Extrapolation Algorithms for Radar Echoes Based on Conditional Diffusion Models

Conditional Diffusion Models

Abstract: The current inaccurate radar extrapolation results exhibit two typical features: the extension of the extrapolation time leads to increasingly attenuated echo strength, and the prediction performance for strong echoes decreases rapidly. This paper presents a Diffuse Radar Extrapolation Algorithm Driven by Radar Echo Frames (DiffREE) to address these issues. The algorithm employs a conditional coding module to deeply fuse past radar echo frames’ spatial and temporal information. Additionally, it automatically extracts spatiotemporal features of the echoes through the Transformer encoder. These extracted features serve as the condition for the conditional diffusion model, which, in turn, drives the diffusion model to reconstruct the current radar echo frames. The experimental results demonstrate that the method can generate high-precision and high-quality radar forecast frames. Compared with the best baseline algorithm, the proposed method shows significant improvements of 42.2%, 51.1%, 49.8%, and 39.5% in CSI, ETS, HSS, and POD metrics, respectively

Model Structure

Conditional Encoding Architecture

Conditional Diffusion Architecture

GIF Effect Display

Jiangsu Model

Title: Jiangsu Model
Title: Jiangsu Model (Another Scenario)

Nationwide Model

Title: Nationwide Model

Qinghai Model

Title: Qinghai Model
Title: Qinghai Model (Another Scenario)

Sanya Model

Title: Sanya Model

Experimental Comparison

Train and Test

CUDA_VISIBLE_DEVICES=3,4 python runner.py --config configs/weather_round.yml --exp weather_20 --config_mod sampling.subsample=100 -t --ni

CUDA_VISIBLE_DEVICES=1 python runner.py --config configs/weather_round.yml --exp weather_20 --config_mod sampling.subsample=100 sampling.num_frames_pred=100 data.revise=False -vg --ni

wswql/DiffREE