Search

CN-121980485-A - Daily root zone soil humidity prediction method based on causal coupling space-time model

CN121980485ACN 121980485 ACN121980485 ACN 121980485ACN-121980485-A

Abstract

The application provides a daily root zone soil humidity prediction method based on a causal coupling space-time model, which relates to the field of soil humidity monitoring, and comprises the steps of obtaining root zone soil humidity and environment variable data and constructing a data set; the method comprises the steps of extracting causal relation strength between soil humidity and environmental variables of a root zone by adopting a causal discovery algorithm to form priori causal knowledge, constructing a causal attention module containing a dynamic causal weight adjustment mechanism, constructing a space-time shift window fransformer module containing a hierarchical multilevel feature extraction strategy, fusing the causal attention module and the space-time shift window fransformer module, constructing a causal coupling space-time fransformer model, training, inputting new environmental variable data into the trained model, and obtaining a soil humidity prediction result. The technical scheme of the application is suitable for long-period high-precision prediction of the soil humidity of the root zone under complex terrains.

Inventors

  • CHEN NENGCHENG
  • WU TINGTAO
  • XU LEI

Assignees

  • 湖北珞珈实验室
  • 中国地质大学（武汉）

Dates

Publication Date
20260505
Application Date
20251127

Claims (9)

  1. 1. A daily root zone soil humidity prediction method based on a causal coupling space-time model is characterized by comprising the following steps: S1, acquiring soil humidity and environment variable data of a root zone, and constructing a data set; S2, extracting causal relation strength between soil humidity and environmental variables in a root zone by adopting a causal discovery algorithm to form priori causal knowledge; S3, constructing a causal attention module containing a dynamic causal weight adjustment mechanism, and fusing priori causal knowledge and data-driven dynamic dependence; S4, constructing a space-time shift window transducer module containing a hierarchical multilevel feature extraction strategy, and modeling a complex space-time dependency relationship; S5, the causal attention module and the space-time shift window transducer module are fused in series, a causal coupling space-time transducer model is built, and training is conducted through a data set; S6, inputting new environment variable data into the trained causal coupling space-time transducer model to obtain a soil humidity prediction result.
  2. 2. The method for predicting the daily root zone soil humidity based on causal coupling space-time model as recited in claim 1, wherein step S1 comprises: Dividing the data set into a training data set, a verification data set and a test data set, wherein the daily soil humidity data of the next 10 days corresponding to the test data set is used as tag data; The environment variable data comprise surface soil humidity, temperature, precipitation, 10 m horizontal wind speed, 10 m vertical wind speed, downward long wave radiation and downward short wave radiation.
  3. 3. The method for predicting the daily root zone soil moisture based on causal coupling space-time model as recited in claim 1, wherein step S2 comprises: the causal discovery algorithm is a Peter-Clark instantaneous condition independent algorithm, and specifically comprises the following steps: Initializing a parent node set of the soil humidity of the root zone as all environment variables and self-hysteresis variables of the past 1 to 7 days; screening out variables with causal relation with the soil humidity of the root zone under the significance level of 0.05 through partial correlation test; quantifying causal intensities into a static causal weight matrix Wherein Is the total number of variables.
  4. 4. A method for predicting soil moisture in a daily root zone based on a causal coupled space-time model as recited in claim 3, wherein step S3 comprises: the implementation of the causal attention module comprises: Constructing a feature extraction network consisting of a 3D convolution layer, grouping normalization, siLU activation functions and the 3D convolution layer, and outputting a dynamic dependency matrix ; Will static causal weight matrix And (3) with Splicing, processing by the same structural network, and outputting a dynamic causal weight matrix ; Will be accumulated by Hadamard Acting on input features, achieving causal enhancement: Wherein the method comprises the steps of In order to input the characteristics of the feature, To be an enhanced feature; features are further extracted by the 3D convolution layer of the 7 x 7 convolution kernel and the 1 x 1 convolution kernel and residual connections are introduced to preserve the original information.
  5. 5. The method for predicting the daily root zone soil moisture based on causal coupling space-time model as recited in claim 1, wherein step S4 comprises: The hierarchical multi-level feature extraction strategy includes three phases: time mode capture phase-reshaping the input features into Extracting time dependence by adopting a 2-layer shift window transducer structure; Spatial dependent extraction stage, remodelling features to Extracting space dependence by adopting a 6-layer shift window transducer structure; Space-time feature integration stage, reorganizing features into Integrating space-time characteristics by adopting a 2-layer shift window transducer structure; Wherein, the In order to be of the size of the batch, For the length of time it is desirable that, In terms of the spatial height and width, Is the embedding dimension.
  6. 6. The method for predicting the daily root zone soil moisture based on causal coupling space-time model as recited in claim 1, wherein step S5 comprises: the construction of the causal coupling space-time transducer model comprises the following steps: sequentially inputting three-dimensional space-time sequence data, namely the space-time sequence of root zone soil humidity and environment variable data, into a causal attention module and a space-time shift window transform module; reconstructing the high-dimensional features into a three-dimensional output consistent with the input spatial resolution by transposed 3D convolution; and adopting an autoregressive iteration mode, taking the predicted result of the previous day as the input of the next day, and generating the predicted result of the soil humidity day by day of 10 days in the future.
  7. 7. The method for predicting the daily root zone soil moisture based on causal coupling space-time model as recited in claim 1, wherein step S6 comprises: The loss function of the causally coupled space-time transducer model is the mean absolute error: Wherein the method comprises the steps of In order to be able to predict the value, To be a true value of the value, Is the number of samples; adopting an early-stopping strategy, and stopping training when the loss of the verification set is not reduced by 20 continuous rounds; The accuracy evaluation indexes of the causal coupling space-time transducer model comprise average absolute error, unbiased root mean square error, pearson correlation coefficient and peak signal to noise ratio.
  8. 8. An electronic device comprising a processor, a memory, a user interface and a network interface, the memory for storing instructions, the user interface and the network interface for communicating to other devices, the processor for executing the instructions stored in the memory to cause the electronic device to perform the causal coupled spatiotemporal model based method of daily root zone soil moisture prediction as defined in any of claims 1 to 7.
  9. 9. A computer readable storage medium storing instructions which, when executed by a computer, perform the causal coupled spatiotemporal model based daily root zone soil moisture prediction method of any of claims 1 to 7.

