CN110084367A - A kind of Forecast of Soil Moisture Content method based on LSTM deep learning model - Google Patents

Abstract

The invention discloses a soil moisture prediction method based on an LSTM deep learning model. First, soil physicochemical and meteorological data of a target farmland within a period of time are collected; the collected data is preprocessed and divided into a training sample set and a test sample set ;Construct the LSTM deep learning model, train the LSTM deep learning model through the training sample set to obtain the LSTM deep learning model after parameter adjustment, and then verify the adjusted LSTM deep learning model through the test sample set, and finally use the verified LSTM deep learning model. The LSTM deep learning model is used as a soil moisture prediction model; the collected data is input into the soil moisture prediction model, and the soil moisture prediction model finally outputs the soil moisture prediction results in the future. The present invention uses the deep learning method to forecast the soil moisture, saves manpower and material resources, can truly reflect the influence of the early data on the later results, and fully reflects the time series characteristics.

Description

A Soil Moisture Prediction Method Based on LSTM Deep Learning Model

technical field

The invention relates to the field of soil moisture prediction methods, in particular to a soil moisture prediction method based on an LSTM deep learning model.

Background technique

my country is a country with severe drought and water shortage. The per capita water resources are only 1/4 of the world average level, and it is one of the countries with the poorest per capita water resources in the world. The implementation of precision irrigation in the agricultural production process can effectively save water resources and promote the growth of crops. However, there are often insufficient and excessive irrigation in farmland irrigation. Insufficient irrigation leads to stagnant crop growth and low yield, while excessive irrigation can easily lead to poor root development and crop death, resulting in reduced water resources utilization. Unable to achieve high productivity. Therefore, establishing a soil moisture prediction model and carrying out soil moisture prediction can effectively solve the problems of insufficient irrigation and excess irrigation, and are the main technical means to realize precise irrigation of farmland. The soil moisture prediction model predicts the soil moisture content at a certain time in the future based on farmland meteorological data, soil physical and chemical data and past soil moisture content, so as to determine the amount of irrigation water, so as to achieve high and stable yields. At present, there is no mature technical method to effectively predict soil moisture. It is one of the important problems to be solved in agricultural precision production to establish a soil moisture prediction model with strong generalization ability and high prediction accuracy. In 2016, the paper "Research on Soil Moisture Prediction in Qidong City Based on Meteorological Factors" (Anhui Agricultural Sciences, 2016, 44(34): 174-176) proposed to use the stepwise regression method to analyze the soil surface moisture in Qidong City from 2011 to 2014. , bottom moisture and the correlation of meteorological factors (precipitation, temperature, humidity, sunshine, wind) in the same period, screened out the key meteorological factors affecting soil moisture, and established a soil moisture forecast model combined with the empirical formula method. The results show that the meteorological factors affecting soil moisture in Qidong City are mainly precipitation, sunshine and air temperature. The forecast model established based on the forecast model predicts the average relative error of soil moisture within 5% in the next 30 days, and the test effect is ideal. It can more accurately forecast soil moisture in the next 30 days and use it to guide agricultural production.

The 2017 paper "Research on Soil Moisture Prediction Accuracy Based on BP Neural Network - Taking Feidong County as an Example" (Soil Bulletin, 2017, 48(02): 292-297). It is proposed to use BP neural network to predict soil moisture. The main idea is to select the average temperature, average humidity, radiation and rainfall, which have a significant impact on soil moisture, as model input samples to establish a network model. In addition, the soil water content at the beginning of the time period has a great influence on the soil moisture at the end of the time period, so the soil water content at the beginning of the time period will also be used as the input sample of the model.

Patent "A Soil Moisture Prediction Method Based on Moisture Index" (application number: N201810457976.X), the model used in this method is a semi-empirical and semi-theoretical model proposed after years of field experimental research, with simple and easy-to-obtain parameters for practical application specialty. This method proposes to calculate the soil moisture index of the crop root development layer based on the measured soil moisture content of the moisture loss sensitive layer, and then to calculate the measured moisture content of the 20cm and 50cm soil layers, calculate the moisture content of the 50cm soil layer, predict the soil moisture index of the 50cm soil layer, and perform the calculation of the soil moisture index of the 20cm and 50cm soil layers in turn. It has realized the standardization of soil moisture monitoring, forecasting, updating, irrigation time and irrigation quota prediction technology, and built a soil moisture monitoring and forecasting information system. It is convenient for moisture information query, drought disaster assessment and risk management and control, and is suitable for forecasting soil moisture in farmland in the vast plains.

