EfficientRainNet: Leveraging EfficientNetV2 for memory-efficient rainfall nowcasting

Abstract

Rainfall nowcasting is critical for timely weather predictions and emergency responses, particularly in flood-prone areas. Existing models, while accurate, often require substantial computational resources. Addressing this challenge, our study introduces EfficientRainNet, a neural network that leverages mobile inverted residual linear bottleneck blocks for memory-efficient rainfall nowcasting. Our evaluation, conducted over the State of Iowa, demonstrates that EfficientRainNet achieves accuracy comparable to that of the widely adopted encoder-decoder convolutional GRUs, yet with a substantially reduced model complexity, possessing less than 6% of the trainable parameters of the encoder-decoder convolutional GRUs and less than 9% of those of the compared Small Attention UNet. This lightweight design opens the possibility for deployment on edge devices, offering a scalable and accessible solution for real-time rainfall prediction. The results suggest further potential for extending the application of EfficientRainNet across broader regions and varied climatic conditions, harnessing its computational efficiency for widespread climate monitoring and forecasting.