Evapotranspiration inversion method and system for data-scarce river basin

Abstract

The application relates to the technical field of hydrology, water resources and remote sensing information processing, and provides an evapotranspiration inversion method and system for a data-scarce basin, which comprises the following steps: dividing a basin into grid units based on a digital elevation model and constructing a hydrological topological relation graph; constructing a runoff generation parameter generation network to output runoff generation parameters of a target grid unit, and constructing a concentration parameter generation network to output concentration parameters of each grid unit; calculating a preliminary evapotranspiration and a total runoff based on the runoff generation parameters; performing grid-by-grid concentration calculation based on the concentration parameters to obtain a basin outlet runoff process; constructing a deep learning correction network to dynamically correct the preliminary evapotranspiration and output a final evapotranspiration inversion result; and constructing a loss function containing runoff simulation errors and evapotranspiration simulation errors to train the runoff generation parameter generation network, the concentration parameter generation network and the deep learning correction network. In this way, the accuracy of evapotranspiration estimation results is improved.

Description

Technical Field

[0001] This application relates to the field of hydrological and water resources and remote sensing information processing technology, specifically to a method and system for evapotranspiration inversion in data-scarce watersheds. Background Technology

[0002] Evapotranspiration is a core link connecting the Earth's hydrological and energy cycles. Accurate inversion of its spatiotemporal distribution is a crucial foundation for optimal reservoir water resource management, agricultural drought monitoring and early warning, eco-hydrological process analysis, and climate change response research. my country's vast majority of small and medium-sized reservoir basins are located in mountainous and hilly areas with complex topography, strong underlying surface heterogeneity, and extremely scarce hydrological and meteorological observation stations. Typically, only one evaporation station and one runoff station at the basin outlet are deployed, resulting in short observation sequences and inconsistent data quality, making them typical data-scarce basins. How to achieve high spatiotemporal resolution, strong physical consistency, and high generalization ability in evapotranspiration inversion under limited observation conditions is a core scientific and engineering problem urgently needing to be solved in the fields of hydrology, water resources, and remote sensing. Summary of the Invention

[0003] This application aims to address the problem of poor accuracy in existing evapotranspiration estimation schemes, and to provide an evapotranspiration inversion method and system for watersheds with scarce data.

[0004] To solve the above problems, this application is implemented as follows:

[0005] Firstly, this application provides a method for evapotranspiration inversion in data-scarce watersheds, including: Based on digital elevation model data of a data-scarce watershed, the data-scarce watershed is divided into multiple grid cells, and a hydrological topology map of the data-scarce watershed is constructed based on the hydrological topology relationships between the grid cells. A runoff generation parameter generation network is constructed, and the static environmental attributes and dynamic meteorological characteristics of the target grid cell in the plurality of grid cells are input into the runoff generation parameter generation network to output the runoff parameters of the target grid cell; A confluence parameter generation network is constructed, and the grid terrain and land use features of each grid cell are input into the confluence parameter generation network to output the confluence parameters of each grid cell. Under the condition of coupling the current soil moisture content of the target grid cell, the preliminary evapotranspiration of the target grid cell in the current time period is calculated based on the runoff parameters; Calculate the total runoff of the target grid cell based on the soil infiltration capacity, real-time precipitation intensity, and runoff parameters. Using the confluence parameters of each grid cell, the calculation order is determined according to the hydrological topology diagram. The confluence calculation is performed grid by grid from upstream to downstream. The surface runoff, interflow and groundwater runoff generated by the multiple grid cells are collected to the unique outlet section of the data-scarce watershed to obtain the outlet runoff process of the data-scarce watershed. A deep learning correction network is constructed, and the preliminary evapotranspiration, the current soil moisture content, the dynamic meteorological characteristics, and the measured evapotranspiration residual of the grid cell where the evaporation station is located are input into the deep learning correction network to dynamically correct the preliminary evapotranspiration and output the final evapotranspiration inversion result. A multi-scale, multi-objective loss function is constructed, which includes runoff simulation error and evapotranspiration simulation error. The runoff simulation error is the error between the outlet runoff process and the measured runoff at the hydrological station, and the evapotranspiration simulation error is the error between the evapotranspiration value of the final evapotranspiration inversion result in the grid cell where the evaporation station is located and the measured evapotranspiration value at the evaporation station. The flow generation parameter generation network, the flow sinking parameter generation network, and the deep learning correction network are trained based on the multi-scale multi-objective loss function to obtain the trained flow generation parameter generation network, the flow sinking parameter generation network, and the deep learning correction network. The final evapotranspiration inversion result and the total yield are used to update the soil moisture state of the target grid cell to obtain the corrected soil moisture content. The corrected soil moisture content is used to calculate the preliminary evapotranspiration of the next time period, which is the period after the current time period.

