guidelines for Nowcasting Techniques [1]. The last part matters: a nowcast begins with an accurate picture of what is happening now, because the best predictor of the next hour is the last hour.

In practice, the useful skill of most methods is concentrated in the first few hours, as the evidence below shows. The WMO guidelines stress that nowcasting is applied to weather on the mesoscale (the scale of storms and storm clusters) and local scales over very short periods. They therefore emphasise rapidly updated, high-resolution observations [1].

Nowcasting and forecasting: what is the difference?

The difference between nowcasting and forecasting is mainly one of time and scale, and of method.

Numerical weather prediction solves the equations of the atmosphere from an estimate of its present state. A model needs time to settle after it starts, and it has difficulty placing a single storm at the right place and hour. The WMO guidelines note that blending observations with a model helps overcome the limits of NWP in the first three hours, caused by its "spin-up time" and its inability to give the location-specific timing that nowcasting needs [1]. A nowcast avoids that problem by starting from what the radar and satellites actually see.

How is a nowcast made?

Nowcasting methods run from very simple to complex [1]. Three families cover most of them.

Extrapolation of radar and satellite images

A weather radar scans the sky every few minutes and maps where rain is falling and how intensely. The oldest idea in nowcasting is to find how the rain pattern moved between two recent scans, then slide it forward in time at the same speed and direction. This is known as Lagrangian extrapolation, "Lagrangian" meaning that the pattern is followed along with the moving air rather than watched from a fixed point [3].

Software estimates the motion by tracking features from one image to the next, a technique called optical flow. An open-source library for this, with ensemble methods that give a probability rather than a single answer, is described by Pulkkinen and co-authors [3]. The same ideas apply to satellite images where radar is missing. A 2024 study over the Maritime Continent, where radar coverage is incomplete and blocked by mountains, used satellite images and found a deterministic optical-flow nowcast skilful up to four hours at spatial scales of 10 km and coarser [4].

Numerical weather prediction

High-resolution models run at kilometre scale can predict storms explicitly and are rerun frequently with fresh observations. The WMO guidelines list three obstacles for model-based nowcasting: radars are too sparse in most places to feed the model with enough data, operational centres have limited computer power dedicated to nowcasting, and model errors grow rapidly at convection-permitting scales [1].

Blending the two

Blending combines extrapolated observations with model output to produce a seamless 0-6 hour nowcast. The WMO guidelines call it one of the preferred methods, with the weight shifting from the extrapolation to the model as lead time increases [1]. Large systems such as squall lines and mesoscale convective systems can be tracked and extrapolated up to about three hours. Small single-cell storms, which usually last less than an hour, are too small and short-lived to be blended with model fields [1].

How far ahead does a nowcast stay useful?

Not very far, and it depends on the weather. Studies agree that extrapolation beats model forecasts at the start and loses its advantage after a few hours:

• Extrapolation of radar images typically outperforms NWP forecasts in the first 2-5 hours, depending on weather conditions, domain and the model used [3].
• A Colorado study of a high-resolution model against radar extrapolation found that model nowcasts of precipitation amount up to 2 hours ahead were less skilful than extrapolation of radar data at a scale of 120 km [2].
• In the pysteps evaluation, reliable and sharp nowcasts for light rain were obtained up to about 2 hours, with a noticeable loss of sharpness after 3 hours [3].

The reason is physical. Extrapolation assumes that the rain keeps its shape and intensity as it moves. Real storms grow and decay. New cells form along outflow boundaries and terrain-driven winds that radar does not see until rain has already started. The Colorado study attributes model errors to the inability to resolve such narrow low-level convergence lines [2]. Nowcasts are therefore better for the movement of existing rain than for the birth of a new storm.

The scale of the feature also matters. In the satellite study, probabilistic nowcasts did better at scales of 60 km or less, and the deterministic method at around 200 km [4]. A forecast for "this district" holds up longer than a forecast for "this village".

Nowcasting in India: IMD's service

The India Meteorological Department (IMD) issues nowcasts of severe weather, mainly thunderstorms and heavy rain. IMD's own monsoon FAQ says its meteorological centres issue nowcasts for heavy rainfall during the monsoon season with a validity of up to 6 hours, and notes that nowcasting an event as small as a cloudburst needs a dense radar network or very high-resolution models [5]. IMD scientists have reviewed the state of convective nowcasting over the Indian region [6], and the operational nowcasting of thunderstorms and its verification have been documented in IMD's own journal [7].

A worked example (illustrative)

Suppose radar shows a line of heavy rain 36 km west of a town. Two scans 10 minutes apart show its leading edge moved 8 km east.

