Long-term agricultural drought efficient prediction method based on time sequence convolutional neural network

Abstract

The invention belongs to the technical field of agricultural drought prediction, and particularly relates to a long-term agricultural drought efficient prediction method based on a time sequence convolutional neural network, which comprises the following steps: acquiring original meteorological element data, and screening potential meteorological elements; carrying out normalization processing on the screened meteorological element data to obtain normalized meteorological element data, and extracting meteorological element spatio-temporal features by adopting a convolutional layer combined with a time dimension; constructing a time sequence-based convolutional neural network based on a convolutional layer combined with a time dimension, and establishing a statistical relationship between meteorological element spatial-temporal characteristics and agricultural drought; and inputting meteorological data in a future scene into the convolutional neural network based on the time sequence to carry out long-term agricultural drought efficient prediction. Time and space characteristics of meteorological elements are comprehensively considered, a statistical relationship between the meteorological elements and agricultural drought is effectively established, agricultural drought evolution in a future scene is predicted, and the adverse effect of agricultural drought on social and economic development of a research area can be relieved.

Description

Technical Field

[0001] The invention belongs to the technical field of agricultural drought prediction, and in particular relates to an efficient long-term agricultural drought prediction method based on a time series convolutional neural network. Background Art

[0002] Drought is a complex natural disaster caused by the imbalance of water revenue and expenditure in the water cycle, which has a huge impact on the economy, society, ecology and other aspects. According to the causes and impacts, drought can be divided into meteorological drought based on precipitation indicators, hydrological drought based on surface and groundwater runoff indicators, agricultural drought based on soil moisture indicators, and socioeconomic drought based on water supply and human water demand indicators. Among them, agricultural drought seriously threatens the growth and development of crops, which in turn leads to crop yield reduction. Against the background of global warming and human activities, the frequency and intensity of global agricultural droughts are showing an increasing trend. According to statistics, from 2005 to 2015 alone, agricultural droughts caused losses of more than 29 billion yuan to agricultural production in developing countries.

[0003] The formation mechanism and evolution process of agricultural drought are complex. It is a repetitive and random periodic event. Compared with other extreme events, agricultural drought generally occurs slowly, lasts for a long time and affects a wide area. Agricultural drought is a major threat to agricultural security, water security and ecological security. Its main feature is that soil moisture cannot normally supply the water required for crop growth, resulting in a serious reduction in crop yields.

[0004] Agricultural drought prediction models in existing technologies usually only consider precipitation and temperature, while ignoring the impact of other climate factors such as potential evapotranspiration and wind speed; existing technologies do not take into account that the development of agricultural drought is a slow process and ignore the time dimension information of prediction factors; the technical solution directly using the CNN network in existing technologies has poor simulation effect; the technical solution using the more complex ConvLSTM neural network model has the problem of excessive calculation and low simulation efficiency.

[0005] Therefore, there is an urgent need for an efficient long-term agricultural drought prediction method based on time series convolutional neural network, which can effectively predict the evolution trend of long-term agricultural drought and provide technical support for drought relief and disaster reduction. Summary of the invention

[0006] The purpose of the present invention is to provide a long-term agricultural drought efficient prediction method based on time series convolutional neural network, characterized in that it comprises the following steps:

[0007] Step S1, obtaining the original meteorological element data within a preset time span of the study area, screening potential meteorological elements based on the correlation coefficient of agricultural drought, and obtaining the screened meteorological element data;

[0008] Step S2, normalizing the filtered meteorological element data to obtain normalized meteorological element data, and extracting the spatiotemporal characteristics of the meteorological elements using a convolutional layer combined with the time dimension;

[0009] Step S3, constructing a time series convolutional neural network based on the convolutional layer combined with the time dimension, and establishing a statistical relationship between the spatiotemporal characteristics of meteorological elements and agricultural drought;

[0010] Step S4: input the meteorological data under the future scenario into the time series convolutional neural network in step S3 to perform long-term agricultural drought prediction.

