Non-data watershed runoff prediction method and system based on space-time deep learning

Abstract

The invention discloses a space-time deep learning-based data-free watershed runoff prediction method and system. The method comprises the following steps: collecting static geographic raster data and dynamic hydro meteorological time series data of multiple watersheds, and performing preprocessing; constructing a double-flow deep learning model fusing space and time features; training a double-flow deep learning model based on the multi-region large-sample watershed data to obtain a general hydrological model; evaluating the adaptability of each network layer in the general hydrological model to a target watershed through a layered unfreezing test, and screening out a key adaptive layer; and under a leave-one-out method cross validation framework, layered progressive unfreezing transfer learning is carried out on the general hydrological model based on the key adaptation layer, and data-free drainage basin runoff prediction is realized. According to the method, multi-basin data driving, spatio-temporal feature fusion and a transfer learning mechanism are organically combined, the conversion of a hydrological modeling norm from local adaptation to global generalization is promoted, and a new path is provided for intelligent prediction of data-free basin runoff.

Description

Technical Field

[0001] This invention relates to the field of hydrological forecasting in hydrology, and to a method and system for predicting runoff in data-free watersheds, particularly a method and system for predicting runoff in data-free watersheds based on spatiotemporal deep learning. Background Technology

[0002] Accurate hydrological runoff forecasting is a fundamental support for scientific water resource allocation, flood control and disaster reduction decision-making, and ecological civilization construction. However, existing runoff forecasting methods focus excessively on time series modeling, and spatial characteristics are often simplified to watershed average parameters. Furthermore, they typically adopt a "single-watershed training-single-watershed validation" model, which inevitably leads to over-adaptation to the hydrological characteristics of specific watersheds. Consequently, they perform poorly in cross-watershed tests, severely limiting the predictive ability of models in watersheds without available data.

[0003] Reconciling the modeling contradictions between the universal physical laws of hydrological processes and the specific characteristics of watersheds has become a core challenge in current hydrological research. In recent years, hydrological modeling has gradually shifted from a single-watershed temporal modeling paradigm to a multi-watershed spatiotemporal joint modeling paradigm. Against this backdrop, constructing a deep learning architecture capable of deeply integrating spatial and temporal features, and establishing a transfer learning mechanism applicable to the hydrological field, has become an inevitable approach and cutting-edge direction for overcoming the bottleneck of data-free watershed prediction and improving the generalization ability of models. Summary of the Invention

[0004] To address the aforementioned technical problems, this invention overcomes the shortcomings of existing technologies and provides a method and system for predicting runoff in data-free watersheds based on spatiotemporal deep learning. This invention selects spatial topographic data and meteorological and hydrological time-series data from multiple watersheds to construct a spatiotemporal deep learning model with a dual-flow network architecture. It uses a stratified thawing test to screen key adaptation layers that are more sensitive to runoff prediction and performs progressive thawing and fine-tuning in the target area, thereby accurately predicting runoff in data-free watersheds.

[0005] The technical solution adopted in this invention is as follows:

[0006] A data-free watershed runoff prediction method based on spatiotemporal deep learning includes the following steps:

[0007] S1. Collect static geographic raster data and dynamic hydrological and meteorological time series data from multiple watersheds, and perform cross-watershed global normalization processing;

[0008] S2. Construct a dual-flow deep learning model that integrates spatial and temporal features. The architecture includes a spatial flow network branch and a temporal flow network branch. The spatial flow network branch is used to extract static geospatial features, and the temporal flow network branch is used to process dynamic meteorological time-series data. The spatial feature vector and the temporal feature vector are concatenated and fused through a dual-flow feature fusion layer to output runoff prediction results.

[0009] S3. Train the dual-flow deep learning model based on large sample watershed data from multiple regions to obtain a general hydrological model;

[0010] S4. Through stratified thawing tests, evaluate the adaptability of each network layer in the general hydrological model to the target watershed, and screen out the key adaptation layers;

[0011] S5. Under the leave-one-out cross-validation framework, hierarchical progressive thawing transfer learning is implemented on the general hydrological model based on the key adaptation layer to achieve runoff prediction in watersheds without data.

[0012] Furthermore, the static geographic grid data includes the watershed's digital elevation model, slope, runoff accumulation, and normalized vegetation index; the dynamic hydrological and meteorological time-series data includes hourly precipitation, soil moisture, and potential evapotranspiration of the watershed.

[0013] Furthermore, the cross-basin global normalization process specifically includes: independently calculating the global maximum and minimum values ​​for each time-series feature matrix and performing cross-sample maximum-minimum scaling; for each spatial feature dimension, extracting raster extreme values ​​across the basin and performing unified maximum-minimum scaling.

[0014] Furthermore, the construction of a dual-stream deep learning model that integrates spatial and temporal features specifically includes:

[0015] S2.1. Construct a spatial flow network branch, using a two-branch convolutional neural network to process terrain data and normalized vegetation index respectively, and perform attention-weighted fusion of the two-branch features through a multi-scale feature fusion module;

[0016] S2.2. Integrate a deep feature extraction module based on the Xception architecture into the spatial flow network branch, and use depthwise separable convolution and residual connections to construct multi-layer Block units for deep spatial feature learning;

[0017] S2.3. Construct a time-series network branch, using a long short-term memory network combined with an attention mechanism to process dynamic meteorological time-series data. The long short-term memory network captures temporal dependencies, and the attention mechanism is used to calculate temporal feature weights to generate context vectors.

[0018] S2.4. Construct the dual-flow feature fusion layer, concatenate the spatial feature vector output by the spatial flow network branch with the temporal feature vector output by the temporal flow network branch, and achieve feature fusion and output runoff prediction results through a fully connected layer and a nonlinear activation function.

