CN116881664A - Soil texture prediction method based on local-global dependency relationship - Google Patents
Soil texture prediction method based on local-global dependency relationship Download PDFInfo
- Publication number
- CN116881664A CN116881664A CN202310950055.8A CN202310950055A CN116881664A CN 116881664 A CN116881664 A CN 116881664A CN 202310950055 A CN202310950055 A CN 202310950055A CN 116881664 A CN116881664 A CN 116881664A
- Authority
- CN
- China
- Prior art keywords
- covariates
- soil
- soil texture
- model
- environment
- 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.)
- Granted
Links
- 239000002689 soil Substances 0.000 title claims abstract description 149
- 238000000034 method Methods 0.000 title claims abstract description 37
- 238000012549 training Methods 0.000 claims abstract description 13
- 238000012216 screening Methods 0.000 claims abstract description 4
- 239000004927 clay Substances 0.000 claims description 18
- 239000004576 sand Substances 0.000 claims description 15
- 238000013507 mapping Methods 0.000 claims description 10
- 230000007613 environmental effect Effects 0.000 claims description 9
- 230000006870 function Effects 0.000 claims description 7
- 239000011159 matrix material Substances 0.000 claims description 7
- 239000013598 vector Substances 0.000 claims description 5
- 238000004458 analytical method Methods 0.000 claims description 4
- 230000004913 activation Effects 0.000 claims description 3
- 238000012876 topography Methods 0.000 claims description 3
- 238000013527 convolutional neural network Methods 0.000 description 6
- 230000008859 change Effects 0.000 description 5
- 230000003993 interaction Effects 0.000 description 5
- 238000012360 testing method Methods 0.000 description 5
- 238000013136 deep learning model Methods 0.000 description 4
- 230000008569 process Effects 0.000 description 4
- 238000013135 deep learning Methods 0.000 description 3
- 230000006872 improvement Effects 0.000 description 3
- 238000010801 machine learning Methods 0.000 description 3
- 230000008033 biological extinction Effects 0.000 description 2
- 230000000694 effects Effects 0.000 description 2
- 238000012407 engineering method Methods 0.000 description 2
- 238000011160 research Methods 0.000 description 2
- 238000013528 artificial neural network Methods 0.000 description 1
- 230000009286 beneficial effect Effects 0.000 description 1
- 230000008901 benefit Effects 0.000 description 1
- 238000004364 calculation method Methods 0.000 description 1
- 238000006243 chemical reaction Methods 0.000 description 1
- 235000009508 confectionery Nutrition 0.000 description 1
- 230000002596 correlated effect Effects 0.000 description 1
- 238000002790 cross-validation Methods 0.000 description 1
- 238000011161 development Methods 0.000 description 1
- 238000010586 diagram Methods 0.000 description 1
- 238000011156 evaluation Methods 0.000 description 1
- 238000011835 investigation Methods 0.000 description 1
- 230000007774 longterm Effects 0.000 description 1
- 230000014759 maintenance of location Effects 0.000 description 1
- 238000012986 modification Methods 0.000 description 1
- 230000004048 modification Effects 0.000 description 1
- 238000012544 monitoring process Methods 0.000 description 1
- 238000005215 recombination Methods 0.000 description 1
- 230000006798 recombination Effects 0.000 description 1
- 230000000306 recurrent effect Effects 0.000 description 1
- 230000009467 reduction Effects 0.000 description 1
- 230000008521 reorganization Effects 0.000 description 1
- 230000009897 systematic effect Effects 0.000 description 1
- 238000013076 uncertainty analysis Methods 0.000 description 1
- 238000010200 validation analysis Methods 0.000 description 1
- 238000012795 verification Methods 0.000 description 1
Classifications
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06F—ELECTRIC DIGITAL DATA PROCESSING
- G06F18/00—Pattern recognition
- G06F18/20—Analysing
-
- 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
- G06N3/044—Recurrent networks, e.g. Hopfield networks
- G06N3/0442—Recurrent networks, e.g. Hopfield networks characterised by memory or gating, e.g. long short-term memory [LSTM] or gated recurrent units [GRU]
-
- 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
-
- G—PHYSICS
- G06—COMPUTING; CALCULATING OR COUNTING
- G06Q—INFORMATION AND COMMUNICATION TECHNOLOGY [ICT] SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES; SYSTEMS OR METHODS SPECIALLY ADAPTED FOR ADMINISTRATIVE, COMMERCIAL, FINANCIAL, MANAGERIAL OR SUPERVISORY PURPOSES, NOT OTHERWISE PROVIDED FOR
- G06Q50/00—Information and communication technology [ICT] specially adapted for implementation of business processes of specific business sectors, e.g. utilities or tourism
- G06Q50/10—Services
- G06Q50/26—Government or public services
-
- Y—GENERAL TAGGING OF NEW TECHNOLOGICAL DEVELOPMENTS; GENERAL TAGGING OF CROSS-SECTIONAL TECHNOLOGIES SPANNING OVER SEVERAL SECTIONS OF THE IPC; TECHNICAL SUBJECTS COVERED BY FORMER USPC CROSS-REFERENCE ART COLLECTIONS [XRACs] AND DIGESTS
- Y02—TECHNOLOGIES OR APPLICATIONS FOR MITIGATION OR ADAPTATION AGAINST CLIMATE CHANGE
- Y02A—TECHNOLOGIES FOR ADAPTATION TO CLIMATE CHANGE
- Y02A90/00—Technologies having an indirect contribution to adaptation to climate change
