Multi-source rainfall intelligent fusion method and system fusing geoscience prior and adaptive Gamma loss

Abstract

The invention provides a multi-source rainfall intelligent fusion method and system fusing geoscience prior and adaptive Gamma loss, and relates to the field of meteorological remote sensing data fusion and artificial intelligence rainfall estimation. According to the method, a multi-source observation tensor and geoscience prior feature matrix is constructed, historical rainfall statistics are extracted, Gamma distribution parameters are generated through meta-learning, and an adaptive loss function is constructed; and inputting the fusion features into a deep network to jointly optimize the weight and distribution parameters of the model, outputting a rainfall estimated value and posterior distribution, and performing confidence interval sampling, terrain consistency checking and smoothing to form a final rainfall product. According to the invention, the extreme rainfall estimation precision and generalization capability can be improved.

Description

Technical Field

[0001] The present invention relates to the field of meteorological remote sensing data fusion and artificial intelligence precipitation estimation, and in particular to a multi-source precipitation intelligent fusion method and system integrating geoscience priors and adaptive Gamma loss. Background Art

[0002] Currently, quantitative precipitation estimation has developed a three-dimensional monitoring system comprised of multiple observation sources, including satellite passive microwave / infrared, ground-based Doppler radar, and gridded rain gauge interpolation. Publicly available products, such as IMERG, GPM, and MRMS, provide near-real-time precipitation information at minute-level and hundred-meter-to-kilometer-level temporal and spatial resolutions, and are widely used in river basin flood forecasting, urban waterlogging warnings, and agricultural drought assessments.

[0003] With the development of deep learning and big data technologies, multi-source precipitation fusion is iteratively moving from traditional statistical weighting methods and Kalman filtering methods to an intelligent fusion paradigm coupled with "data-model-mechanism": ① Introducing high-precision DEM, land use, climate zoning and other geological prior features to improve the quantitative characterization of spatially heterogeneous regions (high mountains and land-sea transition zones); ② Using a probability distribution sensitive loss function to enable the model to explicitly learn the uncertainty and extreme tails of precipitation; ③ Integrating explainable AI and meta-learning to achieve adaptive generalization across different climate scales and precipitation types.

[0004] However, most existing deep fusion methods assume that precipitation errors approximately follow a symmetric distribution (such as Gaussian), and only perform Gamma or log-normal fitting at the output end. This results in insensitivity to skewed, peaked, and thick-tailed medium- and large-scale rainstorm events. At the same time, geoscience mechanism elements such as topographic circulation, monsoon fronts, and land-sea thermal differences are often simplified to static weights or completely ignored, making it difficult to ensure consistent performance across basins and seasons. Summary of the Invention

[0005] In order to overcome the shortcomings of the existing technology, the purpose of the present invention is to provide a multi-source precipitation intelligent fusion method and system that integrates geoscientific priors and adaptive Gamma loss, which can dynamically adjust shape-scale parameters and explicitly embed geoscientific constraints during training, thereby significantly improving the characterization accuracy and generalization ability of extreme precipitation fields.

[0006] To achieve the above object, the present invention provides the following solutions:

[0007] A multi-source precipitation intelligent fusion method that integrates geoscientific priors and adaptive Gamma loss includes:

[0008] Construct a multi-source precipitation observation tensor for the target area;

[0009] Based on the digital elevation model, long-term climate mean, surface dynamic parameters, land use and cover parameters, a geoscience prior feature matrix with the same grid as the multi-source precipitation observation tensor is generated, and the matrix is ​​spliced ​​with the multi-source precipitation observation tensor according to the channel dimension to obtain a fused feature tensor;

[0010] For the historical precipitation samples of the pixels corresponding to the fused feature tensor, a skew distribution statistic is calculated, and based on the skew distribution statistic, a shape parameter k and a scale parameter θ of the Gamma distribution are dynamically generated through a meta-learning strategy to obtain an adaptive Gamma loss function;

[0011] Inputting the fused feature tensor into a deep fusion network with a cross-modal attention mechanism, and using the adaptive Gamma loss function to jointly iteratively optimize the network weights, the shape parameter k, and the scale parameter θ of the deep fusion network to obtain a converged fusion network model;

[0012] The fusion network model is used to infer the fusion feature tensor of the estimated time step, and the fusion precipitation estimation value and the Gamma posterior distribution are output;

[0013] The precipitation confidence interval is obtained based on the posterior distribution, and the final precipitation product is formed after terrain consistency test and spatial smoothing.

[0014] Preferably, constructing a multi-source precipitation observation tensor includes:

[0015] Collecting satellite precipitation inversion data, ground-based radar precipitation estimation data and ground-based rain gauge observation data over the target area under a unified time reference;

[0016] The satellite precipitation inversion data, the ground-based radar precipitation estimation data and the ground rain gauge observation data are subjected to resolution uniformization and pixel alignment to obtain a multi-source precipitation observation tensor.

[0017] Preferably, based on the digital elevation model, long-term climate mean, surface dynamic parameters, land use and cover parameters, a geoscience priori feature matrix with the same grid as the multi-source precipitation observation tensor is generated, and is spliced ​​with the multi-source precipitation observation tensor according to the channel dimension to obtain a fused feature tensor, including:

[0018] Acquiring geoscientific data covering the target area; the geoscientific data includes a digital elevation model, multi-year average precipitation data, multi-year average temperature data, surface wind speed data, surface solar radiation data, a land use classification map, and a surface cover type map;

[0019] The geoscientific data are spatially resampled by bilinear interpolation to make the resolution of the geoscientific data consistent with the multi-source precipitation observation tensor, and grid alignment is performed using a unified geographic coordinate system;

[0020] Calculating terrain indicators based on the digital elevation models respectively;

[0021] Extracting land cover variables based on the land use classification map and the land cover type map; the land cover variables include land cover category codes and normalized difference vegetation indexes;

[0022] constructing four corresponding surface climate variables according to the multi-year average precipitation data, the multi-year average temperature data, the surface wind speed data, and the surface solar radiation data;

[0023] splicing the terrain index, the surface cover variable and the surface classification variable into a fixed-format geoscience prior feature matrix in the channel dimension;

[0024] The geoscience prior feature matrix and the multi-source precipitation observation tensor are spliced ​​pixel by pixel in the channel dimension to obtain a fused feature tensor.

