Soil organic carbon density time sequence spatial distribution prediction method, system and terminal under limited sample condition

Abstract

The invention relates to the technical field of soil organic carbon density prediction, in particular to a soil organic carbon density time sequence space distribution prediction method, system and terminal under the condition of limited samples. Analyzing the interpretation force of each variable on the organic carbon density of the regional soil based on a geographic detector model, fully understanding the soil-environment relationship, and realizing the identification of the optimal environment variable information; researching and developing a soil sample expansion method considering geographical similarity, selecting an optimal sample suitable for modeling, and expanding a limited sample; and finally, different modeling schemes are compared through a system, different space-time prediction models are constructed by relying on a machine learning algorithm, model effects are compared through the system, an optimal scheme is selected, and accurate detection of the soil organic carbon space-time pattern is realized according to the optimal scheme. In addition, the prediction method has the advantages of being short in measurement time, low in cost and capable of being popularized in a large area.

Description

Technical Field

[0001] The present invention relates to the technical field of soil organic carbon density prediction, and particularly to a method, a system and a terminal for predicting the temporal-spatial distribution of soil organic carbon density under the condition of limited samples. Background Art

[0002] Digital soil mapping Figure 1 Generally, it may include the following links: generation of environmental variable information, acquisition of sample point data, establishment and verification of mapping models. However, existing research mostly focuses on the collection of soil samples in small areas in a targeted manner, and the data is difficult to share; moreover, most studies focus on the analysis of data at limited time nodes. Therefore, there are few reports on temporal digital soil mapping for large areas such as provinces, especially for research under the condition of limited samples. The sample points collected in the field are the main data sources for digital soil mapping, and the quantity and layout of the sample points have a great impact on the inference results. Therefore, the limited soil samples bring huge challenges to the temporal-spatial prediction of soil organic carbon density based on digital soil mapping technology, which are mainly reflected in the following aspects:

[0003] 1) Using environmental variables related to natural factors without discrimination according to the Scorpan theory, the spatial heterogeneity of the data is insufficient, which will inevitably affect the mapping effect of soil organic carbon density; it is still unknown how to clarify the explanatory power of each environmental variable for soil organic carbon density (that is, "what are the appropriate environmental variables").

[0004] 2) The size of the sample quantity will have an important impact on the mapping model. Although some scholars have developed transfer models or constructed eigenvectors to measure the similarity between samples, there is a lack of attempts on a large regional scale, and it is still undetermined how to select more available soil samples based on the publicly available data (that is, "what is a reliable sample expansion method").

[0005] 3) On the basis of completing the expansion of soil sample information, fully understanding the relationship between soil organic carbon density and environmental variables under time series, and selecting an appropriate temporal mapping model are the key steps to accurately detect the temporal and spatial changes of soil organic carbon density accordingly. However, the current research lacks in-depth consideration of the long time series of soil organic carbon; it is still unclear how to construct a prediction model for the temporal-spatial distribution of soil organic carbon density suitable for the condition of limited samples (that is, "what is the optimal modeling scheme").

[0006] Therefore, the existing technology still needs to be improved and developed. Summary of the Invention

[0007] In view of the deficiencies of the above-mentioned existing technologies, the purpose of the present invention is to provide a method, system and terminal for predicting the temporal-spatial distribution of soil organic carbon density under the condition of limited samples, aiming to solve the problems that the traditional determination of soil organic carbon content takes a long time and has high economic costs and is difficult to promote on a large scale.

[0008] The technical solution of the present invention is as follows:

[0009] A method for predicting the temporal-spatial distribution of soil organic carbon density under the condition of limited samples, comprising the steps of:

[0010] Obtain the publicly available historical data of soil organic carbon density and the supplemented soil organic carbon density data, and conduct statistical analysis on the soil organic carbon density and the corresponding environmental variables;

[0011] Based on the geographical detector model, analyze the explanatory power of the environmental variables for the regional soil organic carbon density to obtain the soil-environment relationship;

[0012] Based on the soil-environment relationship and geographical similarity, expand the limited samples to obtain a sample data set;

[0013] Using the sample data set, based on the machine learning algorithm and the idea of temporal soil mapping, construct a spatio-temporal prediction model;

[0014] Based on the spatio-temporal prediction model, carry out temporal-spatial prediction mapping of soil organic carbon density to realize the prediction of the temporal-spatial distribution of regional soil organic carbon density.