- Y02A90/10—Information and communication technologies [ICT] supporting adaptation to climate change, e.g. for weather forecasting or climate simulation
Landscapes
- Engineering & Computer Science (AREA)
- Theoretical Computer Science (AREA)
- Physics & Mathematics (AREA)
- Data Mining & Analysis (AREA)
- Business, Economics & Management (AREA)
- General Physics & Mathematics (AREA)
- Life Sciences & Earth Sciences (AREA)
- Artificial Intelligence (AREA)
- General Health & Medical Sciences (AREA)
- Health & Medical Sciences (AREA)
- Evolutionary Computation (AREA)
- General Engineering & Computer Science (AREA)
- Computational Linguistics (AREA)
- Mathematical Physics (AREA)
- Biomedical Technology (AREA)
- Biophysics (AREA)
- Tourism & Hospitality (AREA)
- Software Systems (AREA)
- Molecular Biology (AREA)
- Computing Systems (AREA)
- Development Economics (AREA)
- Bioinformatics & Computational Biology (AREA)
- Bioinformatics & Cheminformatics (AREA)
- Evolutionary Biology (AREA)
- Educational Administration (AREA)
- Computer Vision & Pattern Recognition (AREA)
- Economics (AREA)
- Human Resources & Organizations (AREA)
- Marketing (AREA)
- Primary Health Care (AREA)
- Strategic Management (AREA)
- General Business, Economics & Management (AREA)
- Management, Administration, Business Operations System, And Electronic Commerce (AREA)
Abstract
The application provides a soil texture prediction method based on local-global dependency relationship, which comprises the following steps: fitting the soil texture of the multi-layer soil depth by an equal-area secondary spline method; acquiring a main environment covariate, and screening out a relevant environment covariate with the absolute value of a pearson coefficient of soil texture in the whole soil section of each layer of soil depth being more than 0.02; re-ordering the related environment covariates according to the degree of correlation between the covariates and the soil texture to generate a first model input set; selecting the most relevant environment covariates with highest covariates and soil texture relativity, and sequentially ordering the rest relevant environment covariates according to the sequence with highest adjacent covariates relativity to generate a second model input set; and training a two-layer LSTM model by using the first model input set and the second model input set, establishing a relation model between the soil texture and the environment covariates by applying a full-connection layer FCN to the training result, and predicting the soil texture by using the relation model.
Description
Technical Field
The application relates to the technical field of machine deep learning, in particular to a soil texture prediction method based on local-global dependency relationship.
Background
Accurate and effective soil texture mapping is critical to agricultural development and environmental activities. Current Digital Soil Mapping (DSM) methods lack accurate predictions due to the high degree of spatial heterogeneity of soil texture. Although feature engineering methods have been used to improve drawing accuracy by capturing complex earth-forming relationships, the calculation is time-consuming. Furthermore, the "discrete" features generated by the feature engineering method cannot reflect the interactions (or dependencies) of the environmental covariates. Thus, predictions of soil texture may be affected.
Disclosure of Invention
In order to solve the problems in the prior art, the application provides a new local-global dependency long-short-term memory DL (deep learning) model, firstly, a covariate reconstruction method is designed to generate a plurality of groups of inputs, then a multi-layer long-short-term memory model (LSTM) is adopted to extract the inherent correlation of environment covariates, and finally, the prediction is realized through a full-connection layer, so that the soil texture prediction of different soil depths is improved.
Specifically, the application provides a soil texture prediction method based on local-global dependency relationship, which comprises the following steps:
fitting the soil texture of the multi-layer soil depth by an equal-area secondary spline method;
acquiring a main environment covariate, screening out a related environment covariate with the absolute value of a pearson coefficient of soil texture being more than 0.02 in the whole soil section of each layer of soil depth, and removing redundant environment covariates affecting the prediction performance by using co-linearity analysis;
re-ordering the related environment covariates according to the relevance of the covariates and the soil texture according to the Pearson coefficient to generate a first model input set; selecting the most relevant environment covariates with highest covariates and soil texture relativity, and sequentially ordering the rest relevant environment covariates according to the sequence with highest adjacent covariates relativity to generate a second model input set;
and training two layers of long-short-term memory LSTM models by using the first model input set and the second model input set, establishing a relation model between soil texture and environment covariates by applying a full-connection layer FCN to training results, and predicting the soil texture by using the relation model.
