Understanding water and heat transport through soils is vital for managing soil and groundwater resources, agricultural irrigation, and ecosystem protection. This paper aims to explore the potential application of deep learning methods in simulating water and heat transport processes within soils. It also examines the interactions between soil hydrological processes and environmental factors, including meteorological conditions and groundwater levels. To achieve these, we develop a hybrid model Informer–LSTM by combining two powerful architectures: Informer, a Transformer-based model essentially designed for long-sequence time-series forecasting, and Long Short-Term Memory (LSTM), a neural network that is great at learning short-term patterns in sequential data. The model is applied to field measurements from Henan Township in Ordos, Inner Mongolia, China, for training and testing, to simulate three key variables: soil water content, temperature, and heat flux at different depths in two soil columns with different groundwater levels. Our results confirm that Informer–LSTM is highly effective at simulating the soil water and heat transport. Simultaneously, we evaluate its performance by incorporating various combinations of input data including meteorological data, soil hydrothermal dynamics, and groundwater level. This reveals the relationship between soil hydrothermal processes and meteorological data, as well as coupled processes of soil water and heat transport. Moreover, employing SHapley Additive exPlanations (SHAP) analysis, we identify the most influential factors for predicting heat flux in shallow soils. This research demonstrates that deep learning models are a viable and valuable tool for simulating soil hydrothermal processes in arid and semi-arid regions.

Under the escalating pressures of climate change and freshwater scarcity, understanding how irrigation management alters greenhouse gas dynamics in agricultural ecosystems has attracted increasing attention. This study conducted a comprehensive meta-analysis based on 76 field-derived publications across Chinese coastal and inland regions. We evaluated the impacts of water-saving regimes-including deficit, alternate, and intermittent irrigation-on field emissions of methane (CH ₄ ), nitrous oxide (N ₂ O), and carbon dioxide (CO ₂ ), with random forest (RF) modeling and path analysis employed to disentangle the underlying mechanisms. The results demonstrated that water-saving practices induced a distinct biogeochemical divergence in field agroecosystems, characterized by a significant comprehensive effect of “CH ₄ reduction, N ₂ O promotion, and minor CO ₂ mitigation” (log response ratios, ln RR=-0.48, 0.28, and -0.09, respectively). Specifically, intermittent irrigation exerted the most pronounced impact, with both CH ₄ mitigation (ln RR=-0.51) and N ₂ O promotion (ln RR = 0.52) reaching their peak intensities. Driven by spatial hydrothermal heterogeneity, the Southeast Coastal region exhibited the highest sensitivity in gas flux responses. Conversely, North China showed the lowest risk of N ₂ O promotion (ln RR = 0.12) while maintaining robust mitigation capacity. RF modeling and meta-regression identified geographical region, soil pH, and soil organic matter (SOM) as the core predictors for the variances in CH ₄ ( R ² = 55.2%), N ₂ O ( R ² = 59.4%), and CO ₂ ( R ² = 48.1%) effects, respectively. Crucially, regression models pinpointed neutral-to-alkaline conditions (pH 7.0-7.2) as the critical threshold for N ₂ O responses, beyond which alkaline soil conditions were associated with a transition from N ₂ O promotion to mitigation. Path analysis further confirmed that irrigation modes exerted the strongest direct negative effect on CH ₄ . Soil pH showed a highly significant direct inhibition on N ₂ O, whereas precipitation introduced a notable indirect positive effect on N ₂ O by driving soil acidification; meanwhile, SOM showed a dominant direct contribution to CO ₂ mitigation. In conclusion, the environmental feedback of water-saving irrigation is a product of deep coupling between technical interventions and natural backgrounds. Future mitigation policies must integrate a “smart-adaptation” framework tailored to regional precipitation, soil pH, and SOM matrices, thereby orchestrating a synergy between watershed-scale green agricultural development and carbon neutrality goals.