Physics-Based Deep Spatiotemporal Hyperlocal Radar Nowcasting with a Multi-Variable U-Net for High-Resolution Precipitation Forecasting

1

Physics-Based Deep Spatiotemporal Hyperlocal Radar Nowcasting with a Multi-Variable U-Net for High-Resolution Precipitation Forecasting

arXiv:2607.16080v1 [cs.LG] 17 Jul 2026

Akshay Sunil, Muhammed Rashid, Raja Sekhar Sivaraju, Sushma Nair, and Subimal Ghosh

Abstract—Precipitation nowcasting over the immediate 1090 min period is important for flood management and realtime 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 that limits rapid operational guidance. 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, directionalshear, and vorticity-like channels represent kinematic signatures associated with convergence and boundary interactions without requiring full wind-field retrieval. A high-reflectivity attention module is used to improve sensitivity to convective cores, and physics-guided attribution to examine 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. Index Terms—Nowcasting, Machine Learning, Radar-based Forecasting, Spatio-temporal U-Net, Feature Attribution.

I. I NTRODUCTION CCURATE precipitation prediction is fundamental for reducing weather-related risks in cities, particularly for flood management, aviation operations, emergency response, and transport planning. Within this broader prediction problem, nowcasting focuses on the immediate few minutes to hours ahead, typically up to about 0-90 min, when rainfall systems can evolve rapidly and decisions must be made with minimal lead time [1], [2]. This window is especially important for

A

Akshay Sunil, Subimal Ghosh, and Muhammed Rashid are with the Centre for Climate Studies, Indian Institute of Technology Bombay, Mumbai, India (e-mail: [email protected]; [email protected]; [email protected]). Raja Sekhar Sivaraju and Sushma Nair are with the Regional Meteorological Centre (RMC) in Colaba, Mumbai, India (e-mail: [email protected];[email protected]).

short-duration, high-intensity rainfall spells, which can trigger flash flooding in densely built urban catchments within less than an hour. In coastal megacities such as Mumbai, rapidly evolving monsoon convection, land-sea interactions, and localized cloudbursts can overwhelm stormwater drainage, inundate low-lying roads and underpasses, disrupt transport, and affect aviation operations. The extreme rainfall event of 26 July 2005 in Mumbai underscored the need for neighbourhood-scale nowcast guidance on where intense convective cells are likely to initiate, intensify, and move over the next one to two hours [3]. In addition, extreme rainfall over Mumbai exhibits strong intra-urban spatial variability, with earlier work showing that urban-induced intensification can occur in localized pockets and that rainfall correlations between stations may become statistically insignificant beyond distances of about 10 km [4]. Producing such guidance using radar-assimilated regional numerical models remains operationally challenging because frequent data assimilation, initialization, and model spin-up must be completed within a short decision window. A fast radar-based machine-learning framework therefore offers a practical route for generating localized nowcasts with the latency required for real-time urban response. At nowcast horizons, numerical weather prediction (NWP) faces structural limitations: kilometre-scale convection is highly sensitive to initial-condition errors and model spinup, while microphysics and boundary-layer parameterizations introduce biases. In addition, data assimilation and I/O latency can exceed decision windows for fast-evolving storms [5]. Even convection-allowing models can improve storm morphology yet still exhibit substantial timing and placement errors for convective initiation at 0-3 h lead times [6]. Weather radar remains the backbone of operational nowcasting because it provides high-resolution, near-real-time depictions of storm structure and motion. Classical radar-based approaches, including Tracking Radar Echoes by Correlation [7] and optical-flow extrapolation [8], estimate a motion field and advect reflectivity forward under a Lagrangian persistence assumption. While computationally efficient and skilful at very short lead times, these techniques degrade rapidly beyond 3060 min because they cannot represent convective initiation, growth, or decay [1]. Probabilistic ensemble schemes such as STEPS merge extrapolation with stochastic perturbations and downscaled NWP to quantify uncertainty, but they remain constrained by persistence [9]. Open-source libraries such as pysteps have standardized probabilistic extrapolation, verification, and benchmarking for