As a further illustration of the present application, the soil texture of the multi-layered soil depth comprises 6 standard soil layer soil textures of 0-5cm, 5-15cm, 15-30cm, 30-60cm, 60-100cm and 100-200 cm; the soil texture is soil composition data with the sum of clay, sand and silt content being 100%.
As a further illustration of the present application, the environmental covariates include five soil texture influencing factors, climate, biology, topography, matrix and soil category.
As a further illustration of the present application, the selecting the most relevant environment covariates with the highest relevance between the covariates and the soil texture, and the remaining relevant environment covariates are sequentially ordered according to the order with the highest relevance between the adjacent covariates to generate a second model input set, including:
selecting the most relevant environment covariates with highest covariates and soil texture relativity as the model input data set of the last dimension;
calculating the covariates with the highest correlation degree with the most relevant environment covariates in other covariates to form secondary relevant environment covariates, and taking the secondary relevant environment covariates as a model input data set of the penultimate dimension;
and sequentially calculating covariates with highest correlation degree with the model input data set of the previous dimension, taking the covariates as the model input data set of the current dimension, and finally generating all second model input sets.
As a further illustration of the present application, the LSTM model is capable of calculating a mapping from environmental covariates to soil texture as follows:
i(n)＝σ(W ih h(n-1)+W ix x(n)+W ix c(n-1)+b i ) (1)
f(n)＝σ(W fh h(n-1)+W fx x(n)+W fx c(n-1)+b f ) (2)
o(n)＝σ(W oh h(n-1)+W ox x(n)+W ox c(n)+b o ) (4)
y＝W yh h(n)+b o (6)
where W is the weight matrix of the hidden state for the nth time step, b is the bias, σ (deg.) is the nonlinear activation function, tanh (deg.) is the hyperbolic tangent function,representing the multiplication of two vectors, x (t) and +.>Representing the input and output data, respectively, for the nth time step.
As a further illustration of the present application, the hyper-parameters recommended in the relational model are: epoch is 1000, hidden unit size is 128, sample data is 128, learning rate is 1e-4, time step is 5, LSTM layer number is 2.
Compared with the prior art, the application has the following beneficial technical effects:
the prediction method provided by the application improves the prediction capability of soil texture on the basis of considering inherent environment covariate dependency, so that the improved machine deep learning model provided by the application is helpful for more accurately and economically predicting the soil texture of China, ensures the applicability of the model in monitoring the soil texture change, and provides a new choice in use for a Digital Soil Mapping (DSM) method, in particular to an integrated machine learning model.
Drawings
FIG. 1 is a flow chart of two recombination methods provided by the present application;
fig. 2 is a diagram of an improved predictive model framework provided in accordance with the present application.
Detailed Description
In the following description, for purposes of explanation, numerous specific details are set forth in order to provide a thorough understanding of one or more embodiments. It may be evident, however, that such embodiment(s) may be practiced without these specific details.
Specific embodiments of the present application will be described in detail below with reference to the accompanying drawings.
The soil texture prediction method based on the local-global dependency relationship provided by the application comprises the following steps:
s110: the soil texture of the multi-layer soil depth was fitted by an equal area quadratic spline method.
The soil texture data in the present application is from the second national soil census data (official et al, 2013). Soil profiles with depths of about 1.5-2 meters are typically obtained according to standard field soil investigation methods (Zhang and Gong, 2012).
In order to meet the international soil mapping standard, the equal-area secondary spline function is used for fitting 0-5cm, 5-15cm, 15-30cm, 30-60cm, 60-100cm and 100-200cm of 6 standard soil layers. Soil texture is a component data of clay, sand and silt content totaling 100%. Thus, the logarithmic conversion process is used to address the compositional constraints of clay, sand and silt content. In Table 1, the clay, sand and silt contents range from 0 to 76.73%, 0.01% to 98.31% and 0 to 100%, respectively, with average values of 18.89%, 39.93% and 41.51%, respectively, and standard deviations of 10.97%, 20.98% and 15.41%, respectively, at depths of 0 to 5 cm. With increasing depth, the average clay content tends to increase, while the sand and silt content does not change much. The standard deviation in table 1 shows that the variability of sand content is highest. As depth increases, variability in clay, silt and sand content increases.
Table 1 statistical description of soil texture at different depths (weight percent)
S120: and acquiring main environment covariates, screening out related environment covariates with absolute values of pearson coefficients of soil textures greater than 0.02 in the whole soil section of each layer of soil depth, and removing redundant environment covariates affecting the prediction performance by using co-linearity analysis.