[0006] Secondly, this application provides an evapotranspiration inversion system for data-scarce watersheds, comprising: The module is used to divide the data-scarce watershed into multiple grid cells based on the digital elevation model data of the data-scarce watershed, and to construct a hydrological topology map of the data-scarce watershed based on the hydrological topology relationships between the grid cells. The construction module is used to construct a runoff generation parameter generation network, and inputs the static environmental attributes and dynamic meteorological characteristics of the target grid cell in the plurality of grid cells into the runoff generation parameter generation network to output the runoff parameters of the target grid cell; The construction module is used to construct a confluence parameter generation network, inputting the grid terrain and land use features of each grid cell into the confluence parameter generation network to output the confluence parameters of each grid cell. The calculation module is used to calculate the preliminary evapotranspiration of the target grid cell in the current time period based on the runoff parameters, under the condition of coupling the current soil moisture content of the target grid cell; The calculation module is used to calculate the total runoff of the target grid cell based on the soil infiltration capacity, real-time precipitation intensity, and runoff parameters of the target grid cell. The calculation module is used to use the confluence parameters of each grid cell, determine the calculation order according to the hydrological topology diagram, and perform confluence calculation grid by grid from upstream to downstream. The surface runoff, interflow and groundwater runoff generated by the multiple grid cells are collected to the only outlet section of the data-scarce watershed to obtain the outlet runoff process of the data-scarce watershed. The construction module is used to construct a deep learning correction network, inputting the preliminary evapotranspiration, the current soil moisture content, the dynamic meteorological characteristics, and the measured evapotranspiration residual of the grid cell where the evaporation station is located into the deep learning correction network, dynamically correcting the preliminary evapotranspiration, and outputting the final evapotranspiration inversion result; The construction module is used to construct a multi-scale, multi-objective loss function including runoff simulation error and evapotranspiration simulation error. The runoff simulation error is the error between the outlet runoff process and the measured runoff at the hydrological station, and the evapotranspiration simulation error is the error between the evapotranspiration value of the final evapotranspiration inversion result in the grid cell where the evaporation station is located and the measured evapotranspiration value at the evaporation station. The training module is used to train the flow generation parameter generation network, the flow confluence parameter generation network, and the deep learning correction network based on the multi-scale multi-objective loss function, so as to obtain the trained flow generation parameter generation network, the flow confluence parameter generation network, and the deep learning correction network. The final evapotranspiration inversion result and the total yield are used to update the soil moisture state of the target grid cell to obtain the corrected soil moisture content. The corrected soil moisture content is used to calculate the preliminary evapotranspiration of the next time period, which is the period after the current time period.

[0007] Thirdly, this application provides a terminal device including a processor and a memory, wherein the memory stores a program or instructions executable on the processor, and the program or instructions, when executed by the processor, implement the steps of the method described in the first aspect.

[0008] Fourthly, this application provides a readable storage medium on which a program or instructions are stored, which, when executed by a processor, implement the steps of the method described in the first aspect.

[0009] Fifthly, this application provides a chip including a processor and a communication interface, the communication interface being coupled to the processor, the processor being used to run programs or instructions to implement the steps of the method described in the first aspect.

[0010] In a sixth aspect, this application provides a computer program product stored in a storage medium, which is executed by at least one processor to perform the steps of the method described in the first aspect.

[0011] Compared with existing technologies, this application has the following advantages: Compared to traditional hydrological models that rely on a large amount of ground observation data for parameter calibration, this application only requires runoff observations at the watershed outlet and a small number of evaporation station observations to drive daily evapotranspiration inversion across the entire watershed grid, significantly reducing reliance on the ground observation network and making evapotranspiration monitoring possible in data-scarce areas; by simultaneously incorporating energy balance and water balance into the evapotranspiration calculation, the inversion results conform to both atmospheric energy-driven laws and the watershed water closure principle, overcoming the poor physical interpretability of purely data-driven models; through the runoff parameter generation network and A confluence parameter generation network dynamically generates parameters for each grid cell that are adapted to its environmental characteristics. A deep learning correction network performs closed-loop feedback correction on the initial evapotranspiration, effectively suppressing the accumulation of errors in the physical model and significantly improving the spatiotemporal accuracy of evapotranspiration inversion, thus improving the accuracy of evapotranspiration estimation results. In addition, by constructing a fully differentiable hydrological model based on an automatic differentiation framework, the gradient of the runoff error at the watershed outlet can be backpropagated along the confluence path to each upstream grid cell. This realizes weakly supervised learning that drives the optimization of parameters across the entire watershed through single-point observation, providing a new technical path for hydrological simulation under conditions of scarce data. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in this application or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of this application. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating an embodiment of the evapotranspiration inversion method for data-scarce watersheds provided in this application. Figure 2 This is a schematic diagram of the structure of an evapotranspiration inversion system for data-scarce watersheds provided in an embodiment of this application; Figure 3 This is a schematic diagram of the structure of a terminal device provided in an embodiment of this application. Detailed Implementation

[0014] The technical solutions of this application will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of this application. Based on the embodiments of this application, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of this application.

[0015] The terms "first," "second," etc., used in this application are used to distinguish similar objects and are not necessarily used to describe a specific order or sequence. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not necessarily limited to those steps or units explicitly listed, but may include other steps or units not explicitly listed or inherent to these processes, methods, products, or apparatuses. Additionally, the use of "and / or" in this application indicates at least one of the connected objects, such as A and / or B and / or C, representing seven possibilities: including A alone, B alone, C alone, and the presence of both A and B, both B and C, both A and C, and the presence of A, B, and C.

[0016] In this application, the terms "exemplary" or "for example" are used to indicate that something is an example, illustration, or description. Any embodiment or design described as "exemplary" or "for example" in this application should not be construed as being more preferred or advantageous than other embodiments or designs. Rather, the use of the terms "exemplary" or "for example" is intended to present the relevant concepts in a specific manner.

[0017] The evapotranspiration inversion method for data-scarce watersheds provided in this application will be described below.

[0018] See Figure 1 , Figure 1 This is a flowchart illustrating an embodiment of the evapotranspiration inversion method for data-scarce watersheds provided in this application. Figure 1 The evapotranspiration inversion method shown can be executed by terminal devices such as mobile phones and computers.

[0019] like Figure 1 As shown, the evapotranspiration inversion method for data-scarce watersheds provided in this application may include the following steps: Step 101: Based on the digital elevation model data of the data-scarce watershed, the data-scarce watershed is divided into multiple grid units, and a hydrological topology map of the data-scarce watershed is constructed based on the hydrological topology relationships between the grid units.

[0020] Understandably, before step 101, multi-source data of the data-scarce watershed can be acquired in advance. Multi-source data includes underlying surface data, ground station observation data, and meteorological reanalysis data. After acquiring the multi-source data, it can be processed to unify the coordinate system and spatial resolution of the multi-source data, which will facilitate the subsequent application of the multi-source data.

[0021] For example, the acquired multi-source data is unified into the WGS84 coordinate system and a 1km×1km spatial resolution.

