Soil moisture content and evapotranspiration correction method based on XAJ-LSTM

Abstract

This invention discloses a method for correcting soil moisture content and evapotranspiration based on XAJ-LSTM, comprising: loading basic geographic and hydrological data of the target watershed; building an XAJ model to obtain intermediate physical quantities including soil moisture content and soil evapotranspiration; extracting feature vectors from the intermediate physical quantities and inputting them into the soil moisture LSTM model and the soil evapotranspiration LSTM model respectively, outputting the predicted three-layer soil moisture content and three-layer soil evapotranspiration; calculating the corrected three-layer soil evapotranspiration by combining the soil moisture LSTM model and the XAJ model to obtain the three-layer soil evapotranspiration after fusion weighting; writing back the three-layer soil moisture content predicted by the soil moisture LSTM model to the XAJ model to obtain the corrected three-layer soil moisture content; this invention provides high spatiotemporal resolution state variable support for precise watershed water resource management, drought early warning and forecasting, and hydrological response research under the background of climate change.

Description

Technical Field

[0001] This invention relates to the fields of hydrological modeling and artificial intelligence technology, specifically to a method for correcting soil moisture content and evapotranspiration based on XAJ-LSTM. Background Technology

[0002] With the intensification of global climate change and the deepening impact of human activities, the underlying surface conditions and hydro-meteorological processes in watersheds are undergoing unprecedentedly complex changes. Frequent extreme rainfall events, increased evaporation capacity, and intensified human water use have led to increasingly significant spatiotemporal variability in the watershed's water cycle, placing increasingly stringent demands on the simulation and forecasting accuracy of hydrological models. The Xin'anjiang (XAJ) model, a classic distributed hydrological model developed and validated over decades in my country's humid regions, plays an irreplaceable role in hydrological simulation, flood forecasting, and water resources planning due to its clear runoff generation and confluence physical mechanisms and relatively simple structure. However, facing increasingly complex application scenarios and ever-increasing accuracy requirements, the inherent limitations of traditional XAJ models at the mechanistic level and in their computational framework are gradually becoming apparent, posing a key technical bottleneck to their further development.

[0003] First, the model faces fundamental challenges in simulating the core state variable—soil moisture content. Traditional XAJ models employ a static computational framework based on water storage capacity curves and fixed stratification parameters to deduce soil water dynamics. While this method has a clear physical concept, it inherently fails to capture the complex spatiotemporal heterogeneity of soil moisture in the real world. Soil moisture transport and distribution are influenced by multiple factors, including meteorological forcing, vegetation transpiration, soil properties, and topography, exhibiting high nonlinearity and hysteresis. The simplified linear or empirical relationships in the model, along with fixed upper (WUM), lower (WLM), and deep (WD) water storage capacity parameters, often lead to the systematic accumulation and propagation of simulation errors over long-term continuous simulations. This error accumulation not only distorts the true dynamics of soil moisture but also further reduces the accuracy and reliability of flow simulations by affecting runoff generation calculations and evapotranspiration estimations, a problem particularly prominent in simulations of flood responses after drought periods or the evolution of long-term droughts.

[0004] Secondly, the calculation of the three-layer evapotranspiration, a key process in the model, is highly dependent on a series of empirical parameters, resulting in insufficient adaptability and robustness. Evapotranspiration is a core link connecting the hydrological and energy cycles, and its accurate estimation is crucial for both water and energy balance. Although the three-layer evaporation model in the XAJ model has an ingenious structure, its calculation results are extremely sensitive to empirical parameters such as the evapotranspiration reduction factor (K) and the deep evaporation factor (C). These parameters are usually calibrated using historical data and essentially reflect the average climate and underlying surface conditions during a specific calibration period. When the model is applied to different climate zones, or when the climate background changes significantly between the calibration period and the actual forecast period (such as a persistent drought or wetting trend), these fixed parameters are difficult to adaptively adjust, leading to systematic biases in evapotranspiration estimation. For example, under a warming climate, potential evaporation capacity may increase, and the fixed parameter K may not accurately reflect this change, thus affecting the overall water balance calculation.

[0005] Furthermore, existing methods for calibrating or optimizing hydrological models mostly focus on the fitting accuracy of the final output variables (such as outlet cross-section flow), lacking effective and interpretable dynamic correction mechanisms for internal physical state variables (such as soil moisture content and evapotranspiration). While flow calibration can directly improve the fit between simulated and measured flow, this "black box" or end-point calibration is often achieved by adjusting model parameters or adding error terms. This process may violate hydrological and physical laws, leading to a severe disconnect between the model's internal state (such as soil water storage) and actual conditions. Such distortion of state variables will cause the model to lose physical consistency, significantly reducing its predictive ability, especially its ability to anticipate future extreme events, and failing to provide reliable state data support for applications such as soil moisture-based agricultural drought monitoring.

[0006] In recent years, deep learning technologies, represented by Long Short-Term Memory (LSTM) networks, have demonstrated powerful capabilities in handling complex spatiotemporal sequence problems, providing a new data-driven paradigm for hydrological forecasting. LSTM can automatically learn complex nonlinear mapping relationships between inputs (such as rainfall and meteorological factors) and outputs (such as flow) from large amounts of historical data, and in some cases, it has achieved short-term forecast accuracy superior to traditional physical models. However, a purely data-driven LSTM model is essentially a "black box," as the relationships it learns lack clear physical interpretation. It cannot guarantee that its internal state or intermediate outputs conform to known hydrological physical laws (such as mass conservation and energy conservation), and it is difficult to reliably extrapolate in areas with scarce data or in extreme scenarios exceeding historical ranges. This "black box" characteristic severely restricts its reliability in hydrological operations that require physical interpretability and early warning of extreme scenarios.

