CN112287294B - Space-time bidirectional soil water content interpolation method based on deep learning - Google Patents
Space-time bidirectional soil water content interpolation method based on deep learning Download PDFInfo
- Publication number
- CN112287294B CN112287294B CN202010950190.9A CN202010950190A CN112287294B CN 112287294 B CN112287294 B CN 112287294B CN 202010950190 A CN202010950190 A CN 202010950190A CN 112287294 B CN112287294 B CN 112287294B
- Authority
- CN
- China
- Prior art keywords
- water content
- space
- soil
- soil water
- time
- Prior art date
- Legal status (The legal status is an assumption and is not a legal conclusion. Google has not performed a legal analysis and makes no representation as to the accuracy of the status listed.)
- Active
Links
- 239000002689 soil Substances 0.000 title claims abstract description 106
- XLYOFNOQVPJJNP-UHFFFAOYSA-N water Substances O XLYOFNOQVPJJNP-UHFFFAOYSA-N 0.000 title claims abstract description 70
- 238000000034 method Methods 0.000 title claims abstract description 47
- 238000013135 deep learning Methods 0.000 title claims abstract description 15
- 230000002457 bidirectional effect Effects 0.000 title claims abstract description 9
- 230000007787 long-term memory Effects 0.000 claims abstract description 10
- 238000004088 simulation Methods 0.000 claims description 27
- 238000003062 neural network model Methods 0.000 claims description 18
- 238000012876 topography Methods 0.000 claims description 15
- 238000004422 calculation algorithm Methods 0.000 claims description 12
- 238000001556 precipitation Methods 0.000 claims description 11
- 238000004364 calculation method Methods 0.000 claims description 10
- 238000012549 training Methods 0.000 claims description 10
- 230000015654 memory Effects 0.000 claims description 6
- 238000011176 pooling Methods 0.000 claims description 6
- 238000012545 processing Methods 0.000 claims description 6
- 238000012795 verification Methods 0.000 claims description 6
- 238000012937 correction Methods 0.000 claims description 5
- 238000011478 gradient descent method Methods 0.000 claims description 5
- 238000010606 normalization Methods 0.000 claims description 5
- 238000005070 sampling Methods 0.000 claims description 5
- 230000006403 short-term memory Effects 0.000 claims description 5
- 230000002596 correlated effect Effects 0.000 claims description 3
- 210000003061 neural cell Anatomy 0.000 claims description 3
- 210000002569 neuron Anatomy 0.000 claims description 3
- 229920006395 saturated elastomer Polymers 0.000 claims description 3
- 238000010200 validation analysis Methods 0.000 claims description 3
- 238000013527 convolutional neural network Methods 0.000 claims description 2
- 238000013528 artificial neural network Methods 0.000 abstract description 9
- 230000004927 fusion Effects 0.000 abstract description 7
- 238000013136 deep learning model Methods 0.000 abstract 2
- 238000013459 approach Methods 0.000 abstract 1
- 238000005457 optimization Methods 0.000 abstract 1
- 230000001932 seasonal effect Effects 0.000 abstract 1
- 210000004027 cell Anatomy 0.000 description 9
- 230000006870 function Effects 0.000 description 6
- 238000012544 monitoring process Methods 0.000 description 6
- 230000004913 activation Effects 0.000 description 5
- 238000011160 research Methods 0.000 description 3
- 238000011156 evaluation Methods 0.000 description 2
- 238000010801 machine learning Methods 0.000 description 2
- 238000007637 random forest analysis Methods 0.000 description 2
- 238000004458 analytical method Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000000875 corresponding effect Effects 0.000 description 1
- 230000007547 defect Effects 0.000 description 1
- 238000001514 detection method Methods 0.000 description 1
- 230000007613 environmental effect Effects 0.000 description 1
- 239000000284 extract Substances 0.000 description 1
- 230000014759 maintenance of location Effects 0.000 description 1
- 244000005700 microbiome Species 0.000 description 1
- 230000001373 regressive effect Effects 0.000 description 1
- 230000001105 regulatory effect Effects 0.000 description 1
- 230000000717 retained effect Effects 0.000 description 1
- 239000000126 substance Substances 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F17/00—Digital computing or data processing equipment or methods, specially adapted for specific functions
- G06F17/10—Complex mathematical operations
- G06F17/18—Complex mathematical operations for evaluating statistical data, e.g. average values, frequency distributions, probability functions, regression analysis
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/045—Combinations of networks
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/048—Activation functions
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/04—Architecture, e.g. interconnection topology
- G06N3/049—Temporal neural networks, e.g. delay elements, oscillating neurons or pulsed inputs
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06N—COMPUTING ARRANGEMENTS BASED ON SPECIFIC COMPUTATIONAL MODELS
- G06N3/00—Computing arrangements based on biological models
- G06N3/02—Neural networks
- G06N3/08—Learning methods
Landscapes
- Engineering & Computer Science (AREA)
- Physics & Mathematics (AREA)
- Theoretical Computer Science (AREA)
- Data Mining & Analysis (AREA)
- General Physics & Mathematics (AREA)
- Mathematical Physics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Software Systems (AREA)
