Snowmelt runoff prediction method and system based on direct measurement of snow water equivalent

Abstract

The application discloses a snowmelt runoff prediction method and system based on direct measurement of snow water equivalent, and belongs to the technical field of hydrological monitoring. The system comprises a cooperative observation network composed of an acoustic source array emitting coded acoustic waves and a sensing optical fiber laid on the ground surface; a distributed optical fiber acoustic wave sensing demodulator for receiving a snow layer modulated vibration signal; a fusion inversion calculation unit adopting a physical information neural network algorithm, embedding an acoustic wave fluctuation equation as a constraint into network training, and directly inverting a continuous spatial distribution of the snow water equivalent from the signal; and a runoff prediction module for predicting runoff by using the distributed data. The application breaks through the limitations of traditional point measurement and satellite remote sensing, realizes direct, high-resolution and high-precision surface measurement of the snow water equivalent, fundamentally improves the input data quality of the snowmelt runoff prediction model, and has the characteristics of high precision, strong reliability and cost-effectiveness.

Description

Technical Field

[0001] This invention belongs to the field of cold region hydrological monitoring and water resources management technology, specifically involving the measurement method of key hydrological parameters of snow cover and its application in accurate prediction of snowmelt runoff. Background Technology

[0002] In cold regions, snowmelt is the primary source of river runoff in spring, making accurate forecasting crucial for water resource allocation, flood control and disaster reduction, and ecological research. A core prerequisite for accurate forecasting is acquiring high-precision, high spatiotemporal resolution driving data, with snow water equivalent (SWE, the equivalent depth of liquid water per unit area of ​​snow cover) being the most critical input variable. Current technologies primarily rely on two methods: the first is point-based direct measurement, such as using snow pillows or weighing snow depth gauges for high-precision measurements at a single point; the second is area-based remote sensing indirect inversion, mainly utilizing passive microwave satellite remote sensing data to invert large-scale SWE. Currently, there are studies in academia using active acoustics to measure snow density (e.g., Kinar & Pomeroy, 2015, "SAS2: the system for acoustic sensing of snow"). This study confirmed the physical correlation between sound velocity, attenuation, snow density, and temperature. While DAS monitoring of glacier vibrations or seismic wave signals under snow cover has been applied, it is typically a passive monitoring method. In the field of physical information neural networks, treating the wave equation as a constraint term for acoustic inversion (AcousticFWI) has become a research hotspot in recent years.

[0003] However, the aforementioned existing methods have significant drawbacks, collectively constituting a bottleneck restricting the improvement of prediction accuracy. While point-based measurement methods offer high local accuracy, their spatial representativeness is extremely poor, failing to capture the spatial heterogeneity of snow cover within a watershed caused by factors such as topography, vegetation, and wind speed. Furthermore, the error introduced when interpolating sparse point data into a planar distribution is substantial. Planar remote sensing methods, limited by their physical principles, suffer from inherent defects such as coarse spatial resolution (typically ≥25 km) and severe interference from forest cover, snow moisture, and topographic complexity in their inversion algorithms. In applications in small mountainous watersheds, their accuracy often falls short of hydrological model requirements, with errors generally exceeding 30%. These defects result in inherent ambiguity and structural deficiencies in the snow water equivalent data input into the hydrological model, fundamentally limiting further improvements in model performance.

[0004] Therefore, developing a technology that can achieve direct, continuous, and high spatial resolution acquisition of isometric snow water equivalent, in order to break through the accuracy ceiling of existing prediction models from the data source, has become a key technical problem that urgently needs to be solved in this field, and has important theoretical value and urgent practical needs. Summary of the Invention

[0005] To address the shortcomings of existing snowmelt equivalent measurement technologies that cannot simultaneously achieve directness, high accuracy, and high spatial resolution, this invention provides a direct measurement system and method based on acoustic tomography and distributed fiber optic sensing, which is applied to snowmelt runoff prediction. The aim is to solve the problem of fuzzy input information from the source of the data, thereby achieving accurate and reliable prediction of snowmelt runoff in cold regions.