The environment covariates comprise five soil texture influencing factors including climate, biology, topography, matrix and soil category. Climate data were collected from Modis, woRldClim and chemical-close datasets. Biological data is represented primarily by data sources such as MODIS, GLOBELLAND, global accessibility map and GlobCoveR datasets. The MODIS data is an effective means of exploring biosphere variation, helping to understand the earth climate system, which is distributed at a spatial resolution of 11 km. The GLOBELLAND30 data is an important result of global and local global land cover remote sensing mapping and key technical research projects in China. It is the full coverage of the global land surface with a spatial resolution of 3030 meters. Global reachability map data describes urban or rural population gradients with spatial resolution of 1 km between 2000. GlobCoveR data is a global land cover map developed by the european space agency. The spatial resolution of the GlobCoveR data is chosen to be 11 km. The matrix data was collected from the american geological exploration institute earth science and environmental change science center (GECSC), based on the global lithology map database version 1.1. The relief data is primarily represented by the MERIT Digital Elevation Model (DEM) dataset, which is distributed at a spatial resolution of 100100. It takes into account various errors as an updated on-board DEM. Soil category data are mainly from coordinated world soil databases (spatial resolution 1 km) and soil grids (spatial resolution 250 meters), which are based on a linking method and a machine learning method, respectively.
S130: re-ordering the related environment covariates according to the relevance of the covariates and the soil texture according to the Pearson coefficient to generate a first model input set; and selecting the most relevant environment covariates with highest correlation degree of the covariates and the soil texture, and sequentially ordering the rest relevant environment covariates according to the sequence with highest correlation degree of the adjacent covariates to generate a second model input set.
The selected covariates and the most relevant environment covariates with highest soil texture relativity are sequentially ordered according to the sequence with highest adjacent covariates relativity to generate a second model input set, and the method comprises the following steps:
s131: and selecting the most relevant environment covariates with highest covariates and soil texture correlations as the model input data set of the last dimension.
S132: and calculating the covariates with the highest correlation degree with the most relevant environment covariates in other covariates to form secondary relevant environment covariates, and taking the secondary relevant environment covariates as the model input data set of the penultimate dimension.
S133: and sequentially calculating covariates with highest correlation degree with the model input data set of the previous dimension, taking the covariates as the model input data set of the current dimension, and finally generating all second model input sets.
As the soil texture is produced by complex activities of man and nature over time. In selecting relevant covariates and building predictive models, current research has focused mainly on the correlation between covariates and soil properties. However, it is not reasonable to predict the change of soil properties by only the relevant covariates. Taking soil and sand as an example; researchers have traditionally thought that high soil sand content is often accompanied by low variation in the amount of evapotranspiration. However, many studies have led to the opposite conclusion due to variations in land coverage and other complications. This also demonstrates that predictions made by the model can only be affected by virtue of the relationship between covariates and soil properties. Therefore, it is necessary to further study the potential relationship between the interaction between covariates and the change in soil properties, based on which the present application devised the two covariate reconstruction methods described above.
As shown in fig. 1, the first method is to rank the different covariates according to absolute values of pearson coefficients of soil texture. The most relevant covariates are placed in the last dimension, the last time step in the LSTM model. The second related covariate is placed in the penultimate dimension, the penultimate time step in the LSTM model, and the other covariates are analogized. It is assumed that if there is no correlation between adjacent covariates, the problem of gradient extinction may be faced during back propagation and further affect the prediction performance. To solve this problem, the present application also proposes a second reorganization method in which the present application also places the most relevant covariates (named a) in the last dimension. But instead of selecting the second related covariate, the covariate most related to a among the other covariates is calculated and placed in the penultimate dimension, and so on. In this way, it is ensured that adjacent covariates are correlated and the problem of gradient extinction is alleviated.
S140: and training two layers of long-short-term memory LSTM models by using the first model input set and the second model input set, establishing a relation model between soil texture and environment covariates by applying a full-connection layer FCN to training results, and predicting the soil texture by using the relation model.
The LSTM model is a recurrent neural network that maps variable-length environmental covariates to fixed-length vectors by recursively computing neighboring covariates. The model based can capture the dependency of covariates well. It calculates a mapping from environmental covariates to soil texture as follows:
i(n)＝σ(W ih h(n-1)+W ix x(n)+W ix c(n-1)+b i ) (1)
f(n)＝σ(W fh h(n-1)+W fx x(n)+W fx c(n-1)+b f ) (2)
o(n)＝σ(W oh h(n-1)+W ox x(n)+W ox c(n)+b o ) (4)
y＝W yh h(n)+b o (6)
where W is the weight matrix of the hidden state for the nth time step, b is the bias, σ (deg.) is the nonlinear activation function, tanh (deg.) is the hyperbolic tangent function,representing the multiplication of two vectors, x (t) and +.>Representing the input and output data, respectively, for the nth time step.
According to the above equation, the feature vector in the input parameter of the nth covariate is calculated based on the nth covariate, the global (long-term) correlation from the last covariate, and the local (short-term) correlation. As shown in equation 3, the input data is applied to preserve important features from local dependencies and then mapped to local parts. Equation 2 shows that the mathematical procedure of hiding the state of the nth covariate is similar to the input parameters. It is used to remove useless features from the global dependencies and further obtain the global dependency part in equation 3. Meanwhile, the output mode of the nth argument is generated according to the new unit and then used to update the hidden state in equation 5. Finally, the soil geology is calculated from the updated hidden state in equation 6.
