CN-122019993-A - Groundwater vulnerability spatial interpolation method based on deep learning

CN122019993ACN 122019993 ACN122019993 ACN 122019993ACN-122019993-A

Abstract

The invention discloses a groundwater vulnerability spatial interpolation method based on deep learning, which relates to the technical field of groundwater environment protection and evaluation, and comprises the steps of collecting and preprocessing multi-source data; the method comprises the steps of constructing a composite space weight matrix based on a space topological relation and a distance attenuation principle, constructing a deep learning model, fusing the weight matrix and hydrogeologic parameters as input features, training by adopting a composite loss function, dynamically adjusting the weight parameters by Bayesian optimization and retraining the model by utilizing prediction error distribution, optimizing vulnerability partition boundaries by combining morphological operation and filtering treatment, and finally generating a vulnerability partition map in a grid and vector format according to model prediction results.

Inventors

DU CAIDONG
Huang Guiren
SHANG JIANGTAO
BI TINGTAO
ZHAO XIANGHUA
MEI XIANGYANG

Assignees

云南省生态环境科学研究院

Dates

Publication Date: 20260512
Application Date: 20260108

Claims (10)

1. The underground water vulnerability spatial interpolation method based on deep learning is characterized by comprising the following steps of: Firstly, collecting multi-source data of a research area, wherein the multi-source data comprises water quality monitoring data, geological parameters, land utilization information and meteorological data, and carrying out space coordinate unification, resolution adjustment, missing value filling and standardization treatment on the collected multi-source data; Step two, constructing a GIS space weight matrix, namely defining an adjacent relation based on a space topological relation of a research area, calculating a distance attenuation weight by combining a distance attenuation principle, and fusing the adjacent relation and the distance attenuation weight to construct a composite space weight matrix; Constructing a deep learning model based on a U-Net architecture, fusing the composite space weight matrix obtained in the step two with the multi-source data obtained through pretreatment in the step one to form input features, and training the deep learning model by using the input features, wherein a loss function adopted by training is a composite loss function, and the composite loss function comprises a cross entropy loss item, a space structure similarity loss item and a space autocorrelation loss item; utilizing the deep learning model which is trained in the third step to conduct vulnerability prediction, calculating errors between a prediction result and actual monitoring values, adopting a Bayesian optimization method to dynamically adjust parameters of the composite space weight matrix based on error distribution, re-fusing input features by using the adjusted space weight matrix, re-training the deep learning model, conducting morphological operation and smooth filtering processing on the prediction result of the deep learning model after re-training, and optimizing vulnerability partition boundaries; And fifthly, generating a vulnerability partition map, namely calculating vulnerability indexes of all areas according to the optimized deep learning model prediction result, grading the vulnerability indexes according to a preset grading threshold value, generating a grid-format vulnerability partition map, carrying out post-processing on the grid partition map, and converting the grid partition map into a vector boundary map.
2. The method of spatial interpolation for groundwater vulnerability based on deep learning according to claim 1, wherein in the first step, the data acquisition and preprocessing further comprises: carrying out space alignment and resampling on the multi-source data to a uniform resolution; Filling missing values in the data by adopting an interpolation algorithm; Grading and standardizing the hydrogeologic parameters; the training dataset is augmented by a data enhancement technique that applies random perturbations.
3. The method of spatial interpolation for groundwater vulnerability based on deep learning according to claim 1, wherein in the second step, the constructing a composite spatial weight matrix further comprises: Defining an adjacency of the space unit based on a Queen adjacency or Rook adjacency rule; Calculating the distance attenuation weight between the space units by adopting a distance inverse power function or a Gaussian kernel function; performing linear weighted fusion on the adjacency relation matrix and the distance attenuation weight matrix to generate the composite space weight matrix; and carrying out space autocorrelation verification on the generated composite space weight matrix by calculating the Morand index.
4. The method according to claim 1, wherein in the third step, the deep learning model is a U-Net model of an encoder-decoder structure; The input features are constructed by stitching or multiplying each element of the composite spatial weight matrix with a hydrogeologic parameter vector of a corresponding spatial unit; the hydrogeologic parameters include at least scoring values based on depth groundwater level, net replenishment, aquifer media, soil media, terrain, air-in-air media, hydraulic conductivity, and land utilization type of DRASTIC-LU framework.
5. The depth learning based groundwater vulnerability spatial interpolation method of claim 1, wherein the composite loss function The expression of (2) is: Wherein, the Represents a cross entropy loss term for measuring the difference between the model predictive classification result and the real label, Representing a spatial structure similarity penalty term for constraining spatial structure fidelity of the predicted vulnerability map, Representing a spatial autocorrelation loss term for ensuring that the prediction results conform to the spatial dependence described by the first law of geography, 、、 The weight coefficients of the losses are respectively the weight coefficients of the losses, and the weight coefficients are all adjustable super parameters which are larger than zero.
6. The method of spatial interpolation of groundwater vulnerability based on deep learning according to claim 1, wherein in the fourth step, the calculating the error between the predicted result and the actual monitored value further comprises: calculating the Root Mean Square Error (RMSE) of each space unit; generating a spatial distribution thermodynamic diagram of the prediction error in the GIS platform; performing spatial cluster analysis on error data by adopting a clustering algorithm based on a maximum-minimum distance, and identifying a spatial distribution mode of errors; The distance judgment threshold value parameter in the clustering algorithm based on the maximum-minimum distance is recorded as 。
7. The method for spatial interpolation of groundwater vulnerability based on deep learning according to claim 1, wherein in the fourth step, the parameters of the composite spatial weight matrix are dynamically adjusted by using a bayesian optimization method based on error distribution, specifically: taking an attenuation coefficient of a distance attenuation function related in the construction process of the composite space weight matrix and a fusion proportionality coefficient of the adjacent weight and the distance weight as parameters to be optimized; and iteratively searching the optimal combination of the parameters to be optimized by using a Bayesian optimization framework by taking the overall prediction error of the model on the verification set as an objective function.
8. The method of spatial interpolation of groundwater vulnerability based on deep learning according to claim 1, wherein in the fourth step, morphological operation and smoothing filtering processing are performed on the prediction result, further comprising: smoothing the vulnerability index grid graph by adopting a Gaussian filtering algorithm to inhibit local noise; processing the vulnerability partition boundary by using an open operation and a close operation in mathematical morphology; Setting an area threshold value, and removing the vulnerable plaque of which the area of the connected area is smaller than the threshold value from the partition map; And introducing a space consistency constraint term into a loss function of the model training stage to punish unreasonable vulnerability level mutation between adjacent space units.
9. The method of claim 1, wherein in the fourth step, the dynamic calibration and boundary optimization process is an iterative loop process: in each iteration, sequentially executing the steps of error analysis, space weight parameter adjustment, model retraining and boundary optimization; the iteration termination condition is that the performance evaluation index of the model on the independent test set reaches a preset standard or the iteration number reaches an upper limit; And storing the updated model and the corresponding space weight parameters after each iteration.
10. The method of spatial interpolation of vulnerability of groundwater based on deep learning according to claim 1, wherein in the fifth step, the generating a vulnerability partition map in grid format further comprises: calculating the vulnerability index DI of each pixel according to the output logic value of the optimized model; dividing the continuously distributed vulnerability indexes into high, medium and low vulnerability levels by adopting a natural breakpoint method, an equidistant method or a quantile method; Color is given to different grades and a vulnerability grading thematic map is generated by rendering; and extracting vector boundaries of each level of vulnerability area from the grid partition map through a grid-to-vector algorithm.