[0015] A prediction system for a method for predicting the temporal-spatial distribution of soil organic carbon density under the condition of limited samples, comprising:

[0016] A data acquisition module, configured to obtain the publicly available historical data of soil organic carbon density and the supplemented soil organic carbon density data, and conduct statistical analysis on the soil organic carbon density and the corresponding environmental variables;

[0017] An analysis module, configured to analyze the explanatory power of the environmental variables for the regional soil organic carbon density based on the geographical detector model to obtain the soil-environment relationship;

[0018] A sample expansion module, configured to expand the limited samples based on the soil-environment relationship and geographical similarity to obtain a sample data set;

[0019] A model construction module, configured to use the sample data set to construct a spatio-temporal prediction model based on the machine learning algorithm and the idea of temporal soil mapping;

[0020] A prediction module, configured to carry out temporal-spatial prediction mapping of soil organic carbon density based on the spatio-temporal prediction model to realize the prediction of the temporal-spatial distribution of regional soil organic carbon density.

[0021] A terminal, the terminal includes a memory, a processor, and a program stored on the memory and executable on the processor. When the program is executed by the processor, the steps of the method for predicting the temporal and spatial distribution of soil organic carbon density under limited sample conditions are implemented.

[0022] Beneficial effects: The present invention provides a method, a system, and a terminal for predicting the temporal and spatial distribution of soil organic carbon density under limited sample conditions. The method for predicting the temporal and spatial distribution of soil organic carbon density under limited sample conditions includes the steps of: obtaining the publicly available historical data of soil organic carbon density and the supplemented soil organic carbon density data, and performing statistical analysis on the soil organic carbon density and the corresponding environmental variables; analyzing the explanatory power of the environmental variables for the regional soil organic carbon density based on the geographical detector model to obtain the soil-environment relationship; expanding the limited samples based on the soil-environment relationship and geographical similarity to obtain a sample data set; using the sample data set, based on the machine learning algorithm and the idea of temporal soil mapping, constructing a spatio-temporal prediction model; and carrying out the temporal and spatial prediction mapping of soil organic carbon density based on the spatio-temporal prediction model to realize the prediction of the temporal and spatial distribution of regional soil organic carbon density. Based on the obtained historical data and the supplemented soil organic carbon density data, the present invention completes the screening of environmental variables, analyzes the explanatory power of each variable for the regional soil organic carbon density based on the geographical detector model, fully understands the soil-environment relationship, and realizes the identification of the optimal environmental variable information; develops a method for expanding soil samples considering geographical similarity, selects the optimal samples suitable for modeling and expands the limited samples; finally, by systematically comparing different modeling schemes, relying on the machine learning algorithm to construct different spatio-temporal prediction models, systematically comparing the model effects, selecting the optimal scheme and accordingly realizing the accurate detection of the spatio-temporal pattern of soil organic carbon. This prediction method has the advantages of short measurement time, low cost, and wide area promotion; moreover, this prediction method is conducive to promoting the digital soil mapping technology to the challenging task of limited sample conditions, solving the three problems of "variable screening, sample expansion, and modeling scheme" in digital soil mapping research, and can provide scientific support for enhancing the soil carbon sequestration potential of regional terrestrial ecosystems. Description of the Drawings

[0023] Figure 1 It is a schematic flow chart of a method for predicting the temporal and spatial distribution of soil organic carbon density under limited sample conditions;

[0024] Figure 2 It is a conceptual framework diagram for the quantitative implementation of the geographical similarity law;

[0025] Figure 3 It is a schematic diagram of the principle of the random forest model. Detailed Embodiments

[0026] The present invention provides a method, system and terminal for predicting the temporal and spatial distribution of soil organic carbon density under the condition of limited samples. To make the objectives, technical solutions and effects of the present invention clearer and more definite, the present invention is further described in detail below. It should be understood that the specific embodiments described herein are only used to explain the present invention and are not used to limit the present invention.