The above methods all realize the prediction of the soil moisture change curve at the next moment based on the soil moisture change in the past period of time. But their common problem is that the time series characteristics of the data are not fully taken into account when building the model, and the generalization ability and prediction accuracy need to be improved.

SUMMARY OF THE INVENTION

The purpose of the present invention is to provide a soil moisture prediction method based on the LSTM deep learning model, so as to solve the problem that the prior art soil moisture prediction method does not consider the time series characteristics of the data.

In order to achieve the above object, the technical scheme adopted in the present invention is:

A soil moisture prediction method based on LSTM deep learning model is characterized in that: comprising the following steps:

(1) Collect soil physicochemical and meteorological data of the target farmland for a period of time;

(2), preprocess the soil physicochemical and meteorological data collected in step (1), and then divide the preprocessed data into a training sample set and a test sample set in proportion;

(3) Constructing an LSTM deep learning model, the LSTM deep learning model has an input layer, twenty-five hidden layers and an output layer, and the LSTM deep learning model is trained through the training sample set to obtain the adjusted LSTM Deep learning model, and then verify the adjusted LSTM deep learning model through the test sample set, and finally use the verified LSTM deep learning model as the soil moisture prediction model;

(4), take the soil physicochemical and meteorological data collected in step (1) as the input of the soil moisture prediction model, process the soil physicochemical and meteorological data through the soil moisture prediction model, and finally output the soil moisture in the future by the soil moisture prediction model forecast result.

Described a kind of soil moisture prediction method based on LSTM deep learning model, it is characterized in that: in step (1), adopt linear interpolation method to repair the missing data of collected soil physicochemical and meteorological data, wherein the formula of linear interpolation method is as follows :

In formula (1), i and j represent the values at the i-th time and the j-th time, respectively, and it is required that 0<i<j. x _k and x _k+j are the soil physicochemical and meteorological data collected at time k and k+j respectively, and x _k+i is the soil physicochemical and meteorological data lost at time k+i.

The described soil moisture prediction method based on the LSTM deep learning model is characterized in that: the preprocessing in step (2) is normalization preprocessing, and 85% of the data are used as training samples after the normalization preprocessing set, 15% of the data is used as the test sample set, the formula of normalized preprocessing is

By performing normalization preprocessing on the collected data, the mapping interval of the collected data values is [0, 1]. In the normalization preprocessing formula, x is the original data, x _max , and x _min are respectively The maximum and minimum values in the original data, x _now is the result data after normalization;

Normalized preprocessing is used to eliminate the dimensional influence between indicators and solve the comparability between data indicators. After the original data is preprocessed, each indicator is in the same order of magnitude, which is helpful for model construction.

Described a kind of soil moisture prediction method based on LSTM deep learning model, it is characterized in that: in step (3), the network structure of the LSTM deep learning model of construction is (7,25,1), in the LSTM deep learning model Each hidden layer adopts an LSTM unit with three gates. The three gates of the LSTM unit are a forget gate, an input gate and an output gate. The three gate structures jointly complete the update of the state and output the target value. The hidden layer The data processing process is as follows:

The forget gate determines the degree of forgetting information. First, h ^(t-1) and x ^(t) are read to screen the data, where h ^(t-1) represents the output of the previous memory cell, and x ^(t) represents the input of the current cell, as shown in formula (2):

f ^(t) =σ(W _f ·[h ^(t-1) ,x ^(t) ]+b _f ) (2),

In formula (2), W _f is the weight item, b _f is the bias item, f ^(t) is the forgetting degree of information, σ is the sigmoid function, and the value is between [0, 1]. The sigmoid function outputs a value between 0 and 1 for the update of the cell state C ^(t) , where 1 indicates that the information is completely retained, and 0 indicates that the node data is completely discarded;

The input gate determines that new information is added to the hidden node, where C ^(t-1) is the state of the cell at the previous moment, and i ^(t) is defined as the information to determine the update. The completion of the information addition requires two steps: first, through The sigmoid function of an input gate determines which information needs to be updated; secondly, a vector is generated through a tanh layer, that is, the alternative content a ^(t) for updating, and the two parts are combined to perform a Update, as shown in formulas (3), (4):

i ^(t) =σ(W _i ·[h ^(t-1) ,x ^(t) ]+b _i ) (3),

a ^(t) = tanh(W _c ·[h ^(t-1) ,x ^(t) ]+b _a ) (4),

In formula 3, σ is the _sigmoid function, Wi is the weight item, h ^(t-1) is the final output part at the previous moment, and x ^(t) is the input of the current cell. b _i is the bias term.