[0019] Furthermore, training the dual-flow deep learning model based on large-sample watershed data from multiple regions specifically includes:

[0020] S3.1. Construct a multi-basin training dataset, using the spatial static data of multiple basins as the fixed input of the spatial flow network branch, and dividing the hydrological and meteorological data of each basin into training and test sets according to time series;

[0021] S3.2. Design a multi-basin batch data loading mechanism, using a custom data loader to input topographic data, vegetation data and meteorological time series data into the corresponding branches of the dual-basin deep learning model, and adopt a hybrid precision training and gradient accumulation strategy to optimize computational efficiency;

[0022] S3.3. Configure an adaptive learning rate scheduler, adopt a multi-stage warm-up-cosine decay strategy for updating model parameters, and set weight decay regularization to prevent overfitting.

[0023] S3.4. Establish an automatic hyperparameter search mechanism to perform Bayesian optimization on key hyperparameters such as learning rate, batch size, hidden layer dimension, and dropout rate, and select the optimal parameter combination through parallel training with multiple trials;

[0024] S3.5. Establish a multi-indicator joint evaluation system, adopt an early shutdown mechanism to monitor the hydrological evaluation indicators of the model on the test sets of each watershed, and select and store the optimal model based on the comprehensive performance of the indicators.

[0025] Furthermore, step S4 specifically includes:

[0026] S4.1. Design a hierarchical unfreezing strategy system, dividing the general hydrological model into a fusion layer, a spatial flow output layer, a temporal flow output layer, a spatial flow final convolutional layer, a spatial flow Xception backbone, a spatial flow multi-scale fusion layer, a spatial flow basic CNN, a temporal flow LSTM layer, and a temporal flow attention layer core layer group;

[0027] S4.2. Construct a single-component thawing experiment, thaw each core layer group separately for fine-tuning training, and evaluate the contribution of a single layer group to the adaptability of the target watershed using the controlled variable method;

[0028] S4.3. Design experiments on combined unfreezing strategies, including combined unfreezing of all output layers, combined unfreezing of high-level spatial flow, complete unfreezing of temporal flow, complete unfreezing of spatial flow, and minimum effective set unfreezing strategy, and set up full freeze and full unfreezing as control baselines;

[0029] S4.4. Through a multi-index comprehensive evaluation mechanism, the key adaptation layer combination with the highest parameter efficiency and the strongest generalization ability is selected.

[0030] Furthermore, the hierarchical progressive thawing transfer learning applied to the general hydrological model based on the key adaptation layer within the leave-one-out cross-validation framework specifically includes:

[0031] S5.1. Construct a leave-one-out cross-validation framework, select a single watershed as the target test watershed from N watersheds with available data in the target area, and use the remaining N-1 watersheds as training watersheds to simulate the environment of watersheds without data, and use the parameters of the general hydrological model as the initial weights;

[0032] S5.2. Implement a four-stage, layered, gradual unfreezing strategy, unfreezing the key adaptation layers sequentially.

[0033] S5.3. Perform runoff prediction verification on the target test watershed and output the runoff prediction results for the watershed without data.

[0034] A data-free watershed runoff prediction system based on spatiotemporal deep learning, used to implement the above method, includes:

[0035] Data acquisition and processing module: used to acquire static geographic raster data and dynamic hydrological and meteorological time series data from multiple watersheds, and to perform cross-watershed global normalization processing;

[0036] Dual-stream deep learning architecture construction module: used to build a dual-stream deep learning model that integrates spatial and temporal features. The architecture includes a spatial flow network branch and a temporal flow network branch. The spatial flow network branch is used to extract static geospatial features, and the temporal flow network branch is used to process dynamic meteorological time series data. The spatial feature vector and the temporal feature vector are concatenated and fused through a dual-stream feature fusion layer to output runoff prediction results.

[0037] Model pre-training module: used to train the dual-flow deep learning model based on large sample watershed data from multiple regions to obtain a general hydrological model;

[0038] Key Adaptation Layer Screening Module: Used to evaluate the adaptability of each network layer in the general hydrological model to the target watershed through stratified thawing tests, and to screen out key adaptation layers;

[0039] Transfer learning prediction module: used to implement hierarchical progressive unfreezing transfer learning on the pre-trained general hydrological model based on the key adaptation layer under the leave-one-out cross-validation framework, to achieve watershed runoff prediction without data.

[0040] A computer device, the computer device comprising:

[0041] One or more processors;

[0042] Memory, used to store one or more programs;

[0043] When the one or more programs are executed by the one or more processors, the one or more processors implement the above-described method for predicting runoff in a dataless watershed based on spatiotemporal deep learning.

[0044] A computer-readable storage medium storing computer instructions, characterized in that, when the computer instructions are executed by one or more processors, the one or more processors cause the one or more processors to perform the steps in the above-described method.

[0045] Compared with the prior art, the beneficial effects of the present invention are:

[0046] 1. This invention overcomes the limitations of traditional time series modeling by introducing spatial feature modeling and using joint training with large samples from multiple watersheds, and significantly improves the model's ability to represent spatial heterogeneity.

[0047] 2. This invention innovatively applies a transfer learning mechanism and a hierarchical unfreezing strategy to achieve efficient adaptation of the model to the characteristics of the target watershed, thereby enhancing the accuracy of cross-watershed prediction.

[0048] 3. This invention organically combines multi-basin data-driven approaches, spatiotemporal feature fusion, and transfer learning mechanisms to construct a generalizable and transferable watershed model. This promotes the transformation of hydrological modeling paradigms from local adaptation to global generalization and provides a new path for intelligent prediction of runoff in watersheds without data. Attached Figure Description

[0049] Figure 1 This is a schematic diagram of the method flow in an embodiment of the present invention.

[0050] Figure 2 This is a schematic diagram of a spatiotemporal deep learning model with a two-stream network architecture in an embodiment of the present invention.

[0051] Figure 3 This is a simplified diagram of a spatial flow network branch containing a multi-scale feature fusion module in an embodiment of the present invention.

[0052] Figure 4 This is a multi-stage preheating-cosine decay learning rate scheduling strategy in the embodiments of the present invention.

[0053] Figure 5 This is a graph showing the thawing strategy and corresponding model performance changes in an embodiment of the present invention.

