CN-122020601-A - Causal inference method and causal inference system for agricultural meteorological disaster loss evaluation

CN122020601ACN 122020601 ACN122020601 ACN 122020601ACN-122020601-A

Abstract

A causal inference method and system for evaluating agricultural meteorological disaster loss belong to the field of agricultural meteorological disaster prevention and reduction. The method comprises the steps of obtaining multi-source space-time data comprising time sequence data, static data, disaster data and crop yield observation values, carrying out time sequence processing on the time sequence data to generate depth time sequence characteristics, generating attention weights corresponding to time steps in the time sequence data through a multi-head self-attention mechanism, inputting the depth time sequence characteristics and the static data into a first regression for training, carrying out time step weighting on the time sequence data according to the attention weights, inputting the time sequence data and the static data into a second regression for training, carrying out interpretation analysis on the second regression by adopting an SHAP interpreter, quantifying the influence of environmental characteristics on the yield, taking the disaster data as intervention variables, taking the time sequence data and the static data as the covariates, and carrying out causal effect evaluation according to the intervention variables, the covariates and the crop yield observation values to obtain corresponding crop yield variation under different disaster grades.

Inventors

LIU BUCHUN
CHEN LONG
HAN RUI

Assignees

中国农业科学院农业环境与可持续发展研究所

Dates

Publication Date: 20260512
Application Date: 20251204

Claims (11)

1. The causal inference method for evaluating the loss of the agricultural meteorological disaster is characterized by comprising the following steps of: s1, acquiring multi-source space-time data, wherein the multi-source space-time data comprises time sequence data, static data, disaster data and crop yield observation values, and the time sequence data comprises meteorological data and remote sensing data; S2, carrying out time sequence processing on the time sequence data to generate a depth time sequence characteristic, and generating attention weights corresponding to time steps in the time sequence data through a multi-head self-attention mechanism; S3, inputting the depth time sequence characteristics and the static data into a first regressor for training, simultaneously carrying out time step weighting on the time sequence data according to the attention weight, inputting the time sequence data and the static data into a second regressor for training, adopting a SHAP interpreter to interpret and analyze the second regressor, and quantifying the influence of environmental characteristics on the yield; And S4, taking the disaster data as an intervention variable, taking the time sequence data and the static data as covariates, and carrying out causal effect evaluation according to the intervention variable, the covariates and the crop yield observation values to obtain the corresponding crop yield variation under different disaster grades.
2. The causal inference method for evaluating agricultural meteorological disaster loss according to claim 1, wherein the time-step weighting of the time-series data according to the attention weight in S3 is specifically implemented by multiplying the attention weight corresponding to each time step by the observed values of all time-series data in the time step to generate an attention weighted time-series feature matrix, normalizing the static data to obtain static features, expanding the static features in a time dimension, splicing the static features with the attention weighted time-series feature matrix to obtain attention weighted original input, and flattening the attention weighted original input along the time-step dimension to form a two-dimensional feature matrix for training by the second regressor.
3. The causal inference method of agricultural meteorological disaster damage assessment according to claim 1, wherein the first regressor and the second regressor are trained by adopting a one-year-remaining cross-validation strategy, and an average value of n training result assessment indexes is taken as a final assessment of model performance, wherein n is the total year.
4. The causal inference method for agricultural meteorological disaster damage assessment according to claim 1, wherein causal forest model is adopted to carry out causal effect assessment, so as to obtain individual treatment effect values under different disaster grades, wherein the individual treatment effect values represent corresponding crop yield variation under different disaster grades.
5. The causal inference method of agricultural meteorological disaster damage assessment according to claim 4, wherein the causal forest model is trained by adopting a one-year cross validation strategy before step S4, and a validation set corresponding to disaster types not occurring in the training set is filtered during training.
6. The causal inference method for agricultural meteorological disaster damage assessment according to claim 1, wherein the step S4 further comprises a step of visualization of an average causal effect curve and an individual treatment effect spatial distribution map according to the causal effect assessment result.
7. The causal inference method for agricultural meteorological disaster damage assessment according to claim 1, further comprising preprocessing multi-source spatio-temporal data before S2, wherein the preprocessing comprises the steps of reshaping the meteorological and remote sensing data to form a two-dimensional matrix, setting the two-dimensional matrix as a sample number x a time step, and normalizing the two-dimensional matrix to obtain preprocessed time series data.
8. The causal inference method of agricultural meteorological disaster damage assessment of claim 7, wherein the first regressor and the second regressor are XGBoost regression models.
9. A causal inference system for agricultural meteorological disaster damage assessment employing a causal inference method for agricultural meteorological disaster damage assessment according to any one of claims 1 to 8, characterized in that the system comprises: the system comprises a data acquisition unit, a data processing unit and a data processing unit, wherein the data acquisition unit is used for acquiring multi-source space-time data, the multi-source space-time data comprises time sequence data, static data, disaster data and crop yield observation values, and the time sequence data comprises meteorological data and remote sensing data; The feature extraction unit is connected with the data acquisition unit and is used for carrying out time sequence processing on the time sequence data to generate depth time sequence features and generating attention weights corresponding to time steps in the time sequence data through a multi-head self-attention mechanism; The dual-path regression unit comprises a first regression unit and a second regression unit, is connected with the feature extraction unit and is used for inputting the depth time sequence feature and the static data into the first regression unit for training, and simultaneously, carrying out time step weighting on the time sequence data according to the attention weight and inputting the time sequence data and the static data into the second regression unit for training; An interpretation unit, which adopts an SHAP interpreter, is used for performing interpretation analysis on the second regression, and quantifying the influence of environmental characteristics on the yield; the causal evaluation unit is used for taking the disaster data as an intervention variable, taking the time sequence data and the static data as covariates, carrying out causal effect evaluation by combining the crop yield observation values and outputting individual treatment effect values, wherein the individual treatment effect values represent the corresponding crop yield change amounts under different disaster grades.
10. The causal inference system for agricultural meteorological disaster damage assessment according to claim 9, wherein the causal inference system comprises a preprocessing unit, wherein the preprocessing unit is configured to reshape the meteorological and remote sensing data to form a two-dimensional matrix, the two-dimensional matrix is set to be a sample number x a time step, and the two-dimensional matrix is subjected to standardization processing, so as to obtain preprocessed time sequence data.
11. An electronic device comprising a processor and a memory, the memory storing a computer program, characterized in that the processor implements the steps of the method according to any one of claims 1 to 8 when executing the computer program.

