Saeidinia et al. (2026) High-resolution forecasting of soil thermal regimes using different deep learning frameworks under climate change

Identification

• Journal: Scientific Reports
• Year: 2026
• Date: 2026-02-05
• Authors: Mehri Saeidinia, Amir Hamzeh Haghiabi, Mohammad Nazeri Tahroudi, Aliheidar Nasrolahi, Carlo De Michele
• DOI: 10.1038/s41598-026-38496-6

Research Groups

• Department of Water Engineering, Lorestan University, Khorramabad, Iran
• Department of Civil and Environmental Engineering, Politecnico Di Milano, Milan, Italy

Short Summary

This study developed a deep learning framework to downscale soil temperature (5 cm depth) in western Iran under climate change scenarios. A hybrid CNN-LSTM model accurately projected that high-emission pathways (SSP585) cause initial cooling followed by accelerated warming, while low-emission pathways lead to stable, moderate warming.

Objective

• To simulate and predict future soil temperature dynamics in Lorestan province under climate change scenarios, using deep learning models calibrated with observed and projected climate data.
• To assess the performance of deep learning models (LSTM, GRU, CNN, Hybrid CNN-LSTM) as downscaling models.
• To quantify climate change impacts on soil thermal regimes using IPCC AR6 scenarios (SSPs).
• To provide actionable insights for policymakers and farmers to adapt to changing soil conditions, thereby supporting food security and ecosystem resilience.

Study Configuration

• Spatial Scale: Lorestan Province, western Iran (approximately 28,000 km²), utilizing data from 10 synoptic weather stations.
• Temporal Scale:
  • Observed meteorological data: 1980–2020 (daily).
  • Historical climate model validation: 1979–2014.
  • Near-term scenario plausibility assessment: 2015–2020.
  • Future projections: Short-term (2023–2040), Mid-century (2041–2070), End-of-century (2071–2100).

Methodology and Data

• Models used:
  • Deep Learning Models: Convolutional Neural Network (CNN), Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), Hybrid CNN-LSTM.
  • Climate Model: Canadian Earth System Model (CanESM5) from CMIP6.
  • Feature Selection: Random Forest (Gini) importance, Permutation Importance, SHAP (SHapley Additive exPlanations) analysis.
  • Bias Correction: Empirical Quantile Mapping (EQM).
  • Interpolation: Bilinear interpolation for spatial downscaling, linear, time-based, second-order polynomial, and cubic spline interpolation for missing data.
  • Trend Analysis: Mann–Kendall test, Detrended Fluctuation Analysis (DFA).
• Data sources:
  • Observed mean daily soil temperature (5 cm depth) and meteorological data from 10 synoptic stations in Lorestan Province, Iran (Islamic Republic of Iran Meteorological Organization - IRIMO).
  • Climate projections from CanESM5 (CMIP6) for historical (1979–2014) and future (2015–2100) periods under SSP1-1.9 (1.9 W/m²), SSP1-2.6 (2.6 W/m²), SSP2-4.5 (4.5 W/m²), SSP3-7.0 (7.0 W/m²), and SSP5-8.5 (8.5 W/m²) scenarios.
  • CanESM5 predictor variables (26 variables including surface/upper-air dynamics and thermodynamic variables across 1000 hPa, 850 hPa, and 500 hPa pressure levels, e.g., air temperature at 2 m, total precipitation, mean sea-level pressure, geopotential height, wind components, specific humidity).

