Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Precipitation Prediction Using an Ensemble of Lightweight Learners

Xinzhe Li, Sun Rui, Yiming Niu, Yao Liu

TL;DR

This work tackles the challenge of precipitation prediction under sparse high-rain events and heterogeneous spatiotemporal patterns by building an ensemble of lightweight learners atop a WeatherFusionNet backbone augmented with ConvLSTM to model temporal dynamics. A 3-stage training scheme sequentially optimizes the backbone, ensemble heads, and a controller that assigns weights via a rainfall-probability map, enabling targeted fusion of diverse predictions. The approach achieves top performance on the Weather4cast 2023 core test and nowcasting benchmarks, with notable gains in high-precipitation scenarios. This framework offers a practical, satellite-data–driven method for improved rainfall nowcasting and could enhance decision-making in agriculture and related industries.

Abstract

Precipitation prediction plays a crucial role in modern agriculture and industry. However, it poses significant challenges due to the diverse patterns and dynamics in time and space, as well as the scarcity of high precipitation events. To address this challenge, we propose an ensemble learning framework that leverages multiple learners to capture the diverse patterns of precipitation distribution. Specifically, the framework consists of a precipitation predictor with multiple lightweight heads (learners) and a controller that combines the outputs from these heads. The learners and the controller are separately optimized with a proposed 3-stage training scheme. By utilizing provided satellite images, the proposed approach can effectively model the intricate rainfall patterns, especially for high precipitation events. It achieved 1st place on the core test as well as the nowcasting leaderboards of the Weather4Cast 2023 competition. For detailed implementation, please refer to our GitHub repository at: https://github.com/lxz1217/weather4cast-2023-lxz.

Precipitation Prediction Using an Ensemble of Lightweight Learners

TL;DR

This work tackles the challenge of precipitation prediction under sparse high-rain events and heterogeneous spatiotemporal patterns by building an ensemble of lightweight learners atop a WeatherFusionNet backbone augmented with ConvLSTM to model temporal dynamics. A 3-stage training scheme sequentially optimizes the backbone, ensemble heads, and a controller that assigns weights via a rainfall-probability map, enabling targeted fusion of diverse predictions. The approach achieves top performance on the Weather4cast 2023 core test and nowcasting benchmarks, with notable gains in high-precipitation scenarios. This framework offers a practical, satellite-data–driven method for improved rainfall nowcasting and could enhance decision-making in agriculture and related industries.

Abstract

Precipitation prediction plays a crucial role in modern agriculture and industry. However, it poses significant challenges due to the diverse patterns and dynamics in time and space, as well as the scarcity of high precipitation events. To address this challenge, we propose an ensemble learning framework that leverages multiple learners to capture the diverse patterns of precipitation distribution. Specifically, the framework consists of a precipitation predictor with multiple lightweight heads (learners) and a controller that combines the outputs from these heads. The learners and the controller are separately optimized with a proposed 3-stage training scheme. By utilizing provided satellite images, the proposed approach can effectively model the intricate rainfall patterns, especially for high precipitation events. It achieved 1st place on the core test as well as the nowcasting leaderboards of the Weather4Cast 2023 competition. For detailed implementation, please refer to our GitHub repository at: https://github.com/lxz1217/weather4cast-2023-lxz.
Paper Structure (10 sections, 7 figures, 2 tables)

This paper contains 10 sections, 7 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Methods
  3. Framework Training
  4. Stage 1: Backbone Training
  5. Stage 2: Training of Ensemble Learners
  6. Stage 3: Training of Ensemble Controllers
  7. Experiments
  8. Experimental Setting
  9. Results
  10. Conclusions

Figures (7)

  • Figure 1: The proposed baseline network structure "WeatherFusionNet + ConvLSTM". The baseline contains 4 main components: PhyDNet, sat2radnet, U-Net as well as ConvLSTM. The parameters of first two networks are frozen during all the training process.
  • Figure 2: multiple output heads is attached to the U-Net, the ConvLSTM is shared for all heads
  • Figure 3: The controller is implemented as a gate to control the output
  • Figure 4: Backbone Training Stage
  • Figure 5: Training of Ensemble Learners
  • ...and 2 more figures