Although LSTM may well express the correlation of neighboring covariates. But correlations on different scales should also be considered so that the prediction quality can be improved. To this end, the present application also introduces relationships between secondary related environment covariates. Thus, a deep learning model with two LSTM layers and one FCN layer was designed. As shown in FIG. 2, the first layer is identical to the conventional LSTM model, showing local correlation between adjacent covariates. To further provide more potential information, the second LSTM layer representation shows global dependencies between covariates. The input data in the second LSTM layer is generated by the second reassembly method described above.
The prediction accuracy of the prediction method provided by the application is analyzed as follows:
the prediction accuracy is evaluated based on three common metrics as follows; interpretation of the variants (R) 2 ) For verifying the percentage of variation interpreted by the predictive model. Root Mean Square Error (RMSE) is used to observe the ability to predict fluctuations. The gram-copta efficiency (KGE) represents the goodness of fit based on correlation, conditional and systematic bias. These indices are expressed as follows:
wherein y is i And x i The measured and predicted soil characteristics of the I-th section are respectively, and N is the total number of soil sections. In the KGE metric, R is the pearson coefficient between observed and predicted soil geology values. Alpha = sigma x /σ y Wherein σ is x Sum sigma y Standard deviations of predicted and observed soil geological values are represented, respectively. Beta = mu x /μ y Wherein μ is x Sum mu y Representing the average of the predicted and observed soil geology values.
To test the uncertainty of different predictive models, first, in the training process, a 10-fold cross validation method is applied to build a predictive model with optimal super parameters. In the process, an optimized model can be obtained and a deviation between predicted and observed soil characteristics can be generated from each verification intersection. All the deviations were used to fit quantile regression, forest models. Furthermore, a fit quantile regression forest model was used to calculate the percentile for each soil profile. Finally, the uncertainty is evaluated using a combination of Prediction Interval Coverage Probability (PICP) and average prediction interval (MPI). In the uncertainty index, we expect that 90% of the observed soil characteristics can fall within the 90% prediction interval. The index is expressed as follows:
PICP and MPI are respectively taken as confidence level and average width of the prediction interval. When the model reaches a similar PICP, the model with lower MPI has better predictive performance. As described in equation 10, y i Is the soil characteristic observed in the ith section, and N is the total number of soil sections.And->Representing the lower and upper values of the prediction.
The entire soil profile for each layer depth of the present application was randomly divided into training (80%), validation (10%) and test (10%) groups when evaluated by the following test.
Super parameter evaluation:
many super-parameters in the deep learning model, such as the number of samples, epoch, learning rate, size of hidden units, and LSTM layer number, may affect prediction performance. Therefore, the application evaluates the influence of all super parameters, thereby obtaining the optimal prediction model. Soil clay with a soil depth of 0-5cm is exemplified.
The five indexes are used for a test meter2, performance of different superparameters. The larger period (2000) and smaller period (500) result in the model tending to slightly over fit (R 2 0.603) and under-fitting (R) 2 0.603), the parameters cannot converge to an optimal value and further affect the prediction performance. Furthermore, the different size hidden units represent different information content after being encoded in the model. The small information (hidden layer size 64) cannot fully reflect the complex relationship between covariates and soil clay (R 2 0.593). Although an excessive amount of information (hidden unit size of 192) can achieve better fit results in the training set, generalization ability in the invisible test set is poor (R 2 0.603).
The number of samples represents the number of soil sections selected during training in one epoch; setting the sample number (64) too small is that the model hardly converges to the global optimum (R 2 0.610), while too large a sample number (192) results in poor generalization ability (R 2 Is 0.593). Furthermore, the learning rate is an important super-parameter that determines how the model weights are updated based on each prediction error. A lower learning rate (5 e-5) results in model retention (R 2 0.610) or requires a longer time cost to converge, while a higher learning rate (1 e-3) may miss the sweet spot (R 2 0.597) and further results in an unstable training process. Thus, the accuracy (R 2 0.614) is superior to the model with 4 time steps (R 2 0.608) and a model (R) with 6 time steps 2 0.608) the LSTM layer is the core of the model, which describes the local and global dependencies between covariates. The model of only one LSTM layer is the conventional LSTM model. Covariates in the LGD-LSTM model are selected based on correlations between covariates. However, in the conventional LSTM model, covariates are selected according to the correlation between covariates and soil properties. From Table 2 we can see that the number of LSTM layers is 2, which has the best performance (R 2 Is 0.605 corresponds to 3 layers, R 2 0.610 for layer 1). This also demonstrates that capturing global-local correlations between covariates can provide potential information for improved prediction accuracy. Final resultThe super parameters recommended in the relationship model finally designed by the application are as follows: epoch is 1000, hidden unit size is 128, sample data is 128, learning rate is 1e-4, time step is 5, LSTM layer number is 2.
Table 2 0-results for soil clay at soil depths of 5cm with different super parameter sets for LS-LSTM. The best settings in each group are shown in bold.