[0011] The original meteorological element data in step S1 is: the fifth generation atmospheric reanalysis data set of the European Centre for Medium-Range Weather Forecasts;

[0012] The correlation coefficient of agricultural drought in step S1 is: soil moisture state index;

[0013] The screening of potential meteorological elements in step S1 is: using the Pearson correlation coefficient method to screen meteorological elements with strong correlation;

[0014] The potential meteorological elements in step S1 include: precipitation, potential evapotranspiration, relative humidity, surface temperature, radial wind speed, and latitudinal wind speed.

[0015] The step S2 of normalizing the filtered meteorological element data includes:

[0016]

[0017] In the formula, x is the original value and x' is the normalized result.

[0018] Extracting the spatiotemporal features of meteorological elements by combining the convolutional layer with the time dimension in step S2 includes:

[0019]

[0020] Where T is the time step, P is the type of meteorological element, Q and R are the length and width of the data space dimension, γ j,xyz is the output neuron of the jth feature map at (x, y, z), Act is the activation function, b j is the deviation of the jth feature map, W jt,pqr is the weight of the t-th feature map on the convolution kernel (p, q, r), λ t,(x+p)(y+q)(z+r) is the value in the input meteorological element associated with it.

[0021] The step S3 of establishing the statistical relationship between the spatiotemporal characteristics of meteorological elements and agricultural drought includes:

[0022] The spatiotemporal features of meteorological elements are extracted by combining the convolution layer with the time dimension; the extracted meteorological element features are integrated and output through the fully connected layer; the fully connected layer is:

[0023] γ=Act(Wλ+b) (3)

[0024] In the formula, γ is the output value, λ is the input value, W is the weight, and b is the bias.

[0025] The construction of the time series convolutional neural network in step S3 includes:

[0026] The normalized meteorological element data were randomly divided into training set, validation set and test set according to the ratio of 6:2:2;

[0027] The model structure is set to 3 layers, the first layer is the input layer, the number of neurons is set to 30, the input data is the meteorological data screened in step S1, the second layer is the fully connected layer, the third layer is the output layer, the output data is the agricultural drought index, the cycle is set to 100 times, the size of the convolution kernel is set to 3*3, the optimization algorithm is selected as Adam, and the activation function is selected as tanh;

[0028] The hyperparameters of the time series convolutional neural network model were adjusted, and three hyperparameter combinations of sample size (16, 32 and 64), learning rate (0.001, 0.005 and 0.01) and time step (3, 6, 9 and 12) were screened based on the correlation coefficient.

[0029] The beneficial effects of the present invention are:

[0030] The invention provides an efficient long-term agricultural drought prediction method based on a time series convolutional neural network. Based on the existing convolutional neural network model, the time dimension is considered, and a time series-convolutional neural network model is constructed. The model can effectively extract the spatiotemporal characteristics of meteorological elements, establish the statistical relationship between meteorological elements and agricultural drought, and predict the evolution trend of long-term agricultural drought in future scenarios. In order to avoid the influence of model hyperparameters, the model hyperparameters are adjusted by a grid search method, and the three hyperparameter combinations of sample number, learning rate and time step are screened based on the correlation coefficient, and the optimal hyperparameter combination is selected for subsequent simulation. The convolution layer combined with the time dimension extracts the spatiotemporal characteristics of meteorological elements in one step, which can effectively simulate the temporal and spatial evolution process of agricultural drought, and avoids the complex operation of ConvLSTM using convolution and memory neurons to process spatial and temporal information respectively. On the basis of ensuring the simulation effect, a relatively simple network structure is used to greatly reduce the amount of calculation and significantly improve the simulation efficiency.

[0031] Taking into account the relationship between the spatiotemporal characteristics of meteorological elements and agricultural drought, and compared with other artificial intelligence prediction methods in the existing technology, this method has higher operating efficiency, can quickly and effectively predict the evolution trend of long-term agricultural drought, and provide technical support for drought relief and disaster reduction. Establishing relevant prediction models for the risk of agricultural drought in different periods can effectively reduce drought losses and ensure the sustainable development of economic, social and environmental ecosystems. BRIEF DESCRIPTION OF THE DRAWINGS

[0032] Figure 1 A schematic diagram of a flow chart of a method for efficiently predicting long-term agricultural drought based on a time series convolutional neural network according to the present invention;

[0033] Figure 2 This is a schematic diagram of prediction results of the long-term agricultural drought efficient prediction method according to an embodiment of the present invention. DETAILED DESCRIPTION

[0034] The present invention provides a method for efficiently predicting long-term agricultural drought based on a time series convolutional neural network. The present invention is further described in detail below in conjunction with the accompanying drawings.

