Precipitation nowcasting has attracted considerable attention for its role in real-time weather prediction. Operational systems have traditionally relied on two major sources of guidance. NWP models provide physically grounded atmospheric evolution but can be less responsive at very short lead times because of initialization, spin-up, and computational constraints^[1]. Radar echo extrapolation methods, including TREC-based motion estimation^[2] and enhanced optical-flow techniques^[3], directly infer echo displacement from recent observations and are therefore widely used for immediate forecasts. In practice, seamless nowcasting systems often blend extrapolation with NWP so that observation-driven forecasts dominate early lead times and model guidance contributes increasingly at longer lead times^[4].

Learning-based methods offer an alternative by training neural models to infer precipitation evolution from historical radar sequences. In recent years, data-driven nowcasting has evolved from early sequence modeling approaches toward more expressive architectures that aim to better capture precipitation motion, local structure, and uncertainty.

Existing nowcasting models mainly fall into three categories. Recurrent architectures, such as ConvLSTM^[5], TrajGRU^[6], PredRNN^[7], PredRNN++^[8], and MIM^[9], model temporal evolution through hidden-state transitions and are effective for motion-aware prediction. Encoder-decoder approaches, including U-Net^[34], RainNet^[35], and related radar forecasting models^[36], preserve spatial detail through multi-scale feature extraction and skip connections. More recent large-context and generative methods, such as MetNet^[20], DGMR^[29], NowcastNet^[30], and DiffCast^[31], further improve realism and forecasting quality by enlarging the receptive context or modeling uncertainty explicitly.

Despite this progress, accurately identifying salient echo regions and maintaining coherent multi-frame evolution remain challenging, especially for efficient radar-based nowcasting systems. These issues are particularly relevant when a model must preserve sharp local precipitation structures without sacrificing temporal stability across future predictions.

Attention mechanisms have become an important tool for improving spatiotemporal prediction models by emphasizing informative regions and suppressing irrelevant responses. In precipitation nowcasting, this capability is particularly valuable because key echo structures often occupy only a limited portion of the scene, while surrounding areas may contain clutter or weak background patterns. Attention-based designs have been explored in both recurrent and non-recurrent forecasting models. For example, SwinLSTM^[10] incorporates transformer-style attention into recurrent prediction, Tau^[18] introduces temporal attention for efficient spatiotemporal predictive learning, and Earthformer^[19] further demonstrates the effectiveness of space-time attention mechanisms in Earth system forecasting.

Beyond global or transformer-style attention, lightweight plug-in modules have also been widely studied for efficient feature refinement. CBAM^[33] sequentially applies channel attention and spatial attention, while ECA-Net^[37] shows that efficient channel attention alone can provide meaningful gains with minimal additional complexity. Compared with heavier attention mechanisms, such modules are attractive for radar nowcasting because they can enhance localized precipitation features, storm boundaries, and high-value echo regions without substantially increasing computation.

Similarity-based objectives from representation learning encode structure by preserving similarity among related samples rather than relying solely on reconstruction. In time-series settings, Zhang et al.^[38] examined what makes such objectives effective for forecasting, and SimTS^[39] showed that simple similarity-based objectives improve forecasting-oriented sequence representations. These studies indicate that similarity constraints can help sequential models learn smoother and more discriminative temporal dynamics.

Temporal consistency has also been used as an explicit regularizer in video and sequence modeling. Lai et al.^[40] enforce frame-to-frame consistency to stabilize per-frame processed video; temporal cycle-consistency learning^[41] aligns embeddings across time through a differentiable consistency objective, and TCR has been applied to improve stability in domain-adaptive video segmentation^[42]. A common theme across these works is that constraining the relationship between neighboring frames encourages temporally coherent outputs, which is particularly relevant for radar echo forecasting, where small inconsistencies between adjacent predictions may accumulate and lead to unrealistic precipitation evolution.