- General Engineering & Computer Science (AREA)
- Molecular Biology (AREA)
- Health & Medical Sciences (AREA)
- General Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- Artificial Intelligence (AREA)
- Computational Linguistics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Computing Systems (AREA)
- Computational Mathematics (AREA)
- Pure & Applied Mathematics (AREA)
- Mathematical Optimization (AREA)
- Mathematical Analysis (AREA)
- Evolutionary Biology (AREA)
- Databases & Information Systems (AREA)
- Algebra (AREA)
- Probability & Statistics with Applications (AREA)
- Operations Research (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The invention discloses a space-time bidirectional soil water content interpolation method based on deep learning, which comprises the following steps: constructing two deep learning models, namely a long-term memory LSMT and a full convolution neural network FCNN; taking the parameters of the underlying surfaces of remote sensing soil and aquatic products, meteorological elements, vegetation soil and the like as input data, respectively constructing an LSTM model from a time dimension by grids, constructing an FCNN model from a space dimension by days, and reconstructing soil water content data; and combining a Bayesian network weight self-learning model, and carrying out optimization fusion on soil water data obtained by two sets of different interpolation approaches in time and space. The invention fully considers the space-time three-dimensional characteristics of the data, integrates the seasonal change rule of the soil water on the time course and the geographic characteristics of the space, effectively establishes the internal relation between meteorological and underlying surface elements and the soil water content by using the deep learning model, and obviously improves the interpolation precision of the soil water content.
Description
Technical Field
The invention relates to a soil water content interpolation method, in particular to a space-time bidirectional soil water content interpolation method based on deep learning.
Background
Soil water is an important component of the earth surface system, provides an important environment for the growth of vegetation and microorganisms in the land ecosystem, and plays a key role in regulating the exchange of moisture, energy and substances in the soil-vegetation-atmosphere system. The monitoring of the soil moisture content in a large range is an important component part of hydrologic process research and flood and drought disaster analysis, and the inversion of the soil water environment in a regional scale and even in a global range is an essential parameter in land process mode research, so that the method is significant in improving regional and global climate mode prediction, predicting regional dry and wet conditions and monitoring environmental disaster research.
The method for measuring the water content of the soil mainly comprises contact site monitoring and non-contact remote sensing observation data inversion. The site monitoring is not affected by the atmosphere and vegetation, and the data precision is high; however, the number of sites is not large, a continuous and wide-range spatial data set cannot be formed, and meanwhile, the spatial variability of soil, topography and vegetation coverage is difficult to express. The defects are overcome greatly by remote sensing monitoring, the observation coverage range is large, the duration time is long, the earth surface change can be monitored dynamically in real time, the planar characteristics of the data can better describe the spatial distribution of the soil moisture content, and the spatial heterogeneity information is captured, so that the method is a main method for monitoring the spatial-temporal distribution and the change of the soil moisture in a large area; but the remote sensing data is difficult to meet the inversion of the water content of soil in a vegetation coverage area (the vegetation canopy cannot be penetrated), and is deeply influenced by the cloud coverage and the detection depth of a sensor, so that the problem of data loss exists. The common soil moisture content interpolation method mainly comprises a geostatistical spatial interpolation method and a machine learning algorithm, and comprises the following steps: inverse Distance Weighting (IDW), common Kriging (Kriging) and Regressive Kriging (RK), radial Basis Function (RBF), artificial Neural Network (ANN), random Forest (RF), etc.
The earth statistical spatial interpolation method only interpolates the soil water content from the spatial distance, does not consider the comprehensive influence of other factors on the soil water content, has low precision, and can greatly reduce the precision when the data quantity is small and the degree of the missing is large; the machine learning algorithm is commonly used for simulating and processing a system with a plurality of influence factors and complex relationship, can simulate and interpolate the soil moisture content from aspects of weather, vegetation, soil, topography and the like, has higher precision, is a common method for interpolating the soil moisture content at present, but can not well reflect the spatial distribution characteristics and time variation trend of the soil moisture content at the same time.