[0035] like Figure 1 The embodiment of the present invention shown discloses a method for efficiently predicting long-term agricultural drought based on a time series convolutional neural network, comprising the following steps:

[0036] Step S1, obtaining the original meteorological element data within a preset time span of the study area, screening potential meteorological elements based on the correlation coefficient of agricultural drought, and obtaining the screened meteorological element data;

[0037] Step S2, normalizing the filtered meteorological element data to obtain normalized meteorological element data, and extracting the spatiotemporal characteristics of the meteorological elements using a convolutional layer combined with the time dimension;

[0038] Step S3, constructing a time series convolutional neural network based on the convolutional layer combined with the time dimension, and establishing a statistical relationship between the spatiotemporal characteristics of meteorological elements and agricultural drought;

[0039] Step S4: input the meteorological data under the future scenario into the time series convolutional neural network in step S3 to perform long-term agricultural drought prediction.

[0040] To address the deficiencies of the prior art, an efficient long-term agricultural drought prediction method based on a time series convolutional neural network is provided in this embodiment. Based on the existing convolutional neural network model, the time dimension is taken into consideration to construct a time series-convolutional neural network model. The model can effectively extract the spatiotemporal characteristics of meteorological elements, establish a statistical relationship between meteorological elements and agricultural drought, and predict the evolution trend of long-term agricultural drought under future scenarios.

[0041] The original meteorological element data in step S1 is: the fifth generation atmospheric reanalysis data set of the European Centre for Medium-Range Weather Forecasts;

[0042] The correlation coefficient of agricultural drought in step S1 is: soil moisture state index;

[0043] The screening of potential meteorological elements in step S1 is: using the Pearson correlation coefficient method to screen meteorological elements with strong correlation;

[0044] The potential meteorological elements in step S1 include: precipitation, potential evapotranspiration, relative humidity, surface temperature, radial wind speed, and latitudinal wind speed.

[0045] In this embodiment, actual evaporation, bare soil evaporation, potential evapotranspiration, relative humidity, total precipitation, 2m surface temperature, 10m zonal wind speed and 10m radial wind speed in the study area from 1950 to 2022 are collected. The data are from the fifth generation atmospheric reanalysis data set ERA5 of the European Center for Medium-Range Weather Forecasts. The Pearson correlation coefficient is used to screen meteorological elements related to agricultural drought. The Soil Moisture Conditional Index (SMCI) is used to represent agricultural drought. The calculation formula is as follows:

[0046]

[0047] In the formula, θ max and θ min Represent the maximum and minimum values ​​of soil moisture in the same period of the historical period, θ i Represents the soil moisture state in period i. SMCI standardizes soil moisture into the range of 0-1. The closer to 1, the wetter the soil state, and the closer to 0, the drier the soil moisture.

[0048] The step S2 of normalizing the filtered meteorological element data includes:

[0049]

[0050] In the formula, x is the original value and x' is the normalized result.

[0051] Extracting the spatiotemporal features of meteorological elements by combining the convolutional layer with the time dimension in step S2 includes:

[0052]

[0053] Where T is the time step, P is the type of meteorological element, Q and R are the length and width of the data space dimension, γ j,xyzis the output neuron of the jth feature map at (x, y, z), Act is the activation function, b j is the deviation of the jth feature map, W jt,pqr is the weight of the t-th feature map on the convolution kernel (p, q, r), λ t,(x+p)(y+q)(z+r) is the value in the input meteorological element associated with it.