[0054] Figure 6 This is a schematic diagram of the data-free watershed runoff prediction results using progressive hierarchical transfer learning in an embodiment of the present invention. Detailed Implementation

[0055] The technical solution of the present invention will be further described clearly and in detail below with reference to the accompanying drawings and specific examples.

[0056] Example 1

[0057] like Figure 1 As shown, a data-free watershed runoff prediction method based on spatiotemporal deep learning includes the following steps:

[0058] S1. Collect and preprocess static geographic raster data and dynamic hydrological and meteorological time-series data of the watershed; wherein, the static geographic raster data includes the watershed's digital elevation model, slope, runoff accumulation, and normalized vegetation index, and the dynamic hydrological and meteorological time-series data includes hourly precipitation, soil moisture, and potential evapotranspiration of the watershed. The preprocessing method is to perform global normalization on similar characteristics of multiple watersheds. The specific collection and preprocessing steps are as follows:

[0059] S1.1. Extract watershed topographic raster files from the digital elevation model based on watershed surface features using a mask, calculate the arctangent of the maximum elevation difference and the raster spacing to obtain watershed slope raster files, and use the flow direction matrix method to obtain watershed runoff accumulation raster files;

[0060] S1.2. After preprocessing steps such as atmospheric correction, cloud masking, and noise filtering, a normalized vegetation index raster file is obtained;

[0061] S1.3. Obtain atmospheric driving field products and soil moisture products from the land surface data assimilation system. Extract hourly raster data of precipitation and potential evapotranspiration from the atmospheric driving field products. Aggregate hourly soil moisture at different depth layers to obtain hourly average soil moisture raster data. Extract values ​​within the watershed by masking and calculate the mean to obtain hourly precipitation, soil moisture and potential evapotranspiration time series data for each watershed.

[0062] S1.4 Calculate the global extrema independently for each time series feature matrix, perform cross-sample max-min scaling, and extract raster extrema across watersheds for each spatial feature dimension (such as DEM) and perform uniform max-min scaling to maintain the relative differences of features between different watersheds and eliminate the influence of dimensions.

[0063] S2. Construct a two-stream deep learning model that integrates spatial and temporal features. The specific steps are as follows:

[0064] S2.1. Construct a spatial flow network branch and use a two-branch convolutional neural network to extract static geospatial features. One branch processes terrain data including digital elevation model, slope and runoff accumulation, and the other branch processes normalized vegetation index. Attention-weighted fusion of the two-branch features is performed through a multi-scale feature fusion module.

[0065] S2.2. Design a deep feature extraction module based on the Xception architecture, using depthwise separable convolution and residual connections to construct multi-layer Block units, and integrate this module in the spatial flow branch for deep spatial feature learning;

[0066] S2.3. Construct a time-series network branch, using a Long Short-Term Memory (LSTM) network combined with an attention mechanism to process dynamic meteorological time-series data. The LSTM captures temporal dependencies, and the attention mechanism is used to calculate the weights of temporal features and generate context vectors.

[0067] S2.4. Construct a dual-flow feature fusion layer, which concatenates the spatial feature vector output by the spatial flow with the temporal feature vector output by the temporal flow. The feature fusion is achieved through a fully connected layer and a nonlinear activation function, and the runoff prediction results are output.

[0068] S3. Train the dual-flow deep learning model based on large-sample watershed data from multiple regions to obtain a general hydrological model. The specific steps are as follows:

[0069] S3.1. Construct a multi-basin training dataset for large watershed areas with abundant watershed data. Use the spatial static data of multiple watersheds as the fixed input of the spatial branch of the dual-flow network. Use the time series segmentation method to divide the hydrological and meteorological data of each watershed into training set and test set.

[0070] S3.2. Design a multi-basin batch data loading mechanism. Use a custom data loader to input topographic data, vegetation data and meteorological time series data into the corresponding branches of the dual-stream network respectively, and use a mixed precision training and gradient accumulation strategy to optimize computational efficiency.

[0071] S3.3. Configure an adaptive learning rate scheduler, adopt a multi-stage warm-up-cosine decay strategy and learning rate optimization algorithm to update model parameters, and set weight decay regularization to prevent overfitting;

[0072] S3.4. Establish an automatic hyperparameter search mechanism to perform Bayesian optimization on key hyperparameters such as learning rate, batch size, hidden layer dimension, and dropout rate, and select the optimal parameter combination through parallel training with multiple trials.

[0073] S3.5. Establish a multi-indicator joint evaluation system, adopt an early shutdown mechanism to monitor hydrological evaluation indicators such as root mean square error and Nash efficiency coefficient of the model on test sets in various watersheds, and select and store the optimal model for subsequent fine-tuning based on the comprehensive performance of the indicators.

[0074] S4. The adaptability of each network layer in the general hydrological model to the target watershed is evaluated through stratified thawing tests, and key adaptation layers are selected. The specific steps are as follows:

[0075] S4.1. Design a hierarchical unfreezing strategy system, dividing the dual-stream network model into core layers such as fusion layer, spatial stream output layer, temporal stream output layer, spatial stream final convolutional layer, spatial stream Xception backbone, spatial stream multi-scale fusion layer, spatial stream basic CNN, temporal stream LSTM layer and temporal stream attention layer;

[0076] S4.2. Construct a single-component thawing experiment, thaw each layer group separately for fine-tuning training, evaluate the contribution of a single layer group to the adaptive performance of the target watershed using the controlled variable method, and identify key feature extraction and decision layers;

[0077] S4.3. Design experiments on combined unfreezing strategies, including combined unfreezing of all output layers, combined unfreezing of high-level spatial flow, complete unfreezing of temporal flow, complete unfreezing of spatial flow, and unfreezing of minimum effective set, and set up full freeze and full unfreezing as control baselines;

[0078] S4.4. Establish a multi-index comprehensive evaluation mechanism to select the key adaptation layer combination with the highest parameter efficiency and the strongest generalization ability through performance comparison analysis.

