Source-linked AI summary
SmaAt-UNet: Precipitation Nowcasting using a Small Attention-UNet Architecture
Kevin Trebing, Tomasz Stanczyk, Siamak Mehrkanoon
TL;DR
Short-term weather nowcasting requires responsive forecasts from recent observations, whereas numerical weather prediction is computationally demanding. The paper proposes SmaAt-UNet, a UNet-based model with attention and depthwise-separable convolutions, and evaluates it on precipitation and cloud-cover data. It reports comparable performance to larger models using one quarter of the original UNet’s parameters, with implications for smaller and more energy-efficient deployment.
Problem
Numerical weather prediction is less suitable for forecasts from minutes to 6 hours, motivating responsive nowcasting from the latest observations.
Method
SmaAt-UNet extends UNet with attention modules and depthwise-separable convolutions for precipitation and cloud-cover nowcasting.
Results
A quarter of the original UNet’s parameter size is achieved while maintaining comparable performance.
Takeaways & Limitations
The smaller model supports potential smartphone deployment and more resource-efficient precipitation forecasting.
Takeaways & Limitations
MSE-guided predictions can be blurry because the loss favors expected pixel values across multiple possible future outcomes.
Abstract
from arXiv · showhide
Weather forecasting is dominated by numerical weather prediction that tries to model accurately the physical properties of the atmosphere. A downside of numerical weather prediction is that it is lacking the ability for short-term forecasts using the latest available information. By using a data-driven neural network approach we show that it is possible to produce an accurate precipitation nowcast. To this end, we propose SmaAt-UNet, an efficient convolutional neural networks-based on the well known UNet architecture equipped with attention modules and depthwise-separable convolutions. We evaluate our approaches on a real-life datasets using precipitation maps from the region of the Netherlands and binary images of cloud coverage of France. The experimental results show that in terms of prediction performance, the proposed model is comparable to other examined models while only using a quarter of the trainable parameters.
1. Introduction
Numerical weather prediction provides accurate forecasts but is computationally expensive and less suitable for minute-to-hour nowcasting. The paper proposes SmaAt-UNet, a compact attention-based CNN for precipitation and cloud-cover forecasting.
- Numerical weather prediction models use atmospheric physics but require complex computations and substantial computing power.They are commonly run on expensive supercomputers.
- Nowcasting targets forecasts from minutes to 6 hours using the latest observational weather data.This makes nowcasting more responsive than numerical weather prediction for rapidly changing atmospheric conditions.
- Data-driven neural networks learn from historical weather data rather than calculating the atmosphere’s underlying physics.Examples of learned inputs include past wind speed and precipitation maps.
- SmaAt-UNet predicts precipitation on a high-resolution grid 30 minutes ahead and is also applied to cloud-cover nowcasting.Its precipitation inputs are radar images showing accumulated rainfall over time.
- SmaAt-UNet combines the UNet architecture with attention modules and depthwise-separable convolutions.These design choices target an efficient convolutional architecture for image-based weather nowcasting.
- A quarter of the original UNet’s parameter size is achieved while maintaining comparable performance.The smaller model could support precipitation forecasting on small computation units and reduce training and computational demands.
2. Related Work
Prior precipitation-nowcasting work includes recurrent models with explicit memory, UNet-based models, and the persistence baseline. SmaAt-UNet extends UNet with encoder attention while using depthwise-separable convolutions to reduce parameters.
- LSTM-based nowcasting models retain input information across multiple time steps, supporting time-series prediction.Convolutional and spatiotemporal LSTM variants target spatial and temporal correlations in image sequences.
- UNet-based precipitation models concatenate past precipitation maps and predict future rain-intensity classes without explicit memory modeling.Earlier implementations classified four or many rain-intensity categories one hour ahead.
- Persistence predicts the next image by copying the sequence’s last input image.It is difficult to outperform when consecutive observations are separated by only a few minutes.
- Attention mechanisms amplify wanted signals and suppress unwanted ones by weighting channels and spatial regions.SmaAt-UNet uses CBAMs that apply channel attention followed by spatial attention.
- SmaAt-UNet is reported as the first UNet-based architecture to include the CBAM mechanism.
- Depthwise-separable convolutions split convolution into depthwise and pointwise operations to reduce parameter counts.This reduces model complexity while retaining the basic convolutional processing structure.
3. Methods
The method extends UNet with channel-and-spatial attention and depthwise-separable convolutions, while using multiscale encoder-decoder processing to predict pixel-level values.
- UNet architecture: UNet uses an encoder-decoder structure with max-pooling, double convolutions, bilinear upsampling, skip-connections, and a final 1 × 1 convolution.The encoder halves image size while doubling feature maps; the decoder reverses this process and outputs one feature map.
- SmaAt-UNet modifications: SmaAt-UNet adds CBAM attention to the encoder and replaces regular convolutions with depthwise-separable convolutions.These are the model’s two modifications to the original UNet architecture.