[0025] Preferably, the terrain indicators include altitude, slope, aspect and distance from the sea.

[0026] Preferably, the expression of the geoscience priori characteristic matrix is:

[0027] G=[E,C,L]

[0028] Among them, G represents the geoscience prior feature matrix used for network input after fusion, which is formed by splicing according to the channel dimension and has the dimension of H×W×D g , where H and W are the height and width of the grid image of the target area, respectively, D g is the number of geoscience priori channels; E = [A, B, C, D] represents the terrain variable submatrix, where A is the altitude, B is the slope, C is the sine value encoding after slope aspect conversion, and D is the distance from the sea; C = [E, F, G, H] represents the climate variable submatrix, where E is the multi-year average precipitation data, F is the multi-year average temperature data, G is the multi-year average surface wind speed data, and H is the multi-year average surface solar radiation data; L = [I, J] represents the surface cover variable submatrix, where I is the Normalized Difference Vegetation Index (NDVI) and J is the land class code, which expresses the land use type through numerical mapping.

[0029] Preferably, for the historical precipitation samples of the pixels corresponding to the fused feature tensor, the skew distribution statistics are calculated, and based on the skew distribution statistics, the shape parameter k and scale parameter θ of the Gamma distribution are dynamically generated through a meta-learning strategy to obtain an adaptive Gamma loss function, including:

[0030] For each pixel position corresponding to the fused feature tensor, extracting measured precipitation values ​​at multiple historical time steps to construct a historical precipitation sample sequence;

[0031] Calculating statistics describing skewed distribution characteristics for each of the historical precipitation sample sequences; the statistics include: mean, variance, skewness coefficient, and kurtosis coefficient;

[0032] The skewness statistic is input into a parameter generation subnetwork; the parameter generation subnetwork is a pre-trained shallow neural network that regresses the shape parameter k and scale parameter θ of the Gamma distribution when a small number of samples are available;

[0033] The probability density function expression of the Gamma distribution is constructed using k and θ, and then the log-likelihood loss term between the precipitation estimate and the measured value is constructed;

[0034] The log-likelihood loss term and the mean square error loss are weighted according to a preset ratio to form the final adaptive Gamma loss function.

[0035] Preferably, the calculation formula of the adaptive Gamma loss function is:

[0036]

[0037] in, represents the adaptive Gamma loss function, which is used to measure the skewness error between the predicted precipitation value and the measured precipitation value under the Gamma distribution assumption; y represents the actual observed precipitation value, represents the precipitation estimate output by the model; k represents the shape parameter of the Gamma distribution dynamically generated by the meta-learning module based on the historical sample distribution, which controls the skewness of the distribution curve; θ represents the scale parameter of the Gamma distribution matched with k, which controls the expansion and contraction width of the distribution; Γ(k) represents the Gamma function, which normalizes the probability density function so that its integral is 1; λ represents the weight coefficient of the mean square error loss term, which is used to balance the likelihood loss and the traditional error.

[0038] Preferably, the fused feature tensor is input into a deep fusion network with a cross-modal attention mechanism, and the network weights, the shape parameter k and the scale parameter θ of the deep fusion network are jointly iteratively optimized using the adaptive Gamma loss function to obtain a converged fusion network model, including:

[0039] The constructed fusion feature tensor is fed as input into a deep fusion network equipped with a cross-modal attention mechanism. The deep fusion network consists of an encoder, a cross-channel attention module, and a decoder. The encoder is used to extract the joint feature representation of multi-source precipitation and geoscientific priors, and the decoder is used to restore the high-resolution precipitation field output. The cross-modal attention module is set between the encoding layer and the decoding layer of the network. The cross-modal attention module is used to dynamically adjust the fusion ratio of each modal feature by calculating the attention weights between different data sources to enhance the response strength and spatial consistency of the precipitation signal.

[0040] Inputting the precipitation estimate currently output by the deep fusion network and the measured precipitation sample into the adaptive Gamma loss function, wherein the adaptive Gamma loss function depends on the model prediction value, the measured value, the shape parameter k and the scale parameter θ of the Gamma distribution;

[0041] Constructing a joint optimization objective function, taking the adaptive Gamma loss function as the target, performing end-to-end backpropagation optimization on the weight parameters in the deep fusion network and k and θ in the parameter generation subnetwork, and iteratively updating all model parameters using an adaptive learning rate strategy;

[0042] When the adaptive Gamma loss function meets the preset convergence condition on the validation set or the number of iterations reaches the upper limit, the final parameters of the fusion network are output to form a converged fusion network model.

[0043] Preferably, obtaining a precipitation confidence interval based on the posterior distribution and forming a final precipitation product after terrain consistency check and spatial smoothing includes:

[0044] The parameters of the posterior distribution and Perform no less than 30 Monte Carlo random samplings to obtain a precipitation sample set;

[0045] The 5th percentile value and the 95th percentile value are calculated based on the precipitation sample set and used as the lower confidence limit P5 and the upper confidence limit P 95 , and [P5,P 95 ] is defined as the 90% confidence interval;

[0046] The grid elevation and two-dimensional terrain gradient field were extracted using the digital elevation model, and a terrain reference precipitation field was constructed according to the empirical dry adiabatic precipitation lapse rate.