The proposed method is built upon SimVP, which learns spatial and temporal dependencies from sequential radar echo images. As shown in Figure 1, the model takes a sequence of historical radar frames as input and predicts multiple future precipitation frames. The framework contains three main components: an encoder for hierarchical spatial feature extraction, a spatiotemporal module for latent dynamics modeling, and a decoder for reconstructing the future sequence. CBAM is inserted into the feature transformation pipeline to refine informative channel-wise and spatial responses. During training, the predicted sequence is supervised by a pixel-wise reconstruction loss and an auxiliary temporal consistency loss between adjacent predictions.

More specifically, let the input radar sequence be $$ X \in \mathbb{R}^{B \times T \times C \times H \times W} $$, where $$ B $$ is the batch size, $$ T $$ is the number of historical frames, and $$ C $$, $$ H $$, and $$ W $$ denote the channel number, height, and width, respectively. The input is first reshaped into $$ BT $$ individual frames and processed by a spatial encoder. The encoder consists of $$ N_S $$ stacked convolutional blocks with alternating strides generated by the pattern $$ [1, 2, 1, 2, \ldots] $$, so that spatial downsampling is introduced in every second block. Each block uses a $$ 3 \times 3 $$ convolution, followed by Group Normalization and a LeakyReLU activation. This design keeps the backbone lightweight while still expanding the receptive field through progressive spatial compression.

Formally, after reshaping $$ X $$ into $$ \tilde{X} \in \mathbb{R}^{BT \times C \times H \times W} $$, the spatial encoder produces

(1) $$ \begin{equation} E^{(l)} = \phi_l(E^{(l-1)}), \quad l=1, \ldots, N_S, \end{equation} $$

where $$ E^{(0)}=\tilde{X} $$ and $$ \phi_l(\cdot) $$ denotes the $$ l $$-th spatial encoding block. The final latent feature is $$ E^{(N_S)} \in \mathbb{R}^{BT \times C' \times H' \times W'} $$, and the first-stage feature $$ E^{(1)} $$ is retained as a shallow skip representation for decoding.

After spatial encoding, the latent tensor is rearranged back to the sequence form and passed to a middle spatiotemporal translator. The latent sequence is reshaped from $$ \mathbb{R}^{B \times T \times C' \times H' \times W'} $$ to $$ \mathbb{R}^{B \times (TC') \times H' \times W'} $$ so that temporal interaction can be modeled in a channel-compressed latent space. Each translation block first applies a $$ 1 \times 1 $$ convolution to mix channel information and then uses multiple grouped convolution branches with kernel sizes $$ \{3, 5, 7, 11\} $$ to capture multi-scale dynamics. The outputs of these branches are summed to obtain a richer latent representation. A U-shaped latent encoder-decoder with skip connections across depth is then used to aggregate both local and large-scale motion patterns without relying on recurrent units or self-attention.

Let

(2) $$ \begin{equation} Z^{(0)} = \text{Reshape}\left(E^{(N_S)}\right) \in \mathbb{R}^{B \times (TC') \times H' \times W'}. \end{equation} $$

For each multi-scale translation block, the hidden feature is computed as

(3) $$ \begin{equation} Z^{(m)} = \sum\limits_{k \in \{3, 5, 7, 11\}} g_k\left(\psi\left(Z^{(m-1)}\right)\right), \end{equation} $$

where $$ \psi(\cdot) $$ denotes the $$ 1 \times 1 $$ channel mixing convolution and $$ g_k(\cdot) $$ denotes a grouped convolution with kernel size $$ k $$. The latent encoder-decoder structure further introduces skip connections, so the decoder-side latent update can be written as

(4) $$ \begin{equation} \hat{Z}^{(m)} = h_m\left([\hat{Z}^{(m-1)}; Z^{(M-m)}]\right), \end{equation} $$

where $$ [\cdot;\cdot] $$ denotes channel concatenation and $$ h_m(\cdot) $$ is the $$ m $$-th decoder-side multi-scale block.

Finally, the decoder mirrors the encoder by using the reversed stride schedule to progressively recover spatial resolution. The first-stage encoder feature is preserved as a shallow skip connection and is concatenated with the last decoder feature before the final reconstruction block. A $$ 1 \times 1 $$ readout convolution then maps the hidden representation back to the predicted radar frames. Overall, the backbone comprises a spatial encoder, a multi-scale latent translator, and a skip-connected spatial decoder, providing an efficient basis for the subsequent attention and temporal-regularization enhancements.