In formula 4, tanh is the tanh function, W _c is the weight item, h ^(t-1) is the final output part at the previous moment, and x ^(t) is the input of the current cell. b _a is the bias term.

When updating the old cell state in formula (4), update the content a ^(t-1) updated at the previous moment to the content a ( ^t ) updated at the moment. Multiply the cell state C ^(t-1) at the previous moment by f ^(t) in the forget gate, and add i ^(t) *a ^(t) to update the cell state, as shown in formula (5) shown:

C ^(t) = f ^(t) * C ^(t-1) + i ^(t) * a ^(t) (5),

In formula (5), * represents the Hadamard product, that is, the product of the corresponding positions of the matrix. C ^(t-1) is the cell state at the previous moment, and a ^(t) is the new content. C ^(t) is the new memory state.

The output gate determines the output item. First, based on the state of the memory cell, a sigmoid function is run to determine which information of the memory cell will be output; secondly, the state of the memory cell is processed through tanh to obtain a value between -1 and 1, and multiply this value by the output of the output gate, as shown in Equation (6) and Equation (7):

In formula (6), o ^(t) is what information is output, h ^(t-1) , x ^(t) are the output at the previous moment and the input at this moment. W _o is a weight term, and b _o is a bias term.

In formula (7), h ^(t) is the part of the final output. Multiply the o ^(t) obtained in formula (6) by the value of the current new memory state through the tanh function to achieve the ability to remember the long-term dependent information of the sequence. Effect.

Compared with the prior art, the present invention has the advantages that: the present invention utilizes a deep learning algorithm and adopts a soil moisture prediction method based on a long-term and short-term memory model. Compared with traditional methods, the use of deep learning methods to forecast soil moisture does not require real-time manual measurement, which saves manpower and material resources. In addition, the use of LSTM units can truly reflect the impact of early-stage data on later-stage results, fully reflect the temporal characteristics, improve the prediction efficiency and accuracy, and have a high generalization ability. The soil moisture prediction method based on LSTM deep learning model has good application value.

Description of drawings

FIG. 1 is a flow chart of the implementation of the present invention.

FIG. 2 is a schematic diagram of the LSTM unit used in the present invention.

Figure 3 is a schematic diagram of the activation function.

Figure 4 is a schematic diagram of the activation function.

Figure 5 is a graph of predictions for the test set (soil moisture).

Detailed ways

The present invention will be further described below in conjunction with the accompanying drawings and embodiments.

As shown in Figure 1, the concrete realization process of the present invention is as follows:

1. Monitor the meteorological data and soil physical and chemical data of the farmland. Data was collected every 30 minutes. The amount of data collected for about 3 months, a total of more than 4000 data records, for some missing data, use the linear interpolation method to repair, as shown in formula (1):

In formula (1), i and j represent the values at the i-th time and the j-th time, respectively, and it is required that 0<i<j. x _k and x _k+j are the soil physicochemical and meteorological data collected at time k and k+j respectively, and x _k+i is the soil physicochemical and meteorological data lost at time k+i.

2. Data preprocessing: Before model training, the collected farmland meteorological data and soil physical and chemical data need to be normalized. The so-called normalization process is to map the data to the [0, 1] or [-1, 1] interval. It is guaranteed that the input data of different data ranges play the same role. The normalization processing formula adopted in the present invention is: Where x is the original data to be normalized, x _max and x _min are the maximum and minimum values in the original data, respectively, and x _now is the data after normalization.

The normalized data is divided into two parts: training sample set and test sample set data. The proportion of data is 85% and 15%, which are used for training and testing of the LSTM deep learning model.