[0022] The aforementioned underlying surface data includes: 30m resolution digital elevation model data, which can be resampled to 1km using bilinear interpolation for extracting topographic features; 1km resolution World Soil Database data, used to extract parameters such as soil texture, saturated hydraulic conductivity, field capacity, and wilting coefficient for each grid; 1km resolution MODIS NDVI data, which can be used to generate monthly vegetation cover data using the maximum value synthesis method; and 1km resolution land use data, which can be reclassified into five major categories: cultivated land, forest land, grassland, water area, and construction land.

[0023] The aforementioned ground station observation data includes: daily precipitation data from all precipitation stations within the basin, which can be interpolated to a 1km grid using the inverse distance weighting method; daily runoff data from the hydrological station at the basin outlet, which serves as runoff labels for model training; and daily evaporation data from the evaporation pans of a single evaporation station, converted into water surface evaporation, which can be used as evapotranspiration training labels.

[0024] The meteorological reanalysis data mentioned above includes hourly net surface radiation, 2m air temperature, 10m wind speed, and 2m relative humidity data, which are aggregated into daily averages and resampled to a 1km grid.

[0025] It is understandable that the specific numerical parameters mentioned above can be set based on different application scenarios, and no specific limitations are made here.

[0026] In some embodiments, the elevation increment method can be used to fill depressions in the digital elevation model data to eliminate false depressions; and the D8 single-flow direction algorithm can be used to calculate the water flow direction and cumulative runoff of each grid.

[0027] In some embodiments, data-scarce watersheds can be divided into Each grid cell is used to construct a grid-level upstream and downstream topological connection matrix; and a hydrological topological relationship map of the data-scarce watershed can be constructed based on the hydrological topological relationship between the grid cells. The hydrological topological relationship map is used to determine the calculation order from upstream to downstream and is used for subsequent unified flow calculation across the entire grid.

[0028] Step 102: Construct a runoff generation parameter generation network. Input the static environmental attributes and dynamic meteorological characteristics of the target grid cell in the plurality of grid cells into the runoff generation parameter generation network to output the runoff parameters of the target grid cell.

[0029] The aforementioned target grid cell can be understood as any one of multiple grid cells.

[0030] The aforementioned static environmental attributes include: topographic features such as elevation, slope, aspect, topographic humidity index, and upstream catchment area; soil features such as sand content, silt content, clay content, saturated hydraulic conductivity, and field water holding capacity; and vegetation features such as multi-year average vegetation coverage and land use type coding.

[0031] The aforementioned dynamic meteorological characteristics include meteorological information such as average precipitation, average temperature, and average net radiation.

[0032] In some embodiments, the flow parameter generation network may employ a CNN-Transformer hybrid architecture, enabling it to simultaneously capture spatial neighborhood correlations and temporal dependencies.

[0033] In some embodiments, the aforementioned runoff generation parameters include improved generalized complementary model parameters and runoff generation calculation parameters, wherein the improved generalized complementary model parameters include , , The parameters for generating runoff calculation include the water storage capacity curve index. Soil matrix potential Poor soil moisture content The weighted adjustment coefficient of the runoff generation mechanism Threshold for determining the production mechanism Maximum free water storage capacity of the basin Free water storage capacity curve index soil outflow coefficient and the outflow coefficient of groundwater runoff .

[0034] The static environmental attributes of the target grid cell, such as elevation, slope, topographic humidity index, sand content, silt content, clay content, saturated hydraulic conductivity, field water holding capacity, vegetation cover, and land use type code, as well as the dynamic meteorological characteristics of the target grid cell, such as daily average precipitation, average temperature, and average net radiation, can be used as input features for the runoff parameter generation network to output the runoff parameters of the target grid cell.

[0035] In some embodiments, the runoff generation parameter generation network includes a spatial feature extraction module, a temporal feature extraction module, a feature fusion layer, a parameter output layer, and a physical constraint layer; wherein: The spatial feature extraction module can use a 3-layer 2D CNN model with a kernel size of 3×3, a stride of 1, padding of 1, and output channels of 32, 64, and 128 per layer, respectively, to extract spatial features within a 3×3 neighborhood. The temporal feature extraction module can employ a 2-layer Transformer encoder with 8 attention heads and 128 hidden layer dimensions to capture the temporal dependencies of meteorological sequences. The feature fusion layer is used to concatenate spatial and temporal features into a 256-dimensional fused feature. The parameter output layer can use a 2-layer fully connected network to output 12 flow parameters; The physical constraint layer can use the Sigmoid activation function to map the output to the [0,1] interval, and then use a linear transformation to map it to the physical reasonable range of each parameter.

[0036] Step 103: Construct a confluence parameter generation network. Input the grid terrain and land use features of each grid cell into the confluence parameter generation network to output the confluence parameters of each grid cell.

[0037] In some embodiments, the grid terrain includes topographic features such as elevation, slope, aspect, and upstream catchment area, and the land use features can be land use type encoding.

[0038] In some embodiments, the above-mentioned flow parameters include the storage constant of the grid cell. Flow proportion factor , soil runoff coefficient and underground runoff confluence coefficient .

[0039] In some embodiments, the confluence parameter generation network may adopt a three-layer fully connected architecture, including a first layer, a second layer, a third layer, and a physical constraint layer, wherein: The first layer has a 5-dimensional input and a 64-dimensional output, using the ReLU activation function and a Dropout rate of 0.2. The second layer has a 64-dimensional input and a 32-dimensional output, using the ReLU activation function and a Dropout rate of 0.2. The third layer has a 32-dimensional input and a 4-dimensional output, using the Sigmoid activation function. The physical constraint layer is used to map the output to the physically reasonable range of each bus parameter.

[0040] Step 104: Under the condition of coupling the current soil moisture content of the target grid cell, calculate the preliminary evapotranspiration of the target grid cell in the current time period based on the runoff parameters.