Prediction accuracy analysis:
as shown in tables 3 to 5, the predicted performance of the soil texture varies from depth to depth. According to all precision indexes (R 2 RMSE, KGE), the proposed deep learning model generally performs better than other predictive models. The improved LGD-LSTM model of the present application provides more accurate predictions than other LGD-LSTM models. On average, R 2 About 5.66% increase, about 4.88% decrease in RMSE, and about 2.06% increase in KGE. For soil silt prediction, R 2 About 2.64% increase, about 2.41% decrease in RMSE, and about 0.59% increase in KGE. This suggests that conventional DSM methods typically only consider the relationship between covariates and soil properties, whereas the present application introduces interactions between covariates that can provide more potential information that can affect soil property changes and further enhance predictive power. For soil clay prediction, the improved LGD-LSTM model provided by the application is generally R 2 RMSE and KGE were 0.672 (4.02% increase over LSTM), 6.517% (3.55% decrease over LSTM) and 0.742 (1.48% increase over LSTM), respectively. For soil sand prediction, the average R of the improved LGD-LSTM model provided by the application 2 RMSE and KGE were 0.725 (0.97% increase over LSTM), 10.882% (1.36% decrease over LSTM) and 0.767 (equal to LSTM), respectively. For soil silt prediction, the application provides a general R of an improved LGD-LSTM model 2 RMSE and KGE were 0.661 (1.85% increase over LSTM), 8.956% (1.69% decrease over LSTM) and 0.681 (1.49% increase over LSTM), respectively. This shows that the improved LGD-LSTM model provided by the application has greater prediction advantages than other prediction models after considering the local and overall correlation between covariates and soil properties. It can thus be concluded that interactions between covariates play a greater role in the texture mapping of soil than local and global dependencies between covariates and soil properties.
It is well known that RF models are frequently applied and accepted in DSM, CNN models also exhibit reasonable performance in DSM. Furthermore, from tables 3 to 5, the improved LGD-LSTM model proposed by the present application has a significant improvement compared to the RF and CNN models. Soil clay prediction R compared to LSTM model 2 The average improvement is 17.69%, the average reduction of RMSE is 12.39%, and the average improvement of KGE is 6.46%; soil sand prediction R 2 On average 13.10% increase, 12.50% decrease in RMSE on average, 2.82% increase in KGE on average, soil silt predictive R 2 Average increase 16.58%, average decrease 11.52% RMSE, average increase 4.77% KGE; furthermore, the prediction ability of the CNN model is worse than the RF model, because of inter-channel redundancy, CNN-based models have limited flexibility in DSM. The improved LGD-LSTM model proposed by the present application provides more options for the application of DSM, in particular integrated machine learning models.
Uncertainty analysis:
first, all models were tested for uncertainty using PICP and MPI with a confidence of 90%. Table 2 shows that the PICP of all models was close to the expected 90% confidence level. For soil clay prediction, the PICP of the improved LGD-LSTM model provided by the application is respectively 0.894 to 0.899 at all depths, the LGD-LSTM model is also 0.894 to 0.899, the LSTM model is 0.8999 to 0.913, the CNN model is 0.852 to 0.901, and the RF model is 0.876 to 0.901. For all depths, the highest PICP is implemented by the LSTM model. The LSTM model has relatively highest confidence compared to other predictive models, and consistent and stable soil clay can be predicted. According to MPI, for soil clay prediction, the improved LGD-LSTM model proposed by the present application predicts MPI ranging from 26.519% to 35.076%, LGD-LSTM model ranging from 27.826% to 34.151%, LSTM model ranging from 27.875% to 35.771%, CNN model ranging from 28.681% to 36.667%, and RF model ranging from 27.727% to 33.572, respectively. In general, the RF model produced an average value of MPI that was lower than other soil clay predictive models. LGD-LSTM has the narrowest MPI in all models for soil sand and silt predictions. Furthermore, in addition to soil clay and sand predictions at soil depths of 100-200cm, increases in MPI are typically accompanied by increases in soil depth. Overall, considering the interactions between covariates in the texture mapping of soil according to PICP and MPI can improve predictive ability and create relatively low uncertainty in predictions.
Table 3 accuracy of soil clay predictive model at all soil depths. The best index in the model is shown in bold.
/>
Table 4 is the same as table 3 but for soil sand prediction.
/>
Table 5 is the same as Table 3, but for soil silt prediction.
/>
It will be evident to those skilled in the art that the application is not limited to the details of the foregoing illustrative embodiments, and that the present application may be embodied in other specific forms without departing from the spirit or essential characteristics thereof.
The present embodiments are, therefore, to be considered in all respects as illustrative and not restrictive, the scope of the application being indicated by the appended claims rather than by the foregoing description, and all changes which come within the meaning and range of equivalency of the claims are therefore intended to be embraced therein. Any reference signs in the claims shall not be construed as limiting the claim concerned.
Furthermore, it is evident that the word "comprising" does not exclude other elements or steps, and that the singular does not exclude a plurality. A plurality of units or means recited in the system claims can also be implemented by means of software or hardware by means of one unit or means. The terms second, etc. are used to denote a name, but not any particular order.