[0006] The solution to the technical problem of this invention is as follows: a snowmelt runoff prediction system based on direct measurement of snow water equivalent, comprising: a collaborative observation network, which includes at least one sound source array for transmitting coded acoustic signals to the snow layer, and sensing optical fibers laid on or shallowly buried on the surface of the monitoring area; a distributed optical fiber acoustic wave sensor demodulator, which is connected to the sensing optical fiber, for demodulating and acquiring the acoustic vibration signal modulated by the snow layer; a fusion inversion calculation unit, which is communicatively connected to the distributed optical fiber acoustic wave sensor demodulator, for analyzing and outputting continuous distribution data of snow layer density and snow water equivalent along the sensing optical fiber path based on the acoustic vibration signal, through an inversion algorithm that fuses physical prior constraints and data driving; and a runoff prediction module, which receives the continuous distribution data of snow water equivalent, meteorological driving data, and underlying surface characteristic data, and outputs snowmelt runoff prediction results based on a preset runoff prediction model.

[0007] Preferably, the frequency range of the coded acoustic signal emitted by the sound source array is 20Hz to 200Hz.

[0008] Preferably, the encoded acoustic signal is a linear frequency modulated signal or a pseudo-random binary sequence encoded signal.

[0009] Preferably, the sensing optical fiber is deployed in the monitoring area in such a way that it forms at least one closed loop or cross network.

[0010] Preferably, the inversion algorithm executed by the fusion inversion calculation unit is a physical information neural network algorithm. This algorithm inverts the acoustic vibration signal by embedding the physical equation describing the propagation law of sound waves in the snow layer as a constraint term into the loss function of the neural network; wherein, the weight coefficient λ of the physical equation constraint term in the loss function ranges from 0.4 to 0.6.

[0011] Preferably, the fusion inversion calculation unit further includes an adaptive signal separation module, which performs time-frequency analysis on the acoustic vibration signal, separates the direct wave and at least one reflected wave, and generates an enhanced feature vector based on the travel time difference and relative amplitude attenuation of the direct wave and the reflected wave, and inputs it into the inversion algorithm.

[0012] Preferably, the runoff prediction model is a hydrological model based on physical mechanisms, a data-driven model based on machine learning, or a model that combines both.

[0013] A method for predicting snowmelt runoff based on direct measurement of snow water equivalent is adopted, comprising the following steps: S1: transmitting coded acoustic signals to the snow layer in the monitoring area through a collaborative observation network, and receiving the acoustic vibration signals modulated by the snow layer using laid sensing optical fibers; S2: demodulating and acquiring the acoustic vibration signals using a distributed optical fiber acoustic wave sensor demodulator; S3: processing the acoustic vibration signals by integrating physical prior constraints and data-driven inversion algorithms to obtain continuous distribution data of snow layer density and snow water equivalent along the sensing optical fiber path; S4: inputting the continuous distribution data of snow water equivalent, meteorological driving data, and underlying surface characteristic data into a preset runoff prediction model to obtain and output the predicted snowmelt runoff results for the monitoring area.

[0014] Preferably, the inversion algorithm is a physical information neural network algorithm, and the loss function of its training process includes a data fitting loss term and a physical equation constraint loss term. The physical equation constraint loss term is calculated by the wave equation describing the propagation of sound waves in the snow layer, and the weight coefficient λ of the two losses ranges from 0.4 to 0.6. Step S3 also includes the step of adaptively separating the acoustic vibration signal to extract the enhanced feature vector.

[0015] The beneficial effects of this invention are as follows: 1. It abandons the traditional roundabout path of point interpolation or indirect inversion, and for the first time achieves in-situ, direct, and continuous planar measurement of snow water equivalent by fusing acoustic tomography and distributed fiber optic sensing. Through the physical information neural network fusion inversion algorithm described in the implementation, the theoretical relative error of SWE inversion is controlled within 10%-15%, far superior to passive microwave remote sensing (>30%). Simultaneously, the spatial resolution reaches meter-level continuity along the optical fiber, and the planar resolution is improved to sub-kilometer level, representing an order-of-magnitude improvement compared to point measurement and satellite remote sensing.

