Scholarlink

<p>Accurate spatio-temporal prediction of discrete variables like drought classes is crucial for impact mitigation. This study introduces the Spatio-Temporal Hybrid Convolutional Neural Network-Cellular Automata (ST-CNN-CA), a novel framework that enhances the classic Cellular Automata-Markov (CA-Markov) model. It replaces the static transition matrix with a dynamic, data-driven function, implemented by a Hybrid CNN, to learn complex transition probabilities from historical map sequences. This allows the model to operate with or without the influence of external driving variables. For validation, the model's performance was tested in western Iran by predicting Standardized Precipitation Evapotranspiration Index (SPEI) classes for the 2020–2024 period under two scenarios: (A) using only historical SPEI data, and (B) incorporating climatic drivers, which were ranked using a Random Forest algorithm. The model's superiority was confirmed by validation against real data and comparison with CA-Markov and ConvLSTM benchmarks (average Kappa: 0.38 vs. 0.08 in the univariate case; 0.70 vs. 0.32 in the multivariate case). Subsequently, utilizing the comprehensive 1967–2024 dataset, spatio-temporal patterns of SPEI were&#xa0;projected&#xa0;for 2025–2040 under both scenarios,&#xa0;as a statistical extrapolation of historical dynamics. Projections for 2025–2040 indicate that normal and moderately dry classes will be most frequent. Furthermore, trend analysis (using the Mann–Kendall test and Sen’s Slope estimator) revealed a significant acceleration in the expansion of drier classes through 2040 compared to the 1967–2024 historical period. These findings underscore the need for implementing proactive drought mitigation strategies in western Iran and highlight the ST-CNN-CA model’s capability for projecting other discrete environmental variables.</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

A novel spatio-temporal hybrid convolutional neural network-cellular automata for projecting discrete drought variables

  • 期刊论文

  • 发表日期：

    2026年4月27日

查看全文
  • Abdol Rassoul Zarei

摘要

Accurate spatio-temporal prediction of discrete variables like drought classes is crucial for impact mitigation. This study introduces the Spatio-Temporal Hybrid Convolutional Neural Network-Cellular Automata (ST-CNN-CA), a novel framework that enhances the classic Cellular Automata-Markov (CA-Markov) model. It replaces the static transition matrix with a dynamic, data-driven function, implemented by a Hybrid CNN, to learn complex transition probabilities from historical map sequences. This allows the model to operate with or without the influence of external driving variables. For validation, the model's performance was tested in western Iran by predicting Standardized Precipitation Evapotranspiration Index (SPEI) classes for the 2020–2024 period under two scenarios: (A) using only historical SPEI data, and (B) incorporating climatic drivers, which were ranked using a Random Forest algorithm. The model's superiority was confirmed by validation against real data and comparison with CA-Markov and ConvLSTM benchmarks (average Kappa: 0.38 vs. 0.08 in the univariate case; 0.70 vs. 0.32 in the multivariate case). Subsequently, utilizing the comprehensive 1967–2024 dataset, spatio-temporal patterns of SPEI were projected for 2025–2040 under both scenarios, as a statistical extrapolation of historical dynamics. Projections for 2025–2040 indicate that normal and moderately dry classes will be most frequent. Furthermore, trend analysis (using the Mann–Kendall test and Sen’s Slope estimator) revealed a significant acceleration in the expansion of drier classes through 2040 compared to the 1967–2024 historical period. These findings underscore the need for implementing proactive drought mitigation strategies in western Iran and highlight the ST-CNN-CA model’s capability for projecting other discrete environmental variables.