Description

Causal inference method and causal inference system for agricultural meteorological disaster loss evaluation Technical Field The invention relates to the field of agricultural meteorological disaster prevention and reduction, in particular to a causal inference method and a causal inference system for agricultural meteorological disaster loss assessment. Background Under the background of aggravated global climate change, drought, flood, frost and other agricultural meteorological disasters frequently occur, and the method forms a serious challenge for grain safety production. The method can accurately evaluate the specific loss of the disaster to the crop yield in time, and is a key basis for formulating disaster reduction measures, carrying out agricultural insurance claims and optimizing policies of disaster countermeasures. In order to quantify the loss of a particular disaster in a year (e.g., 2023), the observed yield of that year in the event of an actual disaster is compared to the virtual yield in the event of an assumption that no disaster has occurred in that year. However, in the actual production process, once a disaster occurs, the counter-fact state of the disaster is not observed directly, which constitutes a fundamental technical problem of disaster loss evaluation. At present, traditional crop yield prediction methods are mainly divided into the following two types, but all have obvious limitations: 1. Depending on crop mechanism models such as WOFOST, this type of method predicts potential yield by modeling the interaction of environmental factors such as light, temperature, water, soil, etc. with crop physiological processes. Although the mechanism of the crop mechanism model is transparent and clear and has strong interpretation, the model parameters and data are difficult to acquire, so the model is not suitable for the regional scale. 2. The method utilizes historical meteorological, remote sensing and crop yield data to construct a prediction model by relying on a pure data driving model, so that complex mechanism modeling is avoided. However, such deep learning methods are often used as "black box" structures, and cannot trace back the intermediate process from the input value to the predicted value, and cannot easily explain the relationship between the environment and the yield. Therefore, how to accurately evaluate the causal influence of agricultural meteorological disasters on crop yield and develop targeted disaster reduction measures is of great importance. Disclosure of Invention Accordingly, a primary objective of the present invention is to provide a causal inference method and system for agricultural meteorological disaster damage assessment, so as to at least partially solve the above-mentioned problems. In order to achieve the above object, as a first aspect of the present application, there is provided a causal inference method for agricultural weather disaster damage assessment, comprising the steps of: s1, acquiring multi-source space-time data, wherein the multi-source space-time data comprises time sequence data, static data, disaster data and crop yield observation values, and the time sequence data comprises meteorological data and remote sensing data; S2, carrying out time sequence processing on the time sequence data to generate a depth time sequence characteristic, and generating attention weights corresponding to time steps in the time sequence data through a multi-head self-attention mechanism; S3, inputting the depth time sequence characteristics and the static data into a first regressor for training, simultaneously carrying out time step weighting on the time sequence data according to the attention weight, inputting the time sequence data and the static data into a second regressor for training, adopting a SHAP interpreter to interpret and analyze the second regressor, and quantifying the influence of environmental characteristics on the yield; And S4, taking the disaster data as an intervention variable, taking the time sequence data and the static data as covariates, and carrying out causal effect evaluation according to the intervention variable, the covariates and the crop yield observation values to obtain the corresponding crop yield variation under different disaster grades. In a possible implementation manner, the time step weighting of the time series data according to the attention weight in S3 specifically includes multiplying the attention weight corresponding to each time step by the observed values of all time series data in the time step to generate an attention weighted time series feature matrix, normalizing the static data to obtain a static feature, expanding the static feature in a time dimension, then splicing the static feature with the attention weighted time series feature matrix to obtain an attention weighted original input, flattening the attention weighted original input along the time dimension to form a two-dimensional featu