[0047] Comparing the median estimate of the fusion network with the terrain reference precipitation field grid by grid, and marking the grid point as inconsistent with the terrain when the difference exceeds a preset threshold;

[0048] For the marked inconsistent topographic points, the weighted average of the grid points with similar elevations in the four-neighborhood window is used to replace them, where the weight is inversely proportional to the elevation difference and spatial distance, and the topographically corrected precipitation field is obtained.

[0049] Applying separable Gaussian convolution smoothing to the terrain-corrected precipitation field to obtain a smoothed precipitation field; the convolution kernel scale of the Gaussian convolution smoothing is adaptively set based on the historical correlation distance of the same resolution to ensure a balance between edge preservation and noise suppression;

[0050] Compare the smoothed precipitation field with the corresponding confidence interval [P5,P 95 ] are packaged together as the final precipitation product.

[0051] A multi-source precipitation intelligent fusion system integrating geoscientific priors and adaptive Gamma loss, including:

[0052] Multi-source observation construction unit, used to construct multi-source precipitation observation tensors in the target area;

[0053] A geoscience priori generation unit is configured to generate a geoscience priori feature matrix on the same grid as the multi-source precipitation observation tensor based on a digital elevation model, a long-term climate mean, surface dynamic parameters, and land use and cover parameters, and to concatenate the matrix with the multi-source precipitation observation tensor according to the channel dimension to obtain a fused feature tensor;

[0054] A gamma parameter learning unit is used to calculate a skew distribution statistic for the historical precipitation samples of the pixel corresponding to the fused feature tensor, and dynamically generate a shape parameter k and a scale parameter θ of the gamma distribution based on the skew distribution statistic through a meta-learning strategy to obtain an adaptive gamma loss function;

[0055] A deep fusion training unit, configured to input the fused feature tensor into a deep fusion network with a cross-modal attention mechanism, and jointly iteratively optimize the network weights, the shape parameter k, and the scale parameter θ of the deep fusion network using the adaptive Gamma loss function to obtain a converged fusion network model;

[0056] A fusion inference unit, configured to use the fusion network model to infer the fusion feature tensor of the estimated time step, and output a fusion precipitation estimate and a Gamma posterior distribution;

[0057] The uncertainty quantification and precipitation generation unit is used to obtain the precipitation confidence interval based on the posterior distribution, and form the final precipitation product after terrain consistency test and spatial smoothing.

[0058] According to the specific embodiments provided by the present invention, the present invention discloses the following technical effects:

[0059] The present invention effectively overcomes the problems of insufficient response to precipitation skew characteristics, lack of topographic and climate mechanism constraints, and weak generalization ability of extreme precipitation estimation in the existing technology by integrating geological priors with an intelligent modeling method of adaptive Gamma loss. Compared with the traditional fusion method based on the Gaussian hypothesis, the present invention not only introduces a dynamically adjustable Gamma distribution modeling mechanism in the training stage to more accurately fit the skewed distribution characteristics of medium-to-high intensity precipitation, but also significantly enhances the model's ability to express physical consistency in complex geomorphic areas (such as mountains, coastal areas, and monsoon transition zones) by introducing high-resolution topography, climate, and land cover prior information. At the same time, the present invention adopts a cross-modal attention mechanism to achieve deep fusion of multi-source precipitation data and geological priors, and combines posterior uncertainty estimation with terrain consistency verification in the output stage to ensure the comprehensive improvement of the fused precipitation product in terms of spatial continuity, physical rationality, and confidence assessment, which has important business application value and scientific research promotion significance. BRIEF DESCRIPTION OF THE DRAWINGS

[0060] In order to more clearly illustrate the embodiments of the present invention or the technical solutions in the prior art, the following briefly introduces the drawings required for use in the embodiments. Obviously, the drawings described below are only some embodiments of the present invention. For ordinary technicians in this field, other drawings can be obtained based on these drawings without paying any creative work.

[0061] Figure 1 A flow chart of a method provided by an embodiment of the present invention;

[0062] Figure 2 A schematic diagram of the system structure provided by an embodiment of the present invention. DETAILED DESCRIPTION

[0063] The following will clearly and completely describe the technical solutions in the embodiments of the present invention in conjunction with the accompanying drawings. Obviously, the described embodiments are only part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of the present invention.

[0064] The purpose of the present invention is to provide an intelligent fusion method and system for multi-source precipitation that integrates geoscientific priors and adaptive Gamma loss, which realizes the deep integration of skew modeling, adaptive optimization and geoscientific priors for multi-source precipitation, and significantly improves the accuracy and uncertainty characterization ability of precipitation estimation in complex areas.

[0065] In order to make the above-mentioned objects, features and advantages of the present invention more obvious and easy to understand, the present invention is further described in detail below with reference to the accompanying drawings and specific embodiments.

[0066] Figure 1 A flow chart of the method provided in the embodiment of the present invention is shown in FIG. Figure 1 As shown, the present invention provides a multi-source precipitation intelligent fusion method integrating geoscience priors and adaptive Gamma loss, comprising:

[0067] Step 100: Construct a multi-source precipitation observation tensor for the target area;

[0068] Step 200: Based on the digital elevation model, long-term climate mean, surface dynamic parameters, and land use and cover parameters, a geoscience priori feature matrix with the same grid as the multi-source precipitation observation tensor is generated, and the matrix is ​​concatenated with the multi-source precipitation observation tensor according to the channel dimension to obtain a fused feature tensor.