The decoder output is therefore obtained as

(5) $$ \begin{equation} D^{(l)} = \varphi_l(D^{(l-1)}), \quad l=1, \ldots, N_S-1, \end{equation} $$

(6) $$ \begin{equation} \hat{Y} = \rho \left( \varphi_{N_S}\left([D^{(N_S-1)}; E^{(1)}]\right) \right), \end{equation} $$

where $$ D^{(0)}=\text{Reshape}^{-1}(\hat{Z}^{(M)}) $$, $$ \varphi_l(\cdot) $$ denotes the $$ l $$-th decoder block, and $$ \rho(\cdot) $$ is the final $$ 1 \times 1 $$ readout convolution. The model output satisfies $$ \hat{Y} \in \mathbb{R}^{B \times T \times C \times H \times W} $$.

The encoder and decoder follow the general structure of an encoder-decoder network, with CBAM embedded to enhance attention to informative features. The encoder gradually downsamples the input sequence and extracts hierarchical representations, while the decoder upsamples the encoded features to generate future radar frames. This design preserves the computational simplicity of SimVP while strengthening the network's sensitivity to salient echo intensity patterns, storm boundaries, and localized precipitation cores.

The encoder uses repeated convolutional blocks, where stride-1 layers preserve the current spatial resolution, and stride-2 layers perform downsampling. The decoder follows the reverse process with upsampling blocks to progressively restore the original resolution. This alternating downsampling and upsampling strategy reduces computation while retaining a direct shallow skip path from the first encoder stage to the final reconstruction stage. Such a design helps recover fine precipitation structures that might otherwise be lost in the latent transformation process.

CBAM is applied sequentially with a channel-attention stage followed by a spatial-attention stage. Given an input feature map $$ x \in \mathbb{R}^{C \times H \times W} $$, the refined feature can be written as

(7) $$ \begin{equation} x' = \mathcal{M}_{\text{channel}}(x) \otimes x, \end{equation} $$

(8) $$ \begin{equation} x'' = \mathcal{M}_{\text{spatial}}(x') \otimes x', \end{equation} $$

where $$ \mathcal{M}_{\text{channel}}(\cdot) $$ denotes channel attention, $$ \mathcal{M}_{\text{spatial}}(\cdot) $$ denotes spatial attention, and $$ \otimes $$ denotes element-wise multiplication.

For channel attention, global average pooling and global max pooling are first applied to $$ x $$, and the pooled descriptors are passed through a shared two-layer bottleneck with a ReLU activation:

(9) $$ \begin{equation} \mathcal{M}_{\text{channel}}(x) = \sigma \left( \text{MLP}(\text{AvgPool}(x)) + \text{MLP}(\text{MaxPool}(x)) \right), \end{equation} $$

where $$ \sigma(\cdot) $$ is the sigmoid function. In this process, the pooled channel descriptors are transformed by a shared bottleneck and then broadcast back to the original spatial resolution.

For spatial attention, channel-wise average pooling and max-pooling are computed on the channel-refined feature $$ x' $$, concatenated along the channel dimension, and processed by a convolution layer:

(10) $$ \begin{equation} \mathcal{M}_{\text{spatial}}(x') = \sigma \left( f^{k \times k} \left( [\text{AvgPool}_{c}(x'); \text{MaxPool}_{c}(x')] \right) \right), \end{equation} $$

where $$ [\cdot ; \cdot] $$ denotes channel concatenation and $$ f^{k \times k} $$ is a convolution with an odd kernel size $$ k $$. This stage uses a convolution to generate the spatial attention map, which is then expanded and multiplied with the input feature to obtain the final CBAM output.

To provide additional temporal regularization during training, a cosine consistency loss is applied across multiple predicted frames. The loss is computed between adjacent outputs so that neighboring predictions remain close in representation space while the mean squared error (MSE) term still anchors each frame to the ground truth. This regularization is particularly useful for multi-step nowcasting, where small inconsistencies introduced in early predictions may accumulate and lead to unrealistic echo evolution later in the sequence.