[0041] Step 105: Calculate the total runoff of the target grid cell based on the soil infiltration capacity, real-time precipitation intensity, and runoff parameters of the target grid cell.

[0042] Step 106: Using the confluence parameters of each grid cell, determine the calculation order according to the hydrological topology diagram, and perform confluence calculation grid by grid from upstream to downstream. The surface runoff, interflow, and groundwater runoff generated by the multiple grid cells are collected to the unique outlet section of the data-scarce watershed to obtain the outlet runoff process of the data-scarce watershed.

[0043] Understandably, it is possible to construct a fully distributed hydrological model with a single grid cell as the computational unit, where all physical processes are implemented using PyTorch tensor operations. This fully distributed hydrological model retains the complete computational graph and gradient information and supports end-to-end training.

[0044] The calculation processes in steps 104, 105, and 106 can all be implemented using a fully distributed hydrological model.

[0045] In some embodiments, the calculation of the preliminary evapotranspiration, the total production flow, and the outlet runoff process are all implemented using automatic differential frame tensor operations, and the formulas used are all continuously differentiable mathematical expressions.

[0046] In some embodiments, the FAO Penman-Monteith formula can be used to calculate potential evapotranspiration. :

[0047] in, Net surface radiation, The air temperature is 2m. The wind speed is 2 m, derived from meteorological reanalysis data; This is the wet / dry constant, an empirical value; For soil heat flux (based on) Estimated using empirical formulas, but approximately 0 on a daily scale). The slope of the saturated water vapor pressure curve. The saturated vapor pressure, The actual water vapor pressure is estimated using the following formula:

[0048]

[0049]

[0050] in, This is relative humidity, based on meteorological reanalysis data.

[0051] The preliminary evapotranspiration can be calculated using an improved generalized complementary formula. :

[0052] in, and They are grid cells exist The soil moisture content at time t is the state variable updated iteratively in the soil water balance model, while the maximum soil moisture content is the extracted soil feature. , , To improve the parameters of the generalized complementary model, the flow generation parameters in step 102 are used to generate the network output for the characteristics of each grid cell.

[0053] Understandably, to meet water balance constraints, evapotranspiration cannot exceed the sum of soil moisture content and precipitation.

[0054] In some embodiments, the dominant runoff generation mechanism can be automatically selected based on the underlying surface characteristics and real-time hydrological conditions of each grid cell, enabling adaptive switching between full-saturation runoff, infiltration-excess runoff, and vertically mixed runoff.

[0055] The adaptive switching process between full-flow, over-permeable flow, and vertically mixed flow includes: (1) The mechanism of runoff generation can be determined by constructing a runoff generation mechanism discrimination index. Taking into account both soil infiltration capacity and rainfall intensity:

[0056] in, The saturated hydraulic conductivity of the soil is derived from the extracted soil characteristics. The grid slope is derived from terrain features extracted from digital elevation model data; The precipitation intensity is derived from interpolated data from precipitation stations; To prevent the minimum value from being divided by zero.

[0057] (2) The calculation of full-storage runoff can be based on the Xin'anjiang model's full-storage runoff theory, taking into account the spatial distribution of soil water storage capacity:

[0058] in, The water storage capacity curve index is generated by the flow generation network in step 102, which outputs the characteristics of each grid cell.

[0059] (3) Calculation of excess permeability flow: The excess permeability flow rate can be calculated based on the improved Green-Ampt infiltration model.

[0060]

[0061] in, Infiltration rate; For soil matrix potential, Due to the difference in soil moisture content, both are output by the runoff generation parameter generation network in step 102 for the characteristics of each grid cell. The cumulative infiltration amount is represented by the state variable that is iteratively updated during the infiltration process.

[0062] (4) For mixed runoff calculation, the weights of the two runoff mechanisms can be dynamically allocated according to the discriminant index:

[0063]

[0064] in, The weighting coefficient for the flow generation mechanism; This is the weighting adjustment coefficient for the runoff generation mechanism. Both thresholds for determining the flow generation mechanism are generated by the flow generation parameter generation network in step 102, which outputs the characteristics of each grid cell. For grid cells in time period Total output flow.

[0065] (5) Water source division: input the total production flow into the free water reservoir, and divide the surface runoff, interflow and groundwater runoff based on the free water storage capacity area distribution curve.

[0066] Free water reservoirs assume the free water storage capacity at various points within the watershed. It follows an exponential distribution, and the distribution curve is as follows:

[0067] in, This represents the maximum free water storage capacity of the basin. The free water storage capacity curve index is generated by the flow generation parameter generation network in step 102 for each grid characteristic.

[0068] Surface runoff The calculation formula is as follows:

[0069] in, Let represent the free water storage capacity of the watershed, and represent the state variables that are iteratively updated during the water source allocation process. The remaining water in the free water storage reservoir will generate interflows according to a fixed outflow coefficient. and underground runoff :

[0070]

[0071] in, The outflow coefficient is the soil outflow coefficient. The outflow coefficients for underground runoff are both output by the runoff generation parameter generation network in step 102 for each grid characteristic.

[0072] In some embodiments, all grid cells perform confluence calculations sequentially from upstream to downstream, according to the calculation order determined by the hydrological topology diagram.

[0073] For surface runoff, the Musking root method can be used for calculation. The Musking root method is based on the linear relationship between storage capacity and inflow / outflow, describing the flood propagation process through continuity equations and storage / discharge equations. For arbitrary grids... i Its water balance equation and storage / release equation are as follows:

[0074]

[0075] in, , and They are grid cells i exist t The water storage capacity, total inflow and outflow at any given time; and They are grid cells i The storage constant and flow proportion factor are both output by the confluence parameter generation network in step 102 for each grid characteristic. The implicit finite difference method is used to discretize the basic equations, yielding the Muskingan calculus formula suitable for time-period calculations:

[0076] in , , Here are the Muskingan calculus coefficients, calculated using the following formula:

[0077] The confluence processes of interflow and groundwater runoff were calculated using the linear reservoir method:

[0078]

[0079] in, and These are the runoff coefficients for interflow and subsurface runoff, respectively, both of which are output by the runoff parameter generation network in step 102 for the runoff characteristics of each grid cell. , where is the unit conversion factor; The area of ​​a grid cell, in units .