Disclosure of Invention
The invention aims to: aiming at the problems, the invention provides a space-time bidirectional soil water content interpolation method based on deep learning, which comprehensively considers the influence of the soil water content in the early stage and the spatial relation of the soil water content among grid points, and remarkably improves the accuracy of interpolation.
The technical scheme is as follows: the technical scheme adopted by the invention is a space-time bidirectional soil water content interpolation method based on deep learning, which comprises the following steps:
step one: and collecting meteorological elements, vegetation parameters, soil information, topography and topography conditions and ESA CCI data of the remote sensing soil water content, carrying out normalization processing, and dividing the data set into a training set and a verification set according to the proportion.
The meteorological elements refer to precipitation, air temperature, vapor emission, wind speed, sunlight, air humidity and atmospheric pressure, and the precipitation comprises current precipitation and early-stage precipitation; the vegetation parameters mainly comprise normalized vegetation indexes NDVI, leaf area indexes LAI and vegetation types; the soil information mainly comprises porosity, field water holding capacity, saturated soil water content and withering water content; the topography includes elevation, gradient and slope direction.
The normalization processing is to eliminate the dimensional influence among indexes, improve the solving convergence speed and improve the model training efficiency, and the calculation formula is as follows:
wherein x is i 、o i The original parameter value and the normalized parameter value are respectively; maxx i Is the maximum value of the parameter class; minx i Is the parameter class minimum.
Preferably, the data set is randomly split into training and validation sets in a ratio of 7:3.
Step two: and respectively constructing a long-term memory model and a full convolution neural network model.
The long-term and short-term memory neural network model consists of an input layer, an implicit layer and an output layer, wherein each neuron has 3 gate structures to simulate the memory process of the neural cells along with the time change: the first stage is forgetting gate, which decides which information needs to be forgotten from the cell state; the second stage is an input gate that determines which new information can be deposited into the cell state; the third stage is the output gate, which determines the output of information.
The full convolution neural network model adopts the structure of FCN-16s, the deconvolution step length is 16, and the up-sampling adopts the maximum pooling in the reverse pooling; the full convolution neural network model converts the tail full connection layer into a 1 multiplied by 1 convolution layer, outputs the same size output as the original input through deconvolution, up-sampling and skip structure, and retains the space information of the original input.
Step three: and inputting the normalized meteorological elements, vegetation parameters and remote sensing soil water content data into a long-short-period memory model time by time sequence, and operating to obtain the soil water content of each grid point of the time-by-time space.
Step four: and (3) inputting the normalized meteorological elements, vegetation parameters, soil information, topography conditions and the sequence of the remote sensing soil water content data arranged by the space grid to a full convolution neural network model, and operating to obtain the soil water content of the space grid-by-space grid point long-time sequence.
Step five: and respectively comparing the soil water content data set obtained by simulation of the two models with the original remote sensing soil water content and site soil water content in a verification manner, and evaluating the accuracy of the two algorithms.
The simulation precision of the two model algorithms is calculated by adopting a correlation coefficient and a root mean square error, and the calculation formula of the correlation coefficient CC is as follows:
in which W is sim (i) The simulated soil water content of the ith month; w (W) obs (i) The measured soil water content of the ith month;is the average value of the measured soil water content; />Is an average value of the water content of the simulated soil; n is a number of data;
the calculation formula of the root mean square error RMSE is as follows:
in which W is sim (i) The simulated soil water content of the ith month; w (W) obs (i) The measured soil water content of the ith month; n is a number of data.
Step six: and obtaining weights of interpolation results of the two models by a weight self-learning method based on the Bayesian network.
The weight self-learning method is an algorithm based on a Bayesian network, and continuously approximates to a weight value through a reverse error propagation and gradient descent method, and comprises the following specific steps of: (1) subjectively assigning a weight as an initial weight; (2) constructing a Bayesian network, and setting a target error value; (3) Continuously correcting the weight by a back propagation method and a gradient descent method until the error is smaller than the target value; (4) outputting the weight value.
Step seven: and carrying out fusion of the spatial interpolation results according to the correlation coefficient and the root mean square error correction weight.