[0054] In this embodiment, the spatiotemporal characteristics of meteorological elements are extracted in one step by combining the convolution layer of the time dimension, which can effectively simulate the temporal and spatial evolution process of agricultural drought, avoiding the complex operation of ConvLSTM using convolution and memory neurons to process spatial and temporal information respectively. Therefore, the long-term agricultural drought efficient prediction method based on time series convolutional neural network disclosed in the present invention can use a relatively simple network structure on the basis of ensuring the simulation effect, greatly reduce the amount of calculation, and significantly improve the simulation efficiency.

[0055] The step S3 of establishing the statistical relationship between the spatiotemporal characteristics of meteorological elements and agricultural drought includes:

[0056] The spatiotemporal features of meteorological elements are extracted by combining the convolution layer with the time dimension; the extracted meteorological element features are integrated and output through the fully connected layer; the fully connected layer is:

[0057] γ=Act(Wλ+b) (3)

[0058] In the formula, γ is the output value, λ is the input value, W is the weight, and b is the bias.

[0059] The construction of the time series convolutional neural network in step S3 includes:

[0060] In this embodiment, the training set, validation set, and test set are randomly divided in a ratio of 6:2:2;

[0061] The model structure is set to 3 layers, the first layer is the input layer, the number of neurons is set to 30, the input data is the meteorological data screened in step S1, the second layer is the fully connected layer, the third layer is the output layer, the output data is the agricultural drought index, the cycle is set to 100 times, the size of the convolution kernel is set to 3*3, the optimization algorithm is selected as Adam, and the activation function is selected as tanh;

[0062] The hyperparameters of the time series convolutional neural network model were adjusted, and three hyperparameter combinations of sample size (16, 32 and 64), learning rate (0.001, 0.005 and 0.01) and time step (3, 6, 9 and 12) were screened based on the correlation coefficient.

[0063] In order to avoid the influence of model hyperparameters, the grid search method is used to adjust the model hyperparameters. The three hyperparameter combinations of sample size, learning rate and time step are screened based on the correlation coefficient, and the optimal hyperparameter combination is selected for subsequent simulation.

[0064] The future climate scenarios involved in step S4 include:

[0065] Three shared socioeconomic pathways (SSP) emission scenarios, including SSP245, SSP370 and SSP585. Among them, SSP245 is an updated RCP4.5 scenario based on the moderate development SSP2, with radiative forcing stabilized at 4.5W / m2 in 2100. 2 ; SSP370 is an updated RCP7.0 scenario based on the local development of SSP3, with radiative forcing stabilized at 7.0 W / m in 2100 2 ; SSP585 is an updated RCP8.5 scenario based on the conventional development of SSP5, with radiative forcing stabilized at 8.5 W / m in 2100 2 In order to verify the effectiveness of the long-term agricultural drought efficient prediction method based on time series convolutional neural network disclosed in the present invention, the traditional deep learning method convolutional neural network CNN and the latest spatiotemporal deep learning method convolutional long short-term memory neural network ConvLSTM are introduced.

[0066] Model comparison analysis, using Pearson correlation coefficient, mean absolute error, mean square error and running time to compare the effectiveness of the proposed method and CNN and ConvLSTM in efficient simulation of long-term agricultural drought.

[0067] Based on the screened meteorological elements, appropriate global climate models and future scenarios (Shared Socioeconomic Pathways SSP) were selected to analyze the evolution of agricultural drought under different future scenarios.

[0068] Table 1 Comparison of the results of CNN, ConvLSTM and the efficient agricultural drought prediction method proposed by the present invention

[0069]

[0070] As shown in Table 1, compared with the traditional models CNN and ConvLSTM, the convolutional neural network model (T-CNN) developed by the present invention considering the time dimension can simulate agricultural drought more effectively. In particular, compared with the more complex ConvLSTM model, the T-CNN developed by the present invention can reduce the average running time from 52 seconds to 6 seconds under the premise of similar values ​​of indicators such as Pearson correlation coefficient, mean absolute error, and mean square error, proving that the long-term agricultural drought efficient prediction method based on time series convolutional neural network disclosed by the present invention has better operating efficiency and higher application prospects.

[0071] like Figure 2 As shown, in the future scenario, with the increase of radiation forcing (SSP245 to SSP585), the agricultural drought index value decreases significantly, that is, the intensity of agricultural drought increases, which is consistent with previous research results, indicating that the prediction method disclosed in the present invention has high accuracy.