[0007] Currently, hybrid modeling paradigms combining physical models with data-driven models have become a cutting-edge research direction. However, existing coupling attempts often suffer from insufficient integration, either being simple one-way serial processing (such as using LSTM to post-process the physical model output) or only making local adjustments to a single state variable (such as only correcting soil water or only correcting evapotranspiration). This "semi-coupled" mode fails to establish a two-way feedback channel between physical mechanisms and data learning, and cannot achieve synergistic optimization and dynamic mutual feedback between the two closely coupled and mutually influential core processes of soil moisture estimation and evapotranspiration calculation. A more ideal evapotranspiration estimation should rely on a more accurate soil moisture state, while a more accurate dynamic prediction of soil moisture cannot be separated from a reasonable evapotranspiration consumption estimate. Most existing methods have severed this inherent synergistic relationship.

[0008] Furthermore, in the early stages of model building, the parameter calibration of traditional physical models (such as the dozen or so key parameters of XAJ) heavily relies on manual trial and error or simple automated optimization algorithms (such as SCE-UA). These algorithms are inefficient in their search, easily get trapped in local optima in high-dimensional parameter spaces, and are computationally expensive. This inefficient parameter optimization process is in stark contrast to the rapidly evolving trend of increasing model complexity.

[0009] In conclusion, the field of hydrological simulation is currently at a critical crossroads. On the one hand, physical models have clear mechanisms but are limited by static parameters and simplified processes; on the other hand, data-driven models are flexible and powerful but lack physical constraints and interpretability. Neither approaching the problem independently, nor simply combining them superficially, can fundamentally meet the urgent needs of the new era for high-precision, high-reliability hydrological simulations that provide reliable physical state information. Summary of the Invention

[0010] The purpose of this invention is to provide a method for correcting soil moisture content and evapotranspiration based on XAJ-LSTM. This method constructs a correction framework of "dual LSTM + fusion weights," in which two LSTM networks learn the temporal variation patterns of soil moisture content and evapotranspiration respectively. Then, the two correction results are fused through preset weights, and finally, the fused result is fed back into the Xin'anjiang model to complete the hydrological process simulation. This achieves coordinated, online, and dynamic correction of soil moisture content and evapotranspiration states within the XAJ model, thereby significantly improving the accuracy and physical reliability of runoff simulation and forecasting.

[0011] To achieve the above functions, this invention designs a soil moisture content and evapotranspiration correction method based on XAJ-LSTM, which executes the following steps S1-S7 to complete the synchronous correction of soil moisture content and evapotranspiration in three soil layers:

[0012] Step S1: Load the basic geographic and hydrological data of the target watershed, including the daily rainfall of each rain gauge station, the daily evaporation of each hydrological station, the measured daily flow, the station name and its corresponding sub-watershed control area and area ratio;

[0013] Step S2: Build the XAJ model, including a three-layer evapotranspiration calculation module, a full-storage runoff generation module, a three-source water division module, and a slope confluence and river confluence module; run the XAJ model to obtain intermediate physical quantities including date, site, evapotranspiration of the three soil layers, soil moisture content of the three soil layers, runoff of the three water sources, and total flow at the outlet section of the final watershed.

[0014] Step S3: Construct a soil moisture content LSTM model, extract feature vectors from the intermediate physical quantities output by the XAJ model, input them into the soil moisture content LSTM model, and output the predicted soil moisture content of the three layers in the next time period.

[0015] Step S4: Construct a soil evapotranspiration LSTM model, extract feature vectors from the intermediate physical quantities output by the XAJ model, input them into the soil evapotranspiration LSTM model, and output the predicted evapotranspiration of the three soil layers for the current time period.

[0016] Step S5: Calculate the corrected three-layer soil evapotranspiration by combining the soil moisture content LSTM model and the XAJ model. Then, perform weighted fusion on the corrected three-layer soil evapotranspiration and the three-layer soil evapotranspiration predicted by the soil evapotranspiration LSTM model to obtain the fused weighted three-layer soil evapotranspiration.

[0017] The soil moisture content of the three layers predicted by the LSTM model is written back into the XAJ model and iterated to obtain the corrected soil moisture content of the three layers.

[0018] Step S6: For the target watershed, call the soil moisture LSTM model and the soil evapotranspiration LSTM model. The final output includes the original three-layer soil moisture content and three-layer soil evapotranspiration calculated by the XAJ model, the corrected three-layer soil evapotranspiration, the corrected three-layer soil moisture content, the three-layer soil evapotranspiration after fusion weight, and the fusion weight.

[0019] Step S7: Preset rate periodic and validation period, optimize and evaluate the soil moisture content LSTM model and the soil evapotranspiration LSTM model.