3. Model structure: As shown in Figure 2, the present invention adopts the LSTM deep learning model with 7 input layers, 25 hidden layers and 1 output layer, and the hidden layer units of the LSTM deep learning model all use LSTM (Long Short Term Memory). ) unit, this structure can well solve the problem that the subsequent nodes have a reduced ability to perceive the previous time nodes, overcome the shortcomings of the conventional time series model structure, and can handle time series data well. LSTM is a unit called memory cell which has 3 gate structure.

The forget gate determines the extent to which information is forgotten. Read h ^(t-1) and x ^(t) , h ^(t-1) represents the output of the previous memory cell, x ^(t) represents the input of the current cell, and completes the data screening process, such as Formula (2) shows:

f ^(t) =σ(W _f ·[h ^(t-1) ,x ^(t) ]+b _f ) (2),

In formula (2), W _f is the weight item, b _f is the bias item, f ^(t) is the forgetting degree of information, σ is the sigmoid function, and the value is between [0, 1]. The sigmoid function outputs a value between 0 and 1 for the update of the cell state C ^(t) , where 1 indicates that the information is completely retained, and 0 indicates that the node data is completely discarded;

The input gate determines that new information is added to the hidden nodes. C ^(t-1) is the cell state at the last moment, and i ^(t) is the part that determines the information update. Completing the information addition needs to include two steps: first, determine which information needs to be updated through the sigmoid function of an input gate; second, a tanh layer generates a vector, which is the alternative content a ^(t) to be updated. These two parts are combined to update the cell state, as shown in Equation (3) and Equation (4):

i ^(t) =σ(W _i ·[h ^(t-1) ,x ^(t) ]+b _i ) (3),

a ^(t) = tanh(W _c ·[h ^(t-1) ,x ^(t) ]+b _a ) (4),

In formula 3, σ is the _sigmoid function, Wi is the weight item, h ^(t-1) is the final output part at the previous moment, and x ^(t) is the input of the current cell. b _i is the bias term.

In formula 4, tanh is the tanh function, W _c is the weight item, h ^(t-1) is the final output part at the previous moment, and x ^(t) is the input of the current cell. b _a is the bias term.

When updating the old cell state in formula (4), update the content a ^(t-1) updated at the previous moment to the content a ( ^t ) updated at the moment. Multiply the cell state C ^(t-1) at the previous moment by f ^(t) in the forget gate, and add i ^(t) *a ^(t) to update the cell state, as shown in formula (5) shown:

C ^(t) = f ^(t) * C ^(t-1) + i ^(t) * a ^(t) (5),

In formula (5), * represents the Hadamard product, that is, the product of the corresponding positions of the matrix. C ^(t-1) is the cell state at the previous moment, and a ^(t) is the new content. C ^(t) is the new memory state.

The output gate determines the output term. First, based on the memory cell state, run a sigmoid layer to determine which information of the memory cell will be output; second, process the memory cell state through tanh (to get a value between -1 and 1) and sum it with the output The outputs of the gates are multiplied. The output term is calculated according to equations (6) and (7).

In formula (6), o ^(t) is what information is output, h ^(t-1) , x ^(t) are the output at the previous moment and the input at this moment. W _o is a weight term, and b _o is a bias term.

In formula (7), h ^(t) is the part of the final output. Multiply the o ^(t) obtained in formula (6) by the value of the current new memory state through the tanh function to achieve the ability to remember the long-term dependent information of the sequence. Effect.

The gate structure uses a sigmoid activation function (as shown in Figure 3):

In the above formula (8), x is used as the input data, and the data vector value is 'compressed' to between [0, 1] by the sigmoid function. If the input value is negative and very large, the value is 0. If the input value If a positive number is very large, it is 1.

When updating the state of the cell, the tanh activation function is used (as shown in Figure 4.):

In the above formula (9), x is used as input data, and is mapped to [-1, 1] through the function f(x).

During network training, the final hidden layer state of the current batch is used as the subsequent initial hidden state (traversing the entire training set in order). Set the batch size to 72. The network structure adopted by the LSTM deep learning model used in the present invention is (7, 25, 1). , and the learning rate is set to 0.01. During training, the error is calculated according to the mean absolute error (MeanAbsoluteError) and used to update the weights according to the backpropagation algorithm.

