Zhou et al. (2026) Impact of soil moisture intrinsic signal structure on agricultural drought forecastability in the Yangtze River Basin
Identification
- Journal: Journal of Hydrology Regional Studies
- Year: 2026
- Date: 2026-11-16
- Authors: Lue Zhou, Gengxi Zhang, Thian Yew Gan, Huimin Wang, Jinbai Huang, Shuai Zhou, Jin Zhao
- DOI: 10.1016/j.ejrh.2026.104213
Research Groups
- College of Hydraulic Science and Engineering, Yangzhou University, Yangzhou, China
- Department of Civil and Environmental Engineering, University of Alberta, Edmonton, AB, Canada
- School of Water Conservancy and Hydroelectric Power, Hebei University of Engineering, Handan, China
- Power China Huadong Engineering Corporation Limited, Hangzhou, Zhejiang, China
Short Summary
This study investigates the association between intrinsic temporal properties of soil moisture drought signals (persistence-related singularity structure, αmax, and multifractal complexity, Δα) and agricultural drought forecast skill in the Yangtze River Basin. It reveals that these properties are consistently linked to forecast skill and exhibit significant spatial organization, providing an important context for drought forecastability.
Objective
- To examine how agricultural drought forecastability is associated with the intrinsic temporal organization of soil-moisture drought signals.
- To hypothesize that spatial variations in forecast skill are jointly associated with persistence-related scaling (αmax) and multifractal complexity (Δα), and that these relationships vary with forecast horizon and model architecture.
- To investigate the spatial distribution of persistence-related scaling and multifractal complexity across the Yangtze River Basin (YRB).
- To assess how these structural properties are associated with realized forecast skill at different lead times (1, 3, and 6 months).
- To determine if different machine learning models (Support Vector Regression, Long Short-Term Memory, Temporal Convolutional Network) respond consistently to the same intrinsic signal structure across forecast horizons.
Study Configuration
- Spatial Scale: Yangtze River Basin (YRB), approximately 1.8 million square kilometers, spanning 24°N–36°N and 90°E–123°E. Data were sampled on a 0.5° × 0.5° grid.
- Temporal Scale: Monthly data from January 1950 to January 2025 (75 years). Forecasts were made at 1-, 3-, and 6-month lead times.
Methodology and Data
- Models used:
- Multifractal Detrended Fluctuation Analysis (MF-DFA) for characterizing multiscale temporal structure.
- Gaussian copula for constructing the Joint Standardized Soil Moisture Index (JSSI).
- Support Vector Regression (SVR) for drought forecasting.
- Long Short-Term Memory (LSTM) for drought forecasting.
- Temporal Convolutional Network (TCN) for drought forecasting.
- Generalized Additive Models (GAMs) for analyzing relationships between multifractal descriptors and forecast skill.
- Data sources:
- ERA5-Land reanalysis data (European Centre for Medium-Range Weather Forecasts - ECMWF) at a native spatial resolution of 0.1° × 0.1°, sampled to a 0.5° × 0.5° grid.
- Hydroclimatic variables: Total precipitation (m), Potential evaporation (m), Surface net solar radiation (J m⁻²), Surface latent heat flux (J m⁻²), Surface net thermal radiation (J m⁻²), Soil temperature level 1 (K), Soil temperature level 2 (K), Skin temperature (K), 2 m temperature (K), Leaf area index (m² m⁻²), Sub-surface runoff (m), Surface runoff (m).
- Environmental factors: Geopotential and soil type from ERA5-Land.
Main Results
- The Joint Standardized Soil Moisture Index (JSSI) effectively integrates soil moisture anomalies across multiple accumulation scales (1, 3, 6, 9, and 12 months), showing strong dependence with single-scale Standardized Soil Moisture Index (SSMI).
- Multifractal characteristics (maximum singularity strength, αmax, and singularity-spectrum width, Δα) of the JSSI series exhibit pronounced and spatially coherent organization across the YRB, with significant positive spatial autocorrelation (Global Moran's I for Δα = 0.570, αmax = 0.538).
- Forecast skill (Kling–Gupta efficiency, KGE, and Nash–Sutcliffe efficiency, NSE) systematically deteriorates with increasing lead time for all models, with SVR consistently outperforming LSTM and TCN across all horizons.
- Forecast performance shows a recurrent geographical organization across all three architectures, with lower skill in parts of the western and central YRB and comparatively higher skill in the eastern YRB, particularly at longer lead times.
- KGE decomposition reveals that forecast horizon primarily affects the temporal correlation component (r), while model architecture most clearly differentiates the preservation of drought-signal variability (α).
- αmax and Δα are the two multifractal descriptors most consistently associated with realized forecast skill across models and lead times (median absolute Spearman correlation of 0.446 and 0.393, respectively).
- The explanatory contribution of αmax and Δα to forecast skill varies with lead time and model architecture, with maximum explained deviance occurring at the 3-month lead (SVR: 40.92%, LSTM: 28.78%, TCN: 26.85%).
- αmax provides the most consistent positive association with KGE across all models and lead times, acting as a stable organizing axis for forecast skill.
- The response to Δα is more model- and lead-dependent, converging at the 3-month lead (all models showing broadly increasing KGE with Δα) but diverging at 1- and 6-month leads.
- Cross-model consistency of forecastability landscapes is strongest at the 3-month lead, with pairwise surface correlations of 0.993–0.999, indicating nearly identical ordering of the joint (αmax, Δα) domain.
- At the 6-month lead, the SVR forecastability landscape diverges from those of LSTM and TCN, which remain closely aligned.
- Multifractal structure (αmax and Δα) provides additional forecast-relevant information beyond conventional 1-month persistence, particularly at longer lead times (e.g., increasing explained deviance by 4.40–10.72 percentage points at 6 months).
- Spatial variation in αmax and Δα is systematically associated with environmental factors: geopotential, long-term mean Leaf Area Index (LAI), and soil type jointly explain 36.62% of the spatial variation in αmax and 43.35% in Δα. Mean LAI provides the largest additional contribution to αmax, while geopotential contributes most to Δα.
Contributions
- Explicitly links multifractal signal structure (αmax and Δα) with spatial, forecast horizon, and model architecture-dependent variations in realized agricultural drought forecast skill.
- Demonstrates that multiscale temporal organization, characterized by αmax and Δα, provides additional forecast-relevant information beyond conventional short-term soil moisture memory, particularly for intermediate and longer lead times.
- Establishes the (αmax, Δα) domain as a useful structural basis for interpreting regional differences in agricultural drought forecast performance.
Funding
- China Postdoctoral Science Foundation (Grant No. 2024M752711)
- National Natural Science Foundation of China (Grant No. 52509036)
Citation
@article{Zhou2026Impact,
author = {Zhou, Lue and Zhang, Gengxi and Gan, Thian Yew and Wang, Huimin and Huang, Jinbai and Zhou, Shuai and Zhao, Jin},
title = {Impact of soil moisture intrinsic signal structure on agricultural drought forecastability in the Yangtze River Basin},
journal = {Journal of Hydrology Regional Studies},
year = {2026},
doi = {10.1016/j.ejrh.2026.104213},
url = {https://doi.org/10.1016/j.ejrh.2026.104213}
}
Original Source: https://doi.org/10.1016/j.ejrh.2026.104213