[0020] Beneficial effects: Compared with the prior art, the advantages of the present invention include:

[0021] This invention designs a soil moisture content and evapotranspiration correction method based on XAJ-LSTM, which truly integrates the essence of both physical mechanisms and data-driven approaches to construct a novel hydrological model with dynamic, collaborative, and interpretable correction capabilities for internal state variables (especially soil moisture content and evapotranspiration). This model not only pursues improved accuracy in effluent flow simulation but also strives to ensure the physical rationality and state authenticity of the hydrological processes within the model, thereby providing more solid and reliable technical support for refined water resource management, accurate flood and drought early warning, and climate change impact assessment. Attached Figure Description

[0022] Figure 1 This is a flowchart of a soil moisture content and evapotranspiration correction method based on XAJ-LSTM provided in an embodiment of the present invention;

[0023] Figure 2 This is a schematic diagram of the Tunxi River Basin stations provided according to an embodiment of the present invention;

[0024] Figure 3 This is an architecture diagram of the XAJ model provided according to an embodiment of the present invention;

[0025] Figure 4 This is an architecture diagram of the XAJ-LSTM model provided according to an embodiment of the present invention. Detailed Implementation

[0026] The present invention will be further described below with reference to the accompanying drawings. The following embodiments are only used to more clearly illustrate the technical solution of the present invention, and should not be used to limit the scope of protection of the present invention.

[0027] The soil moisture content and evapotranspiration correction method based on XAJ-LSTM provided in this embodiment of the invention refers to... Figure 1 Perform the following steps S1-S7 to complete the synchronous correction of the soil moisture content and evapotranspiration of the three soil layers:

[0028] Step S1: Load the basic geographic and hydrological data of the target watershed, including the daily rainfall of each rain gauge station, the daily evaporation of each hydrological station, the measured daily flow, the station name and its corresponding sub-watershed control area and area ratio;

[0029] In the embodiments, reference is made to Figure 2 The Tunxi River Basin was selected as the study area, and daily hydrological and meteorological data from January 1, 2008 to December 31, 2023 were used. The period from 2008 to 2018 was the model calibration period, and the period from 2019 to 2023 was the model validation period. The specific steps of step S1 are as follows:

[0030] Step S1.1: Collect daily rainfall data from various rain gauge stations, daily evaporation data from hydrological stations, and measured daily flow data from each station in the target watershed;

[0031] In this embodiment, watershed digital elevation model (DEM) data is imported, using USGS 30m resolution DEM data; flooding extraction is performed based on the DEM data to determine the watershed boundary and generate a watershed mask file; the locations of 18 rain gauge stations and 1 hydrological station are determined; and the latitude and longitude information of the target watershed and each station is extracted.

[0032] Step S1.2: Identify the date columns in each file and convert them to a uniform type;

[0033] Step S1.3: Extract the site name, sub-basin area and area ratio, and perform inner join matching;

[0034] In this embodiment, a model data input set is constructed with "date" as the primary key and includes multi-dimensional information such as rainfall, evaporation, flow rate, and station control area;

[0035] Step S1.4: Read the site flow results output by the first XAJ model and compare them with the total flow results, and extract the evapotranspiration of the three soil layers, the soil moisture content of the three soil layers, the runoff of each water source and the outlet flow.

[0036] Step S1.5: Divide all the acquired data into a preset periodicity and a verification period.

[0037] Step S2: Build the XAJ model, including a three-layer evapotranspiration calculation module, a full-storage runoff generation module, a three-source water division module, and a slope confluence and river confluence module; run the XAJ model to obtain intermediate physical quantities including date, site, evapotranspiration of the three soil layers, soil moisture content of the three soil layers, runoff of the three water sources, and total flow at the outlet section of the final watershed.

[0038] Reference Figure 3 The specific steps of step S2 are as follows:

[0039] Step S2.1: Construct a three-layer evapotranspiration calculation module to calculate the evapotranspiration of the upper, lower, and deep soil layers in the target watershed based on the current soil moisture content, rainfall, and evaporation capacity of the upper, lower, and deep soil layers respectively.

[0040] The calculation logic of the three-layer evapotranspiration calculation module considers the vertical distribution characteristics of soil moisture and calculates the actual evapotranspiration through a physical-driven approach. Key parameters involved include: upper tensional water storage capacity WUM (mm²), lower tensional water storage capacity WLM (mm²), deep evapotranspiration coefficient C (dimensionless), and evapotranspiration conversion factor K (dimensionless). The potential evapotranspiration EM (mm²) is usually obtained by multiplying the evaporation pan observation value by the conversion factor K, i.e., EM = K * E0, where E0 is the evaporation pan observation value. The actual three-layer evapotranspiration calculation follows the following order and formula:

[0041] When the storage of upper-layer tension water is sufficient, the upper-layer evapotranspiration EU is determined by the potential evapotranspiration capacity, and its expression is:

[0042] ;

[0043] If the storage of the upper layer is consumed completely while the storage of the lower layer of soil can still meet the evaporation demand, the calculation of the lower-layer evapotranspiration EL needs to consider the proportional relationship between the actual storage of the lower layer and the storage capacity, and its expression is:

[0044] ;

[0045] If the storage of the lower layer is also insufficient to meet the remaining evaporation demand, it is necessary to further call on the deep storage for evapotranspiration. At this time, the deep-layer evapotranspiration ED is jointly determined by a fixed deep-layer evapotranspiration coefficient and the remaining evaporation capacity, and its expression is:

[0046] ;

[0047] Where: WL is the soil moisture content of the lower layer (mm), EU is the upper-layer evaporation (mm), EL is the lower-layer evaporation (mm), ED is the deep-layer evaporation (mm), and EM is the potential evapotranspiration (mm).