[0080] Step 107: Construct a deep learning correction network. Input the preliminary evapotranspiration, the current soil moisture content, the dynamic meteorological characteristics, and the measured evapotranspiration residual of the grid cell where the evaporation station is located into the deep learning correction network to dynamically correct the preliminary evapotranspiration and output the final evapotranspiration inversion result.

[0081] The final evapotranspiration inversion result and the total yield are used to update the soil moisture state of the target grid cell to obtain the corrected soil moisture content. The corrected soil moisture content is used to calculate the preliminary evapotranspiration of the next time period, which is the period after the current time period.

[0082] In some embodiments, a deep learning correction network can be constructed based on the U-Net coupled LSTM model, and the preliminary evapotranspiration results of the generalized complementary model can be dynamically corrected using the physical state and meteorological data of the previous period.

[0083] In some embodiments, the deep learning correction network employs a long short-term memory network architecture, including: The encoder, consisting of multi-layer convolutional downsampling modules, is used to extract the spatial structural features of the evaporation field; The bottleneck layer flattens the spatial features output by the encoder into a sequence and inputs it into a long short-term memory network to capture the temporal evolution of the evaporation field. The decoder consists of multiple deconvolution upsampling modules and establishes skip connections with the corresponding layers of the encoder to restore spatial resolution and output the evaporation correction value for each grid cell. The output layer adds the evaporation correction value to the initial evaporation amount to obtain the final evaporation inversion result.

[0084] In some embodiments, the encoder described above may employ 4 layers of convolutional downsampling with a kernel size of 3×3 and a stride of 2. The number of output channels per layer is 32, 64, 128, and 256, respectively. Each layer is followed by BatchNorm and ReLU activation functions. The bottleneck layer is used to flatten the spatial features output by the encoder into a sequence, which is then input into a 1-layer LSTM network with a hidden layer dimension of 256 to capture temporal variation features. The decoder described above may employ 4 layers of deconvolutional upsampling with a kernel size of 3×3 and a stride of 2. The number of output channels per layer is 128, 64, 32, and 1, respectively. It is skipped to the corresponding layer of the encoder. The output layer may employ the Tanh activation function to output the evapotranspiration correction value for each grid cell, and its range may be limited to [-5, 5] mm / d.

[0085] Step 108: Construct a multi-scale, multi-objective loss function that includes runoff simulation error and evapotranspiration simulation error.

[0086] The runoff simulation error is the error between the outlet runoff process and the runoff measured at the hydrological station, and the evapotranspiration simulation error is the error between the evapotranspiration value of the final evapotranspiration inversion result in the grid cell where the evaporation station is located and the evapotranspiration value measured at the evaporation station.

[0087] Step 109: Train the flow generation parameter generation network, the flow confluence parameter generation network, and the deep learning correction network based on the multi-scale multi-objective loss function to obtain the trained flow generation parameter generation network, the flow confluence parameter generation network, and the deep learning correction network.

[0088] In some embodiments, training the runoff generation parameter generation network, the sinking parameter generation network, and the deep learning correction network based on the multi-scale multi-objective loss function to obtain the trained runoff generation parameter generation network, the sinking parameter generation network, and the deep learning correction network includes: An automatic differential framework is used to calculate the gradients of the multi-scale, multi-objective loss function on the trainable parameters in the runoff generation network, the runoff generation network, and the deep learning correction network. The gradient of the runoff simulation error is backpropagated through differentiable physical processes in the runoff calculation, runoff generation, and evapotranspiration calculation processes to adjust the runoff generation and runoff generation parameters of all grid cells in the data-scarce watershed. The gradient of the evapotranspiration simulation error is backpropagated through the deep learning correction network to update its network weights, and through the connection between the deep learning correction network and the preliminary evapotranspiration calculation process, it is backpropagated to the preliminary evapotranspiration calculation process to adjust the runoff generation parameters of the grid cell where the evaporation station is located. Based on the calculated gradient, the flow generation parameter generation network, the flow sinking parameter generation network, and the deep learning correction network are trained to obtain the trained flow generation parameter generation network, the flow sinking parameter generation network, and the deep learning correction network.

[0089] In some embodiments, the training strategy for the above model includes: (1) Data set partitioning, such as dividing the dataset into training set, validation set and test set; (2) Phased training: In the first phase (pre-training), the evapotranspiration correction network can be fixed, and only the parameter generation network and hydrological model can be trained and iterated for 50 rounds; In the second phase (joint training), all parameters are unfrozen, end-to-end joint training is carried out and iterated for 150 rounds. (3) The AdamW optimizer can be used, with an initial learning rate of 0.001, which decays to 0.5 every 30 rounds; the weight decay coefficient is 1e-5. (4) Early stopping mechanism: when the validation set loss no longer decreases for 20 consecutive rounds, training is terminated early to save the optimal model weights and complete the training of the model.

[0090] After training is completed, input the meteorological data for the validation period and evaluate the daily runoff and evapotranspiration calculated by the model.

[0091] Compared with the prior art, this application has the following beneficial effects: Traditional hydrological models rely on a large amount of ground observation data for parameter calibration, which makes them difficult to apply in data-scarce watersheds. This application only requires watershed outlet runoff observation and a small number of evaporation station observations to drive the daily evapotranspiration inversion of the entire watershed grid, which greatly reduces the dependence on the ground observation network and makes evapotranspiration monitoring in data-scarce areas possible.

[0092] By incorporating both energy balance and water balance into evapotranspiration calculations, the inversion results conform to both atmospheric energy-driven laws and the watershed water closure principle, overcoming the shortcomings of poor physical interpretability of pure data-driven models.