[0016] 2. The sensing element utilizes passive optical fiber, which is interference-resistant, withstands extreme environments, requires no maintenance, and ensures high system reliability. Implementation results show that compared to deploying traditional automated dot matrix networks with similar information density, the total cost of ownership (TCO) of this invention can be reduced by 30%-40% over five years. The introduced lightweight CNN adaptive signal separation module can automatically identify signals from complex snow layer structures, improving the automation and robustness of feature extraction and reducing the need for manual intervention.

[0017] 3. The improved signal separation and feature enhancement methods for layered snow structures in the embodiments enable the system to not only invert SWE (Snow Wave Evolution) but also simultaneously acquire information about the internal interfaces of the snow layer, expanding its application potential in scenarios such as avalanche risk assessment. The continuous and high-precision SWE data provided offers an unprecedented reliable data source for the calibration and validation of subsequent hydrological models and climate change research, with a clear path to practical application of the technology. Attached Figure Description

[0018] Figure 1 System architecture and workflow diagram.

[0019] Figure 2 : Schematic diagram of the principle of the fusion inversion algorithm (physical information neural network).

[0020] Figure 3 : Schematic diagram of front-end signal processing flow (including adaptive signal separation).

[0021] Figure 4 Flowchart of a method for accurate prediction of snowmelt runoff in cold regions.

[0022] Figure 5 : Figure 2 The internal working principle diagram of the physical information neural network.

[0023] The diagram is labeled as follows: 101, Collaborative Observation Network; 101a, Sound Source Array; 101b, Sensor Fiber Optic Network; 102, DAS Demodulator; 103, Fusion Inversion Calculation Unit; 104, Runoff Prediction Module. Detailed Implementation

[0024] The prediction method and system provided by the present invention will be described in detail below with reference to the accompanying drawings, so that those skilled in the art can implement the present invention. It should be understood that the specific embodiments described herein are for illustrative purposes only and are not intended to limit the scope of protection of the present invention.

[0025] Example 1: In the fields of hydrology, water resource management, and disaster prevention and mitigation in cold regions, accurate prediction of snowmelt runoff is crucial. The core bottleneck lies in the difficulty of accurately, timely, and over a large scale acquiring the most critical input parameter driving the prediction model: snow water equivalent (SWE). Snow water equivalent is the equivalent liquid water depth per unit area after snowmelt, and it is a direct physical quantity determining the amount of water produced by snowmelt. Developing a technology that can achieve direct, continuous, and high spatial resolution area SWE acquisition, thereby overcoming the accuracy bottleneck of existing prediction models at the data source, has become a key technical problem urgently needing to be solved in this field.

[0026] Reference Figure 1 The accurate prediction system for snowmelt runoff in cold regions provided in this embodiment of the invention mainly includes a collaborative observation network 101, a distributed optical fiber acoustic wave sensor (DAS) demodulator 102, a fusion inversion calculation unit 103, and a runoff prediction module 104. The data flow is as follows: acoustic vibration signal → digital signal → feature data → inversion result → prediction result.

[0027] The collaborative observation network 101 consists of a sound source array 101a and a sensing fiber optic network 101b, jointly deployed in the target monitoring watershed. The sound source array 101a comprises multiple (e.g., 3-5) fixed sound source stations, deployed at key locations (such as high points) around the monitoring watershed. The core of each sound source station is a transducer capable of emitting coded sound waves at specific frequencies. In this invention, the operating frequency range of the sound sources is selected as 20Hz to 200Hz. This range is determined based on the following non-obvious technical considerations: sound waves with frequencies below 20Hz have excessively long wavelengths, resulting in insufficient vertical spatial resolution and difficulty in distinguishing shallow snow layers; sound waves with frequencies above 200Hz attenuate too quickly in snow-covered media (following the classical sound wave attenuation model), limiting the effective detection depth and preventing penetration of thick snow layers. This frequency band represents an optimal balance between penetration capability and resolution, requiring a combination of theoretical and experimental determination, and is not common knowledge in the field. The encoded signal is preferably a linear frequency modulated signal or a pseudo-random binary sequence code. These signals have good autocorrelation characteristics, which are beneficial for effective extraction through cross-correlation processing in strong environmental noise. The sensing fiber optic network 101b uses standard single-mode communication fiber as a distributed sensor. Its deployment method is preferably to form at least one closed loop or cross-network (such as...). Figure 1 As shown in the diagram, the optical fibers are laid along key paths such as river channels and ridges. This deployment method forms a multi-angle acoustic tomography observation network, which is beneficial to improving the reliability and spatial resolution of the inversion results. The optical fibers can be laid directly on the ground surface or shallowly buried (depth <10cm) for fixation.

