Extending reliable nowcasting of extreme precipitation could provide critical additional time for warnings and emergency response during high-impact events such as flash floods. Radar-based generative machine-learning models have enabled skilful hyperlocal precipitation nowcasting, but accurate prediction of intense precipitation remains confined to the first few hours. Because storm-scale structure is predictable for longer than individual cells, a natural strategy is to predict that structure while generatively modelling only the uncertain local growth, decay, reorganisation and initiation of storms. Here we present Microsoft Weather Nowcast (MW-Nowcast), a six-hour ensemble radar nowcasting model that jointly learns a deterministic predictor to capture organised precipitation structure shared across ensemble members, and a generator to produce diverse local residuals around this shared prediction. Across independent test data from the United States, Europe and China, MW-Nowcast achieves higher detection skill than leading methods for heavy and extreme precipitation throughout the 6 h horizon. For the most intense rainfall, MW-Nowcast doubles the available warning time across all three regions, delivering 6 h forecasts with skill previously limited to 3 h for the leading generative baseline. A cost-loss decision analysis shows that MW-Nowcast retains substantial value for a broad range of applications even at 4-6 h, where alternative methods offer little benefit. These additional hours can give forecasters and emergency managers the time to warn and act before extreme rainfall strikes, helping to protect lives and property.
Figures & tables
Figure 1 : MW-Nowcast for 6 h ensemble precipitation nowcasting. A history encoder provides the deterministic and residual decoders with a shared representation of the observed radar sequence H . The deterministic decoder produces the common full-horizon forecast S1:T , while the flow-matching residual decoder transforms Gaussian noise Z(k) into sampled residuals through an ODE solver. At each lead time t , the sampled residual is added to St to obtain Xt(k) ; different noise draws yield the forecast ensemble.
Figure 2 : Case study of a US tornado event over the 6 h forecast horizon. The case shows a convective storm over Walthall County, Mississippi, on 15 March 2025 that spawned an EF4 tornado; the forecast is initialised at 17:30 UTC. a , Geographic and radar context, with the forecast domain outlined in orange. b , Forecast diagnostics: CSIN in a 5×5 grid-cell neighbourhood at rain-rate thresholds of 32 and 64 mm h -1 , CRPS, CRPSS, and radially averaged PSD at lead times of 3 and 6 h. “Deterministic” denotes the output of the deterministic decoder of MW-Nowcast . c , Observed and predicted rain-rate fields at hourly lead times from 1 to 6 h. Rows show observations, PySTEPS, SimVPv2, NowcastNet and MW-Nowcast . d , CSIN and POD at thresholds of 32 and 64 mm h -1 for the four MW-Nowcast ensemble members and their PMM summary. e , The deterministic prediction and four ensemble members at lead times of 2, 4 and 6 h.
Figure 3 : Case study of the Beijing Miyun mountainous extreme-rainfall event over the 6 h forecast horizon. Extreme rainfall over the northern Miyun district of Beijing and adjacent Hebei during the late-July 2025 episode; forecast initialised at 11:00 UTC on 26 July 2025 (19:00 local time). Panel descriptions are as in Fig. 2 .
Figure 4 : Quantitative evaluation of 6 h precipitation nowcasts. Curves compare MW-Nowcast , NowcastNet, SimVPv2 and PySTEPS. a , Critical Success Index in a 5×5 grid-cell neighbourhood over the 6 h forecast horizon for the US, EU and CN test sets at rain-rate thresholds of 16, 32 and 64 mm h -1 ; higher values are better. b , grid-point Continuous Ranked Probability Score (CRPS) of the four-member ensembles for the three regions; lower values are better. c , radially averaged power spectral density (PSD) over the US at lead times of 2, 4 and 6 h, with the observed spectrum shown for reference. d , relative economic value (REV) over the US as a function of cost–loss ratio at lead times of 2, 4 and 6 h for rain-rate exceedances of 16 and 32 mm h -1 .