2

radar nowcasting [2]. Recent probabilistic and generative approaches, including DGMR, produce sharper ensembles with improved spatial structure and categorical skill relative to traditional extrapolation methods [10], [11]. Recent advances in deep learning have shifted radar nowcasting from extrapolation-based methods toward data-driven spatiotemporal prediction. Early models such as ConvLSTM demonstrated improved skill by jointly learning spatial and temporal evolution [12], while encoder-decoder architectures such as U-Net preserved fine-scale morphology through multiscale representations and skip connections [13]. Subsequent developments, including TrajGRU, PredRNN, attention-based encoder-forecasters, and large-context models such as MetNet and MetNet-2, further improved storm evolution, temporal coherence, and forecast skill across extended lead times [14]– [17]. More recently, physics-informed nowcasting frameworks have incorporated continuity constraints, transport-aware evolution, and conservation principles to improve long-lead stability, probabilistic calibration, and representation of convective evolution. Representative examples include NowcastNet, LUPIN, and ClimODE, which integrate physical constraints through complementary deterministic, stochastic, or continuous-time formulations [18]–[20]. Collectively, these studies demonstrate the benefits of embedding physical consistency within deep-learning nowcasting systems. Current radar nowcasting frameworks often rely on reflectivity alone and miss the kinematic signatures of convergence, shear, and boundary interactions that precede convective growth. Also, existing deep learning models frequently lack scientific interpretability, which limits operational trust. To address these gaps, in this study, we develop a radar-only deep learning nowcasting framework tailored to Mumbai’s monsoon convection. We construct multi-elevation, multi-variable input tensors that combine reflectivity, Doppler radial velocity, and physically motivated proxy predictors derived from spatial gradients of radial velocity. In contrast to reflectivity-only encoder-decoder baselines, this design injects dynamically informative cues into the learned representation while avoiding the need for full wind retrieval from single-Doppler measurements. The forecasting core is an encoder-decoder U-Net that maps the most recent radar volume scan to a 12-step composite-reflectivity forecast sequence at 7.5-min intervals (up to 90 min). To improve scientific interpretability and operational trust, we additionally apply physics-guided feature attribution to assess whether learned sensitivities concentrate on meteorologically meaningful precursor structures, such as boundary signatures and convergence-like patterns. In the present analysis, our goal is to develop a radar-based nowcasting framework for Mumbai using a machine learning approach. The proposed method is computationally efficient, requires minimal processing effort, and can be rapidly updated with each new radar scan generated every 7.5 minutes, offering strong potential for real-time applications. Building on this framework, we address four key objectives: (i) to demonstrate that multi-elevation, multi-variable inputs improve skill relative to reflectivity-only baselines across standard thresholds and lead times; (ii) to evaluate whether radial-velocity-gradient

proxy features improve representation of convective growth and moderate-to-heavy echoes; (iii) to show that attribution maps provide meteorologically consistent explanations that are useful for practitioners; and (iv) to verify that the gains are robust across temporally disjoint events and feasible within a 7.5-min operational update cycle. II. M ETHODS This section details the end-to-end methodology underpinning the radar-based nowcasting system. We first construct physics-aware inputs by transforming multi-elevation Doppler radar measurements into multi-channel tensors, including reflectivity, radial velocity, and radial-velocity-gradient kinematic proxy fields, and we define multi-step compositereflectivity targets at fixed lead times. We then describe a feed-forward U-Net that maps the most recent multi-elevation radar volume scan to a sequence of future reflectivity fields, augmented with a high-reflectivity attention mechanism to emphasize convective cores. The learning objective combines intensity-weighted reconstruction with structural similarity and threshold-aware terms to align training with both spatial realism and exceedances skill at operational reflectivity thresholds. We conclude with data preparation, channel-wise normalization, and temporally disjoint training/validation/test splits. A. Radar data & preprocessing Radar data for this study are sourced from the Doppler weather radar operated by the India Meteorological Department (IMD) at the Veravali site near Mumbai (≈19.1° N, 72.9° E). The C-band volume-scan radar typically performs 10 elevation sweeps per volume, providing reflectivity and Doppler radial-velocity measurements over the Mumbai metropolitan region and adjacent coastal/offshore areas. IMD surveillance extends to roughly 400-450 km; however, beam broadening, attenuation, and increasing sampling volume degrade effective resolution with distance. Accordingly, echoes within 250 km are retained and mapped to a 1 km × 1 km Cartesian grid. Preprocessing follows a standard pipeline for converting polar radar volumes into quality-controlled Cartesian products suitable for deep learning. Reflectivity and radial-velocity fields undergo quality control, including ground-clutter and anomalous-propagation suppression, masking of low signal-tonoise gates, and radial-velocity dealiasing using IMD operational procedures. Data at each elevation are then interpolated from polar coordinates onto a Cartesian grid using bilinear interpolation at 1 km spacing in both x and y, yielding a 501 x 501 domain (≈ 500 km × 500 km); scans with incomplete azimuthal coverage or severe artefacts are discarded [2], [11], [21]. Reflectivity is clipped to a physically plausible range (e.g., -5 to 55 dBZ) to limit outliers, while radial velocity is scaled to a fixed interval prior to feature derivation. For model input, each 2-D Cartesian field is resampled to 128 × 128 to match the network resolution [22]. Finally, volume scans are aligned to a regular 7.5-min temporal grid via linear interpolation in time, and sequences with gaps exceeding one scan interval are excluded from training and evaluation.

3

B. Physics-Aware Feature Engineering

C. Network Architecture

The forecasting framework employs physics-aware feature engineering to incorporate kinematic information from Doppler radar observations. This is motivated by theoretical and observational evidence that low-level convergence, shear, and rotation influence convective initiation, organization, and storm evolution [23]–[25]. For each radar elevation, six input channels are derived: horizontal reflectivity Z, radial velocity Vr , radial-velocity magnitude |Vr |, a divergence-like radialvelocity-gradient proxy S, a directional shear proxy θ, and a vorticity-like proxy ω.