[0072] The above description is only a preferred specific implementation of the present invention and is not intended to limit the present invention. The methods, data, etc. involved in the present invention may also be specifically formulated according to different research areas. Any modification, equivalent replacement, improvement, etc. made within the scope of the claims of the present invention shall be within the scope of protection of the present invention.

Claims

1. An efficient long-term agricultural drought prediction method based on time series convolutional neural network, characterized in that: The steps include: Step S1, obtaining the original meteorological element data within a preset time span of the study area, screening potential meteorological elements based on the correlation coefficient of agricultural drought, and obtaining the screened meteorological element data; Step S2, normalizing the filtered meteorological element data to obtain normalized meteorological element data, and extracting the spatiotemporal characteristics of the meteorological elements using a convolutional layer combined with the time dimension; Step S3, constructing a time series convolutional neural network based on the convolutional layer combined with the time dimension, and establishing a statistical relationship between the spatiotemporal characteristics of meteorological elements and agricultural drought; Step S4: input the meteorological data under the future scenario into the time series convolutional neural network in step S3 to perform long-term agricultural drought prediction.

2. The long-term agricultural drought efficient prediction method based on time series convolutional neural network according to claim 1 is characterized in that: The original meteorological element data in step S1 is: the fifth generation atmospheric reanalysis data set of the European Centre for Medium-Range Weather Forecasts; The correlation coefficient of agricultural drought in step S1 is: soil moisture state index; The screening of potential meteorological elements in step S1 is: using the Pearson correlation coefficient method to screen meteorological elements with strong correlation; The potential meteorological elements in step S1 include: precipitation, potential evapotranspiration, relative humidity, surface temperature, radial wind speed, and latitudinal wind speed.

3. The long-term agricultural drought efficient prediction method based on time series convolutional neural network according to claim 1 is characterized in that: The step S2 of normalizing the filtered meteorological element data includes: In the formula, x is the original value and x' is the normalized result.

4. The long-term agricultural drought efficient prediction method based on time series convolutional neural network according to claim 1 is characterized in that: Extracting the spatiotemporal features of meteorological elements by combining the convolutional layer with the time dimension in step S2 includes: Where T is the time step, P is the type of meteorological element, Q and R are the length and width of the data space dimension, γ j,xyz is the output neuron of the jth feature map at (x, y, z), Act is the activation function, b j is the deviation of the jth feature map, W jt,pqr is the weight of the t-th feature map on the convolution kernel (p, q, r), λ t,(x+p)(y+q)(z+r) is the value in the input meteorological element associated with it.

5. The long-term agricultural drought efficient prediction method based on time series convolutional neural network according to claim 1 is characterized in that: The step S3 of establishing the statistical relationship between the spatiotemporal characteristics of meteorological elements and agricultural drought includes: The spatiotemporal features of meteorological elements are extracted by combining the convolution layer with the time dimension; the extracted meteorological element features are integrated and output through the fully connected layer; the fully connected layer is: γ=Act(Wλ+b) (3) In the formula, γ is the output value, λ is the input value, W is the weight, and b is the bias.

6. The long-term agricultural drought efficient prediction method based on time series convolutional neural network according to claim 1 is characterized in that: The construction of the time series convolutional neural network in step S3 includes: The normalized meteorological element data were randomly divided into training set, validation set and test set according to the ratio of 6:2:2; The model structure is set to 3 layers, the first layer is the input layer, the number of neurons is set to 30, the input data is the meteorological data screened in step S1, the second layer is the fully connected layer, the third layer is the output layer, the output data is the agricultural drought index, the cycle is set to 100 times, the size of the convolution kernel is set to 3*3, the optimization algorithm is selected as Adam, and the activation function is selected as tanh; The hyperparameters of the time series convolutional neural network model were adjusted, and three hyperparameter combinations of sample size (16, 32 and 64), learning rate (0.001, 0.005 and 0.01) and time step (3, 6, 9 and 12) were screened based on the correlation coefficient.

Images