Symbol
Meaning
Value
T0,T
Numbers of history and future frames
9,36
K
Number of ensemble members
4
N
Number of Heun intervals
20
S,s,c
Radar-grid side, feature-grid side and feature width
128,32,256
g
Number of AFNO channel groups
8
ME
Number of repeated TSF blocks in each context backbone
3
Table 1: Architecture and discrete-inference notation.
Figure 5 : Network architecture of MW-Nowcast . The history encoder maps H to Chist , and the deterministic decoder produces S1:T . At discrete ODE state (rn,τn) , the residual decoder uses the shared history context to predict vn . In panel c, Tin=T0 for the deterministic decoder and Tin=T for the residual decoder. The dashed temporal-upsample block is applied only in the deterministic decoder and is bypassed as an identity operation in the residual decoder. The Heun solver iterates from r0(k)=Z(k) to R1:T(k)=rN(k) . Symbols, shapes and repeated-module counts are defined in Table 1 .
Table 2 : Regional radar datasets used in this study. All datasets are converted to a common 9-frame input and 36-frame forecast protocol at 10 min cadence.
Extended Data Fig. 1 : US flash-flood precipitation forecast. The case is located in Tipton County, Tennessee, on 19 June 2025, and the forecast is initialised at 09:10 UTC. The observed sequence shows an organised heavy-precipitation band moving from the northwestern part of the forecast patch towards the southeast. The band remains relatively continuous during the first half of the 6 h window, before the main precipitation shifts towards the southern to southeastern part of the patch and locally weakens at later lead times. Embedded intense cores repeatedly reorganise during this evolution, making the case sensitive to rain-band displacement, core retention and late-lead structural evolution. All Extended Data case figures use the same five-panel layout: a , Geographic and radar context, with the forecast domain outlined in orange. b , Forecast diagnostics: CSIN in a 5×5 grid-cell neighbourhood at rain-rate thresholds of 32 and 64 mm h -1 , CRPS, CRPSS, and radially averaged PSD at lead times of 3 and 6 h. “Deterministic” denotes the output of the deterministic decoder of MW-Nowcast . c , Observed and predicted rain-rate fields at hourly lead times from 1 to 6 h. Rows show observations, PySTEPS, SimVPv2, NowcastNet and MW-Nowcast . d , CSIN and POD at thresholds of 32 and 64 mm h -1 for the four MW-Nowcast ensemble members and their PMM summary. e , The deterministic prediction and four ensemble members at lead times of 2, 4 and 6 h.
Extended Data Fig. 2 : US thunderstorm-wind convective precipitation forecast. The case is located in Rockland County, New York, on 14 July 2025, and the forecast is initialised at 19:30 UTC. The observed sequence shows convective precipitation moving generally eastward and becoming more organised into a coherent rain band during the first half of the forecast window. At later lead times, the precipitation coverage contracts, and the embedded intense cores weaken and become more fragmented. This case tests whether a model can retain rain-band structure during convective organisation while representing the subsequent contraction and decay of high-intensity cores. Panel descriptions are as in Extended Data Fig. 1 .
Extended Data Fig. 3 : US hail-associated convective precipitation forecast. The case is located in Douglas County, Nebraska, on 24 April 2025, and the forecast is initialised at 21:20 UTC. The observed sequence shows a convective echo field moving generally eastward and expanding across the forecast patch, with intense precipitation organised as several compact local cores rather than a single continuous rain band. These cores repeatedly develop, split and reorganise during the 6 h window, making the case sensitive to convective-core initiation, displacement and multi-core structural retention. Panel descriptions are as in Extended Data Fig. 1 .
Extended Data Fig. 4 : Mountainous heavy-rainfall forecast over western Hebei. The forecast patch is centred near Yi County, Baoding, and is initialised at 14:50 UTC on 24 July 2025. The observed sequence shows persistent heavy precipitation that remains active within the forecast patch and gradually expands, while the main rain area retains broad spatial continuity. Peak intensity weakens at later lead times, but embedded local intense cores continue to reorganise. This case tests whether a model can retain rain-area extent and structure in a persistent mountainous rainfall setting while representing the weakening and reorganisation of local intense cores. Panel descriptions are as in Extended Data Fig. 1 .