The forecasting model is a feed-forward spatio-temporal U-Net, designed to predict the full reflectivity evolution sequence from a single input time slice. This architecture is chosen for its ability to simultaneously handle the spatial complexity of storm structures and the temporal evolution of convective systems, which is crucial for accurate short-term precipitation nowcasting. The encoder-decoder backbone consists of convolution-batch normalization-LeakyReLU blocks, with two downsampling stages to capture multi-scale spatial features and bilinear upsampling in the decoder to reconstruct high-resolution predictions. Skip connections are employed to preserve high-frequency storm structures across scales, which ensures that fine-grained details (such as the edges of convective cells) are maintained throughout the network. We developed a U-Net-based encoder-decoder model to forecast short-term radar reflectivity. The network ingests the most recent 128 × 128 gridded fields at time t0 of radar reflectivity (dBZ) and Doppler radial velocity (m s−1 ) sampled at the native radar scan cadence (≈ 7 min). To enrich the dynamical information available to the model, additional kinematic predictors are computed from the radial-velocity fields and concatenated with the raw inputs prior to encoding, including velocity magnitude |v|, a divergence-like gradient proxy ∂v/∂y, and a direction proxy derived from the gradient orientation defined by atan2(∂v/∂y, ∂v/∂x)/π. The encoder comprises two convolutional stages with 3 × 3 kernels and increasing channel depth (64 and 128). Each stage is followed by batch normalization and LeakyReLU activation, and max pooling is used for spatial downsampling. A 256-channel bottleneck layer captures compact, high-level representations of the input fields. The decoder reconstructs full resolution using bilinear upsampling followed by 3 × 3 convolutions, and integrates skip connections from the corresponding encoder stages to preserve fine-scale spatial structure. A final 1 × 1 convolution produces the forecast sequence, yielding 12 future reflectivity frames (≈ 90 min lead time at scan-time intervals). To emphasize intense precipitation regions that are most relevant for convective nowcasting, we incorporate a highreflectivity attention (HRA) module at the bottleneck and after the final decoding stage. Given a reflectivity feature representation, the attention map A is computed using global spatial pooling followed by a two-layer gating function:

Xe = [Ze , Vr,e , |Vr,e |, Se , θe , ωe ].

(1)

where Xe denotes the multi-channel input tensor for elevation e. The radial-velocity magnitude is defined as q 2 +V2 , (2) |Vr,e | = Vr,e θ,e where Vθ,e is the tangential velocity component at elevation e. The divergence-like radial-velocity-gradient proxy is computed as ∂Vr,e ∂Vr,e + . (3) Se = ∂x ∂y The directional shear proxy is defined as   ∂Vr,e ∂Vr,e 1 , . θe = atan2 π ∂y ∂x

(4)

which represents the orientation of the local radial-velocity gradient normalized to [-1,1]. The vorticity-like radialvelocity-gradient proxy is computed as ωe =

∂Vr,e ∂Vr,e − . ∂y ∂x

(5)

Spatial gradients are estimated on the Cartesian radar grid using central differences: ∂Vr,e Vr,e (i, j + 1) − Vr,e (i, j − 1) = . ∂x i,j 2∆x

(6)

∂Vr,e Vr,e (i + 1, j) − Vr,e (i − 1, j) = . ∂y i,j 2∆y

(7)

where ∆x and ∆y are the Cartesian grid spacing. Where necessary, a light smoothing filter is applied to Vr before differencing to reduce noise amplification. Together, these channels provide a compact representation of reflectivity structure and radial-velocity-derived kinematic signatures associated with convergence, divergence, shear, and rotation, all of which are relevant to convective initiation and storm growth. The six channels are constructed independently for each radar elevation and then stacked across Ne lev elevation angles, giving a total input depth of 6Nelev . The resulting multi-elevation tensor is resampled to a 128 × 128 spatial grid before being supplied to the network. The prediction targets are column-maximum reflectivity fields at 12 future lead times, sampled every 7.5 min and extending to a maximum lead time of 90 min.

A = σ W2 ReLU(W1 Z̄) .

(8)

where Z̄ is the spatially averaged reflectivity feature map, W1 and W2 are trainable parameters implemented as fully connected layers or equivalently 1 × 1 convolutions, and σ denotes the sigmoid activation. The attended feature tensor is then obtained through residual reweighting: F̂ = F ⊙ (1 + αA)

(9)

where F is the input feature tensor, ⊙ denotes element-wise multiplication, and αcontrols the strength of attention-based amplification. In this study, α is set to 1.2. By selectively

4

Fig. 1. U-Net architecture for radar-based weather nowcasting. Multi-elevation radar reflectivity and Doppler radial-velocity fields from the most recent volume scan (≈ 7 min scan cadence) are provided as input and augmented through feature engineering to derive kinematic descriptors, including velocity magnitude, divergence, direction, and vorticity. The network employs a two-level encoder-decoder structure with skip connections to preserve multiscale spatial information, while a bottleneck layer captures compact representations. During decoding, bilinear upsampling and feature concatenation progressively restore spatial resolution. A high-reflectivity attention module is applied after the final decoding stage to emphasize regions associated with intense precipitation. The model outputs T future reflectivity frames corresponding to a 90-min forecast horizon at approximately scan-time (≈ 7 min) intervals.

amplifying features associated with high reflectivity, the HRA mechanism increases the network’s sensitivity to convective cores and improves the reconstruction and prediction of rapidly evolving storm structures. D. Loss Functions Training minimizes a composite objective that balances amplitude fidelity, spatial structure, and operational skill at precipitation thresholds: L = λiw LiwMSE + λSSIM (1 − SSIM) + λthr Lthr .