Based on the model accuracy evaluation result (correlationCoefficient and root mean square error), the weight calculated by the Bayesian network weight self-learning method is corrected, the purpose of the correction is to correct interpolation results with low simulation precision and high weight, and the specific steps are as follows: (1) Selecting grid points with poorer simulation results (the correlation coefficient CC is smaller than 25% quantiles thereof or the RMSE is larger than 75% quantiles thereof) and the weight is larger than or equal to a set threshold value; (2) Comparing whether the interpolation result weight w is positively correlated with the simulation precision or not according to the grids screened in the previous step, and if not, entering the next step without correcting the grid weight coefficient; (3) For grid m needing to correct the weight coefficient, searching grid j similar to the simulation precision of grid m, and meeting the requirement of |CC tm -CC tj |≤δ 1 And |CC sm -CC sj |≤δ 1 Wherein delta 1 To set a threshold for correlation coefficient, or |rmse tm -RMSE tj |≤δ 2 And |RMSE sm -RMSE sj |≤δ 2 Wherein delta 2 To set the threshold for root mean square error, let w tm ＝w tj ，w sm ＝w sj If the number of the grids j meeting the condition is not unique, let w tm 、w sm Respectively equal to the average value of the space weight coefficients when each grid is used.
The fusion of the space-time interpolation results of the site soil moisture content refers to the fusion of the time and space interpolation results of the soil moisture content, and the calculation formula of the soil moisture content W is as follows: w=0.5× (α t ×y t +α s ×y s )+0.5×(β t ×y t +β s ×y s ) The method comprises the steps of carrying out a first treatment on the surface of the Wherein y is t Time-based interpolation results obtained for long-term and short-term memory neural network models, ys is space-based interpolation results obtained for full convolution neural network models, alpha t And alpha s Weights, beta, for two interpolation results at different points in time for each grid point t And beta s Is the weight of two interpolation results of different grid points at the same time point.
The beneficial effects are that: compared with the prior art, the method comprehensively analyzes the influence of meteorological factors, vegetation, soil and topography factors on the water content of the soil from the space-time dimension, considers the influence of the water content of the soil in the early stage on the current time, and simultaneously considers the spatial relation of the water content of the soil among grid points, so that the model simulation result of the water content of the soil is more accurate. The invention adopts the long-short-period memory neural network and the full convolution neural network which are commonly used in deep learning, effectively extracts the characteristics of each influencing element, and deeply analyzes the internal relation between the characteristics and the soil water content, thereby achieving a better simulation result. The weight self-learning method based on the Bayesian network can acquire the weights of the two algorithms through self-learning, and provides an effective path for space-time fusion of interpolation results. The method and the device remarkably improve the interpolation precision of the soil water content, and have important significance for interpolation filling of missing soil water content data.
Drawings
Fig. 1 is a flow chart of a space-time bidirectional soil moisture content interpolation method based on deep learning according to the invention.
Detailed Description
The technical scheme of the invention is further described below with reference to the accompanying drawings and examples.
The flow chart of the space-time bidirectional soil water content interpolation method based on the deep learning is shown in figure 1. The method comprises the following steps:
step one: and collecting meteorological elements, vegetation parameters, soil information, topography and topography conditions and ESA CCI data of remote sensing soil water content, carrying out normalization processing, and splitting the data set into a training set and a verification set.
The meteorological elements mainly refer to precipitation (including current precipitation and early precipitation), air temperature, vapor emission, wind speed, sunlight, air humidity and atmospheric pressure; the vegetation parameters mainly comprise normalized vegetation indexes NDVI, leaf area indexes LAI and vegetation types; the soil information mainly comprises porosity, field water holding capacity, saturated soil water content and withering water content; the topography includes elevation, gradient and slope direction.
The data set is randomly split into a training set and a verification set according to the proportion of 7:3, and the representativeness and consistency of the data distribution in the training set are ensured.
The normalized formula in the first step is:
wherein x is i 、o i The original parameter value and the normalized parameter value are respectively; maxx i Is the maximum value of the parameter class; minx i Is the parameter class minimum.
Step two: respectively constructing a long-term memory neural network model LSMT and a full convolution neural network model FCNN;
the long-term memory neural network model is composed of an input layer, an implicit layer and an output layer, and is different from the traditional neural network in that each neuron has 3 gate structures to simulate the memory process of the neural cell changing along with time, namely the cell state: the first stage is forgetting gate, which decides which information needs to be forgotten from the cell state; the next stage is the input gate, which determines which new information can be deposited into the cell state; the last stage is the output gate, which determines the output of information.