Extended Data Fig. 5 : Hunan Shimen mountainous heavy-rainfall forecast. This forecast window forms part of the May 2026 heavy-rainfall episode in Shimen County, Hunan, during which public reports documented six deaths and ten missing people. The forecast is initialised at 08:20 UTC on 17 May 2026. The observed sequence shows a persistent heavy-precipitation system moving generally eastward and expanding across the forecast patch. Embedded intense cores remain strong through the 6 h window and locally reorganise as the rain area expands. This case tests whether a model can retain high-intensity cores and coherent rain-area structure in a moving and expanding mountainous rainfall system. Panel descriptions are as in Extended Data Fig. 1 .
Extended Data Fig. 6 : Guizhou Guiding–Majiang heavy-rainfall forecast. This forecast window forms part of the 15–20 May 2026 rainstorm episode in Guizhou; an official disaster summary reported 19 people dead or missing across Guiding, Majiang and several other counties. The forecast is initialised at 22:10 UTC on 14 May 2026. The observed sequence shows a heavy-precipitation system moving generally eastward and gradually contracting after the middle of the forecast window. Although precipitation coverage decreases, embedded intense cores persist and locally reorganise. This case tests whether a model can retain the placement, intense cores and late-lead organisation of a moving heavy-rainfall system during contraction and structural adjustment. Panel descriptions are as in Extended Data Fig. 1 .
Extended Data Fig. 7 : European lightning-associated convective precipitation forecast. The case is located in southern Germany on 23 May 2024, and the forecast is initialised at 12:44 UTC. The observed sequence shows organised convective precipitation moving generally eastward and expanding through the first two-thirds of the forecast window. The precipitation area then contracts modestly, while embedded intense cores persist and locally reorganise along the main rain area. This case tests whether a model can retain rain-area continuity and intense-core structure in a moving and expanding convective system while representing the later modest contraction. Panel descriptions are as in Extended Data Fig. 1 .
Extended Data Fig. 8 : European tornado-associated convective precipitation forecast. The case is located in southern France on 3 March 2024, and the forecast is initialised at 00:30 UTC. The observed sequence shows a slowly moving precipitation system that expands through the 6 h window. Intense precipitation is organised mainly as multiple compact cores that repeatedly develop, split and reorganise, rather than as a stable single rain band. This case tests whether a model can represent local core initiation, displacement and morphology changes within a slowly moving but expanding convective system. Panel descriptions are as in Extended Data Fig. 1 .
Extended Data Fig. 9 : European heavy-precipitation forecast over southwestern France. The case is located in southwestern France on 8 June 2024, and the forecast is initialised at 18:00 UTC. The observed sequence shows a slowly moving heavy-precipitation area that expands substantially through the 6 h window, gradually forming a broader and more continuous rain region. Embedded intense cores continue to reorganise within the main rain area rather than rapidly moving out of the forecast patch. This case tests whether a model can retain rain-area extent, spatial continuity and local intense-core evolution in a slowly moving and steadily expanding heavy-precipitation system. Panel descriptions are as in Extended Data Fig. 1 .
Extended Data Fig. 10 : Regional rain-rate distributions and high-intensity-tail recovery. a–c , Conditional rain-rate densities pooled over 4,000 held-out events in the United States (US), Europe (EU) and China (CN), respectively. Densities use effective-rain pixels with 0.1≤R<128 mm h -1 . Blue solid curves show the deterministic branch, orange solid curves show the Probability Matched Mean (PMM) ensemble summary, and black dashed curves show ground truth. The shaded region marks the high-intensity tail, R≥16 mm h -1 , corresponding to the lower high-rain threshold used in the regional quantitative evaluation. Insets report pooled pixel–time exceedance probability conditional on the displayed rain-rate range. Across all three regions, PMM recovers substantially more high-intensity probability mass than the deterministic branch and closely approaches the observed tail.