[0048] Step S2.2: Construct a full-storage runoff generation module, calculate the net rainfall based on the corrected total evapotranspiration, and determine whether runoff occurs in combination with the current total soil moisture content, and calculate the total runoff depth;

[0049] Based on the basin storage capacity curve, calculate the total runoff depth and its impervious component, and its expression is as follows:

[0050] ;

[0051] ;

[0052] ;

[0053] The runoff generation calculation takes the net rainfall PE as the core driving factor, where PE = P - E, P is the rainfall in a time period, and E is the actual evapotranspiration. According to different combinations of the net rainfall PE and the basin storage state, the runoff generation is divided into the following three situations:

[0054] When the net rainfall PE ≤ 0, it indicates that there is no effective runoff generation. At this time, the total runoff depth R is:

[0055] ;

[0056] When the net rainfall PE > 0 and satisfies the condition PE + A < WMM, it indicates that only a partial area of the basin has full-storage runoff generation. At this time, the total runoff depth R is calculated as follows:

[0057] ;

[0058] When net rainfall PE > 0 and the condition PE + A ≥ WMM is met, it indicates that the entire cross-section of the watershed has been filled, and all subsequent net rainfall is converted into runoff. At this time, the total runoff depth R is calculated using the following formula:

[0059] ;

[0060] In the formula: F is the total area of ​​the watershed; f represents the area within the watershed where the vadose zone water storage capacity is less than or equal to WM; WM' is the vadose zone water storage capacity value at each point within the watershed; WMM is the maximum value of the vadose zone water storage capacity; IM is the proportion of the impermeable area of ​​the watershed to the total area; B is the tensile water storage capacity distribution curve index, whose value reflects the spatial non-uniformity of the vadose zone water storage capacity of the watershed; WM is the average tensile water storage capacity of the watershed; W is the average soil moisture content of the watershed at the beginning of the calculation period; A is the vertical coordinate value of the average water storage capacity of the watershed corresponding to W; E0 is the evapotranspiration capacity of the watershed; E is the actual evapotranspiration of the watershed; P is the rainfall during the calculation period.

[0061] Step S2.3: Construct a three-source water division module. Based on the free water capacity, distribution index and outflow coefficient, divide the total runoff of the target watershed into surface runoff RS, interflow RI and groundwater runoff RG, and calculate the runoff volume of surface runoff, interflow and groundwater runoff.

[0062] Key parameters include: surface soil free water storage capacity SM (mm), the power of the curve EX characterizing the spatial non-uniformity of free water storage capacity, the daily flow coefficient KI (1 / day) of free water to soil midflow, and the daily flow coefficient KG (1 / day) of free water to groundwater runoff.

[0063] Define the maximum point free water storage capacity MS within the watershed, which is determined by the average free water storage capacity SM and the power of the distribution curve EX:

[0064] ;

[0065] Calculate the ordinate AU of the free water storage capacity curve corresponding to the current average surface free water storage S. This value reflects the water storage status of the free water reservoir:

[0066] ;

[0067] The generation rates of interflow RI and groundwater runoff RG are determined by the current free water storage S, the runoff-generating area ratio FR, and their respective outflow coefficients, and the calculation formulas are as follows:

[0068] ;

[0069] ;

[0070] The calculation logic of surface runoff RS depends on the interaction between net rainfall PE and the current state of the free water storage reservoir, which is specifically divided into two cases:

[0071] When net rainfall PE ≤ 0 and PE + AU < MS, it indicates that the free water reservoir is not full and there is no net rainfall replenishment or there are losses. At this time, the calculation of surface runoff RS needs to consider the regulation effect of the reservoir:

[0072] ;

[0073] When net rainfall PE > 0 and PE + AU ≥ MS, it indicates that the free water reservoir is full, and the excess net rainfall will directly form surface runoff:

[0074] ;

[0075] In the formula: S is the initial free water storage in the surface soil layer during a period (mm); MS is the maximum free water storage capacity at a single point in the basin (mm); FR is the proportion of the runoff generation area; AU is the ordinate value of the free water storage capacity curve corresponding to the free water storage S (mm); F is the basin area (km²). This module realizes the reasonable division of total runoff at the vertical level through the above formula.

[0076] Step S2.4: Construct a hillslope runoff concentration and river channel runoff concentration module, use the linear reservoir method to calculate the hillslope runoff concentration to obtain the total inflow into the river network, and further use the Muskingum method to calculate the river channel runoff concentration to obtain the total flow at the outlet section of the last basin;

[0077] Use the linear reservoir method to calculate the hillslope runoff concentration:

[0078] The concentration time of surface runoff on the slope is extremely short, and its regulation effect can be ignored. It is usually considered that it instantaneously enters the river network, then:

[0079] ;

[0080] Among them, QS is the surface runoff outflow (m³ / s), U is the unit conversion coefficient used to convert the runoff depth (mm) into flow (m³ / s), and the calculation formula is:

[0081] ;

[0082] In the formula: F is the area of the sub-basin or calculation unit (km²), and Δt is the length of the calculation period (hours).

[0083] Interflow moves laterally in soil pores with a slow speed, and the linear reservoir is used to simulate its regulation process. The calculation formula (recursive form) of interflow outflow QI is:

[0084] ;

[0085] In the formula, CI is the subsurface flow recession coefficient (0 < CI < 1), which reflects the delay degree of subsurface flow concentration; QI(t - 1) is the subsurface flow discharge at the outlet of the previous time period; RI(t) is the subsurface flow depth generated in this time period.