The forgetting gate takes the output of the previous layer and the sequence data to be input in the layer as inputs, and obtains the output as f through an activation function sigmoid t 。f t The output of (1) is equal to [0,1 ]]The interval indicates the probability that the cell state of the previous layer is forgotten, 1 is "complete retention", and 0 is "complete rejection". The input gate comprises two parts, the first part uses a sigmoid activation function and outputs as i t The second part uses the tanh activation function and outputs as Two output multiplications for the input gates indicate how much new information is retained. Updating information of cell state to C t . The output gate is used to control how much of the cell state of the layer is filtered. First, a sigmoid activation function is used to obtain a [0,1 ]]Interval value o t Then, the cell state C t Processed by tanh activation function and then processed by o t Multiplication, i.e. the output h of the layer t 。
The mathematical expression is as follows:
f t ＝σ(W f ×[h t-1 ，x t ]+b f )
i t ＝σ(W i ×[h t-1 ，x t ]+b i )
o t ＝σ(W o ×[h t-1 ，x t ]+b o )
h t ＝o t ×tanh(C t )
wherein h is t-1 For the output of the upper layer, x t For this layer of input, W, b is the weight and bias.
The full convolution neural network model is different from the classical convolution neural network that adopts the full connection layer to output the feature vector with fixed length, the FCNN converts the tail full connection layer into the 1X 1 convolution layer, and outputs the output with the same size as the original input through deconvolution, up-sampling and skip structure, so that the spatial information of the original input is reserved.
The full convolution neural network adopts the structure of FCN-16s, and the deconvolution step length is 16. The upsampling uses the largest pooling of the inverse pooling.
Step three: inputting the normalized meteorological elements, vegetation parameters and ESA CCI data into an LSMT model time by time sequence, and operating to obtain the soil water content of each grid point of the time-by-time space;
step four: the normalized meteorological elements, vegetation parameters, soil information, topography conditions and ESA CCI data are input into the FCNN model in the sequence of space grid by space grid arrangement, and the soil water content of the space grid by space grid point long-time sequence is operated and obtained;
step five: respectively verifying and comparing the soil water content data set obtained by simulation of the two models with the original remote sensing soil water content and the soil water content of each site, and evaluating the accuracy of the two algorithms;
the simulation accuracy in the fifth step is verified through the correlation coefficient CC and the root mean square error RMSE.
The correlation coefficient CC is used for reflecting the degree of closeness of the correlation between the simulation result and the measured result, the closer the value is to 1, the higher the degree of correlation between the simulation result and the measured result is, the higher the simulation precision is, and the calculation formula is as follows:
in which W is sim (i) The simulated soil moisture content for month i; w (W) obs (i) The measured soil moisture content for month i;is the average value of the measured soil water content; />Is an average value of the simulated soil moisture content; n is a number of data.
The root mean square error RMSE is used for reflecting the discrete degree between the simulation result and the actually measured result, the smaller the value is, the smaller the discrete degree between the simulation result and the actually measured result is, the higher the simulation precision is, and the calculation formula is as follows:
in which W is sim (i) The simulated soil moisture content for month i; w (W) obs (i) The measured soil moisture content for month i; n is a number of data.
Step six: and obtaining weights of interpolation results of the two models by a weight self-learning method based on the Bayesian network.
The weight self-learning method is an algorithm based on a Bayesian network, and continuously approximates weight values through a reverse error propagation and gradient descent method, and comprises the following specific steps: (1) subjectively assigning a weight as an initial weight; (2) constructing a Bayesian network, and setting a target error value; (3) Continuously correcting the weight by a back propagation method and a gradient descent method until the error is smaller than the target value; (4) outputting the weight value.
Step seven: and carrying out fusion of the spatial interpolation results according to the correlation coefficient and the root mean square error correction weight.
According to the model precision evaluation result (correlation coefficient and root mean square error), the weight calculated by the Bayesian network weight self-learning method is corrected, and the purpose is to correct the interpolation result with low simulation precision and high weight, which comprises the following specific steps: (1) Grid points with poorer simulation results (the correlation coefficient CC is smaller than 25% quantile thereof or the root mean square error RMSE is larger than 75% quantile thereof) and the weight is larger than or equal to 0.6 are selected; (2) For the grids screened in the previous step, whether the interpolation result weight w is positively correlated with the simulation precision is compared, for example, the time simulation precision is higher than the space simulation precision and the time simulation result weight is larger (CC) t ≥CC s And w is t ≥w s ) If the grid weight coefficients are inconsistent, the grid weight coefficients are considered to be unreasonable, and the grid weight coefficients are required to be corrected; (3) For unreasonable weight coefficients, searching grids meeting the conditions, for example, a certain unreasonable weight grid m, and CC thereof tm ＜CC sm But w is tm ≥w sm If the weight coefficient needs to be corrected, traversing the search grid j in all grids to meet the requirement of |CC tm -CC tj I is less than or equal to 0.1 and CC sm -CC sj I is less than or equal to 0.1, let w tm ＝w tj ，w sm ＝w sj It is noted that if the number of j is not unique, let w ti Equal to the mean of the weight coefficients thereof. For the root mean square error RMSE, the threshold value of the traversal grid is set to be 0.01, i.e. the |rmse is satisfied tm -RMSE tj I is less than or equal to 0.01 and I is RMSE sm -RMSE sj And (5) transplanting corresponding weight coefficients if the I is less than or equal to 0.01.