[0027] Those skilled in the art of the present technology can understand that, unless otherwise defined, all terms (including technical terms and scientific terms) used herein have the same meaning as the general understanding of those of ordinary skill in the art to which the present invention belongs. It should also be understood that terms such as those defined in a general dictionary should be understood to have a meaning consistent with the meaning in the context of the prior art, and will not be interpreted in an idealized or overly formal sense unless specifically defined as here.

[0028] In the global carbon cycle, the soil carbon pool is 5 times that of the forest and other vegetation carbon pools and 3 times that of the atmospheric carbon pool. 60% of the carbon in the soil carbon pool exists in the soil in the form of organic matter. The huge soil carbon storage plays an important role in the carbon cycle of the terrestrial ecosystem. Therefore, as the core regulatory link for carbon emission or carbon sequestration in the terrestrial ecosystem, enhancing its carbon sink capacity is an important path to reduce carbon emissions, achieve carbon neutrality, and mitigate global climate change. Soil organic carbon density is a commonly used indicator to characterize soil organic carbon storage. Detailed and accurate soil information is an important basis for estimating soil carbon storage. However, due to the long time-consuming, high cost and low efficiency of obtaining a large number of soil sampling points, it is difficult to carry out the method of estimating carbon storage by ground surveys in a large spatial range for a long time, and it is extremely difficult to obtain soil organic carbon density information with a large volume, long time series and high quality. Therefore, exploring an efficient and accurate way to detect soil organic carbon density under the condition of limited samples is one of the difficulties and hotspots in current research.

[0029] Aiming at the realistic needs of the coordinated and synergistic goals of terrestrial ecosystem protection, restoration and carbon sequestration and sink enhancement, with the goal of "predicting the temporal and spatial distribution of soil organic carbon density under the condition of limited samples", based on the soil organic carbon spatio-temporal modeling method with a limited number, the present invention mainly solves the problems of the time and economic costs of traditional soil organic carbon content determination, which are difficult to promote on a large scale, and proposes a method, system and terminal for predicting the temporal and spatial distribution of soil organic carbon density under the condition of limited samples.

[0030] Based on this, as Figure 1 shown, the present invention provides a method for predicting the temporal and spatial distribution of soil organic carbon density under the condition of limited samples, including the steps of:

[0031] Step S10: Obtain the publicly available historical data of soil organic carbon density and the supplemented soil organic carbon density data, and conduct statistical analysis on the soil organic carbon density and the corresponding environmental variables;

[0032] Step S20: Analyze the explanatory power of the environmental variables for the regional soil organic carbon density based on the geographical detector model to obtain the soil-environment relationship;

[0033] Step S30: Expand the limited samples based on the soil-environment relationship and geographical similarity to obtain a sample dataset;

[0034] Step S40: Use the sample dataset to construct a spatio-temporal prediction model based on the machine learning algorithm and the idea of temporal soil mapping;

[0035] Step S50: Conduct temporal and spatial prediction mapping of the soil organic carbon density based on the spatio-temporal prediction model to achieve the prediction of the temporal and spatial distribution of the regional soil organic carbon density.

[0036] In this embodiment, based on the obtained historical data and the supplemented soil organic carbon density data, the screening of environmental variables is completed, and the explanatory power of each variable for the regional soil organic carbon density is analyzed based on the geographical detector model to fully understand the soil-environment relationship and realize the identification of the optimal environmental variable information; develop a soil sample expansion method considering geographical similarity, select the optimal samples suitable for modeling and expand the limited samples; finally, by systematically comparing different modeling schemes, rely on the machine learning algorithm to construct different spatio-temporal prediction models, systematically compare the model effects, select the optimal scheme and accordingly accurately explore the spatio-temporal pattern of soil organic carbon. This prediction method has the advantages of short measurement time, low cost and wide area promotion; moreover, this prediction method is conducive to promoting the digital soil mapping technology to the challenging task of limited sample conditions, solving the three problems of "variable screening, sample expansion, and modeling scheme" in digital soil mapping research, and can provide scientific support for enhancing the soil carbon sequestration potential of regional terrestrial ecosystems.

[0037] Specifically, the method for predicting the temporal and spatial distribution of soil organic carbon density under limited sample conditions mainly includes the identification method of optimal environmental variable information, the soil sample expansion method based on geographical similarity, and the prediction of the temporal and spatial distribution of soil organic carbon density.

