Scholarlink

<p>Addressing the critical challenge of predicting discrete climate variables in water-scarce regions, this study introduces a novel autoregressive framework: the Spatio-Temporal Deep Learning-Enhanced Cellular Automata (ST-Deep-CA). This model fundamentally advances traditional Cellular Automata-Markov (CA-Markov) approaches by replacing their static, memoryless transition matrix with a deep Multilayer Perceptron (MLP). The MLP learns complex, non-linear transition rules from a pixel’s entire spatio-temporal context, including its temporal history, spatial neighborhood, and the influence of external driver variables. To demonstrate its efficacy, we applied ST-Deep-CA to predict climatic conditions (based on a modified United Nations Environment Programme or UNEP aridity index) in Iran’s northern half for 2025–2039, using data from 24 stations (1967–2024). Two scenarios were tested: (A) predicting climate class alone, and (B) incorporating climatic drivers (temperature, humidity, etc.), whose importance was ranked using Random Forest (RF) and Multiple Linear Regression (MLR). The model’s predictive accuracy was rigorously validated for the 2020–2024 period against actual data using confusion matrix, Kappa (KI), NRMSE, and MAE metrics. Results confirmed the model’s robust capability. The accuracy of the model in Scenario B was significantly higher than in Scenario A, highlighting the benefit of including driver variables (e.g., Kappa improved from a range of 0.55–0.61 in Scenario A to 0.70–0.80 in Scenario B). The climate projections revealed a significant aridification trend, with a notable contraction of humid regions and expansion of arid zones. This study not only provides a powerful new tool for spatio-temporal forecasting but also delivers critical insights for climate change management in vulnerable regions, recommending the ST-Deep-CA model for broader applications in predicting variables like drought.</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

Introducing a novel spatio-temporal deep learning–enhanced cellular automata framework for discrete climate variable prediction

  • 期刊论文

  • 发表日期：

    2026年1月8日

查看全文
  • Abdol Rassoul Zarei

摘要

Addressing the critical challenge of predicting discrete climate variables in water-scarce regions, this study introduces a novel autoregressive framework: the Spatio-Temporal Deep Learning-Enhanced Cellular Automata (ST-Deep-CA). This model fundamentally advances traditional Cellular Automata-Markov (CA-Markov) approaches by replacing their static, memoryless transition matrix with a deep Multilayer Perceptron (MLP). The MLP learns complex, non-linear transition rules from a pixel’s entire spatio-temporal context, including its temporal history, spatial neighborhood, and the influence of external driver variables. To demonstrate its efficacy, we applied ST-Deep-CA to predict climatic conditions (based on a modified United Nations Environment Programme or UNEP aridity index) in Iran’s northern half for 2025–2039, using data from 24 stations (1967–2024). Two scenarios were tested: (A) predicting climate class alone, and (B) incorporating climatic drivers (temperature, humidity, etc.), whose importance was ranked using Random Forest (RF) and Multiple Linear Regression (MLR). The model’s predictive accuracy was rigorously validated for the 2020–2024 period against actual data using confusion matrix, Kappa (KI), NRMSE, and MAE metrics. Results confirmed the model’s robust capability. The accuracy of the model in Scenario B was significantly higher than in Scenario A, highlighting the benefit of including driver variables (e.g., Kappa improved from a range of 0.55–0.61 in Scenario A to 0.70–0.80 in Scenario B). The climate projections revealed a significant aridification trend, with a notable contraction of humid regions and expansion of arid zones. This study not only provides a powerful new tool for spatio-temporal forecasting but also delivers critical insights for climate change management in vulnerable regions, recommending the ST-Deep-CA model for broader applications in predicting variables like drought.