The fusion of space-time interpolation results is to consider time and space factors at the same time, utilize a weight self-learning method based on Bayesian network, take site soil moisture content as verification, calculate the two interpolation results according to time sequence and space sequence to obtain the weight w of each space grid point at each time point, and obtain the weights alpha of the time and space interpolation results at different time points at each grid point after correction t And alpha s Weights beta of different grid points at the same time point t And beta s . Then the soil moisture content w=0.5× (α) for different nodes of different grids at different times t ×y t +α s ×y s )+0.5×(β t ×y t +β s ×y s ). Wherein y is t Time-based interpolation results, y, obtained for long-term and short-term memory neural network LSTM s The spatial-based interpolation results obtained for the fully convolutional neural network FCNN.
Claims (6)
1. A space-time bidirectional soil water content interpolation method based on deep learning is characterized by comprising the following steps:
step one: collecting meteorological elements, vegetation parameters, soil information, topography and topography conditions and remote sensing soil water content data, carrying out normalization processing, and dividing a data set into a training set and a verification set according to a proportion;
step two: respectively constructing a long-term memory neural network model and a full convolution neural network model;
step three: inputting the normalized meteorological elements, vegetation parameters and remote sensing soil water content data into a long-short-period memory model time by time sequence, and operating to obtain the soil water content of each grid point of the time-by-time space;
step four: the normalized meteorological elements, vegetation parameters, soil information, topography conditions and the sequence of the remote sensing soil water content data arranged by space grids are input into a full convolution neural network model, and the soil water content of the space grids is operated and obtained;
step five: respectively verifying and comparing the soil water content data sets obtained by simulation of the two models with the original remote sensing soil water content and site soil water content, and calculating model accuracy of two model algorithms;
step six: acquiring weights of interpolation results of two models by a weight self-learning method based on a Bayesian network;
step seven: correcting the weight according to the model precision of the two model algorithms, and fusing the time-space interpolation results;
the long-term and short-term memory neural network model in the second step consists of an input layer, an hidden layer and an output layer, wherein each neuron has 3 gate structures to simulate the memory process of the neural cells along with the time change: the first stage is forgetting a door; the second stage is an input gate; the third stage is an output gate;
the full convolution neural network model in the second step adopts the structure of FCN-16s, and the up-sampling adopts the maximum pooling in the reverse pooling; the full convolution neural network model converts the tail full connection layer into a 1 multiplied by 1 convolution layer, outputs the same size output as the original input through deconvolution, up-sampling and skip structure, and retains the space information of the original input;
the weight self-learning method based on the Bayesian network in the step six obtains the weights of the interpolation results of the two models, and the method comprises the following steps: (1) subjectively assigning a weight as an initial weight; (2) constructing a Bayesian network, and setting a target error value; (3) Continuously correcting the weight by a back propagation method and a gradient descent method until the error is smaller than the target value; (4) outputting a weight value;
the model accuracy correction weight according to the two model algorithms described in the seventh step includes the following steps: (1) Selecting grid points with poor simulation precision and weight greater than or equal to a set threshold value; (2) Judging whether the interpolation result weight w is positively correlated with the simulation precision according to the grids screened in the previous step, if so, not correcting the weight w, otherwise, entering the next step; (3) For the grid m needing to correct the weight coefficient, searching a grid j with similar simulation precision to the grid m, namely meeting the requirement of |CC tm -CC tj |≤δ 1 And |CC sm -CC sj |≤δ 1 Wherein delta 1 To set a threshold for correlation coefficient, or |rmse tm -RMSE tj |≤δ 2 And |RMSE sm -RMSE sj |≤δ 2 Wherein delta 2 To set the threshold for root mean square error, let w tm ＝w ti ，w sm ＝w sj If the number of the grids j meeting the condition is not unique, let w tm 、w sm Respectively equal to the average value of the space weight coefficients of each grid;
the step seven of fusing the space-time interpolation results refers to fusing the time and space interpolation results of the soil water content, and the calculation formula of the soil water content W is as follows: w=0.5× (α t ×y t +α s ×y s )+0.5×(β t ×y t +β s ×y s ) The method comprises the steps of carrying out a first treatment on the surface of the Wherein y is t Time-based interpolation results, y, obtained for long-term and short-term memory neural network models s Space-based interpolation result, alpha, for a full convolutional neural network model t And alpha s Weights, beta, for two interpolation results at different points in time for each grid point t And beta s Is the weight of two interpolation results of different grid points at the same time point.