(10)

where SSIM denotes the Structural Similarity Index Measure, which quantifies the similarity between predicted and observed reflectivity fields in terms of luminance, contrast, and structural consistency. This formulation combines an amplitude-focused reconstruction term, a structural similarity term, and a threshold-aware term, so that training is aligned with both pixel-wise accuracy and operational verification metrics. The weighting coefficients λiw , λSSIM , and λthr are chosen such that no single component dominates the gradient,

yielding forecasts that retain realistic storm morphology while maintaining skill at application-relevant reflectivity thresholds. Through the reflectivity-dependent weighting in LiwMSE , the loss places larger penalties on errors in moderate and high reflectivity regions (≥ 25–35 dBZ), so the optimizer concentrates capacity on convective cells and heavy rain rather than weak, low-impact echoes. To emphasize convective echoes, the reconstruction error is weighted by reflectivity in dBZ. The weights are computed in physical reflectivity space, while the errors are evaluated in the model’s normalized space for numerical stability. Let Zt (x, y) and Ẑt (x, y) denote the observed and predicted reflectivity fields at lead time t = 1, . . . , T . The reflectivity-aware weight is defined as  α0 , Z < 15 dBZ,    α , 15 ≤ Z < 25 dBZ, 1 w(Z) = (11)  α2 , 25 ≤ Z < 35 dBZ,    α3 , Z ≥ 35 dBZ with α = {1.0, 1.3, 1.8, 2.5}.

5

The intensity-weighted mean squared error is then computed over all forecast lead times and spatial grid points as

LiwMSE =

T W H
1 XXX w ZtdBZ (x, y) T HW t=1 x=1 y=1  2 × Ẑtnorm (x, y) − Ztnorm (x, y) .

(12)

where “dBZ” and “norm” denote the physical and normalized domains, respectively. The affine normalization is inverted only for computing the reflectivity-dependent weight w(·). E. Training Configuration Training data are drawn from Doppler radar observations collected over Mumbai between May and August 2023, with an 80–10–10 split for training, validation, and testing. Input variables are standardized by channel to ensure consistent scaling across the dataset. During training, data augmentation techniques such as random cropping and horizontal flipping are applied to increase model robustness to spatial variability. The model is trained using the Adam optimizer (learning rate 1 × 10−4 ) with early stopping based on validation SSIM to prevent overfitting. Additionally, a learning rate scheduler is used to adjust the learning rate during training to enhance convergence and optimize training efficiency. The model runs on an NVIDIA GPU (CUDA-enabled) with a batch size of 8 and generates reflectivity predictions every 7.5 min up to 90 min ahead. Validation performance is continuously monitored using metrics such as SSIM and the Critical Success Index (CSI), ensuring that the model not only fits well but also generalizes effectively. Test data, withheld from training and validation, are used to assess the model’s ability to generalize to unseen radar scans and to evaluate performance on a realworld test set. F. Evaluation metrics Model performance was assessed using both categorical and continuous verification metrics. Categorical skill was evaluated at reflectivity thresholds of ≥ 10, ≥ 20, and ≥ 30 dBZ using the Critical Success Index (CSI), Equitable Threat Score (ETS), frequency bias, Probability of Detection (POD), False Alarm Ratio (FAR), and False Positive Rate (FPR). Let H, M , F , and CN denote hits, misses, false alarms, and correct negatives, respectively. These metrics are defined as H H +M +F

(13)

H − Hr H + M + F − Hr

(14)

(H + M )(H + F ) H + M + F + CN

(15)

Bias =

H +F H +M

(16)

POD =

H H +M

(17)

CSI = ETS = Hr =

FAR =

F H +F

(18)

FPR =

F F + CN

(19)

Continuous performance was evaluated using the domainaveraged root-mean-square error (RMSE) and spatial correlation of composite reflectivity: v u N u1 X (Pi − Oi )2 RMSE = t N i=1

(20)

PN r = qP

i=1 (Pi − P̄ )(Oi − Ō)

N 2 i=1 (Pi − P̄ )

qP

N 2 i=1 (Oi − Ō)

(21)

where Pi and Oi denote predicted and observed reflectivity values at grid point i, respectively; P̄ and Ō are their spatial means; and N is the total number of grid cells. To assess scale-dependent structural realism, radially averaged power spectral density (PSD) was compared between forecasts and observations at selected lead times. Observed event support was also reported to provide context on threshold-dependent sample prevalence.

III. R ESULTS This section reports the evaluation of the proposed radar nowcasting framework on an independent test set. We summarize performance across reflectivity thresholds using categorical and continuous verification metrics, examine how skill changes with lead time relative to a persistence baseline, and assess scale dependent structure using spectral and neighbourhood verification. Representative case diagnostics are also provided to illustrate dominant error modes.

A. Overall nowcasting performance and intensity dependence The spatiotemporal U-Net was evaluated on an independent test dataset using standard categorical and continuous verification metrics across reflectivity thresholds of ≥ 10, ≥ 20, and ≥ 30 dBZ. Unless otherwise noted, reported values are aggregated across all 12 forecast steps (7.5-90 min). The model shows strongest skill for weaker and moderate echoes, with performance gradually decreasing as the reflectivity threshold increases, consistent with the greater difficulty of predicting compact and rapidly evolving convective cores. Across thresholds, the model maintains useful categorical skill while falsealarm behaviour remains controlled, indicating good discrimination between precipitating and non-precipitating regions despite pronounced class imbalance. Overall, the results indicate robust performance for weak-to-moderate precipitation, with reduced skill for the most intense echoes.