Description

Daily root zone soil humidity prediction method based on causal coupling space-time model Technical Field The application relates to the field of soil humidity monitoring, in particular to a daily root zone soil humidity prediction method based on a causal coupling space-time model. Background Soil moisture is a key variable regulating moisture and energy flux exchange between land and the atmosphere, and is an important component of global hydrologic cycle. In 2010, soil humidity was listed as a key climate variable for studying the global climate system due to its broad importance. The surface soil humidity (about 0-5 cm) mainly affects the processes of surface evaporation, near-surface energy balance and the like, while the deep soil humidity (about 0-100 cm) dominates the underground processes, including plant root system water absorption and groundwater supply, and is important to the long-term water resource availability and ecological system sustainability. The space-time prediction of soil humidity, namely the process of estimating the future spatial distribution and time variation of the soil humidity through a prediction model, has important significance for decision support in the applications of crop yield optimization, irrigation planning, drought management and the like. Soil moisture prediction methods can be broadly divided into two types, process driven models and data driven models. The process driven model simulates the dynamic change in soil humidity by solving a physical equation describing the terrestrial gas interactions (e.g., a lichalz equation). Although such models make an important contribution in soil moisture prediction, the inherent nonlinear characteristics thereof and the limitation that the factors such as topography, soil properties, land utilization and the like are difficult to comprehensively consider lead to a large uncertainty of a prediction result. With the increasing abundance of multi-source heterogeneous observation data, data-driven machine learning models continue to develop, covering traditional statistical methods, machine learning algorithms, and deep learning models. The deep learning model is believed to be capable of directly extracting the relationship of input features to target variables from observed data without detailing complex physical processes. The classical deep learning model for soil humidity prediction comprises a time prediction model such as a long-term and short-term memory network and a space-time prediction model such as a convolution long-term and short-term memory network. However, pure deep learning models may face the problem of capturing spurious correlations, etc., while incorporating a priori knowledge (including causal relationships) into the deep learning model provides a viable approach to solving these problems. The causal relationship reflects the directionality and meaningful interaction between variables, beyond simple correlation, and can reveal the inherent mechanism of dynamic changes of the drive system. The influence of the historical variable on the target variable can be captured more accurately by embedding the causal relationship, so that the prediction accuracy is improved. The existing research shows that the causal information is integrated into the deep learning model through causal feature selection, structural optimization, loss function constraint and other modes, so that the model performance can be improved. For example, based on causal feature selection of the independent causal reasoning method of the Peter-Clark instantaneous conditions, prediction accuracy of the deep learning model on irrigation water consumption can be effectively improved, causal information is fused into the long-short-period memory network model for wind speed prediction through introduction of a neighborhood threshold value, the result shows that the causal long-period memory network model remarkably improves the prediction accuracy, causal information is fused into a loss function to restrict the long-period memory network model to predict wetland methane emission, and research discovers that the model performance can be effectively improved through embedding causal relations into the prediction model. However, most of the existing researches focus on time prediction, neglect the importance of space dimension, and lack systematic researches for fusing causal reasoning and a space-time prediction model. Disclosure of Invention The invention aims to solve the problems that most of the existing researches focus on time prediction, neglect the importance of space dimension and lack systematic researches which integrate causal reasoning and a space-time prediction model, and provide a daily root zone soil humidity prediction method based on a causal coupling space-time model. The above object of the present application is achieved by the following technical solutions: S1, acquiring soil humidity and environment variable data of a ro