Scholarlink

<p>High-resolution projections of daily temperature extremes are essential for climate-resilient development and informed regional adaptation planning. However, conventional statistical downscaling methods often overlook inherent stochasticity and physical consistency. To address these limitations, we introduce SMT‑HCSD (Stochastic Multi‑Target Hard‑Constrained Spatial Deep Learning), a perfect‑prognosis framework and architecture that simultaneously predicts minimum and maximum temperature, providing probabilistic, uncertainty‑aware, and physically consistent projections. SMT‑HCSD achieves this by training a single model on both targets and applying a hard constraint to ensure Tmin <InlineEquation ID="IEq1"> <EquationSource Format="TEX">\(\:\le\:\)</EquationSource> </InlineEquation> Tmax. Validation against historical observations demonstrates that the framework and architecture accurately reproduce the seasonal cycle, spatial distribution, and extremes of daily temperature. We applied SMT‑HCSD to downscale three CMIP6 global models (MPI-ESM1-2-HR, NorESM2-MM, and MRI-ESM2-0) over the Emilia-Romagna region (Italy) under both low (SSP1-2.6) and high (SSP3-7.0) emission scenarios. By integrating probabilistic modelling, physical constraints, and multi-target learning, SMT-HCSD advances statistical downscaling and produces actionable, reliable high-resolution climate information. The results underscore both the urgency of global emissions reductions and the need for regionally tailored adaptation and resilience strategies in Emilia-Romagna. SMT‑HCSD thus provides a robust and transparent basis for proactive climate policy and risk management.</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

SMT-HCSD: stochastic multi-target hard-constrained spatial deep learning framework and architecture for perfect prognosis CMIP6 daily temperature downscaling

  • 期刊论文

  • 出版模式：

    开放获取

  • 发表日期：

    2026年6月22日

查看全文
  • Wondesen Teshome Bekele,
  • Marco D’Oria

摘要

High-resolution projections of daily temperature extremes are essential for climate-resilient development and informed regional adaptation planning. However, conventional statistical downscaling methods often overlook inherent stochasticity and physical consistency. To address these limitations, we introduce SMT‑HCSD (Stochastic Multi‑Target Hard‑Constrained Spatial Deep Learning), a perfect‑prognosis framework and architecture that simultaneously predicts minimum and maximum temperature, providing probabilistic, uncertainty‑aware, and physically consistent projections. SMT‑HCSD achieves this by training a single model on both targets and applying a hard constraint to ensure Tmin \(\:\le\:\) Tmax. Validation against historical observations demonstrates that the framework and architecture accurately reproduce the seasonal cycle, spatial distribution, and extremes of daily temperature. We applied SMT‑HCSD to downscale three CMIP6 global models (MPI-ESM1-2-HR, NorESM2-MM, and MRI-ESM2-0) over the Emilia-Romagna region (Italy) under both low (SSP1-2.6) and high (SSP3-7.0) emission scenarios. By integrating probabilistic modelling, physical constraints, and multi-target learning, SMT-HCSD advances statistical downscaling and produces actionable, reliable high-resolution climate information. The results underscore both the urgency of global emissions reductions and the need for regionally tailored adaptation and resilience strategies in Emilia-Romagna. SMT‑HCSD thus provides a robust and transparent basis for proactive climate policy and risk management.