[0069] Step 300: Calculate the skew distribution statistics for the historical precipitation samples of the pixels corresponding to the fused feature tensor, and dynamically generate the shape parameter k and scale parameter θ of the Gamma distribution based on the skew distribution statistics through a meta-learning strategy to obtain an adaptive Gamma loss function;

[0070] Step 400: Input the fused feature tensor into a deep fusion network with a cross-modal attention mechanism, and use an adaptive Gamma loss function to jointly iteratively optimize the network weights, shape parameter k, and scale parameter θ of the deep fusion network to obtain a converged fusion network model;

[0071] Step 500: Use the fusion network model to infer the fused feature tensor of the estimated time step, and output the fused precipitation estimate and the Gamma posterior distribution;

[0072] Step 600: Obtain the precipitation confidence interval based on the posterior distribution, and form the final precipitation product after terrain consistency test and spatial smoothing.

[0073] Specifically, step 100 of this embodiment first sets a unified time base for the target area. It then imports hourly data from satellite precipitation swaths, ground-based radar volume scans, and ground-based rain gauges, and records the respective observation times. To account for time differences, this embodiment synchronizes all observations to the same timestamp through interpolation and flattening, limiting temporal drift to no more than half a time step. If a rain gauge is missing, it uses estimates from neighboring stations or geostationary satellites for simple interpolation. The original and interpolated data are each labeled with their source to facilitate subsequent reliability verification. After time alignment, this embodiment uses a unified map coordinate system to project satellite, radar, and rain gauge data onto the highest-resolution radar grid (e.g., 0.01° × 0.01°). Satellite data is interpolated to match grid points using multi-layer interpolation, while rain gauge data is spread to surrounding grid points using a weighted average. A digital elevation model is used to generate radar obscuration and satellite line-of-sight obscuration masks, marking obscured pixels as missing. For example, this embodiment uses a unified time base of either the hour or a 5-minute time step.

[0074] After completing spatiotemporal registration, this implementation aggregates three types of precipitation observations at each grid point, constructing a three-channel observation tensor in the order of satellite, radar, and rain gauge. A quality indicator matrix is ​​also generated simultaneously, indicating whether each channel's data is raw, interpolated, or missing. This tensor-quality matrix combination provides complete multi-source information and explicit data credibility for the deep fusion stage, enabling the model to remain robust in complex terrain or areas with sparsely populated sites.

[0075] Optionally, in step 200 of this embodiment, seven types of geoscience data covering the target area are first obtained, including a digital elevation model (DEM), a multi-year average precipitation field, a multi-year average temperature field, a multi-year average surface wind speed field, a multi-year average surface solar radiation field, a land use classification map, and a surface cover type map. To ensure spatial consistency, all geoscience data use the same map projection and resolution as the multi-source precipitation observation tensor: bilinear interpolation is performed on continuous raster data (precipitation, temperature, wind speed, radiation), and nearest neighbor interpolation is performed on discrete classification data (land use, cover type), and then grid alignment is performed. Subsequently, four terrain indicators of altitude, slope, aspect, and distance from the sea are calculated based on the DEM; surface category codes are extracted based on the land use and cover type map, and the normalized difference vegetation index (NDVI) is calculated using the red-near infrared band reflectance; four climate variable layers are generated based on the multi-year average climate field, thereby forming three types of prior information: terrain, climate, and surface cover.

[0076] During the data organization phase, the three types of prior information are spliced ​​into a fixed-format geoscience prior feature matrix according to the channel dimension. The terrain variable submatrix contains sinusoidal codes for elevation, slope, aspect, and distance from the sea; the climate variable submatrix contains four layers: multi-year average precipitation, temperature, wind speed, and solar radiation; and the land cover variable submatrix contains two layers: NDVI and numerical surface category codes. The spatial size of this matrix is ​​exactly the same as the multi-source precipitation observation tensor, with the height and width corresponding to the number of rows and columns of the target area grid, respectively, and the number of channels equal to the sum of the ten prior variables. Through this hierarchical and ordered channel design, the network can explicitly distinguish different types of physical priors and specifically learn the mapping relationship between each channel and precipitation distribution during training.

[0077] Finally, the geoscientific prior feature matrix is ​​concatenated with the multi-source precipitation observation tensor in a one-to-one correspondence along the channel dimension, resulting in a fused feature tensor that serves as the input to the deep fusion network. This concatenation process ensures that each pixel carries both the three-source precipitation estimate and ten types of geoscientific prior information, providing complete, high-resolution joint feature support for the subsequent cross-modal attention mechanism. It also avoids information loss caused by the spatiotemporal misalignment between prior and observed data, fundamentally improving the physical consistency and prediction accuracy of extreme precipitation fields.

[0078] In this embodiment, to construct an adaptive Gamma loss function that can reflect the skewed characteristics of precipitation, we first extract the measured precipitation values ​​of each pixel position in the fused feature tensor at multiple historical time steps to form a historical precipitation sample in the form of a time series. The number of samples can be flexibly set based on the availability of observational data, such as selecting daily or hourly measured values ​​for the past year. Subsequently, a statistical analysis is performed on each historical precipitation sample sequence to calculate its mean, variance, skewness coefficient, and kurtosis coefficient, four types of statistics describing the skewed distribution morphology. These statistics are used to characterize characteristics such as the concentration, dispersion, distribution asymmetry, and tail thickness of precipitation at that location.