[0093] By using a runoff generation parameter generation network and a runoff parameter generation network, parameters adapted to the environmental characteristics of each grid cell are dynamically generated. A deep learning correction network is used to perform closed-loop feedback correction on the initial evapotranspiration, which effectively suppresses the accumulation of errors in the physical model and significantly improves the spatiotemporal accuracy of evapotranspiration inversion.

[0094] A fully differentiable hydrological model is constructed based on an automatic differentiation framework, enabling the gradient of the runoff error at the watershed outlet to propagate backward along the confluence path to each upstream grid. This achieves weakly supervised learning that drives the optimization of parameters across the entire watershed through single-point observations, providing a new technical approach for hydrological simulation under conditions of scarce data.

[0095] See Figure 2 , Figure 2 This is a schematic diagram of the structure of an evapotranspiration inversion system for data-scarce watersheds provided in an embodiment of this application. Figure 2 As shown, system 200 includes: The construction module 201 is used to divide the data-scarce watershed into multiple grid cells based on the digital elevation model data of the data-scarce watershed, and to construct a hydrological topology map of the data-scarce watershed based on the hydrological topology relationships between the grid cells. The construction module 201 is used to construct a runoff generation parameter generation network, and input the static environmental attributes and dynamic meteorological characteristics of the target grid cell in the plurality of grid cells into the runoff generation parameter generation network to output the runoff parameters of the target grid cell; The construction module 201 is used to construct a confluence parameter generation network, inputting the grid terrain and land use features of each grid cell into the confluence parameter generation network to output the confluence parameters of each grid cell. The calculation module 202 is used to calculate the preliminary evapotranspiration of the target grid cell in the current time period based on the runoff parameters, under the condition of coupling the current soil moisture content of the target grid cell. The calculation module 202 is used to calculate the total runoff of the target grid cell based on the soil infiltration capacity, real-time precipitation intensity and runoff parameters of the target grid cell; The calculation module 202 is used to use the confluence parameters of each grid cell, determine the calculation order according to the hydrological topology diagram, and perform confluence calculation grid by grid from upstream to downstream, so as to collect the surface runoff, interflow and groundwater runoff generated by the multiple grid cells to the only outlet section of the data-scarce watershed, and obtain the outlet runoff process of the data-scarce watershed. The construction module 201 is used to construct a deep learning correction network, inputting the preliminary evapotranspiration, the current soil moisture content, the dynamic meteorological characteristics, and the measured evapotranspiration residual of the grid cell where the evaporation station is located into the deep learning correction network, dynamically correcting the preliminary evapotranspiration, and outputting the final evapotranspiration inversion result. The construction module 201 is used to construct a multi-scale, multi-objective loss function including runoff simulation error and evapotranspiration simulation error, wherein the runoff simulation error is the error between the outlet runoff process and the measured runoff at the hydrological station, and the evapotranspiration simulation error is the error between the evapotranspiration value of the final evapotranspiration inversion result in the grid cell where the evaporation station is located and the measured evapotranspiration value at the evaporation station. Training module 203 is used to train the flow generation parameter generation network, the flow confluence parameter generation network and the deep learning correction network based on the multi-scale multi-objective loss function, so as to obtain the trained flow generation parameter generation network, the flow confluence parameter generation network and the deep learning correction network; The final evapotranspiration inversion result and the total yield are used to update the soil moisture state of the target grid cell to obtain the corrected soil moisture content. The corrected soil moisture content is used to calculate the preliminary evapotranspiration of the next time period, which is the period after the current time period.

[0096] Optionally, the training module is specifically used for: An automatic differential framework is used to calculate the gradients of the multi-scale, multi-objective loss function on the trainable parameters in the runoff generation network, the runoff generation network, and the deep learning correction network. The gradient of the runoff simulation error is backpropagated through differentiable physical processes in the runoff calculation, runoff generation, and evapotranspiration calculation processes to adjust the runoff generation and runoff generation parameters of all grid cells in the data-scarce watershed. The gradient of the evapotranspiration simulation error is backpropagated through the deep learning correction network to update its network weights, and through the connection between the deep learning correction network and the preliminary evapotranspiration calculation process, it is backpropagated to the preliminary evapotranspiration calculation process to adjust the runoff generation parameters of the grid cell where the evaporation station is located. Based on the calculated gradient, the flow generation parameter generation network, the flow sinking parameter generation network, and the deep learning correction network are trained to obtain the trained flow generation parameter generation network, the flow sinking parameter generation network, and the deep learning correction network.

[0097] Optionally, the deep learning correction network adopts a long short-term memory network architecture, including: The encoder, consisting of multi-layer convolutional downsampling modules, is used to extract the spatial structural features of the evaporation field; The bottleneck layer flattens the spatial features output by the encoder into a sequence and inputs it into a long short-term memory network to capture the temporal evolution of the evaporation field. The decoder consists of multiple deconvolution upsampling modules and establishes skip connections with the corresponding layers of the encoder to restore spatial resolution and output the evaporation correction value for each grid cell. The output layer adds the evaporation correction value to the initial evaporation amount to obtain the final evaporation inversion result.

[0098] Optionally, the calculation of the preliminary evapotranspiration, the total production flow, and the outlet runoff process are all implemented using automatic differential frame tensor operations, and the formulas used are all continuously differentiable mathematical expressions.

[0099] The evapotranspiration inversion system for data-scarce watersheds provided in this application can achieve the goals outlined in this application. Figure 1 The various processes in the method embodiments, and the ways to achieve the same beneficial effects, will not be repeated here to avoid repetition.

[0100] like Figure 3 As shown, this application also provides a terminal device, including a processor 301 and a memory 302. The memory 302 stores a program or instructions that can run on the processor 301. When the program or instructions are executed by the processor 301, they implement the various steps of the above-described embodiment of the evapotranspiration inversion method for data-scarce watersheds and achieve the same technical effect. To avoid repetition, they will not be described again here.