[0079] S5. Under the leave-one-out cross-validation framework, hierarchical progressive thawing transfer learning is implemented on the general hydrological model based on the key adaptation layer to achieve runoff prediction in data-free watersheds. The specific steps are as follows:

[0080] S5.1. Construct a leave-one-out cross-validation framework, select a single watershed from the N watersheds with data in the target area as the target test watershed, and the remaining N-1 watersheds as training watersheds to simulate the environment of watersheds without data, and initialize the weights of the dual-stream network with the pre-trained model parameters.

[0081] S5.2. Implement a four-stage hierarchical progressive unfreezing strategy, unfreezing the fusion layer in sequence to learn the watershed-specific feature weights, the spatial output layer and the temporal output layer to adapt to the spatiotemporal feature distribution of the target watershed, the attention mechanism layer to optimize temporal information extraction, and the LSTM layer to achieve temporal pattern adaptation.

[0082] S5.3. Perform runoff prediction verification on the target test watershed and output high-precision runoff prediction results for watersheds without data.

[0083] Example 2

[0084] To achieve runoff prediction in data-free watersheds, this embodiment first selects multiple sub-watersheds within a watershed as source domains, collecting their static geographic data and dynamic meteorological data to construct and pre-train a general dual-flow spatiotemporal deep learning model. Subsequently, eight sub-watersheds in another region are selected as target domains. Under a leave-one-out cross-validation framework, key adaptation layers are selected through stratified thaw testing. Data from seven of these watersheds are then used to progressively transfer learn the pre-trained model, thereby achieving high-precision runoff prediction for the remaining target watershed. This method effectively utilizes general hydrological knowledge from the source domains and quickly adapts to the characteristics of the target watershed, providing an accurate and efficient solution for runoff prediction in data-free areas.

[0085] S1. Collect and preprocess static geographic raster data and dynamic hydrological and meteorological time-series data of the watershed; the static geographic raster data of the watershed includes the digital elevation model, slope, runoff accumulation and normalized vegetation index of the watershed, and the dynamic hydrological and meteorological time-series data of the watershed includes hourly precipitation, soil moisture and potential evapotranspiration of the watershed. The preprocessing method is to perform global normalization processing on similar features of multiple watersheds.

[0086] S1.1. The digital elevation model uses ALOS (Advanced Land Observing Satellite) 12.5-meter DEM data. After splicing, cropping, and projection transformation, the slope and runoff accumulation of the corresponding watershed are calculated using formulas.

[0087] Slope calculation formula: In a digital elevation model (DEM) raster, the elevation of the center pixel is... The elevations of the eight surrounding pixels are respectively The pixel resolution is (Unit: meters), then the slope (in degrees) can be calculated using the following formula:

[0088]

[0089]

[0090]

[0091] In the formula For slope, For the elevation of the center pixel, Elevation of pixels within a 3×3 window ( (centered) Pixel resolution (meters) These represent the rates of elevation change in the east-west and north-south directions, respectively. arctangent function

[0092] The formula for calculating the cumulative sinking volume is as follows: For each cell, the cumulative sinking volume is defined as the number of all upstream cells (including itself), which can be calculated recursively.

[0093]

[0094] In the formula For pixels The cumulative amount of the confluence, Represents all flows The upstream pixel set is usually determined by algorithms such as D8, and all upstream pixels can be recursively traversed.

[0095] S1.2. The Normalized Difference Vegetation Index (NDVI) was obtained using the MOD13Q1.061 dataset. This product uses a 16-day composite period and has a spatial resolution of 250 meters. The data underwent preprocessing steps such as atmospheric correction, cloud masking, and noise filtering, and multi-year averages were calculated.

[0096] S1.3. Precipitation and soil moisture data in the hydrometeorological data were extracted from CLDAS (CMA Land Surface Data Assimilation System), and potential evapotranspiration was extracted from ERA5 (ECMWF Reanalysis v5). Hourly data were extracted from both products and cropped and spatially averaged according to the watershed extent to obtain hourly time-series data at the watershed scale. The soil moisture data from CLDAS was vertically divided into five layers (0-5, 0-10, 10-40, 40-100, 100-200 cm). To vertically integrate the data and reflect the overall response of soil moisture, the soil moisture data at the 0-100 cm depth was weighted and averaged according to soil layer thickness, as shown in the following formula:

[0097] Assume the soil moisture content of each layer is . (unit: Each layer has a thickness of (unit: Then the average soil moisture at a depth of 0-100cm The calculation formula is:

[0098]

[0099] S1.4. Perform global normalization processing on all the above data for the same characteristics, assuming the first... The first of the first watersheds The original feature value of the class is The global minimum value of this type of feature across all watersheds is The global maximum value is Then the normalized eigenvalues The calculation formula is:

[0100]

[0101] Among them, the first The global minimum value of a class feature is , No. Global maximum value of class features .

[0102] S2. Construct a two-stream deep learning model that integrates spatial and temporal features, such as... Figure 2 As shown, the specific steps are as follows:

[0103] S2.1. Construct a spatial flow network branch, employing a two-branch convolutional neural network to extract static geospatial features. One branch processes terrain data, including digital elevation models, slope, and runoff accumulation, while the other branch processes the normalized vegetation index. An attention-weighted fusion module is then used to fuse the features from both branches, such as... Figure 3 As shown.

[0104] Since the spatial resolution of topographic data (DEM, slope, cumulative flow direction) is 12.5m, while the resolution of vegetation data (NDVI) is 150m, a two-branch design is adopted for processing these two types of data. Large kernel convolution (5×5) and a large stride (stride=4) are used to quickly downsample the high-resolution topographic data to capture large-scale spatial structure; small kernel convolution (3×3) and a smaller stride (stride=2) are used to process the low-resolution vegetation data to preserve detailed information. The features of the two branches are aligned and fused through a multi-scale feature fusion module, and a joint attention mechanism dynamically adjusts the branch contributions. The multi-scale fusion formula in spatial flow is expressed as follows:

[0105] Feature downsampling:

[0106] Feature splicing:

[0107] Attention weights:

[0108] Weighted fusion:

[0109] in, and These are feature maps of high-resolution terrain data and low-resolution vegetation data, respectively. and This represents the spatial dimensions of the low-resolution feature map. This represents a bilinear interpolation method for downsampling high-resolution features. This indicates concatenation along the channel dimension. Represents a 1×1 convolution. Representatives criticized normalization. Indicates ReLU activation. For channel dimension normalization, This indicates element-wise multiplication. Multiscale fusion of spatial flow components has the following advantages: (1) Multiscale complementarity: fusing high-resolution details with low-resolution global information; (2) Adaptive weighting: attention mechanism dynamically adjusts the branch contribution of spatial flow; (3) Spatial consistency: bilinear interpolation ensures the continuity of feature alignment at different resolutions.

