Simulating frozen soil infiltration under late-autumn irrigation in salinized irrigation areas using deep learning

Water scarcity and soil salinization have seriously constrained to agricultural production in the Hetao Irrigation areas. Late-autumn irrigation can be expected to serve as the promising potential of the water-saving application, in order to integrate the salt leaching with the moisture conservation in the soil freezing period. Among them, the frozen soil infiltration is one of the most critical physical procedures to directly control the distribution of the irrigation water within the soil profile and the efficiency of salt leaching. In this study, a systematic investigation was implemented to clarify the frozen soil infiltration and salt redistribution under the coupled multiple factors—freeze-thaw cycles, initial water content, and tillage practices. Indoor experiments of the soil column freezing-infiltration were conducted to simulate the late-autumn irrigation scenario in the irrigation areas. Local field data was collected from the observation on the agricultural practices. Some treatments included the initial soil water contents (20%, 25%, and 30%) with/without tillage and freeze-thaw cycles (10 cycles). The dynamic profiles were systematically monitored on the soil temperature, moisture, and salt content. There were some variations in the infiltration rate and desalination rate. A systematic investigation was implemented to explore the influence of the key late-autumn irrigation on the frozen soil infiltration and its coupled water-heat-salt mechanisms. Furthermore, the comparison was made on the prediction performance of three conventional infiltration models (Horton, Philip, and Green-Ampt) and two deep learning models (LSTM, and LSTM-Attention). The optimal model was selected to simulate the frozen soil infiltration under complex conditions. The results indicate that the freeze-thaw cycles (FTCs), tillage practices, and initial soil water content were concurrently regulated the frozen soil infiltration and desalination using coupled water-heat-salt dynamics. Three stages were sequentially found in the frozen soil infiltration: the initial rapid infiltration, phase-transition and stable infiltration. Compared with the unfrozen soil, the pore ice blockage was reduced the average infiltration rate. While the FTCs were significantly enhanced the final infiltration of the soil under identical moisture and tillage condition. The soil structure was also improved the pore volume and connectivity. The initial water content was primarily governed the duration of the phase-transition stage; The higher moisture levels also prolonged the ice-water phase transition, thereby reducing the average infiltration rate during this stage. Tillage was accelerated the initial infiltration, where the surface porosity was enhanced the infiltration effect of the FTCs. Consequently, thermohydraulic exchange was intensified to effectively shorten the duration of the phase-transition stage. The FTCs increased the overall desalination efficiency by 25.63%-46.04%, compared with the unfrozen conditions. More uniform salt leaching was promoted over the soil layers (the desalination rates reached 69.49%~89.26%). The spatial heterogeneity of salinity was significantly reduced after FTCs. A coupling effect was also observed between initial water content and tillage on the desalination efficiency under high water content conditions (W3). The tillage also decreased the desalination efficiency regardless of the FTCs, due primarily to the high formation of the surface crust. In contrast, the peak desalination efficiency was achieved with the combination of the tillage and moderate water content (W2). Furthermore, the accuracy of the conventional models decreased significantly in the prediction of the frozen soil infiltration: The Philip model performed poorly (R²= 0.140), and the Green-Ampt model parameters exhibited physically ambiguous abnormal shifts (the model parameter for the effective pore volume, \theta_s-\theta_l decreased from 0.757 to 0.503). Although the Horton model was maintained the relatively high accuracy (R²= 0.886), its relative root mean square error increased to 16.330%, indicating the limited applicability under the frozen environments. The LSTM-Attention model was also used to balance its powerful capability for the nonlinear fitting and temporal feature capture. Excellent performance was achieved with R²= 0.999 and RRMSE= 1.240%, compared with the conventional models and the standard LSTM. Physical mechanisms can be expected to integrate into deep learning models, in order to overcome their "black box" limitations. Their application can also be extended into the regional scales using distributed modeling. Thereby, the finding can provide the reliable tools for the precise regulation of water and salt in the irrigation areas.