The temporal consistency loss is defined as

(11) $$ \begin{equation} \mathcal{L}_{\text{tc}} = \sum\limits_{i=1}^{T-1} \text{CosineEmbeddingLoss}(\hat{Y}_i, \hat{Y}_{i+1}), \end{equation} $$

where $$ \hat{Y}_i $$ and $$ \hat{Y}_{i+1} $$ are predictions at time steps $$ i $$ and $$ i+1 $$, respectively, and $$ T $$ is the total number of prediction steps. The cosine embedding loss can be written as

(12) $$ \begin{equation} \text{CosineEmbeddingLoss}(x, y) = 1 - \frac{x \cdot y}{\|x\| \|y\|}, \end{equation} $$

where $$ x $$ and $$ y $$ are flattened feature vectors from two adjacent predictions.

The overall training objective combines the MSE loss with the temporal consistency loss:

(13) $$ \begin{equation} \mathcal{L}_{\text{total}} = \mathcal{L}_{\text{MSE}} + \lambda \cdot \mathcal{L}_{\text{tc}}, \end{equation} $$

where $$ \lambda $$ controls the relative contribution of TCR. In all main experiments, $$ \lambda $$ is set to 0.1 unless otherwise specified.

The reconstruction term preserves pixel-level fidelity to the target radar field, while the temporal consistency term regularizes the trajectory of the predicted sequence in feature space. As a result, the model is encouraged not only to minimize frame-wise errors, but also to generate temporally coherent future evolution.

Following prior work, the continuous radar sequences are divided into multiple events. Each event contains 25 frames, and the first 5 frames are used to predict the subsequent 20 frames. For SEVIR, this setting corresponds to forecasting the next 100 min from the previous 25 min of observations because the temporal interval is 5 min per frame. For Shanghai-Radar, where the frame interval is 6 min, the same protocol corresponds to forecasting the next 120 min from the previous 30 min of observations. All frames are resized to $$ 128 \times 128 $$ pixels before training and evaluation.

The model is implemented in PyTorch. All models are trained on a server equipped with an Intel(R) Xeon(R) Gold 6448H @ 2.40 GHz CPU, eight NVIDIA A40 GPUs, and 1 TB of RAM. We train each model for 200 epochs using the AdamW optimizer with an initial learning rate of $$ 1\times10^{-3} $$, a weight decay of 0.01, a cosine learning-rate decay schedule, and a total batch size of $$ 4\times8=32 $$. The temporal consistency weight is set to $$ \lambda=0.1 $$ for the main results. All compared methods follow the same input-output protocol to ensure a fair comparison.

The models are assessed using CSI, HSS, bias score (BS), LPIPS, SSIM, temporal gradient error (TGE), and MSE. For threshold-based metrics, TP, FP, TN, and FN are computed after binarizing the predicted and ground-truth radar fields with the same intensity threshold. For Shanghai-Radar, we use reflectivity thresholds of 20, 30, 35, and 40 dBZ, with 35 dBZ as the primary strong-convection threshold. For SEVIR, we follow common VIL evaluation practice and use thresholds of 16, 74, 133, 160, 181, and 219. Unless otherwise specified, the main categorical analysis reports the mean score across the corresponding threshold set. Among these metrics, the pooled CSI variants (CSI-POOL4 and CSI-POOL16) are additionally reported in the main quantitative comparison, whereas BS and TGE are used in the ablation analysis to examine forecast bias and temporal coherence, respectively.

(14) $$ \begin{equation} \text{CSI} = \frac{TP}{TP + FN + FP}, \end{equation} $$

where CSI is also known as the threat score (TS),

(15) $$ \begin{equation} \text{HSS} = \frac{2(TP \cdot TN - FN \cdot FP)}{(TP + FN)(FN + TN) + (TP + FP)(FP + TN)}, \end{equation} $$

(16) $$ \begin{equation} \text{BS} = \frac{TP + FP}{TP + FN}, \end{equation} $$

where BS measures whether a model over-forecasts or under-forecasts event frequency. A value closer to 1 indicates a better balance between predicted and observed events. CSI-POOL4 and CSI-POOL16 are pooled CSI variants computed after applying spatial max-pooling with kernel sizes 4 and 16, respectively. These pooled scores relax small spatial displacement errors and evaluate whether the predicted precipitation event occurs in the neighboring region.