Finally, it should be noted that the above-mentioned embodiments are merely for illustrating the technical solution of the present application and not for limiting the same, and although the present application has been described in detail with reference to the preferred embodiments, it should be understood by those skilled in the art that modifications and equivalents may be made to the technical solution of the present application without departing from the spirit and scope of the technical solution of the present application.
Claims (6)
1. A soil texture prediction method based on local-global dependency relationship, the method comprising:
fitting the soil texture of the multi-layer soil depth by an equal-area secondary spline method;
acquiring a main environment covariate, screening out a related environment covariate with the absolute value of a pearson coefficient of soil texture being more than 0.02 in the whole soil section of each layer of soil depth, and removing redundant environment covariates affecting the prediction performance by using co-linearity analysis;
re-ordering the related environment covariates according to the relevance of the covariates and the soil texture according to the Pearson coefficient to generate a first model input set; selecting the most relevant environment covariates with highest covariates and soil texture relativity, and sequentially ordering the rest relevant environment covariates according to the sequence with highest adjacent covariates relativity to generate a second model input set;
and training two layers of long-short-term memory LSTM models by using the first model input set and the second model input set, establishing a relation model between soil texture and environment covariates by applying a full-connection layer FCN to training results, and predicting the soil texture by using the relation model.
2. The local-global dependency-based soil texture prediction method according to claim 1, wherein the soil texture of the multi-layered soil depth comprises 6 standard soil layer soil textures of 0-5cm, 5-15cm, 15-30cm, 30-60cm, 60-100cm and 100-200 cm; the soil texture is soil composition data with the sum of clay, sand and silt content being 100%.
3. The local-global dependency-based soil texture prediction method according to claim 1, wherein the environmental covariates comprise five soil texture influencing factors of climate, biology, topography, matrix and soil category.
4. The method for predicting soil texture based on local-global dependency relationship according to claim 1, wherein the selecting the most relevant environment covariates with highest correlation degree of the soil texture, and the remaining relevant environment covariates are sequentially ordered in the order of highest correlation degree of adjacent covariates to generate a second model input set, comprising:
selecting the most relevant environment covariates with highest covariates and soil texture relativity as the model input data set of the last dimension;
calculating the covariates with the highest correlation degree with the most relevant environment covariates in other covariates to form secondary relevant environment covariates, and taking the secondary relevant environment covariates as a model input data set of the penultimate dimension;
and sequentially calculating covariates with highest correlation degree with the model input data set of the previous dimension, taking the covariates as the model input data set of the current dimension, and finally generating all second model input sets.
5. The local-global dependency-based soil texture prediction method according to claim 1, wherein the LSTM model is capable of calculating a mapping from environmental covariates to soil texture as follows:
i(n)＝σ(W ih h(n-1)+W iχ χ(n)+W iχ c(n-1)+b i ) (1)
f(n)＝σ(W fh h(n-1)+W fx x(n)+W fχ c(n-1)+b f ) (2)
o(n)＝σ(W oh h(n-1)+W ox x(n)+W ox c(n)+b o ) (4)
y＝W yh h(n)+b o (6)
where W is the weight matrix of the hidden state for the nth time step, b is the bias, σ (deg.) is the nonlinear activation function, tanh (deg.) is the hyperbolic tangent function,representing the multiplication of two vectors, χ (t) and +.>Representing the input and output data, respectively, for the nth time step.
6. The local-global dependency relationship-based soil texture prediction method according to claim 1, wherein the super parameters recommended in the relationship model are: epoch is 1000, hidden unit size is 128, sample data is 128, learning rate is 1e-4, time step is 5, LSTM layer number is 2.