Model
T+2 h
T+4 h
T+6 h
Mean
16
32
64
16
32
64
16
32
64
16
32
64
MW-Nowcast
0.284
0.214
0.160
0.197
0.129
0.085
0.095
0.055
0.032
0.256
0.194
0.147
Latent
0.216
0.147
0.099
0.155
0.100
0.062
0.072
0.038
0.021
0.204
0.144
0.103
OneStage
0.239
0.168
0.118
0.164
0.096
0.058
0.078
0.039
0.021
0.225
0.164
0.121
TwoStage
0.233
0.174
0.129
0.163
0.114
0.079
0.073
0.044
0.027
0.218
0.169
0.130
EDM-ODE
0.247
0.181
0.129
0.187
0.126
0.085
0.075
0.046
0.027
0.235
0.178
0.134
Supplementary Table 1 : CSIN 5 comparison of MW-Nowcast ablation variants on the held-out US test set. Neighbourhood Critical Success Index with a 5×5 neighbourhood is reported at rain-rate thresholds of 16, 32 and 64 mm h -1 for representative lead times T+2 h, T+4 h and T+6 h, together with the arithmetic mean of the unsmoothed per-lead scores over all 36 lead times from T+10 min to T+6 h. Higher is better. The best displayed value in each column is shown in bold.
Supplementary Fig. 1 : Visual comparison of MW-Nowcast ablation variants for a held-out US strong-precipitation case. The forecast is initialised at 09:00 UTC on 13 February 2025 for a patch centred near 29.4 ∘ N, 87.5 ∘ W. The figure includes the input context and the subsequent 6 h forecast evolution. Rows compare the observed fields, MW-Nowcast and the comparison variants Latent , OneStage , TwoStage , EDM-ODE and EDM-SDE . The case illustrates how the ablated design choices affect the retention of compact intense precipitation cores and organised storm structure over the forecast horizon.
Supplementary Fig. 2 : Flow-matching construction of the probabilistic residual for a US hail case. The case was reported at 23:07 UTC on 19 October 2024 near 33.87 ∘ N, 104.59 ∘ W, with the radar clip spanning 20:07 UTC on 19 October to 05:07 UTC on 20 October over a patch of approximately 31.29–36.45 ∘ N and 107.17–102.01 ∘ W. For each lead time, the top row shows the probabilistic residual evolving along the flow-matching trajectory from a Gaussian sample at τ=0.00 to its terminal state at τ=1.00 , and the bottom row shows the prediction formed by adding the evolving probabilistic residual to the deterministic prediction. The right-hand column gives the deterministic prediction (top) and the ground-truth observation (bottom).
Supplementary Fig. 3 : Flow-matching construction of the probabilistic residual for a European thunderstorm case. The radar clip spans 07:50–15:50 UTC on 15 May 2024 over the Alpine region, with the patch centred near 46.04 ∘ N, 8.67 ∘ E and covering approximately 43.38–48.50 ∘ N and 6.31–11.43 ∘ E. Layout follows Supplementary Fig. 2 : the top row of each panel traces the probabilistic residual along the flow-matching trajectory and the bottom row shows the resulting prediction, with the deterministic prediction and ground-truth observation in the right-hand column.
Supplementary Fig. 4 : Ensemble divergence of the probabilistic residual for a US hail case. The case was reported at 23:07 UTC on 19 October 2024 near 33.87 ∘ N, 104.59 ∘ W. Four members (seeds 1–4) are integrated along the flow from independent Gaussian samples. Section A : the probabilistic residual for each member at flow positions τ=0.75 (top row) and τ=1.00 (bottom row). Section B : the per-pixel ensemble mean (top) and standard deviation (bottom) of the probabilistic residual at τ=1.00 . The standard deviation localises along precipitation cores and edges, showing where the probabilistic residual contributes the most uncertainty.