6

Fig. 2. Radar nowcasting verification diagnostics for an extreme precipitation event on 25 September 2024 (lead = 85 min). This specific high-impact case was selected as a stress test for the model; on this day, rapid convective intensification delivered over 100 mm of rainfall in a few hours, triggering an IMD Red Alert and severe urban flooding in Mumbai. Composite reflectivity fields, Z (dBZ), are shown over the study domain with coastline overlay and latitude/longitude axes. The top row shows (a) the last input frame at t0 , (b) the verifying observation at t0 + 85 min, and (c) the model prediction for the same valid time; all panels use an identical reflectivity color scale. The bottom row presents (d) the spatial error field, ∆Z = Zpred − Ztruth (dBZ), where positive (negative) values indicate overprediction (underprediction); (e) marginal reflectivity distributions over finite-overlap pixels for the truth and prediction; and (f) the pixel-wise truth–prediction joint density (hexbin; log-scaled counts) for pixels with truth ≥ 10 dBZ, with the 1:1 reference line. The footer reports categorical verification at the 10 dBZ exceedance threshold (hits, misses, false alarms, and correct negatives) and derived scores (POD, FAR, FPR, CSI, and frequency bias), together with continuous intensity errors (MAE and MSE).

B. Spectral characteristics (PSD) and loss of small-scale power To assess whether forecasts preserve spatial structure across spatial scales, we compare the radially averaged power spectral density (PSD) of the predicted fields against the verifying observations at selected lead times shown in Figure 4, spanning approximately 8, 31, 62, and 85 min. At 8 min, the model spectrum is close to the observed spectrum at larger wavelengths, while showing reduced power at shorter wavelengths, indicative of smoothing and attenuation of finescale variability. With increasing lead time (≈31, ≈62, and

≈85 min), the deficit in short-wavelength power becomes more pronounced, consistent with progressive weakening of sharp gradients and compact convective features. The optical-flow baseline generally retains more short-wavelength power than the model, although both methods diverge increasingly from the observed spectrum as lead time increases. Overall, the PSD diagnostics complement the categorical metrics by indicating that forecast degradation is associated primarily with loss of small-scale structure and intensity damping, rather than the introduction of spurious variability.

7

Fig. 3. Radially averaged power spectral density (PSD) of composite reflectivity (10logR) as a function of wavelength for the observations, optical-flow extrapolation, and the proposed model at selected lead times of approximately 8, 31, 62, and 85 min.

C. Lead time dependence of categorical skill To further examine the temporal robustness of the proposed framework, we next evaluate how categorical skill varies with forecast lead time across weak, moderate, and intense reflectivity thresholds. Figure 5 shows that the proposed model maintains useful categorical skill throughout the 90min forecast horizon, with 90-min CSI values of 0.437, 0.332, and 0.193 for the ≥ 10, ≥ 20, and ≥ 30 dBZ thresholds, respectively. Forecast skill decreases gradually with increasing lead time, indicating stable temporal behaviour rather than an abrupt loss of predictability. For ≥ 20 dBZ, CSI declines from 0.501 at 7.5 min to 0.387 at 60 min, and to 0.332 at 90 min. ETS (Equitable Threat Score, which corrects CSI for random hits) follows the same pattern, decreasing from 0.484 at 7.5 min to 0.367 at 60 min and 0.314 at 90 min. Frequency bias for ≥ 20 dBZ remains above unity at all lead times (for example, 1.168 at 7.5 min and 1.228 at 90 min), indicating persistent overprediction of areal coverage for moderate echoes even as CSI decreases. For ≥ 10 dBZ, CSI decreases from 0.596 at 7.5 min to 0.496 at 60 min, and to 0.437 at 90 min, with bias remaining close to unity (approximately 1.07–1.10), indicating only slight overcoverage. In contrast, intense echoes (≥ 30 dBZ) show the strongest degradation and clear underprediction at long lead times. CSI decreases from 0.423 at 7.5 min to 0.283 at 60 min, and to 0.193 at 90 min. Frequency bias drops from 0.723 at 7.5 min to 0.492 at 90 min. This confirms that the primary limitation at extended lead times is the representation of compact high-reflectivity cores. D. Continuous verification against a persistence baseline We next examine lead time dependence using continuous verification against a persistence baseline (Fig. 5), based on domain-averaged RMSE and spatial correlation of composite reflectivity. Persistence is most competitive at the shortest horizon: at 7.5 min, it achieves RMSE 2.315 dBZ and correlation 0.897, compared with 3.411 dBZ and 0.846 for the model. This short-lead advantage of persistence is expected and consistent with the literature, since no learned model can improve upon the trivial first-step extrapolation before storm structures have had time to evolve. By approximately 31 min, the model becomes competitive with persistence, with slightly