Priority Applications (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310950055.8A CN116881664B (en) | 2023-07-31 | 2023-07-31 | Soil texture prediction method based on local-global dependency relationship |
Applications Claiming Priority (1)
| Application Number | Priority Date | Filing Date | Title |
|---|---|---|---|
| CN202310950055.8A CN116881664B (en) | 2023-07-31 | 2023-07-31 | Soil texture prediction method based on local-global dependency relationship |
Publications (2)
| Publication Number | Publication Date |
|---|---|
| CN116881664A true CN116881664A (en) | 2023-10-13 |
| CN116881664B CN116881664B (en) | 2024-02-02 |
Family
ID=88268093
Family Applications (1)
| Application Number | Title | Priority Date | Filing Date |
|---|---|---|---|
| CN202310950055.8A Active CN116881664B (en) | 2023-07-31 | 2023-07-31 | Soil texture prediction method based on local-global dependency relationship |
Country Status (1)
| Country | Link |
|---|---|
| CN (1) | CN116881664B (en) |
Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180172659A1 (en) * | 2016-12-16 | 2018-06-21 | Farmers Edge Inc. | Classification of Soil Texture and Content by Near-Infrared Spectroscopy |
| CN114386332A (en) * | 2022-01-17 | 2022-04-22 | 长春师范大学 | Soil humidity prediction method and device based on EDC-LSTM model and storage medium |
| CN115017798A (en) * | 2022-05-13 | 2022-09-06 | 东华大学 | Soil plough layer moisture temperature prediction method based on deep learning |
| CN115616194A (en) * | 2022-11-03 | 2023-01-17 | 中科合肥智慧农业协同创新研究院 | Soil organic matter prediction method based on auxiliary information |
| CN116307086A (en) * | 2023-01-13 | 2023-06-23 | 中国环境科学研究院 | Method and device for predicting organic carbon in soil |
| CN116433068A (en) * | 2023-02-22 | 2023-07-14 | 农业农村部环境保护科研监测所 | Causal machine learning method for evaluating soil versatility |
-
2023
- 2023-07-31 CN CN202310950055.8A patent/CN116881664B/en active Active
Patent Citations (6)
| Publication number | Priority date | Publication date | Assignee | Title |
|---|---|---|---|---|
| US20180172659A1 (en) * | 2016-12-16 | 2018-06-21 | Farmers Edge Inc. | Classification of Soil Texture and Content by Near-Infrared Spectroscopy |
| CN114386332A (en) * | 2022-01-17 | 2022-04-22 | 长春师范大学 | Soil humidity prediction method and device based on EDC-LSTM model and storage medium |
| CN115017798A (en) * | 2022-05-13 | 2022-09-06 | 东华大学 | Soil plough layer moisture temperature prediction method based on deep learning |
| CN115616194A (en) * | 2022-11-03 | 2023-01-17 | 中科合肥智慧农业协同创新研究院 | Soil organic matter prediction method based on auxiliary information |
| CN116307086A (en) * | 2023-01-13 | 2023-06-23 | 中国环境科学研究院 | Method and device for predicting organic carbon in soil |
| CN116433068A (en) * | 2023-02-22 | 2023-07-14 | 农业农村部环境保护科研监测所 | Causal machine learning method for evaluating soil versatility |
Non-Patent Citations (1)
| Title |
|---|
| 李文耀;魏楠;黄丽娜;上官微;: "土壤数据集对全球陆面过程模拟的影响", 气候与环境研究, no. 05 * |
Also Published As
| Publication number | Publication date |
|---|---|
| CN116881664B (en) | 2024-02-02 |
Similar Documents
| Publication | Publication Date | Title |
|---|---|---|
| Bevilacqua et al. | Quantifying volcanic hazard at Campi Flegrei caldera (Italy) with uncertainty assessment: 1. Vent opening maps | |
| Villani et al. | Repeatable site and path effects on the ground‐motion sigma based on empirical data from southern California and simulated waveforms from the CyberShake platform | |
| Li et al. | An extended cellular automaton using case‐based reasoning for simulating urban development in a large complex region | |
| Smith et al. | Identifying influences on model uncertainty: an application using a forest carbon budget model | |
| Anderson et al. | LIDAR density and linear interpolator effects on elevation estimates | |
| Stewart-Oaten et al. | Temporal and spatial variation in environmental impact assessment | |
| Danciu et al. | The 2014 Earthquake Model of the Middle East: ground motion model and uncertainties | |
| Veregin | Developing and testing of an error propagation model for GIS overlay operations | |
| Douglas | Capturing geographically-varying uncertainty in earthquake ground motion models or what we think we know may change | |
| Ganora et al. | An approach to estimate nonparametric flow duration curves in ungauged basins | |
| Tromans et al. | Probabilistic seismic hazard assessment for a new-build nuclear power plant site in the UK | |
| Chevalier et al. | Data integration methods to account for spatial niche truncation effects in regional projections of species distribution | |
| Werner et al. | Retrospective evaluation of the five-year and ten-year CSEP-Italy earthquake forecasts | |
| Tao et al. | Deep learning optimizes data-driven representation of soil organic carbon in Earth system model over the conterminous United States | |
| Zafarani et al. | Applicability of different ground-motion prediction models for northern Iran | |
| Broennimann et al. | Package ‘ecospat’ | |
| Chevallier et al. | Probabilistic global maps of the CO2 column at daily and monthly scales from sparse satellite measurements | |
| Verkaik et al. | GLOBGM v1. 0: a parallel implementation of a 30 arcsec PCR-GLOBWB-MODFLOW global-scale groundwater model | |
| Visini et al. | An updated area-source seismogenic model (MA4) for seismic hazard of Italy | |
| Vemula et al. | Neural network-based subduction ground motion model and its application to New Zealand and the Andaman and Nicobar Islands | |
| Lee | Engineering Characterization of Earthquake Ground Motions. | |
| CN116881664B (en) | Soil texture prediction method based on local-global dependency relationship | |
| Gates et al. | Estimating ancient biogeographic patterns with statistical model discrimination | |
| Haining | Data problems in spatial econometric modeling | |
| Chen et al. | Dynamically weighted ensemble of geoscientific models via automated machine-learning-based classification |
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 |