[0079] After obtaining the above-mentioned skew statistics, this embodiment sends them as input to a pre-trained parameter generation subnetwork. This subnetwork is a shallow neural network with a simple structure and has the ability to model nonlinear mapping relationships under small sample conditions. Through offline training, the subnetwork can map statistical features into corresponding Gamma distribution parameters, namely shape parameter k and scale parameter θ, and quickly generate the Gamma distribution expression corresponding to each pixel position during the model operation. The introduction of this network solves the limitations brought about by traditional parameter fixing or manual setting, so that each spatial position can adaptively obtain a suitable probability distribution structure under different climate types or precipitation backgrounds.

[0080] This embodiment further constructs a corresponding Gamma distribution probability density function based on the k and θ generated above, and uses this as a basis to construct the log-likelihood loss term between the precipitation estimate and the measured value. In order to balance numerical accuracy and probability consistency, this likelihood loss term is weightedly combined with the traditional mean square error loss to form the final adaptive Gamma loss function, which is used as the objective function for deep fusion network training. This loss function can dynamically adapt to the precipitation distribution characteristics of different regions and time periods, while maintaining regression stability while strengthening the model's response to extreme tail events, thereby improving the overall precipitation estimation accuracy and uncertainty modeling capabilities.

[0081] Furthermore, in this embodiment, the deviation between the predicted precipitation value and the measured precipitation value is measured by an adaptive Gamma loss function, which consists of two parts: one is the log-likelihood term constructed based on the Gamma distribution, which is used to reflect the degree of fit of the predicted value under the target probability distribution; the other is the traditional mean square error term, which is used to constrain the numerical accuracy of the prediction result. In the loss function, the actual observed precipitation value is directly obtained from the historical observation data, and the precipitation estimate output by the model comes from the forward reasoning result of the deep fusion network. The shape parameter and scale parameter are dynamically generated by the meta-learning module. During the training phase, the module performs regression reasoning based on the statistical characteristics of the historical precipitation sequence of each pixel (including mean, variance, skewness and kurtosis) to ensure that it matches the precipitation distribution morphology of the current spatial location. The Gamma function is used to normalize the probability density expression. It is a fixed mathematical function term and does not require training to determine. The weight coefficient of the mean square error loss is set by hyperparameters before training and is tuned according to the performance of the model on the validation set. The purpose is to strike a balance between the distribution consistency of the log-likelihood and the numerical accuracy of the mean square error. The above parameters have clear sources and are reproducible, ensuring that the loss function is fully open and feasible.

[0082] Furthermore, in this embodiment, the constructed fusion feature tensor is first input into a deep fusion network with a cross-modal attention mechanism. The fusion network consists of three parts: an encoder, a cross-channel attention module, and a decoder. The encoder is responsible for extracting the joint spatial expression of multi-source precipitation information and geological prior information contained in the fusion feature tensor, and the decoder restores the encoded high-dimensional feature map to a high-resolution precipitation estimation map corresponding to the input grid. A cross-modal attention module is set between the encoder and the decoder. The module calculates the attention weights of feature channels extracted from different data sources (such as satellites, radars, rain gauges, terrain, climate, etc.), and dynamically controls the contribution ratio of each channel to the fusion result, thereby enhancing the expression ability of key modes in heavy precipitation areas and improving the spatial consistency and physical interpretability of precipitation estimation results.

[0083] After the fusion network completes forward inference, the output precipitation estimates and the corresponding measured precipitation samples are fed into an adaptive Gamma loss function for loss evaluation. This loss function relies not only on the difference between the model's predicted and measured values, but also on the shape and scale parameters of the Gamma distribution provided by the parameter generation subnetwork, dynamically adjusting the tolerance for skewness errors. During the loss calculation process, the network simultaneously optimizes the weight parameters of the backbone network and the internal weights of the parameter generation subnetwork through a backpropagation mechanism, achieving simultaneous learning of precipitation distribution morphology and spatial characteristics, ensuring the generalization of the training process to different regions and precipitation types.

[0084] This embodiment adopts a joint optimization strategy, takes the above-mentioned adaptive Gamma loss function as the only objective function, and performs end-to-end joint training on the deep fusion network backbone and the Gamma parameter generation sub-network. An adaptive learning rate mechanism is introduced during the training process to adjust the learning step size according to the loss reduction rate to improve training efficiency and convergence stability. When the loss function reaches the preset convergence threshold on the validation set, or the number of training rounds reaches the upper limit, all the parameters of the current network are output to form a final converged fusion network model. This training process ensures that the model can autonomously learn the optimal multi-source fusion mechanism and skew modeling capabilities without relying on fixed distribution assumptions, and has a good foundation for practical deployment.

[0085] In step 500 of this embodiment, after the fusion network model completes training and converges, the fused feature tensor of the time step to be estimated is fed into the network as input for forward reasoning. The fusion network uses the learned multi-source precipitation feature expression capabilities and the geoscientific prior condition response mechanism to output the fused precipitation estimate for each grid point in the corresponding time step, and simultaneously outputs the Gamma distribution shape parameters and scale parameters predicted by the parameter generation subnetwork. Based on this set of parameters, a Gamma posterior distribution can be constructed for each pixel to describe the model's expected probability distribution of precipitation at that location, thereby providing not only a point value estimate but also a distribution basis for subsequent uncertainty quantification and confidence interval assessment. This process achieves spatially continuous, multi-source combined, and physically constrained precipitation estimation output with high accuracy and interpretability.

