An Evolutionary-Algorithm-Driven Efficient Temporal Convolutional Network for Radar Image Extrapolation

Peiyang Wei, Changyuan Fan, Yuyi Wang, Tianlong Li, Jianhong Gan, Can Hu, Zhibin LiBiomimetics 2026 · DOI

TL;DR

This paper tackles the challenge of accurately predicting future radar weather patterns by developing an "Evolutionary-Algorithm-Driven Efficient Temporal Convolutional Network" (E-HEOA). By combining novel evolutionary optimization techniques with a ConvLSTM2D deep learning architecture, the E-HEOA significantly improves the fidelity and structural similarity of extrapolated radar images, setting new benchmarks for short-term weather forecasting.

What Problem Does This Paper Address?

Accurate short-term weather forecasting relies heavily on predicting future radar echo patterns, a process known as radar image extrapolation. Traditional methods often struggle with image degradation and artifacts, limiting their reliability. While deep learning models like ConvLSTM2D have shown promise in capturing spatiotemporal features, their performance is highly sensitive to hyperparameter choices (like learning rate and dropout rate). Manually tuning these hyperparameters is time-consuming and prone to local optima. This paper addresses the need for an automated, efficient, and robust hyperparameter optimization strategy to enhance the accuracy and clarity of radar image extrapolation.

Methods

The core of this research is the E-HEOA (Evolutionary-Algorithm-Driven Efficient Temporal Convolutional Network). This framework combines two key innovations:

  1. Hybrid Metaheuristic Optimizer (E-HEOA): This optimizer merges two evolutionary algorithms:
  2. Egret Swarm Optimization Algorithm (ESOA) with Gaussian Mutation: Inspired by egret foraging behavior, this component focuses on global exploration and handling sudden, nonlinear changes in radar echoes. The Gaussian mutation enhances its ability to escape local optima.
  3. Human Evolutionary Optimization Algorithm (HEOA) with Cauchy Mutation: This algorithm simulates human exploration and development stages, focusing on local refinement and knowledge inheritance. The Cauchy mutation aids in fine-grained exploitation and stability. This hybrid approach aims to leverage the strengths of both global exploration and local refinement for hyperparameter optimization.

  4. Enhanced ConvLSTM2D Architecture: The deep learning model itself is built upon ConvLSTM2D layers, which are designed to effectively preserve spatiotemporal features crucial for weather prediction.

The E-HEOA framework adaptively tunes two critical hyperparameters: the learning rate and the dropout rate for the ConvLSTM2D-CNN model. The optimization process involves initializing populations for both ESOA and HEOA, iteratively updating their positions based on their respective strategies, and then combining their outputs through a weighted summation to produce the final prediction.

Datasets

The study utilizes the HKO-7 radar echo image dataset, a publicly available dataset containing 170,000 radar images. These images cover a 101 km × 101 km area around a target location, with measurements taken at six-minute intervals across four different altitudes. The dataset spans two years. The data is preprocessed through loading, resizing, and normalization before being split into training (80%) and testing (20%) sets.

Key Insights and Ideas

  • Synergy of Global and Local Optimization: The paper demonstrates that combining a global exploration strategy (ESOA with Gaussian mutation) with a local refinement strategy (HEOA with Cauchy mutation) for hyperparameter tuning yields superior results compared to using either algorithm alone. This hybrid approach effectively balances the need to explore a wide range of hyperparameter possibilities with the precision required for fine-tuning.
  • Adaptive Hyperparameter Tuning is Crucial: The research strongly emphasizes that the performance of deep learning models is highly dependent on hyperparameter configuration. Automating this process with sophisticated evolutionary algorithms like E-HEOA is key to unlocking the full potential of these models.
  • ConvLSTM2D's Strength in Spatiotemporal Features: The ConvLSTM2D architecture proves effective in capturing and preserving the complex spatiotemporal dynamics present in radar echo sequences, which is fundamental for accurate extrapolation.
  • Weighted Fusion of Model Outputs: Combining the predictions from models optimized by ESOA and HEOA through a weighted summation (with a specific weight of 0.48 for HEOA) leads to improved overall performance, suggesting that the strengths of each optimized model complement each other.

Key Results

  • State-of-the-Art Performance: The E-HEOA model achieved the lowest Mean Squared Error (MSE) of 19.9989 and the highest Structural Similarity Index (SSIM) of 0.8093 among all evaluated models and optimization algorithms. This significantly outperforms baseline models and other evolutionary algorithms.
  • Improved Prediction Fidelity and Clarity: The extrapolated radar images generated by E-HEOA exhibit superior prediction fidelity, structural similarity, and reduced degradation compared to conventional methods and other benchmark models.
  • Enhanced Robustness: The E-HEOA demonstrates stronger noise resistance, with its performance gap widening over other models as noise intensity increases.
  • Efficient Convergence: The E-HEOA model shows competitive convergence speeds, particularly in the training set, and achieves optimal fitting performance on both training and validation sets.
  • Validation on Large Dataset: The experimental validation on a dataset of 170,000 radar echo samples provides strong evidence for the model's effectiveness and generalizability.

Limitations and Future Work

  • Computational Cost: While the E-HEOA framework is designed for efficiency, the paper acknowledges that hyperparameter optimization can still be computationally intensive, especially when parallel computation is not fully utilized.
  • Future Research Directions:
    • Improving Model Tuning: Further research will focus on investigating adaptive tuning approaches to optimize the joint weights within the E-HEOA framework more independently, aiming for full automation.
  • Exploring Other Radar Imaging Applications: The model's applicability will be explored for other radar imaging tasks, with an emphasis on predicting extreme rainfall events and intense convective echoes.
  • Expanding Scope: The research suggests adapting the model for applications beyond weather forecasting, such as traffic prediction, video prediction, and temperature forecasting, to enhance its generalization capabilities. This involves adjustments in data modeling, network architecture, training strategies, and evaluation metrics.

Summary updated

Abstract

Radar image extrapolation serves as a fundamental methodology in meteorological forecasting, facilitating precise short-term weather prediction through spatiotemporal sequence analysis. Conventional approaches remain constrained by progressive image degradation and artifacts, substantially limiting operational forecasting reliability. This research introduces E-HEOA-an enhanced deep learning architecture with integrated hyperparameter optimization. Our framework incorporates two fundamental innovations: (a) a hybrid metaheuristic optimizer merging a Gaussian-mutated ESOA and Cauchy-mutated HEOA for autonomous learning rate and dropout optimization and (b) embedded ConvLSTM2D modules for enhanced spatiotemporal feature preservation. Experimental validation on 170,000 radar echo samples demonstrates superior performance, demonstrating considerable enhancement in almost all aspects relative to the baseline model while establishing new state-of-the-art benchmarks in prediction fidelity, convergence efficiency, and structural similarity metrics.
BibTeX
@inproceedings{wei2026evolutionary,
  author = {Peiyang Wei and Changyuan Fan and Yuyi Wang and Tianlong Li and Jianhong Gan and Can Hu and Zhibin Li},
  title = {{An Evolutionary-Algorithm-Driven Efficient Temporal Convolutional Network for Radar Image Extrapolation}},
  booktitle = {Biomimetics},
  year = {2026},
  doi = {10.3390/biomimetics11020122},
}
Figure