Supplementary Fig. 5 : Ensemble divergence of the probabilistic residual for a European thunderstorm case. The case occurred on 15 May 2024 over the Alpine region, with the patch centred near 46.04 ∘ N, 8.67 ∘ E. Layout follows Supplementary Fig. 4 : Section A shows four members at τ=0.75 and τ=1.00 , and Section B shows the ensemble mean and standard deviation of the probabilistic residual at τ=1.00 .
Supplementary Fig. 6 : Pointwise detection and false-alarm characteristics across regions. Rows show the US, EU and CN test sets, and columns show rain-rate thresholds of 16, 32 and 64 mm h -1 . a , Probability of detection (POD). b , False alarm ratio (FAR). Both metrics are computed from pointwise threshold exceedances over lead times from T+10 min to T+6 h. Curves compare MW-Nowcast with NowcastNet, SimVPv2 and PySTEPS. Higher values are better for POD, whereas lower values are better for FAR.
Supplementary Fig. 7 : FSS sensitivity to neighbourhood size across regions. a , US; b , EU; c , CN. Within each regional block, the upper row reports FSS for 5×5 and 11×11 grid-cell neighbourhoods, and the lower row reports FSS for 17×17 and 25×25 neighbourhoods. For each neighbourhood size, columns show rain-rate thresholds of 16, 32 and 64 mm h -1 . Curves compare MW-Nowcast with NowcastNet, SimVPv2 and PySTEPS over lead times from T+10 min to T+6 h. Higher values indicate better neighbourhood-scale forecast skill.
Supplementary Fig. 8 : Structural and relative economic value diagnostics for the EU and CN regions. The upper row shows the EU test set and the lower row shows the CN test set. a , Radially averaged power spectral density (PSD) at T+2 h, T+4 h and T+6 h. The observed spectrum is shown together with forecast spectra from MW-Nowcast, NowcastNet, SimVPv2 and PySTEPS; closer agreement with the observed spectrum indicates better preservation of multiscale precipitation structure. b , Standard pixel-wise relative economic value (REV) at the same lead times as a function of cost–loss ratio. Solid and dashed curves denote rain-rate thresholds of 16 and 32 mm h -1 , respectively. Positive REV indicates lower expected expense than the optimal climatological strategy, and higher values indicate greater economic value under the corresponding cost–loss setting.
Region
Model
T+2 h
T+4 h
T+6 h
Mean
16
32
64
16
32
64
16
32
64
16
32
64
US
NowcastNet
0.192
0.144
0.070
0.075
0.037
0.004
0.038
0.016
0.002
0.171
0.129
0.064
SimVPv2
0.043
0.023
0.006
0.005
0.001
0.000
0.000
0.000
0.000
0.065
0.047
0.025
PySTEPS
0.089
0.043
0.022
0.037
0.017
0.009
0.026
0.014
0.007
0.099
0.064
0.045
MW-Nowcast
0.284
0.214
0.160
0.197
0.129
0.085
0.095
0.055
0.032
0.256
0.194
0.147
EU
NowcastNet
0.171
0.123
0.087
0.081
0.051
0.028
0.036
0.022
0.011
0.158
0.118
0.086
Supplementary Table 2 : Regional CSIN 5 comparison of baseline models. Neighbourhood Critical Success Index with a 5×5 neighbourhood is reported at rain-rate thresholds of 16, 32 and 64 mm h -1 for representative lead times T+2 h, T+4 h and T+6 h, together with the mean over all 36 lead times from T+10 min to T+6 h. Values are computed from the same regional evaluation data as the main-text quantitative comparison. Higher is better; the best value within each region and column is shown in bold.
Supplementary Fig. 9 : Convective heavy-precipitation forecast over the US Gulf Coast. The forecast is initialised at 05:30 UTC on 27 October 2025 for a patch centred near 30.50 ∘ N, 86.33 ∘ W. The observed sequence shows convective precipitation developing over the Gulf Coast region and generally expanding eastward to southeastward across the forecast patch. Rainfall intensity increases during the first half of the 6 h window, while the strongest precipitation is organised as multiple compact, discrete cores that merge, split and locally reorganise during the evolution.
