Scholarlink

<p>Deep learning approaches are increasingly being applied to forecasting challenges, but many current methods rely on advanced neural networks that provide only point estimates, without adequately addressing the epistemic and aleatoric uncertainties associated with predictions. Accurately capturing and quantifying both types of uncertainty is essential for assessing the confidence level of model outputs, especially in decision-sensitive contexts. Traditional techniques such as quantile regression, Bayesian neural networks, and Monte Carlo dropout tend to be either computationally intensive or prone to inaccuracies. In this paper, we propose an evidential deep learning (EDL)-based approach for drought forecasting that integrates uncertainty quantification. This method characterizes both aleatoric and epistemic uncertainties, offering a more comprehensive understanding of predictive reliability in relation to drought events. By leveraging the Dirichlet distribution, the model interprets uncertainties as evidence values associated with the neural network outputs. The proposed EDL approach is evaluated across three deep learning architectures-CNN, LSTM, and GAN-applied to datasets from the Horn of Africa and Southwestern Europe. Experimental results show that the EDL method outperforms traditional probabilistic and deterministic models, emphasizing the critical role of uncertainty quantification in enhancing forecasting accuracy.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Evidential uncertainty quantification with multiple deep learning architectures for spatiotemporal drought forecasting

  • 期刊论文

  • 发表日期：

    2025年2月15日

查看全文
  • Ahlem Ferchichi,
  • Mejda Chihaoui,
  • Radhia Toujani,
  • Aya Ferchichi,
  • Fatma Hendaoui

摘要

Deep learning approaches are increasingly being applied to forecasting challenges, but many current methods rely on advanced neural networks that provide only point estimates, without adequately addressing the epistemic and aleatoric uncertainties associated with predictions. Accurately capturing and quantifying both types of uncertainty is essential for assessing the confidence level of model outputs, especially in decision-sensitive contexts. Traditional techniques such as quantile regression, Bayesian neural networks, and Monte Carlo dropout tend to be either computationally intensive or prone to inaccuracies. In this paper, we propose an evidential deep learning (EDL)-based approach for drought forecasting that integrates uncertainty quantification. This method characterizes both aleatoric and epistemic uncertainties, offering a more comprehensive understanding of predictive reliability in relation to drought events. By leveraging the Dirichlet distribution, the model interprets uncertainties as evidence values associated with the neural network outputs. The proposed EDL approach is evaluated across three deep learning architectures-CNN, LSTM, and GAN-applied to datasets from the Horn of Africa and Southwestern Europe. Experimental results show that the EDL method outperforms traditional probabilistic and deterministic models, emphasizing the critical role of uncertainty quantification in enhancing forecasting accuracy.