[0110] S2.2. Design a deep feature extraction module based on the Xception architecture. This module employs depthwise separable convolutions and residual connections to construct multi-layer Block units, and integrates this module into the spatial flow branch for deep spatial feature learning, such as... Figure 3 As shown.

[0111] The Spatial Flow core uses the Xception component, which constructs an efficient deep network through depthwise separable convolutions and residual connections. The formulas for depthwise separable convolutions and residual connections are expressed as follows:

[0112] Depthwise convolution:

[0113] Pointwise convolution:

[0114] Residual connections during dimension matching:

[0115] Residual connections during dimensional transformation:

[0116] in, Indicates the input number of the first... The value of the channel at (i,j) represents each pixel in the spatial data. , These represent the output of depthwise convolution and the output of pointwise convolution, respectively. and These represent the depthwise convolution kernel weights and the pointwise convolution weights, respectively. The radius of the convolution kernel is 1. Input the number of channels. Main branch transformation, The input features are Xception components and residual connections have the following advantages: (1) Efficiency: Depthwise separable convolution reduces computational complexity by about 8 times; (2) Stability: Residual connections alleviate gradient vanishing and improve the training effect of deep networks; (3) Feature enhancement: Directly reuse input information and optimize feature expression.

[0117] S2.3. Construct a temporal network branch, employing a Long Short-Term Memory (LSTM) network combined with an attention mechanism to process dynamic meteorological time-series data. The LSTM captures temporal dependencies, and the attention mechanism calculates temporal feature weights and generates context vectors, such as... Figure 2 As shown.

[0118] Attention-Long Short-Term Memory (LSTM) networks not only overcome the gradient vanishing and gradient exploding problems of traditional RNNs through mechanisms such as forget gates, input gates, output gates, and unit states, but also introduce an attention mechanism to selectively focus on the most relevant moments or features when processing hydrological and meteorological time-series data. This allows them to dynamically weight input information while maintaining long-term memory capabilities, improving their ability to capture key information. The definition of Attention-LSTM is as follows:

[0119]

[0120]

[0121]

[0122]

[0123]

[0124] in, It is time Attention-weighted input and They are time Input gate, forget gate, hidden state, cell state, and output gate. and It is the weight matrix output by the attention mechanism to each gate. and It is the weight matrix from the hidden state of the previous time step to the current gate. and These are the bias vectors for each gate. This represents the sigmoid activation function. This represents the hyperbolic tangent activation function. This indicates element-wise multiplication.

[0125] S2.4. Construct a dual-flow feature fusion layer, concatenating the spatial feature vector output from the spatial flow with the temporal feature vector output from the temporal flow. Feature fusion is achieved through a fully connected layer and a nonlinear activation function, outputting the runoff prediction result, such as... Figure 2 As shown.

[0126] The spatial and temporal output features are concatenated along the channel dimension to form a comprehensive feature vector, which is then fused through a multilayer perceptron and finally passed through a fully connected layer to generate a scalar prediction value. The formula for this part is as follows:

[0127] Feature splicing:

[0128] Multilayer perceptron fusion:

[0129] Loss function:

[0130] in and These represent the output spatial and temporal feature vectors, respectively. and These represent the combined feature vectors after concatenation and fusion, respectively, where... It is also the runoff prediction value output after fusion and dimensionality reduction. and Representing the first The true and predicted runoff values ​​for each sample are used, and the mean squared error is used as the regression loss function.

[0131] The proposed dual-stream Xception-LSTM model effectively addresses the feature extraction and fusion challenges of multi-source heterogeneous hydrological data through its innovative network architecture. The introduction of depthwise separable convolutions and residual connections, multi-scale feature fusion, and attention mechanisms not only optimizes deep feature extraction but also enables spatiotemporal collaborative modeling. This significantly improves computational efficiency while enhancing the accuracy and robustness of runoff prediction. This model provides a new technical approach for runoff prediction and hydrological modeling, possessing significant theoretical value and practical application implications.

[0132] S3. Train a dual-flow deep learning model based on large-sample watershed data from multiple regions to obtain a general hydrological model. The specific steps are as follows:

[0133] S3.1. Constructing multi-basin training datasets for large watershed areas with abundant watershed data:

[0134] Small watersheds in mountainous areas with no available data typically have a drainage area of ​​less than 1000 square kilometers, low flow rates, pronounced seasonality, and are mostly headwater streams. Therefore, suitable watersheds were selected from hydrological yearbooks based on the above criteria to construct a comprehensive dataset containing multiple small watersheds in a specific mountainous region and eight small watersheds in another region. Each watershed... Spatial static data were extracted following the steps described above. and time-series dynamic data In one section, small mountainous watersheds within a specific watershed area were used to pre-train the model, learning general hydrophysical laws. The spatial static data of these watersheds served as globally shared features in the two-flow network, remaining unchanged throughout the training and testing process, acting as fixed inputs for the spatial flow branch; while the temporal dynamic data was segmented according to time series, with the first 80% used as the training set and the last 20% as the test set.

[0135] S3.2. Design a multi-basin batch data loading mechanism:

[0136] A custom data loader was built to handle the separate input of terrain, vegetation, and meteorological time-series data. A mixed-precision training technique was employed, using Automatic Mixed Precision (AMP) to convert some computations in the forward propagation into half-precision floating-point numbers, significantly reducing GPU memory usage. A gradient accumulation strategy was implemented, accumulating gradients in... Parameters are updated after accumulating data across multiple small batches.