In the above formula (11), m is the total number of training data, x ⁽ⁱ⁾ is the data input value, k(x ⁽ⁱ⁾ ) is the predicted output value, and y ⁽ⁱ⁾ is the actual output value. The error value obtained according to the above formula. The sequence in the training sample data is used as the training input, the network model is continuously trained and the parameters are adjusted, and the iteration ends when the number of iterations reaches 300. A stable forecasting model was obtained and used as a forecasting model for soil moisture.

The LSTM model has two hidden states h ^(t) and C ^(t) , and there are many parameters in the model.

(1) The process of the forward propagation process at each sequence index position is:

①Update the forget gate output

f ^(t) =σ(W _f ·[h ^(t-1) ,x ^(t) ]+b _f ),

②Update the two-part output of the input gate

i ^(t) =σ(W _i ·[h ^(t-1) ,x ^(t) ]+b _i ),

a ^(t) =tanh(W _c ·[h ^(t-1) ,x ^(t) ]+b _a ),

③ Update the cell state

C ^(t) = f ^(t) * C ^(t-1) + i ^(t) * a ^(t) ,

④Update the output of the output gate

⑤Update the current sequence index prediction output

y(t)=σ(W _y h ^(t) +c)

(2) Backpropagation algorithm: In order to backpropagate the error, stepwise forward propagation through the gradient δ ^(t ) of the hidden state h ^(t) . Backpropagation of an LSTM with two hidden states h ^(t) and C ^(t) . Define two δ, namely: and

In order to facilitate the derivation, the loss function is divided into two parts, one is the loss l ^(t) at the time t position, and the other is the loss L ^(t+1) after the time t, namely:

And at the last sequence index position τ, its and They are:

Then by Back derivation The gradient of is determined by the output gradient error at time t of this layer and the error greater than time t, namely:

The key to LSTM backpropagation is calculation. Since h ^(t) = o ^(t) *tanh(C ^(t) ) contains a recurrence relation of h in the first term o ^(t) , making the second term tanh(C ^(t) ) more complicated , the tanh function variable can be expressed as:

C ^(t) = C ^(t-1) *f ^(t) +i ^(t) *a ^(t) ,

and The inverse gradient error of , consists of two parts, the previous layer The gradient error of and the gradient error returned from h ^(t) of this layer.

A known and That is, the gradient calculation process of W _f can be obtained, and other parameters are the same as above.

4. Network test (parameter adjustment and optimization). Input the preprocessed training set data into the built LSTM deep learning model, and continuously optimize the parameters to gradually improve the prediction accuracy. In order to prevent overfitting, regularization is required. Finally, the model is obtained, and the output is the prediction result of soil moisture at a certain time in the future. The average relative error of the prediction is less than 0.25%. The prediction result is shown in Figure 5, and the prediction result is good.

The method adopted in the present invention utilizes and integrates the relationship between historical data to a great extent, fully considers the time series characteristics of the data, and builds a reasonable time series model, which can improve the prediction ability of soil moisture and enhance the model's accuracy. Generalization ability has good application value.

Claims (4)

1. a kind of Forecast of Soil Moisture Content method based on LSTM deep learning model, it is characterised in that: the following steps are included:

(1), the soil physical chemistry and meteorological data in a period of time in target farmland are collected；

(2), soil physical chemistry and meteorological data that step (1) is collected into are pre-processed, then again by pretreated data It is divided into training sample set and test sample collection in proportion；

(3), LSTM deep learning model is constructed, the LSTM deep learning model is hidden with an input layer, 25 Layer and an output layer are trained LSTM deep learning model by training sample set to obtain adjusting the LSTM depth after ginseng Learning model, then by test sample collection exchange ginseng after LSTM deep learning model verified, finally with verifying after LSTM deep learning model is as Forecast of Soil Moisture Content model；

(4), the soil physical chemistry and meteorological data collected step (1) pass through soil as the input of Forecast of Soil Moisture Content model Soil moisture content prediction model handles soil physical chemistry and meteorological data, finally by Forecast of Soil Moisture Content model output future time instance Forecast of Soil Moisture Content result.

2. a kind of Forecast of Soil Moisture Content method based on LSTM deep learning model according to claim 1, feature exist In: in step (1), the missing data of collected soil physical chemistry and meteorological data is repaired using linear interpolation method, wherein linearly The formula of interpolation method is as follows:

In formula (1), i and j respectively indicate the i-th moment and jth moment value, it is desirable that 0 < i < j；x_kAnd x_k+jIt is k moment and k+j respectively The soil physical chemistry and meteorological data of moment acquisition, x_k+iFor the soil physical chemistry and meteorological data lost when the k+i moment.