lower RMSE (4.275 dBZ versus 4.310 dBZ) and higher spatial correlation (0.774 versus 0.748). Beyond roughly 35 to 40 min, persistence degrades more rapidly. At approximately 54 min, the model shows lower RMSE (4.416 dBZ versus 4.801 dBZ) and higher correlation (0.768 versus 0.712), and by 85 min the separation increases further, with model RMSE 4.746 dBZ and correlation 0.714 relative to persistence RMSE 5.461 dBZ and correlation 0.629. Overall, the model maintains stronger spatial agreement and lower error at longer horizons, indicating improved representation of evolving storm structure relative to pure persistence. E. Station-Wise Multi-Event Validation To further evaluate operational applicability, the proposed framework was additionally validated against rainfall observations from 35 automatic weather stations across the Mumbai metropolitan region during three independent high-impact rainfall events (17–19 August 2025). The model reproduced the observed spatial distribution of moderate-to-heavy rainfall and showed a gradual reduction in station-wise agreement with increasing lead time. Detailed station-wise comparisons and lead-time correlation analyses are provided in the Supplementary Material (Section S1). Overall, the station-wise validation across multiple extremerainfall days demonstrates that the proposed framework effectively captures moderate-to-heavy rainfall variability at the station scale. The model shows consistent performance across independent high-impact events, with strong agreement in moderate-rainfall regimes and controlled underestimation in extreme cases. These results reinforce the model’s generalization capability and highlight its suitability for urban-scale operational nowcasting, where accurate spatial representation of rainfall distribution is critical. In addition, the lead-time correlation analysis shows that station-wise agreement degrades gradually across successive 15-minute windows rather than collapsing abruptly, indicating stable temporal behaviour across all three validation days. IV. D ISCUSSION The results indicate that fusing multi-elevation radar observations with Doppler radial velocity–derived kinematic proxy channels within a U-Net-style encoder–decoder can

8

Fig. 4. Lead-time evolution of categorical verification metrics for the proposed model and a persistence baseline across reflectivity thresholds of ≥ 10, ≥ 20, and ≥ 30 dBZ. Panels show (a) Critical Success Index (CSI), (b) Equitable Threat Score (ETS), (c) frequency bias, and (d) observed event support as functions of forecast lead time. Solid lines denote the proposed model, and dashed lines denote persistence. The model retains useful skill across the 90-min forecast horizon, with the highest performance at lower reflectivity thresholds and increasing underprediction for intense echoes at longer lead times.

yield robust short-range precipitation nowcasts in a monsoondominated coastal environment. Traditional extrapolation methods are effective at very short lead times, but skill often degrades beyond ∼ 30–60 min because Lagrangian persistence cannot represent convective initiation, intensification, or decay [1], [7], [8]. Deep learning nowcasting models can partially address this limitation by learning nonlinear structure and evolution from radar archives, as demonstrated by ConvLSTM sequence learners and encoder–decoder families such as UNet [12], [13]. Consistent with prior work, our evaluation shows gradual skill degradation with lead time, controlled false-alarm behaviour, and near-neutral frequency bias for

light-to-moderate echoes. These characteristics suggest that the model captures storm-scale organization and displacement while limiting spurious precipitation, a behaviour commonly reported for deterministic radar nowcasts that preserve coherent morphology even as intensity errors grow with forecast horizon [16], [17]. A pronounced intensity dependence remains evident. Skill decreases at higher reflectivity thresholds, and intense convective cores are underpredicted at longer lead times. Such attenuation and smoothing of sharp reflectivity gradients are a widely recognized limitation of deterministic point prediction in radar nowcasting, particularly for rare, compact extremes that evolve

9

Fig. 5. Lead-time evolution of continuous verification metrics for composite reflectivity, comparing the proposed model (solid) with a persistence baseline (dashed): (a) domain-averaged RMSE (dBZ) and (b) spatial correlation between forecast and observation. Persistence is most competitive at the shortest lead times, while the proposed model retains lower RMSE and higher spatial correlation at longer horizons.

rapidly through nonlinear growth and decay processes [1], [10], [16], [17]. Importantly, the dominant error mode in the present study is weakening or missing of intense echoes rather than widespread false echoes, as reflected by persistently low false positive rates and high specificity across thresholds. This conservative error behaviour is operationally meaningful in applications such as urban flood early warning and aviation support, where false alarms can be costly and where reliable spatial guidance is often preferred over aggressive intensity amplification. In addition to categorical verification, the spectral diagnostics provide an objective view of scale-dependent realism. Fig. 3 shows that the radially averaged power spectral density progressively under-represents short-wavelength variance with increasing lead time, indicating smoothing and loss of finescale structure, while larger-scale variability is comparatively better retained. This behaviour is consistent with a known limitation of deterministic encoder–decoder nowcasting, where optimization with pixel-wise objectives tends to damp sharp gradients and compact convective maxima. The PSD result therefore supports the interpretation that the dominant error mode is intensity attenuation and boundary smoothing rather than widespread generation of spurious echoes, and it motivates future extensions that better preserve small-scale variability (e.g., scale-aware losses, multi-scale supervision, or probabilistic/generative formulations). Within the broader methodological landscape, the proposed framework occupies a complementary position. Large-context models such as MetNet and MetNet-2 improve regional consistency through multiscale context aggregation [16], [17]. Generative and physics-constrained approaches, including DGMR, NowcastNet, LUPIN, and ClimODE, focus on probabilistic forecasting and physically constrained evolution to enhance uncertainty representation, calibration, and long-lead forecast skill [10], [18]–[20]. In contrast, the present study emphasizes a compact, deterministic, radar-only architecture that incorporates radial-velocity-gradient proxy channels without requiring full wind retrieval. The ability to preserve mesoscale structure, orientation, and coherent displacement at extended lead times suggests that these proxies regularize the learned