[0137]

[0138] in For gradient accumulation steps, For loss function, It is a two-stream network model.

[0139] S3.3. Configure the adaptive learning rate scheduler:

[0140] A multi-stage warm-up-cosine decay learning rate scheduling strategy is adopted, such as Figure 4 As shown, the learning rate goes through two stages: warm-up-cosine decay and further cosine decay-stabilization-fine-tuning oscillation, and the model parameters are updated in conjunction with the Adam optimizer.

[0141] During the warm-up phase, the learning rate increases linearly:

[0142]

[0143] During the cosine decay phase, the learning rate decays according to the cosine function:

[0144]

[0145] During the fine-tuning oscillation phase, the learning rate oscillates and fine-tunes between the minimum learning rate and a slightly higher learning rate:

[0146]

[0147] Simultaneously set a weight decay regularization term. Prevent overfitting.

[0148] S3.4. Establish an automatic hyperparameter search mechanism:

[0149] Implementing Bayesian optimization based on the Ray framework, adjusting the learning rate. Batch size LSTM hidden layer dimension Number of LSTM layers Dropout ratio The search is performed on key hyperparameters. The ASHA scheduler is used to achieve parallel training across multiple trials. The objective function for hyperparameter search is:

[0150]

[0151] in For the hyperparameter search space, For the test dataset, The coefficient of determination.

[0152] S3.5. Establish a multi-indicator joint evaluation system:

[0153] A comprehensive evaluation system incorporating multiple hydrological assessment indicators, including the Nash efficiency coefficient (NSE), is constructed. An early stopping mechanism is employed to monitor model performance, halting training when an indicator fails to improve for several consecutive rounds. The optimal model weights are selected and saved based on the overall performance of multiple indicators for subsequent transfer learning fine-tuning.

[0154] S4. Through stratified thawing tests, evaluate the adaptability of each network layer in the general hydrological model to the target watershed, and screen out key adaptation layers, such as... Figure 5 As shown, the specific steps are as follows:

[0155] S4.1. Design of a layered unfreezing strategy. To systematically identify the adaptive contribution of each layer to the target watershed and achieve precise fine-tuning rather than blind full-parameter optimization, the dual-stream network model is divided into core layers including a fusion-only layer, a spatial head-only output layer, a temporal head-only output layer, a spatial final convolutional layer, a spatial xception backbone, a spatial fusion-only multi-scale fusion layer, a spatial backbone-only CNN, a temporal LSTM layer, and a temporal attention layer. The core functions of each layer are detailed in [link to documentation]. Figure 5 .

[0156] S4.2. Construct a single-component thawing experiment, thaw each layer group separately for fine-tuning training, evaluate the contribution of a single layer group to the adaptive performance of the target watershed using the controlled variable method, and identify key feature extraction and decision layers.

[0157] Figure 5 Experimental results show that the temporal output layer (temporal_head_only) and fusion layer (fusion_only) are the most critical adaptation layers, each achieving a performance improvement of about 9.7% with only about 0.2% of the parameters. The temporal LSTM layer (temporal_lstm_only) is the next most important, achieving a performance improvement of 9.1% with 2.0% of the parameters. The performance improvement of all spatial flow components is 7.21%, showing strong coupling.

[0158] S4.3. Design experiments on combined unfreezing strategies, including unfreezing all output layers, combining high-level spatial flows, unfreezing the entire temporal flow, unfreezing the entire spatial flow, and unfreezing the minimum effective set. Set up full freeze and full unfreeze as control baselines.

[0159] Figure 5 Experimental results show that the "all_heads" strategy achieves a 9.76% performance improvement with only 1.4% of the parameters, exhibiting the highest parameter efficiency. The "minimal_effective" strategy, while increasing the parameters to 35.5%, achieves the same performance improvement, indicating the redundancy of the final convolutional layer in the spatial flow. The "temporal_complete" strategy, which only unfreezes the temporal flow output layer, achieves a 9.54% improvement with only 2.2% of the parameters, confirming the high adaptability of the temporal flow. However, the "spatial_complete" strategy, requiring 95.4% of the parameters, only achieves a 7.21% improvement, demonstrating the strong coupling and low adaptability of the spatial flow components. While the "unfreeze_all" strategy achieves the highest performance (10.26% improvement), its parameter efficiency is significantly lower than the selected combination strategy.

[0160] S4.4. Establish a multi-index comprehensive evaluation mechanism to select the key adaptation layer combination with the highest parameter efficiency and the strongest generalization ability through performance comparison analysis.

[0161] Based on a comprehensive evaluation of parameter efficiency and performance improvement, a key adaptation layer combination consisting of a fusion layer, temporal and spatial flow output layers, an attention mechanism, and an LSTM layer was determined.

[0162] 4. The method for predicting runoff in a data-free watershed based on spatiotemporal deep learning according to claim 1, characterized in that, in step S5, the specific steps for implementing a hierarchical progressive thawing strategy for the key adaptation layer to fine-tune the model and realize runoff prediction in the data-free watershed are as follows:

[0163] S5.1 Construct a Leave-One-Out Cross-Validation (LOOCV) framework. With runoff prediction in a data-free watershed in another region as the objective, select a single watershed from eight sub-watersheds as the target test watershed, and the remaining seven watersheds as training watersheds. Simulate a data-free watershed environment, and initialize the weights of the dual-stream network with pre-trained model parameters.

[0164] (1) Data segmentation strategy: From the eight watersheds DT, FS, FQ, FT, HJT, KH, XF and YK, one watershed is selected as the test set each time, and the other seven watersheds are selected as the training set, forming eight different training-test combinations to ensure that each watershed has the opportunity to be used as an independent validation target;

[0165] (2) Model initialization mechanism: Before each round of LOOCV experiment, the weights of the dual-flow network are initialized using model parameters pre-trained based on a certain watershed, rather than being randomly initialized. This makes full use of the pre-trained knowledge and ensures the fairness and repeatability of the experiment.

