A daily agricultural drought warning method based on soil moisture

Abstract

The present invention discloses a daily agricultural drought warning technology based on soil moisture. This invention can improve the timeliness of existing remote sensing detection means for agricultural drought monitoring. By training an end-to-end convolutional neural network on daily soil moisture data over the years, the soil moisture content for the next five days can be predicted. The predicted data is subjected to Sen slope estimation with the data of the previous seven days, and finally, the significance of the change trend is judged by the Mann-Kendall method, so as to realize the estimation of the agricultural drought trend for the next five days and achieve the purpose of early warning. The agricultural drought warning model based on neural network prediction of soil moisture proposed by the present invention can achieve efficient prediction of soil water content without introducing complex modules, strategies and techniques through a simple and effective convolutional neural network model, thereby improving the timeliness of disaster early warning and reducing the computational complexity. Compared with the traditional method of warning agricultural drought by remote sensing drought index, it has better spatial consistency and higher spatio-temporal resolution.

Description

Technical Field

[0001] The present invention relates to the field of agricultural drought monitoring in remote sensing applications, and particularly to an agricultural drought prediction technology based on soil moisture data. Background Art

[0002] The formation of drought is a cumulative process, which is not easily detected in the early stage. Once formed, it has the characteristics of a wide affected area and high severity. The history of drought monitoring is long. Traditional methods rely on the cooperation of manual work, meteorological stations and agricultural departments. Through observation stations, actual measurements are carried out, and drought index calculations and grading are performed on the obtained data. However, this method has the characteristics of low efficiency and long time consumption, and does not meet the requirements of real-time applications. With the development of drought monitoring technologies and means, drought monitoring based on remote sensing technology can more macroscopically establish a drought monitoring model based on remote sensing data, summarize the drought situation based on a longer time span, and achieve real-time drought monitoring.

[0003] Soil moisture is a key variable for evaluating agricultural drought. In recent years, with the significant improvement of the spatio-temporal coverage and product accuracy of soil moisture data, agricultural drought monitoring based on soil moisture has gradually attracted more attention. However, most of the soil moisture data used in drought research currently adopt a time scale of more than one month, and the daily soil moisture methods suitable for drought monitoring and forecasting are relatively scarce, which is not conducive to improving the accuracy of drought monitoring and forecasting and evaluating the existing daily drought situation.

[0004] The present invention focuses on effectively monitoring and evaluating the changes in soil moisture by predicting soil moisture. The present invention combines the existing deep learning spatio-temporal sequence data prediction model and trend calculation method, and designs a method for monitoring agricultural drought by predicting daily soil moisture. Summary of the Invention

[0005] In order to effectively monitor the changes in drought disasters through soil moisture, the present invention proposes an agricultural drought remote sensing monitoring technology based on neural network soil moisture prediction. This technology is based on deep learning methods, conducts trend analysis on spatio-temporal sequence data, and proposes a method for daily monitoring of drought.

[0006] The technical solution adopted by the present invention is as follows:

[0007] Step 1: The encoder stacks 4 BasicConv blocks to extract data features;

[0008] Step 2: Input the feature map extracted in Step 1 into the translator, and learn the time evolution through the Translator module. The Translator module consists of 8 Inception modules;

[0009] Step 3: The decoder reconstructs the soil moisture content data using 4 GroupConv2d blocks;

[0010] Step 4: Perform Sen slope estimation on the soil moisture content data predicted in Step 3 and the soil moisture content data from the previous week at the relative time;

[0011] Step 5: Determine the significance of the change trend of the result in Step 4 through the Mann - Kendall method to conduct drought early warning.