2. The deep learning-based spatio-temporal bi-directional soil moisture content interpolation method of claim 1, wherein: the meteorological elements in the first step are precipitation, air temperature, vapor emission, wind speed, sunlight, air humidity and atmospheric pressure, wherein the precipitation comprises current precipitation and early-stage precipitation; the vegetation parameters mainly comprise normalized vegetation indexes, leaf area indexes and vegetation types; the soil information mainly comprises porosity, field water holding capacity, saturated soil water content and withering water content; the topography includes elevation, gradient and slope direction.
3. The deep learning-based spatio-temporal bi-directional soil moisture content interpolation method of claim 1, wherein: the normalization processing in the first step is to eliminate the dimensional influence among indexes, and improve the convergence rate of solving and the training efficiency of the model, and the calculation formula is as follows:
wherein x is i 、o i The original parameter value and the normalized parameter value are respectively; maxx i Is the maximum value of the parameter class; minx i Is the parameter class minimum.
4. The deep learning-based space-time two-way soil moisture content interpolation method according to claim 1, wherein the step one of dividing the data set into the training set and the validation set according to the proportion is to divide the data set into 7:3 into training and validation sets.
5. The deep learning-based spatio-temporal bi-directional soil moisture content interpolation method of claim 1, wherein: in the fifth step, the simulation precision of the two model algorithms is calculated by adopting a correlation coefficient and a root mean square error, and the calculation formula of the correlation coefficient CC is as follows:
in which W is sim (i) The simulated soil moisture content for month i; w (W) obs (i) The measured soil moisture content for month i;is the average value of the measured soil water content; />Is an average value of the simulated soil moisture content; n is a number of data;
the calculation formula of the root mean square error RMSE is as follows:
in which W is sim (i) The simulated soil moisture content for month i; w (W) obs (i) The measured soil moisture content for month i; n is a number of data.
6. The deep learning-based spatio-temporal bi-directional soil moisture content interpolation method of claim 1, wherein the analog precision difference is that the correlation coefficient CC is less than 25% quantile thereof or the root mean square error RMSE is greater than 75% quantile thereof.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010950190.9A CN112287294B (en) | 2020-09-10 | 2020-09-10 | Space-time bidirectional soil water content interpolation method based on deep learning |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202010950190.9A CN112287294B (en) | 2020-09-10 | 2020-09-10 | Space-time bidirectional soil water content interpolation method based on deep learning |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN112287294A CN112287294A (en) | 2021-01-29 |
| CN112287294B true CN112287294B (en) | 2024-02-27 |
Family
ID=74419768
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202010950190.9A Active CN112287294B (en) | 2020-09-10 | 2020-09-10 | Space-time bidirectional soil water content interpolation method based on deep learning |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN112287294B (en) |
Families Citing this family (7)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN113888386B (en) * | 2021-12-03 | 2022-02-15 | 中国科学院、水利部成都山地灾害与环境研究所 | High-resolution time-space seamless earth surface soil moisture estimation method and system |
| CN114461971B (en) * | 2022-01-13 | 2024-04-16 | 桂林理工大学 | Earth surface soil water content inversion method integrating soil physical properties and remote sensing data |
| CN115203639B (en) * | 2022-06-21 | 2023-03-10 | 中国长江三峡集团有限公司 | Irregular grid surface rainfall calculation method and system based on matrix operation |
| CN115544875B (en) * | 2022-09-28 | 2023-09-12 | 中国科学院地理科学与资源研究所 | Soil moisture reconstruction method, device and equipment based on multi-rain cloud area |
| CN115795401B (en) * | 2023-02-08 | 2023-04-21 | 青岛海洋地质研究所 | Multi-data fusion system of marine pasture full-element monitoring sensor |
| CN117216480B (en) * | 2023-09-18 | 2024-06-28 | 宁波大学 | Near-surface ozone remote sensing estimation method for deep coupling geographic space-time information |
| CN117556695B (en) * | 2023-11-11 | 2024-05-14 | 水利部交通运输部国家能源局南京水利科学研究院 | Crop root soil water content simulation method based on deep learning |
Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| CN109635246A (en) * | 2018-12-06 | 2019-04-16 | 西南交通大学 | A kind of multiattribute data modeling method based on deep learning |
| CN110084367A (en) * | 2019-04-19 | 2019-08-02 | 安徽农业大学 | A kind of Forecast of Soil Moisture Content method based on LSTM deep learning model |
| WO2020047739A1 (en) * | 2018-09-04 | 2020-03-12 | 安徽中科智能感知大数据产业技术研究院有限责任公司 | Method for predicting severe wheat disease on the basis of multiple time-series attribute element depth features |
Family Cites Families (1)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US11455807B2 (en) * | 2018-09-20 | 2022-09-27 | Nvidia Corporation | Training neural networks for vehicle re-identification |
-
2020
- 2020-09-10 CN CN202010950190.9A patent/CN112287294B/en active Active
Patent Citations (3)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| WO2020047739A1 (en) * | 2018-09-04 | 2020-03-12 | 安徽中科智能感知大数据产业技术研究院有限责任公司 | Method for predicting severe wheat disease on the basis of multiple time-series attribute element depth features |
| CN109635246A (en) * | 2018-12-06 | 2019-04-16 | 西南交通大学 | A kind of multiattribute data modeling method based on deep learning |
| CN110084367A (en) * | 2019-04-19 | 2019-08-02 | 安徽农业大学 | A kind of Forecast of Soil Moisture Content method based on LSTM deep learning model |
Non-Patent Citations (3)
| Title |
|---|
| 基于贝叶斯融合的时空流异常行为检测模型;陈莹;何丹丹;;电子与信息学报(第05期);全文 * |
| 基于遥感图像的土壤含水量信息提取研究;崔丽霞;王蕾;;唐山学院学报(第06期);全文 * |
| 陆面水文模式Noah-LSM在赣江上游区的适用性研究;田慧敏;袁飞;姚新宇;张利敏;章益棋;;水电能源科学(第05期);全文 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN112287294A (en) | 2021-01-29 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| CN112287294B (en) | Space-time bidirectional soil water content interpolation method based on deep learning | |
| CN110084367B (en) | Soil moisture content prediction method based on LSTM deep learning model | |
| CN111310889B (en) | Evaporation waveguide profile estimation method based on deep neural network | |
| CN112288164B (en) | Wind power combined prediction method considering spatial correlation and correcting numerical weather forecast | |
| CN110232471B (en) | Rainfall sensor network node layout optimization method and device | |
| CN113553764B (en) | Mountain fire prediction method based on deep learning network | |
| CN105243435B (en) | A kind of soil moisture content prediction technique based on deep learning cellular Automation Model | |
| CN108064047B (en) | Water quality sensor network optimization deployment method based on particle swarm | |
| Marofi et al. | Predicting spatial distribution of snow water equivalent using multivariate non-linear regression and computational intelligence methods | |
| CN111783987A (en) | Farmland reference crop evapotranspiration prediction method based on improved BP neural network | |
| CN110020712B (en) | Optimized particle swarm BP network prediction method and system based on clustering | |
| CN115099500B (en) | Water level prediction method based on weight correction and DRSN-LSTM model | |
| CN112163375A (en) | Long-time sequence near-surface ozone inversion method based on neural network | |
| CN107133686A (en) | City-level PM2.5 concentration prediction methods based on Spatio-Temporal Data Model for Spatial | |
| CN116415730A (en) | Fusion self-attention mechanism time-space deep learning model for predicting water level | |
| CN114723188A (en) | Water quality prediction method, device, computer equipment and storage medium | |
| Gauch et al. | Data-driven vs. physically-based streamflow prediction models | |
| Ahmadi et al. | Input data selection for solar radiation estimation | |
| CN118350678B (en) | Water environment monitoring data processing method and system based on Internet of things and big data | |
| CN114004163A (en) | PM2.5 inversion method based on MODIS and long-and-short-term memory network model | |
| CN115877483A (en) | Typhoon path forecasting method based on random forest and GRU | |
| Cui et al. | Bayesian optimization of typhoon full-track simulation on the Northwestern Pacific segmented by QuadTree decomposition | |
| CN112560633B (en) | Plant key matter waiting period time point prediction method and system based on deep learning | |
| CN117494419A (en) | Multi-model coupling drainage basin soil erosion remote sensing monitoring method | |
| CN115392128B (en) | Method for simulating river basin runoff by utilizing space-time convolution LSTM network |
Legal Events
| Date | Code | Title | Description |
|---|---|---|---|
| PB01 | Publication | ||
| PB01 | Publication | ||
| SE01 | Entry into force of request for substantive examination | ||
| SE01 | Entry into force of request for substantive examination | ||
| GR01 | Patent grant | ||
| GR01 | Patent grant |