[0166] (3) Data-free watershed simulation: By completely isolating the data of the target test watershed and using only the data of the other 7 watersheds for model fine-tuning, the prediction scenario in the case of lack of historical observation data in the target watershed is realistically simulated, thereby evaluating the model's generalization ability and practical application potential in data-scarce areas.

[0167] S5.2 Implement a four-stage hierarchical progressive unfreezing strategy, successively unfreezing the fusion layer to learn the watershed-specific feature weights, the spatial output layer and the temporal output layer to adapt to the spatiotemporal feature distribution of the target watershed, the attention mechanism layer to optimize temporal information extraction, and the LSTM layer to achieve temporal pattern adaptation.

[0168] Based on the above unfreezing experiment results, the fusion layer achieved a 9.61% performance improvement with 0.2% parameter reduction, making it the most parameter-efficient and a key decision layer for dual-stream feature fusion. Therefore, it was preferentially chosen as the starting point for fine-tuning. The temporal and spatial output layers respectively undertake the mapping tasks from temporal and spatial features to the final prediction, forming a complete prediction path with the fusion layer. Moreover, the temporal output layer (9.76% improvement) slightly outperformed the spatial components (7.21% improvement), demonstrating the importance of temporal adaptation. Following the hierarchical fine-tuning logic of "decision layer → output layer → feature enhancement layer → basic modeling layer," a hierarchical progressive unfreezing fine-tuning strategy of "fusion layer → spatial and temporal output layers → temporal attention layer → temporal LSTM layer" was finally determined, ensuring that each step is finely adjusted based on the adaptation of the previous layer.

[0169] The total number of fine-tuning rounds is divided into four stages. In each stage, the parameters of the corresponding layer are unfrozen in the order of the four key adaptation layer combinations, and fine-tuning training is performed. At the same time, the model is optimized by continuing to use a two-stage warm-up-cosine decay learning rate scheduling.

[0170] (1) First stage: only unfreeze the parameters of the fusion layer, keep the parameters of the CNN backbone network and LSTM network frozen, and learn the fusion weights of the target watershed specific features;

[0171] (2) Second stage: Unfreeze the parameters of the fusion layer, spatial output layer and temporal output layer to adapt the model to the spatial-temporal characteristic distribution of the target watershed;

[0172] (3) Third stage: Based on the above, unfreeze the parameters of the attention mechanism layer and optimize the adaptive selection mechanism of time-series key information;

[0173] (4) Fourth stage: Unfreeze the LSTM layer parameters, fine-tune the weights of the recurrent neural network for time-series modeling, and achieve complete time-series pattern adaptation.

[0174] S5.3 Perform runoff prediction verification on the target test watershed and output high-precision runoff prediction results for watersheds without data.

[0175] Transfer learning is used to transfer knowledge from general features across multiple watersheds to specific features specific to a single watershed in a pre-trained model. The final output is a high-precision runoff prediction for the target watershed with no data. A hierarchical progressive thawing fine-tuning strategy is then employed to refine the runoff prediction results for the data-free watershed. Figure 6 As shown, the model's prediction accuracy is significantly higher than the average level.

[0176] The present invention proposes a method for inverting the soil moisture movement velocity based on soil moisture distribution, meteorological dynamic data and static soil data. This method can quickly invert the process and spatiotemporal differentiation of soil moisture movement in a three-dimensional watershed. The present invention relies on field measured data and numerical simulation results of physical hydrological models. However, the model scale can cover the entire three-dimensional watershed and has high spatiotemporal accuracy. It provides a practical and feasible solution for strengthening the scale and method of soil moisture monitoring.

[0177] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0178] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0179] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0180] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0181] Other embodiments of this application will readily occur to those skilled in the art upon consideration of the specification and practice of the disclosure herein. This application is intended to cover any variations, uses, or adaptations of this application that follow the general principles of this application and include common knowledge or customary techniques in the art not disclosed herein. The specification and embodiments are to be considered exemplary only, and the true scope and spirit of this application are indicated by the claims.

[0182] It should be understood that this application is not limited to the computational process described above and shown in the accompanying drawings, and various modifications and changes can be made without departing from its scope. The scope of this application is limited only by the appended claims.

Claims

1. A method for predicting runoff in a data-free watershed based on spatiotemporal deep learning, characterized in that, Includes the following steps: S1. Collect static geographic raster data and dynamic hydrological and meteorological time series data from multiple watersheds, and perform cross-watershed global normalization processing; S2. Construct a dual-flow deep learning model that integrates spatial and temporal features. The architecture includes a spatial flow network branch and a temporal flow network branch. The spatial flow network branch is used to extract static geospatial features, and the temporal flow network branch is used to process dynamic meteorological time-series data. The spatial feature vector and the temporal feature vector are concatenated and fused through a dual-flow feature fusion layer to output runoff prediction results. S3. Train the dual-flow deep learning model based on large sample watershed data from multiple regions to obtain a general hydrological model; S4. Through stratified thawing tests, evaluate the adaptability of each network layer in the general hydrological model to the target watershed, and screen out the key adaptation layers; S5. Under the leave-one-out cross-validation framework, hierarchical progressive thawing transfer learning is implemented on the general hydrological model based on the key adaptation layer to achieve runoff prediction in watersheds without data.

2. The method for predicting runoff in a data-free watershed based on spatiotemporal deep learning according to claim 1, characterized in that, The static geographic grid data includes the watershed's digital elevation model, slope, runoff accumulation, and normalized vegetation index; the dynamic hydro-meteorological time series data includes hourly precipitation, soil moisture, and potential evapotranspiration in the watershed.

3. The method for predicting watershed runoff in data-free basins based on spatiotemporal deep learning according to claim 1, characterized in that, The cross-basin global normalization process specifically includes: independently calculating the global maximum and minimum values ​​for each time-series feature matrix and performing cross-sample maximum-minimum scaling; for each spatial feature dimension, extracting raster extreme values ​​across the watershed and performing unified maximum-minimum scaling.