[0086] Furthermore, in this embodiment, to quantify the uncertainty of precipitation estimation results, Monte Carlo random sampling is first performed at least 30 times for each pixel based on the Gamma posterior distribution parameters obtained by fusion network inference to generate a precipitation sample set. Subsequently, the 5th and 95th percentile values ​​of the sample set corresponding to each pixel are calculated and used as the lower and upper confidence limits of the pixel, respectively, to form a 90% confidence interval range. In addition, the median value of each pixel sample set is extracted as a reference point value for subsequent terrain consistency verification. This confidence interval estimation process not only provides the upper and lower limits of the point value prediction results, but also provides a clear quantitative basis for users to understand the uncertainty of the model output.

[0087] After the confidence interval calculation is completed, this embodiment extracts the altitude of each pixel and its two-dimensional terrain gradient information based on the digital elevation model, and constructs a reference precipitation field that matches the terrain based on the empirical dry adiabatic precipitation decrease rate. The fusion network median estimation result is compared with the reference field pixel by pixel. If the difference between the two exceeds the set threshold, the pixel is marked as a "terrain inconsistent" area. For these areas, this embodiment uses the weighted average method of grid points with similar altitudes in the four-neighborhood window to replace and correct their precipitation values. The weight calculation takes into account both spatial distance and altitude differences, thereby obtaining a precipitation field after terrain consistency correction, ensuring that the precipitation value is spatially consistent with the terrain logic.

[0088] After terrain correction is complete, this embodiment performs a separable Gaussian convolution smoothing operation on the terrain-corrected precipitation field to further enhance the spatial smoothness and structural integrity of the precipitation estimate. The convolution kernel scale for this smoothing operation is automatically set based on the spatial correlation distance of the historical precipitation field, suppressing local noise while maintaining the continuity and authenticity of the edge structure. Finally, the smoothed precipitation field is packaged together with the aforementioned 90% confidence interval to form the final precipitation product output, which includes the precipitation estimate for each pixel, upper and lower confidence limits, and terrain consistency information. This provides complete and reliable fused precipitation data support for downstream applications such as meteorology and hydrology.

[0089] Corresponding to the above method, such as Figure 2 As shown, this embodiment also provides a multi-source precipitation intelligent fusion system that integrates geoscience priors and adaptive Gamma loss, including:

[0090] Multi-source observation construction unit, used to construct multi-source precipitation observation tensors in the target area;

[0091] A geoscience priori generation unit is configured to generate a geoscience priori feature matrix on the same grid as the multi-source precipitation observation tensor based on a digital elevation model, a long-term climate mean, surface dynamic parameters, and land use and cover parameters, and to concatenate the matrix with the multi-source precipitation observation tensor according to the channel dimension to obtain a fused feature tensor;

[0092] A gamma parameter learning unit is used to calculate a skew distribution statistic for the historical precipitation samples of the pixel corresponding to the fused feature tensor, and dynamically generate a shape parameter k and a scale parameter θ of the gamma distribution based on the skew distribution statistic through a meta-learning strategy to obtain an adaptive gamma loss function;

[0093] A deep fusion training unit, configured to input the fused feature tensor into a deep fusion network with a cross-modal attention mechanism, and jointly iteratively optimize the network weights, the shape parameter k, and the scale parameter θ of the deep fusion network using the adaptive Gamma loss function to obtain a converged fusion network model;

[0094] A fusion inference unit, configured to use the fusion network model to infer the fusion feature tensor of the estimated time step, and output a fusion precipitation estimate and a Gamma posterior distribution;

[0095] The uncertainty quantification and precipitation generation unit is used to obtain the precipitation confidence interval based on the posterior distribution, and form the final precipitation product after terrain consistency test and spatial smoothing.

[0096] The beneficial effects of the present invention are as follows:

[0097] (1) The present invention introduces a deep fusion mechanism of geological prior features and multi-source precipitation observation information, making full use of basic data such as digital elevation model, climate mean field, surface dynamic parameters and land cover type, thereby improving the model's ability to depict the precipitation formation mechanism in different landforms and climate zones, and achieving enhanced spatial adaptability of precipitation estimation in complex areas such as plateaus, mountainous areas, and coastal areas.

[0098] (2) Based on the skewed statistical characteristics of historical precipitation samples, the present invention dynamically generates the shape parameters and scale parameters of the Gamma distribution through a meta-learning strategy, and introduces a distribution adaptation mechanism in the loss function construction stage, breaking through the traditional method's reliance on symmetric distribution or fixed priors, and significantly improving the model's response accuracy to extreme precipitation events and tail skew errors.

[0099] (3) The cross-modal attention deep fusion network proposed in this paper can automatically perceive the quality differences and physical correlations of various data sources, dynamically adjust the fusion weights of different modal features, effectively alleviate the fusion errors caused by redundancy, inconsistency or missing multi-source data, and improve the robustness and generalization ability of precipitation estimation.

[0100] (4) The present invention combines mechanisms such as posterior distribution sampling, terrain consistency verification, and adaptive spatial smoothing to not only provide precipitation estimates, but also output confidence intervals and terrain rationality assessment results, achieving quantitative uncertainty characterization and physical consistency assurance of the results, providing more complete data support for high-confidence meteorological and hydrological applications.

[0101] The various embodiments in this specification are described in a progressive manner, with each embodiment focusing on the differences from other embodiments. Reference can be made to the common and similar parts between the various embodiments. For the systems disclosed in the embodiments, since they correspond to the methods disclosed in the embodiments, the description is relatively simple, and the relevant parts can be referred to the method description.

[0102] This document uses specific examples to illustrate the principles and implementation methods of the present invention. The above examples are only intended to help understand the method and core concept of the present invention. At the same time, those skilled in the art will find that the specific implementation methods and application scopes may vary based on the concept of the present invention. In summary, the contents of this specification should not be construed as limiting the present invention.