[0028] A distributed optical fiber acoustic sensing (DAS) demodulator 102 is optically connected to a sensing optical fiber 101b. It continuously acquires and demodulates the phase changes of backscattered Rayleigh light caused by acoustic vibration along the fiber at a high sampling rate (e.g., ≥500Hz), converting them into continuous acoustic / vibration time-series signal data along the fiber space (sampling intervals up to 1 meter). When the acoustic waves emitted by the acoustic source array 101a penetrate the snow layer and are modulated, the resulting vibrations are sensed by the optical fiber and recorded by this device.

[0029] The fusion inversion calculation unit 103 is the core of realizing direct measurement. It receives the raw data from the DAS demodulator 102 and executes the improved fusion inversion algorithm: First, it performs signal preprocessing and feature extraction. Using the known sound source emission coding sequence, it performs cross-correlation calculation and delay superposition processing on the data of each sound source-fiber receiving point pair to separate the corresponding signal from the environmental noise with a high signal-to-noise ratio. It also extracts multi-dimensional features such as the travel time, amplitude attenuation spectrum, and center frequency shift of the sound wave signal at each measurement point. Figure 2 and Figure 5This paper visually demonstrates how input features, neural network structure, and physical equation constraints (wave equation) are fused through a loss function and optimized via backpropagation to ultimately output the SWE (Sensitive Wave Equation). Specifically, the core lies in the fusion and inversion algorithm based on a physical information neural network. The construction and training of this algorithm includes the following key steps.

[0030] 1. Network Construction: Construct a deep neural network. The input layer receives the aforementioned multi-dimensional acoustic features and the spatial coordinates of the measurement points, and the output layer predicts the snow density profile ρ below that location. Figure 5 The algorithm's network architecture is shown in detail, especially how the hidden layers achieve the fusion of physical constraints and data features. Among them, 210 is the input layer, 211 is the travel time feature t, 212 is the amplitude feature A, 213 is the spatial feature z / y / z, 214 is the reflected wave feature Δt / ΔA, 220 is the hidden layer, 221 is the feature extraction (hidden layer one), 222 is the nonlinear mapping (hidden layer two), 223 is the physical constraint fusion (hidden layer three), 2231 is the uncertainty unit, 2232 is the physical guidance unit, 230 is the output layer, 231 is the density ρz, 232 is the sound speed cρ, 233 is the attenuation βρ, 240 is the physical constraint, and 250 is the loss synthesis.

[0031] 2. Define physical constraints: The physical laws describing the propagation of sound waves in non-uniform, attenuating media are embedded as strong constraints. Specifically, a simplified form of the sound wave equation is used:

[0032] (1);

[0033] Where P is the sound pressure, c(ρ) is the sound velocity related to the snow density ρ (determined by the empirical formula c(ρ)=aρ+b, where a and b are coefficients obtained through laboratory calibration), and β(ρ) is the density-related attenuation coefficient.

[0034] 3. Design the loss function: The loss function L of the neural network consists of two parts: L = L_data + λL_physics (2); L_data (data loss): calculates the mean square error between the density prediction value output by the network and a coarse initial density model obtained by traditional travel-time tomography inversion. L_physics (physical loss): during the training process, a large number of spatiotemporal coordinate points are randomly sampled in the entire solution domain, and the current predicted ρ field of the network is substituted into the above wave equation to calculate the mean square value of the equation residual.

[0035] λ is a weighting coefficient used to balance data-driven approaches and physical constraints. Extensive simulation experiments have verified that, for the specific medium of snow cover in cold regions, the inversion results achieve optimal accuracy and physical plausibility when the value of λ is between 0.4 and 0.6. This specific range is a preferred finding of this invention.