[0101] It should be noted that the terminal device in this application can be a terminal or other devices besides a terminal. For example, the terminal device can be a mobile phone, tablet computer, laptop computer, etc., and this application does not make any specific limitation.

[0102] This application also provides a readable storage medium storing a program or instructions that, when executed by a processor, implement the various processes of the above-described embodiments of the evapotranspiration inversion method for data-scarce watersheds and achieve the same technical effect. To avoid repetition, these will not be described again here.

[0103] The processor is the processor in the terminal device described in the above embodiments. The readable storage medium includes a computer-readable storage medium, such as a computer read-only memory (Read-Only Memory). Only memory (ROM), random access memory (RAM), magnetic disks or optical disks, etc.

[0104] This application also provides a chip, which includes a processor and a communication interface. The communication interface is coupled to the processor. The processor is used to run programs or instructions to implement the various processes of the above-described embodiments of the evapotranspiration inversion method for data-scarce watersheds, and can achieve the same technical effect. To avoid repetition, it will not be described again here.

[0105] It should be understood that the chip mentioned in this application may also be referred to as a system-on-a-chip, system chip, chip system, or system-on-a-chip, etc.

[0106] This application provides a computer program product stored in a storage medium. The program product is executed by at least one processor to implement the various processes of the above-described embodiments of the evapotranspiration inversion method for data-scarce watersheds, and can achieve the same technical effects. To avoid repetition, it will not be described again here.

[0107] It should be noted that, in this document, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes that element. Furthermore, it should be noted that the scope of the methods and apparatuses in the embodiments of this application is not limited to performing functions in the order shown or discussed, but may also include performing functions substantially simultaneously or in the reverse order, depending on the functions involved. For example, the described methods may be performed in a different order than described, and various steps may be added, omitted, or combined. Additionally, features described with reference to certain examples may be combined in other examples.

[0108] Through the above description of the embodiments, those skilled in the art can clearly understand that the methods of the above embodiments can be implemented by means of software plus necessary general-purpose hardware platforms. Of course, they can also be implemented by hardware, but in many cases the former is a better implementation method. Based on this understanding, the technical solution of this application, in essence, or the part that contributes to the prior art, can be embodied in the form of a computer software product, which is stored in a storage medium (such as a read-only memory). The device includes a number of instructions in a ROM (random access memory), RAM (magnetic disk), or optical disk to cause a terminal (which may be a mobile phone, computer, server, or network device, etc.) to execute the methods described in the various embodiments of this application.

[0109] The embodiments of this application have been described above with reference to the accompanying drawings. However, this application is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of this application without departing from the spirit and scope of the claims, and all of these forms are within the protection scope of this application.

Claims

1. A method for evapotranspiration inversion in data-scarce watersheds, characterized in that, include: Based on digital elevation model data of a data-scarce watershed, the data-scarce watershed is divided into multiple grid cells, and a hydrological topology map of the data-scarce watershed is constructed based on the hydrological topology relationships between the grid cells. A runoff generation parameter generation network is constructed, and the static environmental attributes and dynamic meteorological characteristics of the target grid cell in the plurality of grid cells are input into the runoff generation parameter generation network to output the runoff parameters of the target grid cell; A confluence parameter generation network is constructed, and the grid terrain and land use features of each grid cell are input into the confluence parameter generation network to output the confluence parameters of each grid cell. Under the condition of coupling the current soil moisture content of the target grid cell, the preliminary evapotranspiration of the target grid cell in the current time period is calculated based on the runoff parameters; Calculate the total runoff of the target grid cell based on the soil infiltration capacity, real-time precipitation intensity, and runoff parameters. Using the confluence parameters of each grid cell, the calculation order is determined according to the hydrological topology diagram. The confluence calculation is performed grid by grid from upstream to downstream. The surface runoff, interflow and groundwater runoff generated by the multiple grid cells are collected to the unique outlet section of the data-scarce watershed to obtain the outlet runoff process of the data-scarce watershed. A deep learning correction network is constructed, and the preliminary evapotranspiration, the current soil moisture content, the dynamic meteorological characteristics, and the measured evapotranspiration residual of the grid cell where the evaporation station is located are input into the deep learning correction network to dynamically correct the preliminary evapotranspiration and output the final evapotranspiration inversion result. A multi-scale, multi-objective loss function is constructed, which includes runoff simulation error and evapotranspiration simulation error. The runoff simulation error is the error between the outlet runoff process and the measured runoff at the hydrological station, and the evapotranspiration simulation error is the error between the evapotranspiration value of the final evapotranspiration inversion result in the grid cell where the evaporation station is located and the measured evapotranspiration value at the evaporation station. The flow generation parameter generation network, the flow sinking parameter generation network, and the deep learning correction network are trained based on the multi-scale multi-objective loss function to obtain the trained flow generation parameter generation network, the flow sinking parameter generation network, and the deep learning correction network. The final evapotranspiration inversion result and the total yield are used to update the soil moisture state of the target grid cell to obtain the corrected soil moisture content. The corrected soil moisture content is used to calculate the preliminary evapotranspiration of the next time period, which is the period after the current time period.

2. The method according to claim 1, characterized in that, The step of training the runoff generation network, the sinking parameter generation network, and the deep learning correction network based on the multi-scale multi-objective loss function to obtain the trained runoff generation network, the sinking parameter generation network, and the deep learning correction network includes: An automatic differential framework is used to calculate the gradients of the multi-scale, multi-objective loss function on the trainable parameters in the runoff generation network, the runoff generation network, and the deep learning correction network. The gradient of the runoff simulation error is backpropagated through differentiable physical processes in the runoff calculation, runoff generation, and evapotranspiration calculation processes to adjust the runoff generation and runoff generation parameters of all grid cells in the data-scarce watershed. The gradient of the evapotranspiration simulation error is backpropagated through the deep learning correction network to update its network weights, and through the connection between the deep learning correction network and the preliminary evapotranspiration calculation process, it is backpropagated to the preliminary evapotranspiration calculation process to adjust the runoff generation parameters of the grid cell where the evaporation station is located. Based on the calculated gradient, the flow generation parameter generation network, the flow sinking parameter generation network, and the deep learning correction network are trained to obtain the trained flow generation parameter generation network, the flow sinking parameter generation network, and the deep learning correction network.