[0086] The baseflow moves in a deeper aquifer, with the slowest velocity and the most significant regulating effect. The linear reservoir is also used for simulation. The calculation formula for the baseflow discharge QG is:

[0087] ;

[0088] In the formula: CG is the baseflow recession coefficient (0 < CG < 1, usually CG > CI), which reflects the extremely slow recession characteristic of the baseflow; QG(t - 1) is the baseflow discharge at the outlet of the previous time period; RG(t) is the baseflow depth generated in this time period.

[0089] After the hillslope runoff concentration stage, the total inflow QT(t) into the river network at time t is the sum of the three:

[0090] ;

[0091] Among them: , is the total flow rate (m³ / s) at time t; L is the river network concentration lag time; CS is the river network water flow recession coefficient.

[0092] The Muskingum method is used for river channel runoff calculation:

[0093] The Muskingum method is a commonly used linear river channel flood routing method. It calculates the inflow process at the upstream section into the outflow process at the downstream section through the water balance equation and the linear storage - discharge relationship, and can better reflect the translation and attenuation effects of floods.

[0094] The basic calculation formula of the Muskingum method is:

[0095] ;

[0096] In the formula:

[0097] is the total flow rate (m³ / s) at the end of this time period at the outlet section of the basin, that is, the final output of this module; and are the total inflows (m³ / s) into the river network in this time period and the previous time period respectively; is the flow rate at the outlet section at the end of the previous time period (m³ / s);

[0098] C0, C1, and C2 are the calculation coefficients of the Muskingan method, which are determined by the hydraulic characteristic parameters of the river channel. The calculation formula is as follows:

[0099] ;

[0100] ;

[0101] ;

[0102] And satisfy ;

[0103] in:

[0104] K S is the storage constant of the river section, with the dimension of time (usually hours), representing the propagation time of the flood wave in the river section and reflecting the translation effect; x is the flow proportion factor (dimensionless, usually 0≤x≤0.5), reflecting the regulation and storage capacity of the river section. The smaller the value of x, the stronger the regulation and storage effect and the more obvious the flood flattening; Δt is the calculation period length (which needs to be compared with K). S The units are consistent, usually in hours.

[0105] Step S2.5: Combine the three-layer evapotranspiration calculation module, the full-soil runoff generation module, the three-water source division module, and the slope confluence and river confluence module to form a complete XAJ model. Output 29 intermediate physical quantities, including date, station, three-layer soil evapotranspiration, three-layer soil moisture content, three-water source runoff, and total flow at the end-basin outlet section.

[0106] Step S3: Construct a soil moisture content LSTM model, extract feature vectors from the intermediate physical quantities output by the XAJ model, input them into the soil moisture content LSTM model, and output the predicted soil moisture content of the three layers in the next time period.

[0107] The specific steps of step S3 are as follows:

[0108] Step S3.1: Based on the intermediate physical quantities output by the XAJ model, extract daily evaporation, net rainfall, precipitation, current upper soil moisture content, current lower soil moisture content, and current deep soil moisture content to form a 6-dimensional feature vector;

[0109] Step S3.2: Use the soil moisture content (WU, WL, WD) of the three layers immediately preceding the next time period as the prediction target vector of the model;

[0110] Step S3.3: Construct a soil moisture content LSTM model, input a 6-dimensional feature vector, the hidden layer size of the soil moisture content LSTM model is 64 neurons, the sequence length is 5 days, and the end connection is a fully connected layer to output the predicted three-layer soil moisture content;

[0111] The LSTM model calculation formula is as follows:

[0112] Forgotten Gate:

[0113] ;

[0114] in, For the Gate of Oblivion For the input vector, It is the sigmoid activation function. It is the weight matrix of the forget gate. It is the bias vector of the forget gate. This indicates that two vectors are joined together.

[0115] Input Gate:

[0116] ;

[0117] Where δ is the activation function. For input gate, It is the weight matrix of the input gate. It is the bias vector of the input gate.

[0118] Candidate values:

[0119] ;

[0120] in, Candidate cell state, It is the weight matrix of the candidate cell states. It is the bias vector of the candidate cell state.

[0121] Cell state:

[0122] ;

[0123] in, In cellular state, This represents element-wise multiplication (Hadamard product). It represents the cell state at the previous time step.

[0124] Output gate:

[0125] ;

[0126] Where δ is the activation function. For output gate, It is the weight matrix of the output gate. It is the bias vector of the output gate.

[0127] Hidden state:

[0128] ;

[0129] in, It is in a hidden state.

[0130] Step S3.4: Train the soil moisture content LSTM model using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and an early stopping threshold of 10 rounds; obtain the trained soil moisture content LSTM model.

[0131] Step S3.5: Standardize the input feature vector and the predicted target vector respectively.

[0132] Step S4: Construct a soil evapotranspiration LSTM model, extract feature vectors from the intermediate physical quantities output by the XAJ model, input them into the soil evapotranspiration LSTM model, and output the predicted evapotranspiration of the three soil layers for the current time period.

[0133] Reference Figure 4 The specific steps of step S4 are as follows:

[0134] Step S4.1: Based on the intermediate physical quantities output by the XAJ model, extract rainfall, evaporation capacity, current upper soil moisture content, current lower soil moisture content, current deep soil moisture content, and current soil evapotranspiration to form a 6-dimensional feature vector;

[0135] Step S4.2: Use the evapotranspiration of the three soil layers (EU, EL, ED) in the current time period as the prediction target vector of the model;

[0136] Step S4.3: Construct a soil evapotranspiration LSTM model, using the same network structure as the soil moisture content LSTM model to ensure consistent parameter scale;

[0137] Step S4.4: Train the soil evapotranspiration LSTM model using the same hyperparameter settings as the soil moisture content LSTM model to ensure consistency in the training process; obtain the trained soil evapotranspiration LSTM model.