Description

Groundwater vulnerability spatial interpolation method based on deep learning Technical Field The invention relates to the technical field of groundwater environment protection and evaluation, in particular to a groundwater vulnerability spatial interpolation method based on deep learning. Background The evaluation of the vulnerability of the underground water is a key technology for identifying the potential risk degree of pollution of the underground water system, and is an important scientific basis for protecting and managing the underground water resources. Traditional evaluation methods, such as DRASTIC models, have been widely used to generate vulnerability maps by selecting hydrogeologic parameters for weighted superposition. However, with the increase of the requirements on the evaluation precision and the space detail, the limitations of the traditional method in the aspects of fixed weight distribution, rough boundary characterization and the like are increasingly highlighted. In the prior art, researchers have tried various approaches to improve the objectivity and accuracy of the evaluation. For example, some prior art techniques determine evaluation index weights by using analytic hierarchy process and perform spatial overlay analysis in conjunction with GIS. Although a more scientific weight determination mode is introduced, the weight of the method is fixed in the whole area once determined, and the spatial heterogeneity is not reflected. Other prior art techniques propose to utilize machine learning algorithms such as random forests to predict by learning the relationship between the monitored data and various environmental factors. The method improves the nonlinear fitting capability of the model, but is still an attribute-driven prediction in nature, and cannot be explicitly integrated into and optimize the geographic dependence and interaction relation between space units, so that an effective modeling and post-processing mechanism is lacking for the spatial continuity and smoothness of the vulnerability partition boundary. Therefore, the prior art has the common defect that the model cannot be optimally adjusted according to the space-time distribution characteristics of the prediction errors in the training process because the adaptive dynamic calibration of the space weight parameters cannot be realized in the conventional index superposition model based on fixed weights or the currently introduced machine learning prediction method. The defect causes the final vulnerability partition map to have a lifting space in the aspect of space consistency and boundary rationality, and is difficult to meet the requirement of high-precision underground water resource fine management. In summary, the existing groundwater vulnerability assessment method still lacks an effective solution in terms of how to deeply integrate the spatial geographic constraint and the machine learning model and construct a closed-loop assessment system capable of dynamically calibrating spatial weights and iteratively optimizing partition boundaries. Disclosure of Invention The invention aims to make up the defects of the prior art, provides a groundwater vulnerability spatial interpolation method based on deep learning, and can realize self-adaptive learning of spatial dependency and fine characterization of the vulnerability partition boundary by constructing a spatial weight matrix capable of being dynamically calibrated and deeply integrating the spatial weight matrix into a model training and optimizing process, thereby making up the defects of the prior method in terms of spatial heterogeneity modeling and boundary optimization. The invention provides a groundwater vulnerability spatial interpolation method based on deep learning, which aims to solve the technical problems and comprises the following steps: Firstly, collecting multi-source data of a research area, wherein the multi-source data comprises water quality monitoring data, geological parameters, land utilization information and meteorological data, and carrying out space coordinate unification, resolution adjustment, missing value filling and standardization treatment on the collected multi-source data; Step two, constructing a GIS space weight matrix, namely defining an adjacent relation based on a space topological relation of a research area, calculating a distance attenuation weight by combining a distance attenuation principle, and fusing the adjacent relation and the distance attenuation weight to construct a composite space weight matrix; Constructing a deep learning model based on a U-Net architecture, fusing the composite space weight matrix obtained in the step two with the multi-source data obtained through pretreatment in the step one to form input features, and training the deep learning model by using the input features, wherein a loss function adopted by training is a composite loss function, and the composite loss function comprises a c