Claims

1. A multi-source precipitation intelligent fusion method integrating geoscience priors and adaptive Gamma loss, characterized by: include: Construct a multi-source precipitation observation tensor for the target area; Based on the digital elevation model, long-term climate mean, surface dynamic parameters, land use and cover parameters, a geoscience prior feature matrix with the same grid as the multi-source precipitation observation tensor is generated, and the matrix is ​​spliced ​​with the multi-source precipitation observation tensor according to the channel dimension to obtain a fused feature tensor; For the historical precipitation samples of the pixels corresponding to the fused feature tensor, a skew distribution statistic is calculated, and based on the skew distribution statistic, a shape parameter k and a scale parameter θ of the Gamma distribution are dynamically generated through a meta-learning strategy to obtain an adaptive Gamma loss function; Inputting the fused feature tensor into a deep fusion network with a cross-modal attention mechanism, and using the adaptive Gamma loss function to jointly iteratively optimize the network weights, the shape parameter k, and the scale parameter θ of the deep fusion network to obtain a converged fusion network model; The fusion network model is used to infer the fusion feature tensor of the estimated time step, and the fusion precipitation estimation value and the Gamma posterior distribution are output; The precipitation confidence interval is obtained based on the posterior distribution, and the final precipitation product is formed after terrain consistency test and spatial smoothing.

2. The multi-source precipitation intelligent fusion method integrating geoscience priors and adaptive Gamma loss according to claim 1 is characterized in that: Construct a multi-source precipitation observation tensor, including: Collecting satellite precipitation inversion data, ground-based radar precipitation estimation data and ground-based rain gauge observation data over the target area under a unified time reference; The satellite precipitation inversion data, the ground-based radar precipitation estimation data and the ground rain gauge observation data are subjected to resolution uniformization and pixel alignment to obtain a multi-source precipitation observation tensor.

3. The multi-source precipitation intelligent fusion method integrating geoscience priors and adaptive Gamma loss according to claim 1 is characterized in that: Based on the digital elevation model, long-term climate mean, surface dynamic parameters, land use and cover parameters, a geoscience priori feature matrix with the same grid as the multi-source precipitation observation tensor is generated, and then spliced ​​with the multi-source precipitation observation tensor according to the channel dimension to obtain a fused feature tensor, including: Acquiring geoscientific data covering the target area; the geoscientific data includes a digital elevation model, multi-year average precipitation data, multi-year average temperature data, surface wind speed data, surface solar radiation data, a land use classification map, and a surface cover type map; The geoscientific data are spatially resampled by bilinear interpolation to make the resolution of the geoscientific data consistent with the multi-source precipitation observation tensor, and grid alignment is performed using a unified geographic coordinate system; Calculating terrain indicators based on the digital elevation models respectively; Extracting land cover variables based on the land use classification map and the land cover type map; the land cover variables include land cover category codes and normalized difference vegetation indexes; constructing four corresponding surface climate variables according to the multi-year average precipitation data, the multi-year average temperature data, the surface wind speed data, and the surface solar radiation data; splicing the terrain index, the surface cover variable and the surface classification variable into a fixed-format geoscience prior feature matrix in the channel dimension; The geoscience prior feature matrix and the multi-source precipitation observation tensor are spliced ​​pixel by pixel in the channel dimension to obtain a fused feature tensor.

4. The multi-source precipitation intelligent fusion method integrating geoscience priors and adaptive Gamma loss according to claim 3 is characterized in that: The terrain indicators include altitude, slope, aspect and distance from the sea.

5. The multi-source precipitation intelligent fusion method integrating geoscience priors and adaptive Gamma loss according to claim 4 is characterized in that: The expression of the geoscience priori characteristic matrix is: G=[E,C,L] Among them, G represents the geoscience prior feature matrix used for network input after fusion, which is formed by splicing according to the channel dimension and has the dimension of H×W×D g , where H and W are the height and width of the grid image of the target area, respectively, D g is the number of geoscience priori channels; E = [A, B, C, D] represents the terrain variable submatrix, where A is the altitude, B is the slope, C is the sine value encoding after slope aspect conversion, and D is the distance from the sea; C = [E, F, G, H] represents the climate variable submatrix, where E is the multi-year average precipitation data, F is the multi-year average temperature data, G is the multi-year average surface wind speed data, and H is the multi-year average surface solar radiation data; L = [I, J] represents the surface cover variable submatrix, where I is the Normalized Difference Vegetation Index (NDVI) and J is the land class code, which expresses the land use type through numerical mapping.

6. The multi-source precipitation intelligent fusion method integrating geoscience priors and adaptive Gamma loss according to claim 1 is characterized in that: For the historical precipitation samples of the pixels corresponding to the fused feature tensor, the skew distribution statistics are calculated, and based on the skew distribution statistics, the shape parameter k and scale parameter θ of the Gamma distribution are dynamically generated through a meta-learning strategy to obtain an adaptive Gamma loss function, including: For each pixel position corresponding to the fused feature tensor, extracting measured precipitation values ​​at multiple historical time steps to construct a historical precipitation sample sequence; Calculating statistics describing skewed distribution characteristics for each of the historical precipitation sample sequences; the statistics include: mean, variance, skewness coefficient, and kurtosis coefficient; The skewness statistic is input into a parameter generation subnetwork; the parameter generation subnetwork is a pre-trained shallow neural network that regresses the shape parameter k and scale parameter θ of the Gamma distribution when a small number of samples are available; The probability density function expression of the Gamma distribution is constructed using k and θ, and then the log-likelihood loss term between the precipitation estimate and the measured value is constructed; The log-likelihood loss term and the mean square error loss are weighted according to a preset ratio to form the final adaptive Gamma loss function.