[0036] 4. Training and Output: The network is trained by minimizing the total loss function L. After training, the network can directly output a high-resolution snow density profile ρ(z) from the input acoustic features.

[0037] Snow water equivalent calculation in the snow water equivalent calculation and runoff prediction module 104: The high-resolution snow layer density profile ρ(z) obtained by the fusion inversion calculation unit 103 is combined with the snow depth H and calculated by integrating SWE=∫0 H *ρ(z)*dz (z is the vertical coordinate, unit m; ρ(z) is the instantaneous density at height z, unit kg / m³) -3 The continuous SWE values ​​along the fiber optic path are calculated. The snow depth H can be obtained through static strain inversion caused by the static load of snow accumulation monitored by the DAS system, or provided by the deployed auxiliary ultrasonic snow depth gauges. Runoff prediction: The runoff prediction module 104 receives the above-mentioned continuous SWE distribution data, and simultaneously accesses the watershed's meteorological driving data (temperature, precipitation, radiation) and underlying surface data. These data are input into a pre-trained runoff prediction model (such as the physics-based distributed hydrological model SWAT, or an LSTM-based machine learning model) to output a predicted sequence of snowmelt runoff at the watershed outlet for future periods. The SWE data provided by this invention, as a high-precision initial field, can significantly improve the prediction accuracy of such models.

[0038] Example 2: Based on the basic scheme of Example 1, in order to further improve the inversion capability for complex layered snow cover, a preferred embodiment is described below.

[0039] For adaptive signal separation of layered snow structures, natural snow often exhibits a layered structure including new snow, old snow, deep frost layers, and ice layers. To extract more information, an adaptive signal separation module is added after the feature extraction stage. This module performs time-frequency analysis (such as wavelet transform) on each received signal to identify the direct wave and primary and secondary reflected waves from different layer interfaces. Figure 3 The process is shown to separate the direct wave and reflected wave (R1, R2) from the original DAS signal through time-frequency analysis, calculate the travel time difference Δt and amplitude ratio A, and finally form an enhanced eigenvector.

[0040] The Feature Enhancement and Co-inversion module further calculates the travel time difference (Δt_i) and relative amplitude attenuation (A_i) of each reflected wave relative to the direct wave. These Δt_i directly reflect the depth of the layer interface, and A_i reflects the impedance difference on both sides of the interface. These multipath wave features are added as a new dimension, forming an enhanced feature vector together with the direct wave features, which is then input into the physical information neural network. This feature not only improves the inversion accuracy of SWE but also simultaneously outputs information on the main layer interfaces within the snow layer. This provides crucial data for avalanche risk early warning. Furthermore, by using information from multiple propagation paths for joint constraints, the robustness of the algorithm is significantly enhanced when a single path signal is interfered with.

[0041] The effectiveness of this invention was verified through theoretical analysis and simulation experiments, achieving direct, area-based, and continuous measurement of SWE (Surface Wetted Elevation) with a spatial resolution reaching the sub-kilometer level, representing a significant improvement over point-based measurements and satellite remote sensing. In a typical layered snow cover simulation model, the average relative error of SWE retrieval using the method of this invention is significantly better than that of passive microwave remote sensing under the same conditions. Geophysical exploration techniques such as acoustic tomography, fiber optic DAS sensing, and hydrophysical models are deeply integrated using a physical information neural network, and key parameters (e.g., frequency range 20-200Hz, λ=0.4-0.6) are determined to address the problem of direct SWE measurement in cold regions.

[0042] Example 3: Based on the high-precision continuous distribution data of snowmelt equivalent obtained from the aforementioned system, the accurate prediction method for snowmelt runoff in cold regions in this example is implemented through the following specific steps, forming a complete and operable prediction process, such as... Figure 4 As shown, this forms a complete and operable closed-loop process from data acquisition to prediction output.