Main Results

• Predictor Selection: Surface air temperature (temp) and mid-tropospheric pressure at 500 hPa (p500) were consistently identified as the top two predictors. Mean sea-level pressure (mslp) and geopotential height at 500 hPa and 850 hPa (p5_z, p850) were also consistently in the top five.
• Model Performance (Historical Downscaling): The hybrid CNN-LSTM model generally outperformed other deep learning models, achieving high accuracy with Nash–Sutcliffe Efficiency (NSE) values exceeding 0.86, Root Mean Square Error (RMSE) less than 4.3 °C, and Mean Absolute Error (MAE) as low as 2.71 °C. Kendall’s tau correlation coefficients between observed and simulated soil temperatures ranged from 0.73 to 0.78.
• Near-term Scenario Plausibility (2015–2020): The hybrid CNN-LSTM model successfully reproduced observed recent trends (Kendall’s tau = 0.72–0.80). High-elevation stations (> 1600 m) showed closest agreement with SSP119/SSP126, while low-elevation stations (< 1200 m) aligned best with SSP245/SSP370. Western stations (more arid) favored SSP245/SSP370, and eastern stations (Zagros Mountains influence) favored SSP119/SSP126.
• Future Soil Temperature Projections:
  • Short-term (2023–2040): A statistically significant increasing trend in soil temperature was observed across all scenarios. However, SSP585 showed an initial cooling trend (e.g., -4.11 °C for Sepiddasht, -3.7 °C for Aligudarz) compared to SSP126/SSP245, primarily due to a significant reduction in incoming solar radiation (approximately 11% in some stations). DFA scaling exponents (α) between 1.48 and 1.54 indicated strong long-term persistence.
  • Medium-term (2041–2070): The ensemble mean temperature for SSP126 was 19.21 °C, SSP245 was 19.40 °C, and SSP585 dropped to 18.18 °C. SSP585 continued to show lower mean temperatures than SSP126/SSP245, with reduced temperature fluctuations. SSP585 showed a consistent shift towards positive skewness (more frequent hot days) despite lower mean temperatures.
  • Long-term (2071–2100): The anomalous cooling under SSP585 observed in mid-century completely reversed. Ensemble mean temperature rose sharply to 21.15 °C under SSP585, representing a total warming of 1.39 °C from SSP126. SSP585 exhibited reduced day-to-day temperature fluctuations and a shift towards neutral or negative skewness. SSP585 showed unanimous, intensifying warming trends (slopes 2 to 6 times greater than SSP245), while SSP126 showed no significant trend, indicating stabilization.
• Decadal Warming Rates:
  • SSP126: Rapidly declining warming rates (from 1.23 ± 0.16 °C/decade to 0.28 ± 0.03 °C/decade by late century).
  • SSP245: Variable warming rates, peaking mid-century (0.79 ± 0.20 °C/decade).
  • SSP585: Dramatic transition from significant cooling (-2.75 ± 0.74 °C/decade in 2023–2040) to accelerated warming (0.96 ± 0.18 °C/decade by mid-century, 0.83 ± 0.13 °C/decade by late century). By 2071–2100, SSP585 warms 3.0 times faster than SSP126.

Contributions

• Developed and applied a novel deep learning-based downscaling framework (hybrid CNN-LSTM) for high-resolution soil temperature projections (5 cm depth) under CMIP6 scenarios in a semi-arid, topographically complex region, addressing a gap in existing literature.
• Demonstrated the superior performance of the hybrid CNN-LSTM model in capturing complex spatial and temporal dependencies of soil thermal dynamics compared to other deep learning architectures.
• Provided detailed, station-specific insights into nonlinear soil temperature responses to different SSPs, including an initial cooling phase under high-emission scenarios (SSP585) followed by accelerated warming, which creates a "cooling-then-heating" shock for ecosystems.
• Emphasized the critical influence of emission pathways on the future pace and pattern of soil temperature change, highlighting that high emissions defer but ultimately cause the most intense warming.
• Offered actionable insights for climate adaptation strategies in agriculture and environmental management in vulnerable regions.

Funding

This research did not receive any specific grant from funding agencies in the public, commercial, or not-for-profit sectors.

Citation

@article{Saeidinia2026Highresolution,
  author = {Saeidinia, Mehri and Haghiabi, Amir Hamzeh and Tahroudi, Mohammad Nazeri and Nasrolahi, Aliheidar and Michele, Carlo De},
  title = {High-resolution forecasting of soil thermal regimes using different deep learning frameworks under climate change},
  journal = {Scientific Reports},
  year = {2026},
  doi = {10.1038/s41598-026-38496-6},
  url = {https://doi.org/10.1038/s41598-026-38496-6}
}