(17) $$ \begin{equation} \text{TGE} = \frac{1}{T-1}\sum\limits_{t=1}^{T-1}\left\|(\hat{Y}_{t+1}-\hat{Y}_{t})-(Y_{t+1}-Y_t)\right\|_1, \end{equation} $$

where TGE directly measures whether the temporal change between adjacent predicted frames matches the observed temporal change. Lower TGE indicates better temporal coherence relative to the ground truth.

(18) $$ \begin{equation} \text{LPIPS}(I_p, I_t) = \sum\limits_l w_l \left\| \phi_l(I_p) - \phi_l(I_t) \right\|_2^2, \end{equation} $$

where $$ \phi_l $$ denotes the features extracted from the $$ l $$-th layer of a deep model,

(19) $$ \begin{equation} \text{SSIM}(x, y) = \frac{(2\mu_x\mu_y + C_1)(2\sigma_{xy} + C_2)}{(\mu_x^2 + \mu_y^2 + C_1)(\sigma_x^2 + \sigma_y^2 + C_2)}, \end{equation} $$

and

(20) $$ \begin{equation} \text{MSE} = \frac{1}{N} \sum\limits_{i=1}^{N} (x_i - y_i)^2. \end{equation} $$

The results on the Shanghai-Radar and SEVIR datasets are summarized in Tables 1 and 2. Across most evaluation metrics, the proposed model outperforms the baselines, indicating superior forecasting skill and greater structural fidelity.

On the Shanghai-Radar dataset, the proposed model achieves the best results on all reported metrics. Relative to the SimVP baseline, it raises CSI from 0.3614 to 0.4038 and HSS from 0.5083 to 0.5419, while lowering LPIPS from 0.3691 to 0.2893 and MSE from 29.6566 to 29.4490. The pooled scores improve in parallel (CSI-POOL4 from 0.3803 to 0.4384 and CSI-POOL16 from 0.4257 to 0.5207), and the gains over the strongest competing baselines on CSI and SSIM further indicate that the proposed design better preserves local echo structures and fine-grained precipitation patterns rather than simply smoothing the output.

On the SEVIR dataset, the proposed model again achieves the best performance on CSI, CSI-POOL4, CSI-POOL16, HSS, SSIM, and MSE, while obtaining the second-best LPIPS. It reaches a CSI of 0.2765 and an HSS of 0.3622, exceeding both the SimVP baseline (0.2644 and 0.3411) and the strongest competing baseline, PredRNN++ (0.2663 and 0.3430), and attains the lowest MSE of 445.03. Its LPIPS of 0.3494 trails only Earthformer (0.3458), whose categorical skill is far lower (CSI 0.1882), so the small LPIPS gap does not reflect better forecasting ability. Because SEVIR contains more diverse storm morphologies and broader scene variability than Shanghai-Radar, these results indicate that the proposed framework generalizes effectively beyond a single regional radar distribution.

The representative examples in Figures 2 and 3 illustrate the different visual characteristics of the two datasets. The Shanghai-Radar samples exhibit relatively concentrated regional echo structures, whereas SEVIR contains more diverse large-scale storm morphologies. This contrast highlights the value of evaluating the model on both a regional radar dataset and a more heterogeneous benchmark.

To isolate the contribution of each component, we compare the SimVP baseline, SimVP with CBAM, SimVP with TCR, and the full ConCast model that combines both, as reported in Tables 3 and 4. This component-isolation protocol directly tests whether the attention module, the temporal regularization term, and their combination each contribute to final performance.