[0043] Step S1: Specific implementation of signal excitation and acquisition, starting the collaborative observation network: Within the monitoring basin, the deployed sound source array 101a is controlled to operate sequentially according to a preset timing protocol (e.g., every hour on the hour). Each sound source station emits a predefined coded acoustic signal, preferably a linear frequency modulated signal with a frequency in the range of 20Hz to 200Hz, to optimize penetration and resolution in the snow layer. Signal reception and sensing: During propagation, the emitted sound waves are modulated (including reflection, refraction, scattering, and attenuation) by the physical characteristics (density, layering structure, liquid water content) of the surface snow layer within the basin. The sensing fiber optic network 101b, laid on the surface or shallowly buried, acts as a distributed sensor, continuously sensing and capturing these modulated acoustic vibration signals. The entire process realizes physical data acquisition from active excitation to distributed sensing.

[0044] Step S2: Specific implementation of signal demodulation and acquisition. Signal demodulation: The distributed optical fiber acoustic wave sensor (DAS) demodulator 102 receives the backscattered Rayleigh light signal returned from the sensing fiber in real time. Through coherent demodulation technology, the phase change of the optical signal (caused by acoustic vibration) is converted into a corresponding electrical signal. Data digitization and generation: The above electrical signal is converted from analog to digital to generate a spatiotemporal sequence dataset of acoustic vibration that is continuously distributed along the optical fiber (spatial sampling intervals can reach the meter level) and has high temporal resolution (sampling rate ≥ 500 Hz). This dataset is the direct input for subsequent inversion processing.

[0045] Step S3: Specific implementation of fusion inversion and data generation, feature extraction: The raw DAS data obtained in S2 is first preprocessed (filtering, noise reduction), and then the acoustic feature parameters for each sound source-receiver pair are extracted using a cross-correlation algorithm, mainly including sound wave travel time, amplitude attenuation, center frequency, etc. Execution of the fusion inversion algorithm: The extracted acoustic features and corresponding spatial coordinates are input to the fusion inversion calculation unit 103. This unit runs the core Physical Information Neural Network (PINN) algorithm. The specific implementation of this algorithm includes network construction, physical constraint embedding, inversion calculation, and generation of SWE products. Network construction involves building a deep neural network with acoustic features as input and snow layer density profile as output. Physical constraint embedding involves using the physical equation describing sound wave propagation (the wave equation as described in claim 5) as a mandatory constraint during the network training and inference process, applied to the network optimization process through a loss function term (L_physics), where the physical constraint weight coefficient λ is preferably 0.4-0.6. The inversion calculation, performed by the network through forward propagation, directly outputs high-resolution snow density ρ(z) distribution data along the fiber optic path. The SWE product is generated by combining the inverted density profile ρ(z) with synchronously acquired snow depth H data (which can be provided by DAS static strain inversion or auxiliary sensors), and calculating SWE=∫_0^Hρ(z)dz through integration, ultimately generating a high-spatiotemporal resolution continuous distribution data product of snow water equivalent (SWE). This product is the core achievement of this invention's direct measurement function.

[0046] Step S4: Specific implementation of runoff prediction and result output. The data products obtained in Step S3 are applied to the final prediction, including reading the SWE continuous distribution data generated in S3 as the most accurate initial snowmelt source input for the hydrological model. Meteorological driving data (such as time-series data of temperature, precipitation, and radiation) and underlying surface characteristic data (such as digital elevation model, soil type, and vegetation cover map) for the monitored area are prepared simultaneously. All data are uniformly interpolated or resampled onto the same spatiotemporal grid to ensure data consistency in both spatial and temporal dimensions. The preset runoff prediction model is invoked. This model can be a distributed hydrological model based on physical mechanisms (such as an adaptively modified SWAT model), a data-driven model based on machine learning (such as an LSTM network), or a coupled model of both. The fused multi-source data is input into the prediction model. Based on its inherent algorithm and calibrated parameters, the model simulates or extrapolates the entire process of snowmelt, runoff generation, and confluence. After the model runs, it directly outputs the predicted time-series results of snowmelt runoff at the outlet section of the monitored area for a specific future period (such as the next 24 or 72 hours). The prediction results are stored and published in standard data formats (such as CSV or NetCDF) or as visualization charts (flow process line graphs). Simultaneously, thresholds can be set for early warning alerts.