evolution during organized monsoon convection, consistent with evidence that convergence- and shear-related signatures contain useful information for convective initiation and storm evolution [23], [24], [26], [27]. A further objective of this study is to improve operational trust through interpretability. The attribution analysis links model sensitivity to recognizable storm-scale features, including convergence-line signatures and gust-front-like boundaries, and verifies that the network learns physically meaningful cues rather than artefacts, consistent with recent work in interpretable machine learning for weather and climate applications [28], [29]. Case-based diagnostics (Fig. 2) further show that the model reproduces the location, orientation, and spatial extent of organized systems, with errors primarily arising from intensity damping and modest displacement rather than widespread spurious precipitation. Despite these strengths, underprediction of intense cores remains the primary limitation and motivates clear directions for improvement. Promising pathways include hybrid deterministic–stochastic formulations, explicit evolution constraints, and training objectives that better emphasize rare extremes while preserving the conservative false-alarm behaviour observed here [10], [18]. Incorporating the proposed kinematic proxy representation into probabilistic or generative frameworks may be particularly beneficial, as it offers a mechanism to retain the spatial reliability and dynamical structure captured by the deterministic model while improving representation of high-intensity convection. V. C ONCLUSIONS This study presents a compact, radar-only deep learning framework for short-range precipitation nowcasting tailored to Mumbai’s monsoon convection. The approach fuses multielevation reflectivity with Doppler radial velocity and radialvelocity–gradient proxy channels within a spatiotemporal UNet, and produces 12-step composite-reflectivity forecasts at 7.5 min intervals up to approximately 90 min lead time. The proxy representation is designed to encode convergence- and shear-related kinematic signatures from single-Doppler observations without requiring retrieval of a full two-dimensional

10

wind field, while a high-reflectivity attention mechanism improves sensitivity to convective cores. Evaluation on temporally disjoint events shows that forecast skill degrades gradually with lead time, while false alarms remain controlled across thresholds, indicating stable behaviour beyond the ∼ 30–60 min range where extrapolationbased methods typically deteriorate. The model retains useful categorical skill throughout the forecast horizon, with 90min CSI values of 0.437, 0.332, and 0.193 at ≥ 10, ≥ 20, and ≥ 30 dBZ, respectively. Performance remains intensity dependent, with the strongest skill at lower thresholds and systematic underprediction of compact high-reflectivity cores at longer lead times. Continuous verification further shows that although persistence is strongest at the shortest lead time, the proposed model achieves lower RMSE and higher spatial correlation at longer horizons, indicating improved representation of evolving storm structure. Physics-guided attribution indicates that the network’s sensitivity is concentrated in physically interpretable regions, highlighting kinematic precursor patterns near developing cells and along boundary-like features consistent with convergence lines and gust fronts. Together, these findings support the value of physics-aware, multi-variable radar inputs for improving deterministic nowcast robustness and interpretability in a tropical coastal setting. Future work should focus on improving extreme-core prediction through probabilistic or hybrid deterministic–stochastic formulations, incorporating explicit evolution or continuity constraints, and refining objective functions and sampling strategies to better represent rare, high-intensity convection while preserving the conservative false-alarm characteristics demonstrated here. ACKNOWLEDGMENTS This project has received funding and support from the HDFC ERGO - IIT Bombay Innovation Lab, a partnership between HDFC ERGO General Insurance Company Ltd. and Indian Institute of Technology Bombay committed to support innovation and entrepreneurship in insurance and related areas. R EFERENCES [1] J. W. Wilson, N. A. Crook, C. K. Mueller, J. Sun, and M. Dixon, “Nowcasting thunderstorms: A status report,” Bulletin of the American Meteorological Society, vol. 79, pp. 2079–2099, 1998. [2] S. Pulkkinen, D. Nerini, A. Pérez Hortal, C. Velasco-Forero, U. Germann, A. Seed, and L. Foresti, “Pysteps: an open-source Python library for probabilistic precipitation nowcasting (v1.0),” Geoscientific Model Development, vol. 12, pp. 4185–4219, 2019. [3] S. Mohanty, M. Swain, R. Nadimpalli, K. K. Osuri, U. C. Mohanty, P. Patel, and D. Niyogi, “Meteorological conditions of extreme heavy rains over Mumbai,” Journal of Applied Meteorology and Climatology, vol. 62, pp. 191–208, 2023. [4] S. Paul, S. Ghosh, M. Mathew, A. Devanand, S. Karmakar, and D. Niyogi, “Increased spatial variability and intensification of extreme monsoon rainfall due to urbanization,” Scientific Reports, vol. 8, p. 3918, 2018. [5] P. Bauer, A. Thorpe, and G. Brunet, “The quiet revolution of numerical weather prediction,” Nature, vol. 525, pp. 47–55, 2015. [6] A. J. Clark, W. A. Gallus Jr., M. Xue, and F. Kong, “Convectionallowing models for forecasting convective weather: An overview and status report,” Bulletin of the American Meteorological Society, vol. 97, pp. 1703–1723, 2016.