Supplementary Fig. 10 : Convective heavy-precipitation forecast over the central US. The forecast is initialised at 06:00 UTC on 17 July 2025 for a patch centred near 38.17 ∘ N, 96.93 ∘ W. The observed sequence shows organised convective precipitation over the central US, with the rain area first advancing eastward to southeastward and then extending mainly towards the southern part of the patch. The high-intensity area broadens through the forecast window, but its embedded cores become increasingly fragmented, producing a multi-core structure rather than a single continuous rain band.
Supplementary Fig. 11 : Heavy-precipitation forecast over northern Italy. The forecast is initialised at 02:15 UTC on 7 July 2024 and is associated with a precipitation report near 45.95 ∘ N, 9.09 ∘ E in northern Italy. The observed sequence shows Alpine convective precipitation that remains concentrated near the main rain area rather than translating as a single coherent band. Localised intense echoes are repeatedly renewed and reorganised during the 6 h window, giving the case a compact but intermittently multi-core structure.
Supplementary Fig. 12 : Heavy-precipitation forecast over southern Germany. The forecast is initialised at 12:00 UTC on 3 June 2024 and is associated with a precipitation report near 47.62 ∘ N, 11.22 ∘ E in southern Germany. The observed sequence shows organised heavy precipitation along the Alpine region, with the main precipitation area remaining relatively coherent through the forecast window. Intense echoes persist within this broader system, while local cores are renewed and rearranged along the rain area instead of separating into fully isolated cells.
Supplementary Fig. 13 : Heavy-precipitation forecast over southern China. The forecast is initialised at 18:00 UTC on 17 June 2026 for a patch centred near 21.93 ∘ N, 112.05 ∘ E. The observed sequence shows developing heavy precipitation over southern China, with the rain area expanding mainly towards the eastern to northeastern part of the patch. The intense precipitation area strengthens through the first half of the forecast window and then forms a broader, semi-continuous core region with smaller embedded maxima that reorganise locally.
Supplementary Fig. 14 : Heavy-precipitation forecast over eastern China. The forecast is initialised at 11:10 UTC on 25 May 2026 for a patch centred near 31.77 ∘ N, 117.87 ∘ E. The observed sequence shows an organised precipitation system over eastern China that progresses eastward to southeastward across the forecast patch. The high-intensity precipitation expands and becomes more connected during the first half of the window, before the strongest region shifts downstream and breaks into several embedded cores at later lead times.
Supplementary Fig. 15 : Ablation forecast grid for the US strong-precipitation case initialised at 05:30 UTC on 27 October 2025. The patch is centred near 30.50 ∘ N, 86.33 ∘ W. Rows show the observation and the MW-Nowcast, Latent, OneStage, TwoStage, EDM-ODE and EDM-SDE variants; columns show the input context and forecast lead times through T+6 h.
Supplementary Fig. 16 : Ablation forecast grid for the US strong-precipitation case initialised at 06:00 UTC on 17 July 2025. The patch is centred near 38.17 ∘ N, 96.93 ∘ W. Layout follows Supplementary Fig. 15 : rows give the observation and the MW-Nowcast, Latent, OneStage, TwoStage, EDM-ODE and EDM-SDE variants, and columns give the input context and forecast lead times through T+6 h.