[0047] It should be noted that the above embodiments and accompanying drawings are merely illustrative examples of the core principles and key structures of the present invention. The accompanying drawings are simplified schematic diagrams. This specification focuses on describing the innovative technical means necessary to achieve the purpose of the invention; while conventional design details that can be implemented by those skilled in the art without creative effort are not described in detail, they should be understood as naturally included in the specific implementation of the present invention.

Claims

1. A snowmelt runoff prediction system based on direct measurement of snow water equivalent, characterized by, The application relates to a snowmelt runoff prediction method based on a cooperative observation network. The cooperative observation network comprises at least one sound source array for emitting coded acoustic signals to a snow layer and a sensing optical fiber laid on the ground surface or shallowly buried in a monitoring area; A distributed optical fiber acoustic sensing demodulator is connected to the sensing optical fiber and used for demodulating and acquiring acoustic vibration signals modulated by the snow layer; A fusion inversion calculation unit is communicatively connected to the distributed optical fiber acoustic sensing demodulator and used for analyzing and outputting continuous distribution data of snow layer density and snow water equivalent along the sensing optical fiber path based on the acoustic vibration signals by fusing a physical prior constraint and a data-driven inversion algorithm; A runoff prediction module receives the continuous distribution data of the snow water equivalent, meteorological driving data and underlying surface feature data and outputs a snowmelt runoff prediction result based on a preset runoff prediction model. The frequency range of the coded acoustic signals emitted by the sound source array is 20-200 Hz. The fusion inversion calculation unit further comprises an adaptive signal separation module for performing time-frequency analysis on the acoustic vibration signals, separating out direct waves and at least one reflected wave and generating an enhanced feature vector based on the travel time difference and relative amplitude attenuation of the direct waves and the reflected waves and inputting the enhanced feature vector to the inversion algorithm.

2. The snowmelt runoff forecast system of claim 1, wherein, The coded acoustic signals are linear frequency modulation signals or pseudo-random binary sequence coded signals.

3. The snowmelt runoff forecast system of claim 1, wherein, The sensing optical fiber is laid in the monitoring area in a manner of forming at least one closed loop or cross network.

4. The snowmelt runoff forecast system of claim 1, wherein, The inversion algorithm executed by the fusion inversion calculation unit is a physical information neural network algorithm which embeds a physical equation describing the propagation law of acoustic waves in the snow layer into a loss function of a neural network as a constraint term to perform inversion on the acoustic vibration signals; wherein the weight coefficient lambda of the physical equation constraint term in the loss function ranges from 0.4 to 0.

6.

5. The snowmelt runoff forecast system of claim 1, wherein, The runoff prediction model is a hydrological model based on a physical mechanism, a data-driven model based on machine learning or a model coupled with both.

6. A snowmelt runoff prediction method based on the system of claim 1, characterized by, The method comprises the following steps: S1: emitting coded acoustic signals to the snow layer of the monitoring area through the cooperative observation network and receiving acoustic vibration signals modulated by the snow layer by using the laid sensing optical fiber; S2: demodulating and acquiring the acoustic vibration signals by using the distributed optical fiber acoustic sensing demodulator; S3: processing the acoustic vibration signals by fusing a physical prior constraint and a data-driven inversion algorithm to obtain continuous distribution data of snow layer density and snow water equivalent along the sensing optical fiber path; S4: inputting the continuous distribution data of the snow water equivalent, meteorological driving data and underlying surface feature data into a preset runoff prediction model to obtain and output a snowmelt runoff prediction result of the monitoring area.

7. The snowmelt runoff forecast method of claim 6, wherein, The inversion algorithm is a physical information neural network algorithm, the loss function of the training process of which comprises a data fitting loss term and a physical equation constraint loss term, the physical equation constraint loss term is calculated by a wave equation describing the propagation of acoustic waves in the snow layer, and the weight coefficient lambda of the two loss terms ranges from 0.4 to 0.6; the step S3 further comprises a step of adaptively separating the acoustic vibration signals to extract an enhanced feature vector.

Images