[0138] Step S4.5: Standardize the input feature vector and the predicted target vector respectively.

[0139] Step S5: Calculate the corrected three-layer soil evapotranspiration by combining the soil moisture content LSTM model and the XAJ model. Then, perform weighted fusion on the corrected three-layer soil evapotranspiration and the three-layer soil evapotranspiration predicted by the soil evapotranspiration LSTM model to obtain the fused weighted three-layer soil evapotranspiration.

[0140] The soil moisture content of the three layers predicted by the LSTM model is written back into the XAJ model and iterated to obtain the corrected soil moisture content of the three layers.

[0141] The specific steps of step S5 are as follows:

[0142] Step S5.1: Load the initial XAJ model calculation results and obtain the initial soil moisture content of the three layers;

[0143] Step S5.2: Run the XAJ model to calculate the evapotranspiration and runoff of the three soil layers;

[0144] Step S5.3: Call the soil moisture content LSTM model and the soil evapotranspiration LSTM model respectively to obtain the predicted soil moisture content and soil evapotranspiration of the three layers;

[0145] Using the LSTM model for soil moisture content, and inputting the feature vectors from day [t-4] to [t], predict the soil moisture content WU of the three layers on day t+1. pred WL pred WD pred ;

[0146] Using the LSTM model for soil evapotranspiration, and inputting the feature vectors from day [t-4] to day [t], predict the evapotranspiration (EU) of the three soil layers on day t. pred EL pred ED pred ;

[0147] Step S5.4: Obtain the calculation results of the evapotranspiration of the three soil layers for the three models;

[0148] Evapotranspiration of the three soil layers calculated by the XAJ model: Evapotranspiration of the upper soil layer (EU) according to the XAJ model. orig Evapotranspiration EL of the middle soil layer in the XAJ model orig Deep soil evapotranspiration ED from the XAJ model orig ;

[0149] Corrected evapotranspiration of three soil layers was calculated using a combination of the soil moisture LSTM model and the XAJ model: The upper soil moisture content (WU) was calculated using the soil moisture LSTM model. pred Middle soil moisture content WL pred WD (Moisture content of middle soil layer) pred The corrected upper soil evapotranspiration EU was recalculated within the XAJ model framework. soil Evapotranspiration of middle soil layer EL soil Deep soil evapotranspiration ED soil ;

[0150] Evapotranspiration of three soil layers calculated by the soil evapotranspiration LSTM model: Evapotranspiration (EU) of the upper soil layer directly predicted based on the soil evapotranspiration LSTM model. pred Evapotranspiration of middle soil layer EL pred Deep soil evapotranspiration ED pred ;

[0151] Step S5.5: Perform the fusion calculation, as shown in the following formula:

[0152] ;

[0153] ;

[0154] ;

[0155] in, This refers to the evapotranspiration of the upper soil layer after weighting. This refers to the evapotranspiration of the middle soil layer after weighting. This refers to the deep soil evapotranspiration after weighting.

[0156] Step 5.6: Write back the soil moisture content predicted by the LSTM model to the XAJ model state variables to obtain the corrected soil moisture content for the three layers.

[0157] Step 5.7: Repeat steps 5.2-5.6 to form a closed loop of prediction-writeback-recalculation;

[0158] Step 5.8: Stop when the Nash efficiency coefficient reaches the preset threshold or the maximum number of iterations is reached.

[0159] In this embodiment, the preset threshold for the Nash efficiency coefficient is set to 0.95, and the maximum number of iterations is set to 5.

[0160] Step S6: For the target watershed, call the soil moisture LSTM model and the soil evapotranspiration LSTM model. The final output includes the original three-layer soil moisture content and three-layer soil evapotranspiration calculated by the XAJ model, the corrected three-layer soil evapotranspiration, the corrected three-layer soil moisture content, the three-layer soil evapotranspiration after fusion weight, and the fusion weight.

[0161] The specific steps of step S6 are as follows:

[0162] Step 6.1: Calculate net rainfall based on the evapotranspiration of the three soil layers after weighting, and perform runoff calculation;

[0163] Step 6.2: Divide the total runoff into surface runoff, interflow, and groundwater runoff;

[0164] Step 6.3: Use the Muskingen method to calculate the river confluence and obtain the total discharge at the outlet section of the final basin;

[0165] Step 6.4: Calculate the runoff error and ensure that the water balance error is less than 0.1 mm;

[0166] Step 6.5: Output the original three-layer soil moisture content and three-layer soil evapotranspiration calculated by the XAJ model, the corrected three-layer soil evapotranspiration, the corrected three-layer soil moisture content, the three-layer soil evapotranspiration after fusion weighting, and the fusion weights, in the form of an Excel file.

[0167] Step S7: Preset rate periodic and validation period, optimize and evaluate the soil moisture content LSTM model and the soil evapotranspiration LSTM model.

[0168] The specific steps of step S7 are as follows:

[0169] Step S7.1: Preset rate periodicity and verification period;

[0170] Step S7.2: Regularly train the soil moisture content LSTM model and the soil evapotranspiration LSTM model, and optimize the fusion weights;

[0171] Step S7.3: Evaluate model performance during the validation period and calculate the Nash efficiency coefficient;

[0172] Step S7.4: Draw a comparison chart of the flow process lines during the periodic and validation periods;

[0173] Step S7.5: Analyze the improvement effect of the model accuracy before and after the correction of the soil moisture content and evapotranspiration of the three soil layers.