4. The method for predicting runoff in a data-free watershed based on spatiotemporal deep learning according to claim 1, characterized in that, The construction of the dual-stream deep learning model that integrates spatial and temporal features specifically includes: S2.

1. Construct a spatial flow network branch, using a two-branch convolutional neural network to process terrain data and normalized vegetation index respectively, and perform attention-weighted fusion of the two-branch features through a multi-scale feature fusion module; S2.

2. Integrate a deep feature extraction module based on the Xception architecture into the spatial flow network branch, and use depthwise separable convolution and residual connections to construct multi-layer Block units for deep spatial feature learning; S2.

3. Construct a time-series network branch, using a long short-term memory network combined with an attention mechanism to process dynamic meteorological time-series data. The long short-term memory network captures temporal dependencies, and the attention mechanism is used to calculate temporal feature weights to generate context vectors. S2.

4. Construct the dual-flow feature fusion layer, concatenate the spatial feature vector output by the spatial flow network branch with the temporal feature vector output by the temporal flow network branch, and achieve feature fusion and output runoff prediction results through a fully connected layer and a nonlinear activation function.

5. The method for predicting watershed runoff in data-free basins based on spatiotemporal deep learning according to claim 1, characterized in that, The training of the dual-flow deep learning model based on large-sample watershed data from multiple regions specifically includes: S3.

1. Construct a multi-basin training dataset, using the spatial static data of multiple basins as the fixed input of the spatial flow network branch, and dividing the hydrological and meteorological data of each basin into training and test sets according to time series; S3.

2. Design a multi-basin batch data loading mechanism, using a custom data loader to input topographic data, vegetation data and meteorological time series data into the corresponding branches of the dual-basin deep learning model, and adopt a hybrid precision training and gradient accumulation strategy to optimize computational efficiency; S3.

3. Configure an adaptive learning rate scheduler, adopt a multi-stage warm-up-cosine decay strategy for updating model parameters, and set weight decay regularization to prevent overfitting. S3.

4. Establish an automatic hyperparameter search mechanism to perform Bayesian optimization on key hyperparameters such as learning rate, batch size, hidden layer dimension, and dropout rate, and select the optimal parameter combination through parallel training with multiple trials; S3.

5. Establish a multi-indicator joint evaluation system, adopt an early shutdown mechanism to monitor the hydrological evaluation indicators of the model on the test sets of each watershed, and select and store the optimal model based on the comprehensive performance of the indicators.

6. The method for predicting runoff in a data-free watershed based on spatiotemporal deep learning according to claim 1, characterized in that, Step S4 specifically includes: S4.

1. Design a hierarchical unfreezing strategy system, dividing the general hydrological model into a fusion layer, a spatial flow output layer, a temporal flow output layer, a spatial flow final convolutional layer, a spatial flow Xception backbone, a spatial flow multi-scale fusion layer, a spatial flow basic CNN, a temporal flow LSTM layer, and a temporal flow attention layer core layer group; S4.

2. Construct a single-component thawing experiment, thaw each core layer group separately for fine-tuning training, and evaluate the contribution of a single layer group to the adaptability of the target watershed using the controlled variable method; S4.

3. Design experiments on combined unfreezing strategies, including combined unfreezing of all output layers, combined unfreezing of high-level spatial flow, complete unfreezing of temporal flow, complete unfreezing of spatial flow, and minimum effective set unfreezing strategy, and set up full freeze and full unfreezing as control baselines; S4.

4. Through a multi-index comprehensive evaluation mechanism, the key adaptation layer combination with the highest parameter efficiency and the strongest generalization ability is selected.

7. The method for predicting runoff in a data-free watershed based on spatiotemporal deep learning according to claim 1, characterized in that, Under the leave-one-out cross-validation framework, the hierarchical progressive thawing transfer learning of the general hydrological model based on the key adaptation layer specifically includes: S5.

1. Construct a leave-one-out cross-validation framework, select a single watershed as the target test watershed from N watersheds with available data in the target area, and use the remaining N-1 watersheds as training watersheds to simulate the environment of watersheds without data, and use the parameters of the general hydrological model as the initial weights; S5.

2. Implement a four-stage, layered, gradual unfreezing strategy, unfreezing the key adaptation layers sequentially. S5.

3. Perform runoff prediction verification on the target test watershed and output the runoff prediction results for the watershed without data.

8. A data-free watershed runoff prediction system based on spatiotemporal deep learning, characterized in that, To implement the method of any one of claims 1 to 7, comprising: Data acquisition and processing module: used to acquire static geographic raster data and dynamic hydrological and meteorological time series data from multiple watersheds, and to perform cross-watershed global normalization processing; Dual-stream deep learning architecture construction module: used to build a dual-stream deep learning model that integrates spatial and temporal features. The architecture includes a spatial flow network branch and a temporal flow network branch. The spatial flow network branch is used to extract static geospatial features, and the temporal flow network branch is used to process dynamic meteorological time series data. The spatial feature vector and the temporal feature vector are concatenated and fused through a dual-stream feature fusion layer to output runoff prediction results. Model pre-training module: used to train the dual-flow deep learning model based on large sample watershed data from multiple regions to obtain a general hydrological model; Key Adaptation Layer Screening Module: Used to evaluate the adaptability of each network layer in the general hydrological model to the target watershed through stratified thawing tests, and to screen out key adaptation layers; Transfer learning prediction module: used to implement hierarchical progressive unfreezing transfer learning on the pre-trained general hydrological model based on the key adaptation layer under the leave-one-out cross-validation framework, to achieve watershed runoff prediction without data.

9. A computer device, characterized in that, The computer device includes: One or more processors; Memory, used to store one or more programs; When the one or more programs are executed by the one or more processors, the one or more processors implement the dataless watershed runoff prediction method based on spatiotemporal deep learning as described in any one of claims 1 to 7.

10. A computer-readable storage medium storing computer instructions, characterized in that, When the computer instructions are executed by one or more processors, the one or more processors are caused to perform the steps of the method according to any one of claims 1 to 7.

Images