3. The method according to claim 1, characterized in that, The deep learning correction network adopts a long short-term memory network architecture, including: The encoder, consisting of multi-layer convolutional downsampling modules, is used to extract the spatial structural features of the evaporation field; The bottleneck layer flattens the spatial features output by the encoder into a sequence and inputs it into a long short-term memory network to capture the temporal evolution of the evaporation field. The decoder consists of multiple deconvolution upsampling modules and establishes skip connections with the corresponding layers of the encoder to restore spatial resolution and output the evaporation correction value for each grid cell. The output layer adds the evaporation correction value to the initial evaporation amount to obtain the final evaporation inversion result.

4. The method according to any one of claims 1 to 3, characterized in that, The calculation processes for the preliminary evapotranspiration, the total production flow, and the outlet runoff are all implemented using automatic differential tensor operations, and the formulas used are all continuously differentiable mathematical expressions.

5. An evapotranspiration retrieval system for data-scarce watersheds, characterized in that, include: The module is used to divide the data-scarce watershed into multiple grid cells based on the digital elevation model data of the data-scarce watershed, and to construct a hydrological topology map of the data-scarce watershed based on the hydrological topology relationships between the grid cells. The construction module is used to construct a runoff generation parameter generation network, and inputs the static environmental attributes and dynamic meteorological characteristics of the target grid cell in the plurality of grid cells into the runoff generation parameter generation network to output the runoff parameters of the target grid cell; The construction module is used to construct a confluence parameter generation network, inputting the grid terrain and land use features of each grid cell into the confluence parameter generation network to output the confluence parameters of each grid cell. The calculation module is used to calculate the preliminary evapotranspiration of the target grid cell in the current time period based on the runoff parameters, under the condition of coupling the current soil moisture content of the target grid cell. The calculation module is used to calculate the total runoff of the target grid cell based on the soil infiltration capacity, real-time precipitation intensity and runoff parameters of the target grid cell. The calculation module is used to use the confluence parameters of each grid cell, determine the calculation order according to the hydrological topology diagram, and perform confluence calculation grid by grid from upstream to downstream. The surface runoff, interflow and groundwater runoff generated by the multiple grid cells are collected to the only outlet section of the data-scarce watershed to obtain the outlet runoff process of the data-scarce watershed. The construction module is used to construct a deep learning correction network, inputting the preliminary evapotranspiration, the current soil moisture content, the dynamic meteorological characteristics, and the measured evapotranspiration residual of the grid cell where the evaporation station is located into the deep learning correction network, dynamically correcting the preliminary evapotranspiration, and outputting the final evapotranspiration inversion result; The construction module is used to construct a multi-scale, multi-objective loss function including runoff simulation error and evapotranspiration simulation error. The runoff simulation error is the error between the outlet runoff process and the measured runoff at the hydrological station, and the evapotranspiration simulation error is the error between the evapotranspiration value of the final evapotranspiration inversion result in the grid cell where the evaporation station is located and the measured evapotranspiration value at the evaporation station. The training module is used to train the flow generation parameter generation network, the flow confluence parameter generation network, and the deep learning correction network based on the multi-scale multi-objective loss function, so as to obtain the trained flow generation parameter generation network, the flow confluence parameter generation network, and the deep learning correction network. The final evapotranspiration inversion result and the total yield are used to update the soil moisture state of the target grid cell to obtain the corrected soil moisture content. The corrected soil moisture content is used to calculate the preliminary evapotranspiration of the next time period, which is the period after the current time period.

6. The system according to claim 5, characterized in that, The training module is specifically used for: An automatic differential framework is used to calculate the gradients of the multi-scale, multi-objective loss function on the trainable parameters in the runoff generation network, the runoff generation network, and the deep learning correction network. The gradient of the runoff simulation error is backpropagated through differentiable physical processes in the runoff calculation, runoff generation, and evapotranspiration calculation processes to adjust the runoff generation and runoff generation parameters of all grid cells in the data-scarce watershed. The gradient of the evapotranspiration simulation error is backpropagated through the deep learning correction network to update its network weights, and through the connection between the deep learning correction network and the preliminary evapotranspiration calculation process, it is backpropagated to the preliminary evapotranspiration calculation process to adjust the runoff generation parameters of the grid cell where the evaporation station is located. Based on the calculated gradient, the flow generation parameter generation network, the flow sinking parameter generation network, and the deep learning correction network are trained to obtain the trained flow generation parameter generation network, the flow sinking parameter generation network, and the deep learning correction network.

7. The system according to claim 5, characterized in that, The deep learning correction network adopts a long short-term memory network architecture, including: The encoder, consisting of multi-layer convolutional downsampling modules, is used to extract the spatial structural features of the evaporation field; The bottleneck layer flattens the spatial features output by the encoder into a sequence and inputs it into a long short-term memory network to capture the temporal evolution of the evaporation field. The decoder consists of multiple deconvolution upsampling modules and establishes skip connections with the corresponding layers of the encoder to restore spatial resolution and output the evaporation correction value for each grid cell. The output layer adds the evaporation correction value to the initial evaporation amount to obtain the final evaporation inversion result.

8. The system according to any one of claims 5 to 7, characterized in that, The calculation processes for the preliminary evapotranspiration, the total production flow, and the outlet runoff are all implemented using automatic differential tensor operations, and the formulas used are all continuously differentiable mathematical expressions.

9. A terminal device, characterized in that, It includes a processor and a memory, wherein the memory stores a program or instructions executable on the processor, wherein the program or instructions, when executed by the processor, implement the steps of the method as described in any one of claims 1 to 4.

Images