[0012] Compared with the prior art, the beneficial effects of the present invention are:

[0013] (1) Compared with the traditional remote sensing drought index, it has better spatial consistency with the distribution of soil moisture content;

[0014] (2) Compared with the soil moisture monitoring method that directly retrieves soil moisture, it has higher spatio - temporal resolution;

[0015] (3) The real - time performance of early warning is improved and the complexity is reduced. Description of the Drawings

[0016] Figure 1 It is: Schematic diagram of the SimSMP framework;

[0017] Figure 2 It is: Structure diagram of the BasicConv2d Block;

[0018] Figure 3 It is: Structure diagram of the Translator Block;

[0019] Figure 4 It is: Structure diagram of the Inception Block;

[0020] Figure 5 It is: Structure diagram of the GroupConv2d Block;

[0021] Figure 6 It is: Scatter plot of the correlation coefficient between predicted soil moisture and real soil moisture;

[0022] Figure 7 It is: Mann - Kendall test trend category; Detailed Embodiment

[0023] To make the objectives, technical solutions, and advantages of the present invention clearer, the present invention will be further described in detail below in conjunction with the embodiments and the drawings.

[0024] In this specific embodiment, the method for agricultural drought monitoring includes the following steps:

[0025] The process of predicting soil moisture remote sensing images using the SimSMP network model is as follows. Figure 1 As shown in the figure. The encoder is used to extract spatial features, the translator learns temporal evolution, and the decoder integrates spatio-temporal information to predict future frames.

[0026] Step 1: Extract the feature map using the feature extraction backbone network.

[0027] First, stack the BasicConv2d module 4 times. The module consists of three parts: Con2d, Groupnorm, and LeakyRelu. The kernel size of the convolutional layer is (3,3), and the strides are used alternately as (2,2) and (1,1), with a stride of (2,2) and a padding size of (1,1). GroupNorm divides the image channels into groups and then normalizes within each group, which is set to 2 here. LeakyRelu is the activation function, as shown in Figure 2 the figure.

[0028] After preprocessing the remote sensing image, the input image has a size of 256 in length, 256 in width, and 3 channels. The dimensionality size changes after each stage are 256×256×64, 128×128×64, 128×128×64, and 64×64×64 in sequence. The size of the finally output feature map is 64×64×64.

[0029] The constructed feature extraction structure has a better simulation effect on soil moisture on a specific soil moisture dataset, and can reduce the number of network layers and network parameters, making the model more lightweight and easier to deploy.

[0030] Step 2: Input the feature map extracted in Step 1 into the translator to learn temporal evolution through 8 Inception modules, as shown in Figure 3 the figure.

[0031] The Inception module consists of a Head Conv2d, followed by parallel BasicConv2d operations. To avoid the expression bottleneck, as the size of the feature map decreases, its dimension can be further reduced, and the number of channels can be decreased. Using 1x1Conv2d convolution not only does not affect the model accuracy, but also can accelerate its convergence speed. The parallel connection mechanism of the Inception module can capture multi-scale image features simultaneously, making the network more advantageous when processing targets of different sizes. Group convolution is performed on the input feature map, and then each group is convolved separately, which can reduce the number of parameters and is not easily overfitted. Here, GroupConv2d is set to 8. As shown in Figure 4 the figure. The kernel sizes in the Inception layer are (3,5,7,11) in sequence. After passing through the translator, the size of the output feature map is 64×64×64.

[0032] Step 3: The decoder uses 4 DeConv2d blocks to reconstruct the soil moisture data;

[0033] The DeConv2d block is also constructed in a stacked manner, including the convtranspose2d deconvolution layer, with a kernel size of (3,3), a stride of (2,2) and (1,1) alternating between the padding size of (1,1), and an output padding size of (1,1). GroupNorm divides the image into groups along the channel direction, and then normalizes each group, which is set to 2 here; LeakyRelu is the activation function, such as Figure 5 As shown in Figure 2, four GroupConv2d blocks reconstruct the soil moisture data, with each stage size being 64×64×64, 128×128×64, 128×128×64, and 256×256×64. Finally, 1×1 convolution is used to reduce the dimension and output a predicted image of 256×256×3.

[0034] Figure 6 This is a Pearson correlation coefficient result diagram between the soil moisture data predicted by the method of the present invention and the actual soil moisture content. It can be seen from the figure that the soil moisture data predicted by the method has a high correlation with the actual soil moisture content.