3. a kind of Forecast of Soil Moisture Content method based on LSTM deep learning model according to claim 1, feature exist In: the pretreatment in step (2) is normalization pretreatment, using wherein 85% data as training sample after normalization pretreatment Collection, as test sample collection, normalize pretreated formula is 15% data

By the way that pretreatment is normalized to collected data, make the mapping range [0,1] for the data value collected, normalizing Change in pretreated formula, x is initial data, x_max, x_minMaxima and minima respectively in initial data, x_nowTo return Result data after one change processing；

By normalization pretreatment to eliminate the dimension impact between index, the comparativity between data target, original number are solved According to after pretreatment, each index is in the same order of magnitude, facilitates the building of model.

4. a kind of Forecast of Soil Moisture Content method based on LSTM deep learning model according to claim 1, feature exist In: in step (3), the network structure of the LSTM deep learning model of building is (7,25,1), in LSTM deep learning model Tool is respectively adopted there are three the LSTM unit of door in each hidden layer, and three doors of the LSTM unit are to forget door, input gate respectively And out gate, the update of complete pair state and target value is exported jointly by three doors, the data processing of hidden layer Journey is as follows:

Forget door and determines the degree for forgeing information, first reading h^(t-1)And x^(t)To carry out Screening Treatment, wherein h to data^(t-1)Table That show is the output of a upper memory cell, x^(t)What is indicated is when precellular input, as shown in formula (2):

f^(t)=σ (W_f·[h^(t-1),x^(t)]+b_f) (2),

In formula (2), W_fIt is weight term, b_fIt is bias term, f^(t)It is the forgetting degree of information, σ is sigmoid function, and value is Between [0,1], sigmoid function exports the numerical value between 0 to 1 and is used for cell state C^(t)Update, wherein 1 indicate It indicates to give up this node data completely for information is fully retained, 0；

Input gate determines that new information is added in concealed nodes, wherein C^(t-1)It is the cell state of last moment, defines i^(t)For It determines the information updated, completes information addition and need to include two steps: firstly, passing through the sigmoid function of an input gate Determine which information needs to update；Secondly, by one vector of a tanh layer generation, that is, it is alternative interior for updating Hold a^(t), this two parts is joined together, a update is carried out to the state of cell, as shown in formula (3), (4):

i^(t)=σ (W_i·[h^(t-1),x^(t)]+b_i) (3),

a^(t)=tanh (W_c·[h^(t-1),x^(t)]+b_a) (4),

In formula 3, σ is sigmoid function, W_iFor weight term, h^(t-1)It is the part of last moment final output, x^(t)It is current The input of cell, b_iIt is bias term；

In formula 4, tanh is tanh function, W_cFor weight term, h^(t-1)It is the part of last moment final output, x^(t)It is current The input of cell, b_aIt is bias term；

In formula (4) when new and old cell state, by the content a of last moment update^(t-1)It is updated to the content a updated this moment^(t), the cell state C of last moment^(t-1)With the f in forgetting door^(t)It is multiplied, and adds i^(t)*a^(t), reach update cell state Effect, as shown in formula (5):

C^(t)=f^(t)*C^(t-1)+i^(t)*a^(t)(5),

In formula (5) formula, * indicates Hadamard product, the i.e. product of representing matrix corresponding position, C^(t-1)For the cell of last moment State, a^(t)For new content, C^(t)For new memory state；

Out gate determines output item, is primarily based on memory cell state, runs a sigmoid function to determine memory cell Which information will export；Secondly, memory cell state is handled by tanh, a value between -1 to 1 is obtained, And the value is multiplied with the output of out gate, as shown in formula (6) and formula (7):

In formula (6), o^(t)To export which information, h^(t-1),x^(t)It is expressed as the output of last moment and inputs this moment, W_oFor power Weight item, b_oFor bias term；

H in formula (7)^(t)For the part of final output, by o obtained in formula (6)^(t)Pass through multiplied by current new memory state The value of tanh function achievees the effect that the information for remembeing that sequence relies on for a long time.