Precipitation nowcasting over the immediate 10-90 min period is important for flood management and real-time decision-making in urban regions. Conventional short-range forecasting with high-resolution numerical weather prediction requires frequent data assimilation, model initialization, and spin-up, introducing computational latency. Machine learning provides an alternative by learning storm evolution directly from high-frequency observations and producing forecasts quickly after training. This is particularly relevant for Mumbai, India, where monsoon convection, land-sea interactions, and localized intense rainfall make short-term prediction difficult. Here, we develop a compact radar-only nowcasting framework that combines multi-elevation reflectivity, Doppler radial velocity, and radial-velocity-gradient proxy features within an encoder-decoder U-Net. Using the most recent radar volume scan, the model predicts 12 future composite reflectivity fields at 7.5-min intervals up to 90 min lead time. The derived velocity magnitude, divergence-like, directional-shear, and vorticity-like channels represent kinematic signatures associated with convergence and boundary interactions without requiring full wind-field retrieval. A high-reflectivity attention module improves sensitivity to convective cores, and physics-guided attribution examines whether the learned sensitivities are meteorologically meaningful. The model is trained using Mumbai Doppler radar observations from May to August 2023 and evaluated on temporally independent events. At 90 min lead time, Critical Success Index values are 0.437, 0.332, and 0.193 for ≥10, ≥20, and ≥30 dBZ thresholds, respectively. Compared with persistence, the model gives lower RMSE and higher spatial correlation at longer lead times. Once trained, it runs on a standard computer, generating nowcasts within seconds for real-time use.
Akshay Sunil, Muhammed Rashid, Raja Sekhar Sivaraju +2
Centre for Climate Studies, Indian Institute of Technology Bombay, Mumbai, India · Regional Meteorological Centre (RMC) in Colaba, Mumbai, India
Precipitation nowcasting predicts the spatiotemporal evolution of future radar echoes from historical radar echo sequences, thereby estimating the occurrence, development, and movement of precipitation over the near term. In recent years, deep learning has become an important approach to precipitation nowcasting. Although state-of-the-art models can generally capture the overall spatial distribution of future precipitation, their predictions still exhibit substantial biases in radar echo intensity at individual locations. This observation motivates a more targeted strategy for reducing forecast errors. Instead of regenerating an entire radar echo sequence without spatial constraints, the predicted precipitation structure can be used to guide the refinement of echo intensities at individual locations. This structure-guided refinement directly targets echo intensity biases. Accordingly, we propose FreCast, a two-stage framework for radar echo prediction. The first stage generates an initial forecast of future radar echoes. The second stage uses the spatial structure of the initial forecast as a constraint to further correct intensity biases at individual locations in the first-stage prediction. Experiments on three datasets demonstrate that FreCast achieves consistent improvements across forecast skill metrics. Qualitative results further show that FreCast better preserves rainband continuity and intense precipitation structures at longer lead times.
Heping Fang, Zihuai Yin, Kaicheng Mao +2
Department of Statistics and Data Science, Southern University of Science and Technology, Shenzhen 518055, China · Department of Computer Science and Engineering, Southern University of Science and Technology, Shenzhen 518055, China · Guangdong Provincial Key Laboratory of Brain-Inspired Intelligent Computation, Department of Computer Science and Engineering and the Department of Statistics and Data Science, Southern University of Science and Technology, Shenzhen 518055, China
Proper short-term forecasting of precipitation is crucial in disaster management and preparedness. Nonetheless, the variability and nonlinearity of precipitation make short-term forecasting challenging for meteorologists. Moreover, capturing temporal dependencies in spatiotemporal data is a challenge in precipitation nowcasting. In this article, we introduce a lightweight deep learning model for half-hourly precipitation nowcasting. This model has been designed by incorporating the DenseNet architecture, residual connections, and transformer encoders for effective precipitation nowcasting with reduced model parameters. The North-Eastern region of India has been selected as the area of interest for our study. The region receives the highest precipitation during the months of June-September due to the monsoon season. The proposed model takes the previous five time-steps of half-hourly precipitation as inputs and predicts the precipitation in the next two half-hours. The GPM IMERG precipitation dataset with a 30-minute cadence has been used in this study for training and testing the model. The proposed architecture achieves best MAE of 0.235 millimetres, RMSE of 0.735 millimetres, and KGE score of 0.816 at an interval of 30 minutes.
Gourav Jyoti Kalita, Hidam Kumarjit Singh
Department of Electronics and Communication Technology, Gauhati University, Guwahati-781014, Assam, India