[7] R. E. Rinehart and E. T. Garvey, “Three-dimensional storm motion detection by conventional weather radar,” Nature, vol. 273, pp. 287– 289, 1978. [8] N. E. Bowler, C. E. Pierce, and A. W. Seed, “Development of a precipitation nowcasting algorithm based upon optical flow techniques,” Journal of Hydrology, vol. 288, pp. 74–91, 2004. [9] ——, “Steps: A probabilistic precipitation forecasting scheme which merges extrapolation nowcast and downscaled NWP,” Quarterly Journal of the Royal Meteorological Society, vol. 132, pp. 2127–2155, 2006. [10] S. Ravuri et al., “Skilful precipitation nowcasting using deep generative models,” Nature, vol. 597, pp. 672–677, 2021. [11] A. Augst and M. Hagen, “Interpolation of operational radar data to a regular Cartesian grid exemplified by Munich’s airport radar configuration,” Journal of Atmospheric and Oceanic Technology, vol. 34, pp. 495–510, 2017. [12] X. Shi, Z. Chen, H. Wang, D.-Y. Yeung, W.-K. Wong, and W.-C. Woo, “Convolutional LSTM network: A machine learning approach for precipitation nowcasting,” in Advances in Neural Information Processing Systems, vol. 28, 2015. [13] O. Ronneberger, P. Fischer, and T. Brox, “U-Net: Convolutional networks for biomedical image segmentation,” in Medical Image Computing and Computer-Assisted Intervention – MICCAI 2015, ser. Lecture Notes in Computer Science, vol. 9351, 2015, pp. 234–241. [14] X. Shi, Z. Gao, L. Lausen, H. Wang, D.-Y. Yeung, W.-K. Wong, and W.-C. Woo, “Deep learning for precipitation nowcasting: A benchmark and a new model (TrajGRU),” 2017. [15] Y. Wang, M. Long, J. Wang, Z. Gao, and P. S. Yu, “Predrnn: Recurrent neural networks for predictive learning using spatiotemporal lstms,” in Proceedings of the 31st International Conference on Neural Information Processing Systems (NIPS’17), ser. NIPS’17. Red Hook, NY, USA: Curran Associates, Inc., 2017, pp. 879–888, location: Long Beach, California, USA. [16] C. K. Sønderby, L. Espeholt, J. Heek et al., “MetNet: A neural weather model for precipitation forecasting,” 2020. [17] R. Lam, A. Sanchez-Gonzalez, M. Willson et al., “GraphCast: Learning skillful medium-range global weather forecasting,” 2022. [18] Y. Zhang, M. Long, J. Chen et al., “Skilful nowcasting of extreme precipitation with NowcastNet,” Nature, vol. 619, pp. 526–532, 2023. [19] P. Pavlı́k, M. Výboh, A. Bou Ezzeddine, and V. Rozinajová, “Fully differentiable Lagrangian convolutional neural network for continuity consistent physics informed precipitation nowcasting,” 2024. [20] Y. Verma, M. Heinonen, and V. Garg, “ClimODE: Climate and weather forecasting with physics-informed neural ODEs,” in Proceedings of the International Conference on Learning Representations (ICLR 2024), 2024. [21] K. Ośródka and J. Szturc, “Quality-based generation of weather radar Cartesian products,” Atmospheric Measurement Techniques, vol. 8, pp. 2173–2181, 2015. [22] S. Ochoa-Rodriguez, L.-P. Wang, A. Gires, R. D. Pina, R. ReinosoRondinel, G. Bruni et al., “A review of radar–rain gauge data merging methods and their potential for hydrological applications,” Water Resources Research, vol. 55, pp. 6356–6391, 2019. [23] T. M. Weckwerth and D. B. Parsons, “A review of convection initiation and motivation for IHOP 2002,” Monthly Weather Review, vol. 134, pp. 5–22, 2006. [24] J. R. Mecikalski and K. M. Bedka, “Forecasting convective initiation by monitoring the evolution of moving cumulus in daytime GOES imagery,” Monthly Weather Review, vol. 134, pp. 49–78, 2006. [25] P. Markowski and Y. Richardson, Mesoscale Meteorology in Midlatitudes. Chichester, UK: John Wiley & Sons, 2010. [26] J. R. Mecikalski, W. M. MacKenzie, R. A. Kohrs, C. P. Jewett, T. G. Chronis, and G. J. Jedlovec, “Aviation applications for satellitebased observations of convective initiation,” Bulletin of the American Meteorological Society, vol. 96, pp. 181–196, 2015. [27] Y. Zeng, G. Huang, Q. Meng, J. Sun, and Y. Zhang, “Nowcasting convective storm initiation with machine learning methods,” Weather and Forecasting, vol. 34, pp. 1035–1052, 2019. [28] I. Ebert-Uphoff and K. Hilburn, “Evaluation, tuning, and interpretation of neural networks for working with images in meteorological applications,” Bulletin of the American Meteorological Society, vol. 101, pp. E2149–E2170, 2020. [29] R. Yang, J. Hu, Z. Li, J. Mu, T. Yu, J. Xia, X. Li, A. Dasgupta, and H. Xiong, “Interpretable machine learning for weather and climate prediction: A review,” Atmospheric Environment, vol. 338, p. 120797, 2024.