7. The multi-source precipitation intelligent fusion method integrating geoscience priors and adaptive Gamma loss according to claim 1 is characterized in that: The calculation formula of the adaptive Gamma loss function is: in, represents the adaptive Gamma loss function, which is used to measure the skewness error between the predicted precipitation value and the measured precipitation value under the Gamma distribution assumption; y represents the actual observed precipitation value, represents the precipitation estimate output by the model; k represents the shape parameter of the Gamma distribution dynamically generated by the meta-learning module based on the historical sample distribution, which controls the skewness of the distribution curve; θ represents the scale parameter of the Gamma distribution matched with k, which controls the expansion and contraction width of the distribution; Γ(k) represents the Gamma function, which normalizes the probability density function so that its integral is 1; λ represents the weight coefficient of the mean square error loss term, which is used to balance the likelihood loss and the traditional error.

8. The multi-source precipitation intelligent fusion method integrating geoscience priors and adaptive Gamma loss according to claim 1 is characterized in that: Inputting the fused feature tensor into a deep fusion network with a cross-modal attention mechanism, and using the adaptive Gamma loss function to jointly iteratively optimize the network weights, the shape parameter k, and the scale parameter θ of the deep fusion network to obtain a converged fusion network model, including: The constructed fusion feature tensor is fed as input into a deep fusion network equipped with a cross-modal attention mechanism. The deep fusion network consists of an encoder, a cross-channel attention module, and a decoder. The encoder is used to extract the joint feature representation of multi-source precipitation and geoscientific priors, and the decoder is used to restore the high-resolution precipitation field output. The cross-modal attention module is set between the encoding layer and the decoding layer of the network. The cross-modal attention module is used to dynamically adjust the fusion ratio of each modal feature by calculating the attention weights between different data sources to enhance the response strength and spatial consistency of the precipitation signal. Inputting the precipitation estimate currently output by the deep fusion network and the measured precipitation sample into the adaptive Gamma loss function, wherein the adaptive Gamma loss function depends on the model prediction value, the measured value, the shape parameter k and the scale parameter θ of the Gamma distribution; Constructing a joint optimization objective function, taking the adaptive Gamma loss function as the target, performing end-to-end backpropagation optimization on the weight parameters in the deep fusion network and k and θ in the parameter generation subnetwork, and iteratively updating all model parameters using an adaptive learning rate strategy; When the adaptive Gamma loss function meets the preset convergence condition on the validation set or the number of iterations reaches the upper limit, the final parameters of the fusion network are output to form a converged fusion network model.

9. The multi-source precipitation intelligent fusion method integrating geoscience priors and adaptive Gamma loss according to claim 1 is characterized in that: The precipitation confidence interval is obtained based on the posterior distribution, and the final precipitation product is formed after terrain consistency test and spatial smoothing, including: The parameters of the posterior distribution and Perform no less than 30 Monte Carlo random samplings to obtain a precipitation sample set; The 5th percentile value and the 95th percentile value are calculated based on the precipitation sample set and used as the lower confidence limit P5 and the upper confidence limit P 95 , and [P5,P 95 ] is defined as the 90% confidence interval; The grid elevation and two-dimensional terrain gradient field were extracted using the digital elevation model, and a terrain reference precipitation field was constructed according to the empirical dry adiabatic precipitation lapse rate. Comparing the median estimate of the fusion network with the terrain reference precipitation field grid by grid, and marking the grid point as inconsistent with the terrain when the difference exceeds a preset threshold; For the marked inconsistent topographic points, the weighted average of the grid points with similar elevations in the four-neighborhood window is used to replace them, where the weight is inversely proportional to the elevation difference and spatial distance, and the topographically corrected precipitation field is obtained. Applying separable Gaussian convolution smoothing to the terrain-corrected precipitation field to obtain a smoothed precipitation field; the convolution kernel scale of the Gaussian convolution smoothing is adaptively set based on the historical correlation distance of the same resolution to ensure a balance between edge preservation and noise suppression; Compare the smoothed precipitation field with the corresponding confidence interval [P5,P 95 ] are packaged together as the final precipitation product.

10. A multi-source precipitation intelligent fusion system integrating geoscience priors and adaptive Gamma loss, characterized by: include: Multi-source observation construction unit, used to construct multi-source precipitation observation tensors in the target area; A geoscience priori generation unit is configured to generate a geoscience priori feature matrix on the same grid as the multi-source precipitation observation tensor based on a digital elevation model, a long-term climate mean, surface dynamic parameters, and land use and cover parameters, and to concatenate the matrix with the multi-source precipitation observation tensor according to the channel dimension to obtain a fused feature tensor; A gamma parameter learning unit is used to calculate a skew distribution statistic for the historical precipitation samples of the pixel corresponding to the fused feature tensor, and dynamically generate a shape parameter k and a scale parameter θ of the gamma distribution based on the skew distribution statistic through a meta-learning strategy to obtain an adaptive gamma loss function; A deep fusion training unit, configured to input the fused feature tensor into a deep fusion network with a cross-modal attention mechanism, and jointly iteratively optimize the network weights, the shape parameter k, and the scale parameter θ of the deep fusion network using the adaptive Gamma loss function to obtain a converged fusion network model; A fusion inference unit, configured to use the fusion network model to infer the fusion feature tensor of the estimated time step, and output a fusion precipitation estimate and a Gamma posterior distribution; The uncertainty quantification and precipitation generation unit is used to obtain the precipitation confidence interval based on the posterior distribution, and form the final precipitation product after terrain consistency test and spatial smoothing.

Images