What problem does this paper attempt to address?

The problem that this paper attempts to solve is to improve the accuracy of precipitation forecasting while reducing the consumption of computational resources and their impact on the environment. Specifically: 1. **Accurate precipitation forecasting**: Accurate precipitation forecasting is crucial for decision - making in multiple industries such as agriculture, transportation, and energy production. Extreme weather events (such as storms, floods, and droughts) may bring serious humanitarian and economic losses if they cannot be effectively predicted and prepared for. 2. **Computational resources and environmental impact**: Traditional physics - based numerical weather prediction (NWP) models, although having high precision, require huge computational resources, which brings environmental problems. In recent years, data - driven deep - learning methods have shown potential in weather forecasting, especially for short - term forecasting. However, the training cost of these models increases with the increase in the number of parameters, raising concerns about environmental costs. To solve these problems, the author proposes an efficient deep - learning method, using an improved U - Net architecture for precipitation forecasting. This model aims to provide a low - resource and efficient alternative and can serve as the basis for future weather forecasting research. ### Main contributions: - **Model structure**: The Small Attention U - Net (SmaAt - Unet), an improved U - Net architecture, is adopted, which reduces the amount of computation through depth - wise separable convolution (DSC) and channel - wise attention mechanism (CBAM). - **Data processing**: Pre - processed multi - spectral satellite image data, including operations such as filtering non - rainfall samples, central cropping, and resizing, to improve the learning efficiency of the model. - **Experimental results**: Evaluated on the dataset of the Weather4cast 2023 challenge, the proposed model outperforms the baseline model in terms of the Critical Success Index (CSI), and the training time and computational cost are significantly reduced. In conclusion, this paper aims to balance model performance and energy efficiency, especially in the current situation where deep - learning models are becoming increasingly complex, emphasizing the importance of considering the model's ecological footprint.

Efficient Baseline for Quantitative Precipitation Forecasting in Weather4cast 2023

Efficient spatio-temporal weather forecasting using U-Net

RainAI -- Precipitation Nowcasting from Satellite Data

A Variational U-Net for Weather Forecasting

FourCastNet: Accelerating Global High-Resolution Weather Forecasting using Adaptive Fourier Neural Operators

DiffCast: A Unified Framework via Residual Diffusion for Precipitation Nowcasting

Skilful Precipitation Nowcasting Using NowcastNet

FourCastNet: A Global Data-driven High-resolution Weather Model using Adaptive Fourier Neural Operators

Short-term precipitation prediction using deep learning

DuoCast: Duo-Probabilistic Meteorology-Aware Model for Extended Precipitation Nowcasting

Nowcasting-Nets: Deep Neural Network Structures for Precipitation Nowcasting Using IMERG

MetNet: A Neural Weather Model for Precipitation Forecasting

PuYun: Medium-Range Global Weather Forecasting Using Large Kernel Attention Convolutional Networks

PostRainBench: A comprehensive benchmark and a new model for precipitation forecasting

Generative Modeling of High-resolution Global Precipitation Forecasts

Skilful precipitation nowcasting using deep generative models of radar

A Scalable Real-Time Data Assimilation Framework for Predicting Turbulent Atmosphere Dynamics

Data-driven medium-range weather prediction with a Resnet pretrained on climate simulations: A new model for WeatherBench

Precipitaion Nowcasting using Deep Neural Network

PAUNet: Precipitation Attention-based U-Net for rain prediction from satellite radiance data

WF-UNet: Weather Fusion UNet for Precipitation Nowcasting