[0035] Step 4: Estimating the Sen slope of the predicted data in step 3 and the soil moisture data of the previous week;

[0036] Theil-Sen Median method, also known as Sen slope estimation, is a robust nonparametric statistical trend calculation method. This method is highly efficient and insensitive to measurement errors and outliers, and is often used in trend analysis of long time series data. Its calculation formula is:

[0037]

[0038] Where: x i , x j For time series data, median is the median calculation. β greater than 0 indicates that the time series shows an upward trend; β less than 0 indicates that the time series shows a downward trend.

[0039] Step 5: Use the Mann-Kendall method to determine the significance of the change trend of the results of step 4, so as to provide drought warning;

[0040] The Mann-Kendall test, also known as the Mann-Kendall test, can be used to determine whether there is a mutation in the sequence. If so, the location of the mutation can be determined.

[0041] For the sequence \(X_t = x_1, x_2, \ldots, x_n\), first determine all the pairwise values \((x i , x j , j > i)\) and the magnitude relationship between \(x i \) and \(x j \) (denoted as \(S\)). Make the following assumptions: \(H_0\) is that the data in the sequence are randomly arranged, i.e., there is no significant trend; \(H_1\) is that the sequence has an upward or downward trend. The calculation formula for the test statistic \(S\) is (in the calculation formula of \(Z\), when \(S > 0\), the numerator is \(S - 1\))

[0042]

[0043] where: \(\text{sgn}()\) is the sign function.

[0044] Use the test statistic \(Z\) for trend testing. The calculation method of the \(Z\) value is as follows:

[0045]

[0046] where: The calculation formula for \(\text{Var}\) is:

[0047]

[0048] where: \(n\) is the number of data in the sequence, and \(m\) is the number of data groups that repeat in the sequence.

[0049] Similarly, a two-sided trend test is adopted. At a given significance level, the critical value \(Z_{1 - \alpha / 2}\) is obtained from the normal distribution table. When \(|Z| \leq Z_{-\alpha / 2}\), the null hypothesis is accepted, that is, the trend is not significant; if \(|Z| > Z_{-\alpha / 2}\), the null hypothesis is rejected, that is, the trend is considered significant. In this case, the given significance level = 0.05, then the critical value \(Z_{1 - \alpha / 2} = 1.96\). When the absolute value of \(Z\) is greater than 1.65, 1.96, and 2.58, it means that the trend has passed the significance tests with confidence levels of 90%, 95%, and 99% respectively. The method for judging trend significance is shown in Figure 7 .

[0050] Grade the confidence level statistic to obtain the daily agricultural drought prediction situation in the region. Visualize the results, which can reflect the spatio-temporal variation trend of drought in the region.

[0051] As described above, only the specific embodiments of the present invention are concerned. Any feature disclosed in this specification, unless specifically stated, can be replaced by other equivalent or similar-purpose alternative features; all the features disclosed, or all the steps in any method or process, except for mutually exclusive features or / and steps, can be combined in any way.

Claims

1. A daily agricultural drought warning method based on soil moisture, characterized in that It includes the following steps: Step 1: The encoder stacks 4 BasicConv blocks to extract soil data features; the module consists of three parts: Con2d, Groupnorm, and LeakyRelu; the kernel size of the convolutional layer is (3,3), the strides use (2,2), (1,1) alternately, the stride is (2,2), the padding size is (1,1), GroupNorm divides into groups in the image channel direction and then normalizes within each group, and LeakyRelu is the activation function; Step 2: Input the feature map extracted in Step 1 into the translator, and learn the time evolution through the Translator module, which consists of 8 Inception modules; the Inception module consists of a 1x1Conv2d, followed by parallel BasicConv2d operations to avoid the expression bottleneck; Step 3: The decoder uses 4 DeConv2d blocks to reconstruct the soil moisture content data; the DeConv2d blocks are also constructed in a stacked manner, including the convtranspose2d transposed convolutional layer, with a kernel size of (3,3), strides using (2,2), (1,1) alternately, a padding size of (1,1), and an output padding size of (1,1); GroupNorm divides into groups in the image channel direction and then normalizes within each group; LeakyRelu is the activation function; Step 4: Perform Sen slope estimation on the soil moisture content data predicted in Step 3 and the soil moisture content data one week before the relative time; Step 5: Judge the significance of the change trend of the result in Step 4 through the Mann—Kendall method to conduct drought early warning.

Images