Speed    = 8 km / 10 min = 0.8 km/min = 48 km/h
Arrival  = distance / speed = 36 km / 0.8 km/min = 45 min

The extrapolation says that rain reaches the town in about 45 minutes. That is well inside the window where extrapolation is reliable. Now suppose the hill catchment above the town has a time of concentration (the time for water from the furthest point to reach the outlet) of about one hour; see time of concentration. If the rain falls on the catchment from the moment it starts, the stream can be expected to respond roughly an hour after that. Detection plus travel gives the warning team up to about 45 minutes before rain begins, and the catchment's own lag adds time before the river rises.

The same arithmetic shows the weakness. If a new cell forms 5 km from the town, there is no 45-minute warning. A nowcast based on motion alone sees nothing until rain is already on the radar. That is why small, fast catchments need observations on the ground as well; see flash floods vs river floods: warning time.

Why observations still matter

Nowcasting depends on observation networks. The WMO guidelines recommend a 5-minute update of precipitation fields for small rivers, with 15 minutes sufficient in most cases, and warn that data transmission delays of more than 10-15 minutes from rain gauges could lead to late warnings [1]. Radar sees rain in the air; a rain gauge measures what reaches the ground; a river sensor shows how the catchment has responded. For a broader view of who observes what, see weather radar, rain gauges and satellites.

In our own study of the July 2023 floods in Himachal Pradesh, the flood peak was inflated and sustained by extremely heavy rainfall acting on near-saturated soil, with snowmelt also contributing [8]. That combination is exactly what a rainfall nowcast cannot judge alone: how much of the rain becomes runoff depends on the state of the ground and the river, which has to be measured.

Key points

• Nowcasting is forecasting with local detail, by any method, from now to about 6 hours ahead [1].
• It starts from the latest radar, satellite, lightning and gauge observations, and extrapolates them forward, often blended with a model [1].
• Extrapolation outperforms models in the first few hours, but its skill fades quickly, particularly for small convective storms [1][3].
• Nowcasts predict the movement of existing rain better than the birth of new storms.
• In India, IMD issues nowcasts for heavy rainfall in the monsoon season with a validity of up to 6 hours [5].
• Warnings for fast catchments need river and rain observations as well as forecasts.

Where ClimMaTech fits

A nowcast tells you rain is coming; observations tell you what it did. At ClimMaTech we build monitoring devices for river level, surface velocity and warning dissemination, so that a nowcast can be checked against what the river is actually doing and a warning can reach the people downstream. How these pieces work together is described in our flood warning solution.

References

1. World Meteorological Organization, Guidelines for Nowcasting Techniques, WMO-No. 1198, 2017 edition, Geneva. library.wmo.int
2. J. W. Wilson, D. Megenhardt, J. Pinto, NWP and Radar Extrapolation: Comparisons and Explanation of Errors, Monthly Weather Review 148(12), 4783-4798, 2020. doi:10.1175/MWR-D-20-0221.1
3. S. Pulkkinen et al., Pysteps: an open-source Python library for probabilistic precipitation nowcasting (v1.0), Geoscientific Model Development 12, 4185-4219, 2019. doi:10.5194/gmd-12-4185-2019
4. J. Smith, C. Birch, J. Marsham, S. Peatman, M. Bollasina, G. Pankiewicz, Evaluating pySTEPS optical flow algorithms for convection nowcasting over the Maritime Continent using satellite data, Natural Hazards and Earth System Sciences 24, 567-582, 2024. doi:10.5194/nhess-24-567-2024
5. India Meteorological Department, Frequently Asked Questions (FAQs) on Monsoon, Q24 (forecasts at different scales) and Q46 (cloudbursts), IMD Pune. imdpune.gov.in
6. S. S. Roy, M. Mohapatra, A. Tyagi, S. K. Roy Bhowmik, A review of Nowcasting of convective weather over the Indian region, MAUSAM 70(3), 465-484, 2019. doi:10.54302/mausam.v70i3.227
7. K. Ray, B. K. Bandopadhyay, S. C. Bhan, Operational nowcasting of thunderstorms in India and its verification, MAUSAM 66(3), 595-602, 2015. doi:10.54302/mausam.v66i3.566
8. V. Gupta, B. Syed, A. Pathania, S. Raaj, A. Nanda, S. Awasthi, D. P. Shukla, Hydrometeorological analysis of July-2023 floods in Himachal Pradesh, India, Natural Hazards 120(8), 7549-7574, 2024. doi:10.1007/s11069-024-06520-5