The component-isolation results in Tables 3 and 4 let us attribute the gains to each module separately. Relative to the SimVP baseline, adding CBAM yields a larger improvement in spatial skill and structural fidelity: on Shanghai-Radar, it lifts CSI from 0.3614 to 0.3876 and cuts LPIPS from 0.3691 to 0.3157, clearly exceeding the CSI of 0.3789 obtained by adding TCR alone. Adding TCR instead yields a pronounced reduction in TGE, from 4.86 to 4.31 vs. 4.68 for CBAM on Shanghai-Radar and from 18.92 to 17.35 vs. 18.21 on SEVIR, and brings the BS closer to one. The two effects are therefore largely complementary rather than redundant: CBAM sharpens where echoes are placed, whereas TCR constrains how predictions evolve between adjacent frames. The full model combines both effects and attains the best balanced performance on both datasets, achieving the highest CSI and HSS, the lowest LPIPS and TGE, and a BS closest to one. The SimVP+TCR variant, in particular, improves TGE over both SimVP and SimVP+CBAM, which isolates the contribution of the temporal regularization term independently of the attention module.

Because forecast skill in nowcasting degrades with longer lead times, we report the lead-time-wise performance across all 20 forecast steps in Figure 4. All metrics deteriorate monotonically as the horizon extends, confirming that the later steps are intrinsically harder. On Shanghai-Radar, the proposed model holds a clear and consistent advantage over SimVP in both CSI and HSS throughout the sequence, and the gap widens at the longer lead times where the baseline degrades fastest, while the two MSE curves remain close. On SEVIR, the CSI curves of the two models nearly overlap, but the proposed model retains a small edge in HSS and, more notably, yields a visibly lower MSE that grows with the lead time, indicating that the temporal regularization mainly suppresses the accumulation of pixel-level error at the far horizon.

The temporal consistency weight $$ \lambda $$ in Equation (13) balances the reconstruction term against the temporal regularization term. We analyze its effect on the Shanghai-Radar dataset in Table 5. Skill scores improve as $$ \lambda $$ increases from zero and peak around $$ \lambda=0.1 $$, the value used in the main results. Beyond this point, the BS keeps decreasing below one, and TGE starts to rise again while CSI, HSS, and SSIM degrade, consistent with the concern that an overly strong consistency penalty over-smooths adjacent predictions and freezes the predicted echo evolution. A moderate value of $$ \lambda=0.1 $$ therefore provides the best trade-off between forecast skill and temporal coherence.

Since real-time nowcasting systems require both accuracy and efficiency, we further compare model complexity and inference cost. The results in Table 6 show that the proposed model introduces almost no increase in parameter count, FLOPs, or inference time relative to the SimVP baseline, while increasing GPU memory usage because of the additional attention operations. Even with this overhead, the model stays competitive with most recurrent and transformer-based alternatives, so the added attention does not undermine the deployment advantages of the underlying backbone.

Made substantial contributions to the conception and design of the study, data analysis and interpretation, and drafting and revision of the manuscript: Zhao, P.; Wang, R.; Zhang, Y.; Yang, X.; Li, C.

All authors read and approved the final manuscript.

The data supporting this study are publicly available. The Shanghai-Radar dataset is available from Harvard Dataverse at https://dataverse.harvard.edu/dataset.xhtml?persistentId=doi:10.7910/DVN/2GKMQJ and is also mirrored at https://drive.google.com/file/d/14JB4ElkZKHzqxIGKMFrbnY2P4zcae8RA/view. The SEVIR dataset is available at https://github.com/MIT-AI-Accelerator/neurips-2020-sevir, with an example tutorial at https://github.com/MIT-AI-Accelerator/eie-sevir/blob/master/examples/SEVIR_Tutorial.ipynb. SEVIR can also be downloaded from the public AWS S3 bucket according to the official instructions provided by the dataset authors.

During the preparation of this manuscript, the AI tool ChatGPT (GPT-5.5, released 2026-04-24) was used solely for language editing. The tool did not influence the study design, data collection, analysis, interpretation, or the scientific content of the work. All authors take full responsibility for the accuracy, integrity, and final content of the manuscript.

This paper was funded by Natural Science Foundation of Shandong Province (No.ZR2025MS997).

Zhao, P. is affiliated with China Energy Investment Corporation Co., Ltd., while the other authors have declared that they have no conflicts of interest.

Not applicable.

Not applicable.

© The Author(s) 2026.