[0174] The embodiments of the present invention have been described in detail above with reference to the accompanying drawings. However, the present invention is not limited to the above embodiments. Within the scope of knowledge possessed by those skilled in the art, various changes can be made without departing from the spirit of the present invention.

Claims

1. A method for correcting soil moisture content and evapotranspiration based on XAJ-LSTM, characterized in that, Perform the following steps S1-S7 to complete the synchronous correction of the soil moisture content and evapotranspiration of the three soil layers: Step S1: Load the basic geographic and hydrological data of the target watershed, including the daily rainfall of each rain gauge station, the daily evaporation of each hydrological station, the measured daily flow, the station name and its corresponding sub-watershed control area and area ratio; Step S2: Build the XAJ model, including a three-layer evapotranspiration calculation module, a full-storage runoff generation module, a three-source water division module, and a slope confluence and river confluence module; run the XAJ model to obtain intermediate physical quantities including date, site, evapotranspiration of the three soil layers, soil moisture content of the three soil layers, runoff of the three water sources, and total flow at the outlet section of the final watershed. Step S3: Construct a soil moisture content LSTM model, extract feature vectors from the intermediate physical quantities output by the XAJ model, input them into the soil moisture content LSTM model, and output the predicted soil moisture content of the three layers in the next time period. Step S4: Construct a soil evapotranspiration LSTM model, extract feature vectors from the intermediate physical quantities output by the XAJ model, input them into the soil evapotranspiration LSTM model, and output the predicted evapotranspiration of the three soil layers for the current time period. Step S5: Calculate the corrected three-layer soil evapotranspiration by combining the soil moisture content LSTM model and the XAJ model. Then, perform weighted fusion on the corrected three-layer soil evapotranspiration and the three-layer soil evapotranspiration predicted by the soil evapotranspiration LSTM model to obtain the fused weighted three-layer soil evapotranspiration. The soil moisture content of the three layers predicted by the LSTM model is written back into the XAJ model and iterated to obtain the corrected soil moisture content of the three layers. Step S6: For the target watershed, call the soil moisture LSTM model and the soil evapotranspiration LSTM model. The final output includes the original three-layer soil moisture content and three-layer soil evapotranspiration calculated by the XAJ model, the corrected three-layer soil evapotranspiration, the corrected three-layer soil moisture content, the three-layer soil evapotranspiration after fusion weight, and the fusion weight. Step S7: Preset rate periodic and validation period, optimize and evaluate the soil moisture content LSTM model and the soil evapotranspiration LSTM model.

2. The method for correcting soil moisture content and evapotranspiration based on XAJ-LSTM according to claim 1, characterized in that, The specific steps of step S1 are as follows: Step S1.1: Collect daily rainfall data from various rain gauge stations, daily evaporation data from hydrological stations, and measured daily flow data from each station in the target watershed; Step S1.2: Identify the date columns in each file and convert them to a uniform type; Step S1.3: Extract the site name, sub-basin area and area ratio, and perform inner join matching; Step S1.4: Read the site flow results output by the first XAJ model and compare them with the total flow results, and extract the evapotranspiration of the three soil layers, the soil moisture content of the three soil layers, the runoff of each water source and the outlet flow. Step S1.5: Divide all the acquired data into a preset periodicity and a verification period.

3. The method for correcting soil moisture content and evapotranspiration based on XAJ-LSTM according to claim 1, characterized in that, The specific steps of step S2 are as follows: Step S2.1: Construct a three-layer evapotranspiration calculation module to calculate the evapotranspiration of the upper, lower, and deep soil layers in the target watershed based on the current soil moisture content, rainfall, and evaporation capacity of the upper, lower, and deep soil layers respectively. Step S2.2: Construct a full-storage runoff generation module, calculate net rainfall based on the corrected total evapotranspiration, determine whether runoff generation is possible by combining the current total soil moisture content, and calculate the total runoff depth; Step S2.3: Construct a three-source division module. Based on free water capacity, distribution index and outflow coefficient, divide the total runoff of the target watershed into surface runoff, interflow and groundwater runoff, and calculate the runoff volume of surface runoff, interflow and groundwater runoff. Step S2.4: Construct slope confluence and river confluence modules, use the linear reservoir method to calculate slope confluence and obtain the total inflow into the river network, and further use the Muskingen method to calculate river confluence and obtain the total discharge at the outlet section of the final watershed. Step S2.5: Combine the three-layer evapotranspiration calculation module, the full-soil runoff generation module, the three-water source division module, and the slope confluence and river confluence module to form a complete XAJ model, and output intermediate physical quantities including date, station, three-layer soil evapotranspiration, three-layer soil moisture content, three-water source runoff, and total flow at the end-basin outlet section.

4. The method for correcting soil moisture content and evapotranspiration based on XAJ-LSTM according to claim 1, characterized in that, The specific steps of step S3 are as follows: Step S3.1: Based on the intermediate physical quantities output by the XAJ model, extract daily evaporation, net rainfall, precipitation, current upper soil moisture content, current lower soil moisture content, and current deep soil moisture content to form a 6-dimensional feature vector; Step S3.2: Use the soil moisture content of the three layers immediately preceding the next time period as the prediction target vector of the model; Step S3.3: Construct a soil moisture content LSTM model, input a 6-dimensional feature vector, the hidden layer size of the soil moisture content LSTM model is 64 neurons, the sequence length is 5 days, and the end connection is a fully connected layer to output the predicted three-layer soil moisture content; Step S3.4: Train the soil moisture content LSTM model using the Adam optimizer with a learning rate of 0.001, a batch size of 32, and an early stopping threshold of 10 rounds; obtain the trained soil moisture content LSTM model. Step S3.5: Standardize the input feature vector and the predicted target vector respectively.