- Attention mechanism: CBAM applies attention sequentially across channels and spatial dimensions, with modules placed after the first double convolution and at every encoder.Attention features feed the corresponding upsampling paths through skip-connections while preserving original features in encoder inputs.
- Parameter reduction: Depthwise-separable convolutions replace all original UNet convolutions to reduce the number of parameters, while CBAMs retain regular convolutions.The implementation uses depthwise-separable convolutions with two kernels-per-layer.
- Experimental setup: The compared models are original UNet, UNet with CBAM, UNet with DSCs, and SmaAt-UNet, trained with early stopping and a learning-rate scheduler.Training used a maximum of 200 epochs, although early stopping occurred in all training iterations.
- Model evaluation: The loss is mean squared error, supplemented by threshold-based precision, recall, accuracy, F1-score, CSI, FAR, and HSS evaluation metrics.For precipitation maps, thresholded scores use rainfall greater than 0.5mm/h, while cloud-cover data are already binarized.
4. Experiments
Experiments use radar precipitation maps from the Netherlands and binary cloud-cover images from France, with preprocessing and task-specific sampling choices for nowcasting evaluation.
- Precipitation map dataset: The precipitation dataset contains about 420,000 rain maps at 5-minute intervals from 2016–2019 over the Netherlands and neighboring countries.The maps come from two C-band Doppler weather radar stations operated by KNMI.
- Precipitation preprocessing: The precipitation maps are normalized by the highest training-set value and center-cropped to reduce no-data regions outside radar range.The additional center crop is 288 pixels.
- Precipitation nowcasting task: Each precipitation-map input contains 12 stacked maps representing one hour of observations, and the target is a 288 × 288 rainfall-intensity map 30 minutes later.The task predicts exact rainfall intensities for every pixel.
- Dataset sampling: Training uses NL-50 targets containing at least 50% rainy pixels, while NL-20 with at least 20% rainy pixels provides an additional performance indicator.These subsets address the prevalence of images with little or no rain, though they contain fewer samples than the original dataset.
- Evaluation caveat: The test set may yield worse performance than data resembling the training distribution because predominantly rainy maps can bias models toward predicting more rain.This caveat concerns the precipitation-map test set.
- Cloud-cover dataset: The cloud-cover dataset contains 256 × 256 binary images of France, with four input frames and six ground-truth frames spaced 15 minutes apart.The network predicts the probability of cloud presence at each pixel, without normalization or cropping.
5. Results and Discussion
Across precipitation and cloud-cover nowcasting, SmaAt-UNet achieves performance comparable to UNet-based alternatives while using roughly one quarter of the parameters. Results also show attention maps focus on different input characteristics, although MSE-guided precipitation outputs can be blurry.
- Precipitation map results: Every tested model outperformed the persistence baseline by a large margin on the precipitation map dataset.This is notable because small input-to-target changes can make persistence difficult to beat.
- Precipitation map results: 0.0127 versus 0.0122 MSE: UNet with DSCs performed worse than original UNet but better than UNet with CBAM on the precipitation map dataset.Using DSCs alone reduced the network size to one quarter of the original UNet.
- Qualitative analysis: MSE-guided precipitation predictions were blurry because minimizing mean squared error favors pixelwise expected values across possible future frames.The described example shows this blurriness across all examined models.
- Precipitation map results: SmaAt-UNet achieved second-best performance on six of seven precipitation-map scores while using 1/4 of UNet’s parameters.The original UNet performed best on most scores.
- Precipitation map results: SmaAt-UNet and UNet ranked closely on NL-20, with UNet slightly better in NMSE and both outperforming the other compared models.The lower MSE on NL-20 is attributed to the prevalence of near-zero precipitation values, making NMSE a better comparison metric.
- Attention visualization: Attention feature maps progressively decreased in resolution and focused on different input characteristics, with the final 18 × 18 maps identifying coarse rain patterns.Examples include maps focusing on rain clusters or regions with little to no rain.
- Cloud-cover results: On cloud cover, model scores were comparable; SmaAt-UNet achieved the best Recall while using approximately 1/4 the parameters of UNet-based larger models.UNet with CBAM had the lowest MSE and most best scores, while the overall score differences remained minor.
6. Conclusion
SmaAt-UNet is a smaller, attentive UNet that matches larger architectures on precipitation nowcasting while supporting more resource-conscious deployment.
- SmaAt-UNet is a smaller and attentive version of the UNet architecture.
- SmaAt-UNet performs on par with substantially larger architectures on a precipitation nowcasting task.
- Small, efficient SmaAt-UNet models can support smartphone applications that forecast precipitation using the latest available data at users’ request.
- Energy-efficient architectures such as SmaAt-UNet can reduce carbon footprint and support sustainable machine learning practices.