5. The method for correcting soil moisture content and evapotranspiration based on XAJ-LSTM according to claim 1, characterized in that, The specific steps of step S4 are as follows: Step S4.1: Based on the intermediate physical quantities output by the XAJ model, extract rainfall, evaporation capacity, current upper soil moisture content, current lower soil moisture content, current deep soil moisture content, and current soil evapotranspiration to form a 6-dimensional feature vector; Step S4.2: Use the evapotranspiration of the three soil layers in the current time period as the prediction target vector of the model; Step S4.3: Construct a soil evapotranspiration LSTM model, using the same network structure as the soil moisture content LSTM model to ensure consistent parameter scale; Step S4.4: Train the soil evapotranspiration LSTM model using the same hyperparameter settings as the soil moisture content LSTM model to ensure consistency in the training process; obtain the trained soil evapotranspiration LSTM model. Step S4.5: Standardize the input feature vector and the predicted target vector respectively.

6. The method for correcting soil moisture content and evapotranspiration based on XAJ-LSTM according to claim 1, characterized in that, The specific steps of step S5 are as follows: Step S5.1: Load the initial XAJ model calculation results and obtain the initial soil moisture content of the three layers; Step S5.2: Run the XAJ model to calculate the evapotranspiration and runoff of the three soil layers; Step S5.3: Call the soil moisture content LSTM model and the soil evapotranspiration LSTM model respectively to obtain the predicted soil moisture content and soil evapotranspiration of the three layers; Using the LSTM model for soil moisture content, and inputting the feature vectors from day [t-4] to [t], predict the soil moisture content WU of the three layers on day t+1. pred WL pred WD pred ; Using the LSTM model for soil evapotranspiration, and inputting the feature vectors from day [t-4] to day [t], predict the evapotranspiration (EU) of the three soil layers on day t. pred EL pred ED pred ; Step S5.4: Obtain the calculation results of the evapotranspiration of the three soil layers for the three models; Evapotranspiration of the three soil layers calculated by the XAJ model: Evapotranspiration of the upper soil layer (EU) according to the XAJ model. orig Evapotranspiration EL of the middle soil layer in the XAJ model orig Deep soil evapotranspiration ED from the XAJ model orig ; Corrected evapotranspiration of three soil layers was calculated using a combination of the soil moisture LSTM model and the XAJ model: The upper soil moisture content (WU) was calculated using the soil moisture LSTM model. pred Middle soil moisture content WL pred WD (Moisture content of middle soil layer) pred The corrected upper soil evapotranspiration EU was recalculated within the XAJ model framework. soil Evapotranspiration of middle soil layer EL soil Deep soil evapotranspiration ED soil ; Evapotranspiration of three soil layers calculated by the soil evapotranspiration LSTM model: Evapotranspiration (EU) of the upper soil layer directly predicted based on the soil evapotranspiration LSTM model. pred Evapotranspiration of middle soil layer EL pred Deep soil evapotranspiration ED pred ; Step S5.5: Perform the fusion calculation, as shown in the following formula: ； ； ； in, This refers to the evapotranspiration of the upper soil layer after weighting. This refers to the evapotranspiration of the middle soil layer after weighting. This refers to the deep soil evapotranspiration after weighting. Step 5.6: Write back the soil moisture content predicted by the LSTM model to the XAJ model state variables to obtain the corrected soil moisture content for the three layers. Step 5.7: Repeat steps 5.2-5.6 to form a closed loop of prediction-writeback-recalculation; Step 5.8: Stop when the Nash efficiency coefficient reaches the preset threshold or the maximum number of iterations is reached.

7. The method for correcting soil moisture content and evapotranspiration based on XAJ-LSTM according to claim 1, characterized in that, The specific steps of step S6 are as follows: Step 6.1: Calculate net rainfall based on the evapotranspiration of the three soil layers after weighting, and perform runoff calculation; Step 6.2: Divide the total runoff into surface runoff, interflow, and groundwater runoff; Step 6.3: Use the Muskingen method to calculate the river confluence and obtain the total discharge at the outlet section of the final basin; Step 6.4: Calculate the runoff error and ensure that the water balance error is less than 0.1 mm; Step 6.5: Output includes the original three-layer soil moisture content and three-layer soil evapotranspiration calculated by the XAJ model, the corrected three-layer soil evapotranspiration, the corrected three-layer soil moisture content, the three-layer soil evapotranspiration after fusion weighting, and the fusion weights.

8. The method for correcting soil moisture content and evapotranspiration based on XAJ-LSTM according to claim 1, characterized in that, The specific steps of step S7 are as follows: Step S7.1: Preset rate periodicity and verification period; Step S7.2: Regularly train the soil moisture content LSTM model and the soil evapotranspiration LSTM model, and optimize the fusion weights; Step S7.3: Evaluate model performance during the validation period and calculate the Nash efficiency coefficient; Step S7.4: Draw a comparison chart of the flow process lines during the periodic and validation periods; Step S7.5: Analyze the improvement effect of the model accuracy before and after the correction of the soil moisture content and evapotranspiration of the three soil layers.

Images