[0038] 1) Identification method of optimal environmental variable information: On the basis of integrating existing public data and supplementary data, initially conduct statistical analysis of soil organic carbon density and corresponding environmental variables; then use the geographical detector model to analyze the explanatory power of each variable for the regional soil organic carbon density to fully understand the soil-environment relationship; and on this basis, realize the identification of the optimal environmental variable information, providing a reliable basis for the subsequent expansion of soil samples. That is, it mainly includes the descriptive statistics of soil samples and environmental variables, variable screening based on geographical detectors, and the identification of optimal environmental variable information.

[0039] 2) Soil sample expansion method based on geographical similarity: On the basis of exploring the optimal environmental variable information, the importance weights of each environmental variable are comprehensively considered according to the law of geographical similarity, and the geographical environmental similarity between local sample points and foreign sample points is calculated; then the similarity threshold index is constructed based on the number of samples, model prediction error and accuracy; finally, with its support, the optimal samples suitable for modeling are selected and the limited samples are expanded to provide data support for the subsequent prediction of the spatial distribution of soil organic carbon density time series. That is, it mainly includes the measurement of geographical environmental similarity, the construction of the similarity threshold index and the determination of the optimal modeling sample.

[0040] 3) Prediction of spatial and temporal distribution of soil organic carbon density: Based on the above research, different modeling schemes are designed based on machine learning algorithms and typical ideas of temporal soil mapping; the effects of different models are systematically compared and the best scheme is selected; finally, long-term spatial prediction and mapping of soil organic carbon density are carried out based on this, and the spatial distribution characteristics of uncertainty are given, the reliability of the mapping results is quantified, and the accurate exploration of the temporal and spatial distribution of regional soil organic carbon density is achieved. This mainly includes the spatiotemporal prediction of soil organic carbon density of different modeling schemes, the comparison and optimization of the prediction accuracy of different modeling schemes, and the temporal mapping and uncertainty analysis of soil organic carbon density.

[0041] In this implementation, the scientific goal is to achieve the prediction of the temporal and spatial distribution of soil organic carbon under limited sample conditions. The entire Guangdong Province is used as the experimental area. Combined with historical soil samples and supplementary sample data, the geographic detector model is first used to analyze the interpretability of different variables on regional soil organic carbon density and the interaction between variables to identify the optimal environmental variables; then, based on the law of geographical similarity, the soil samples are expanded to serve subsequent modeling; finally, on this basis, a soil organic carbon time series prediction model is constructed and its reliability is verified to achieve accurate detection of the temporal and spatial information of regional soil organic carbon density.

[0042] In some embodiments, obtaining publicly available historical data on soil organic carbon density and supplementary soil organic carbon density data, and conducting statistical analysis of soil organic carbon density and corresponding environmental variables specifically include: collecting soil sample data sets and conducting physical and chemical analysis, collecting remote sensing data and analyzing it, and selecting environmental factors required for soil organic carbon prediction and mapping.

[0043] In some embodiments, the soil sample dataset includes soil organic carbon density information of the topsoil layer and profile layer of the ecosystem at different periods, and the soil sample dataset needs to remove abnormal and missing values ​​of the soil organic carbon density information;

[0044] In the physical and chemical analysis, the calculation method of soil organic carbon density (SOCD) at different periods is shown in Formula 1:

[0045]

[0046] Among them, the soil organic carbon density of a certain layer of SOCD (kg C m -2 ), H is the soil layer thickness (cm), SOC S is the soil organic carbon content (g kg -1 ), BD represents the soil bulk density (kg m -3 ), and Gr represents the percentage content of gravel (particle size ≥ 2 mm). Further, the above two data are both obtained from SoilGrids250m 2.0.

[0047] Specifically, the dataset needs to include the soil organic carbon density information (kg C m -2 ) of the topsoil (0 - 20 cm) and the profile (0 - 100 cm) levels of the main ecosystems (such as forests, grasslands, shrubs, and farmlands); on this basis, it is necessary to integrate the data and eliminate data anomalies (such as missing adoption periods and coordinate anomalies) and missing values.

[0048] On the basis of a detailed analysis of the existing samples, in order to ensure the continuity and reliability of the long - time - series data, in this embodiment, 1 additional sampling of the topsoil samples (0 - 20 cm) is carried out, with a total of 450 sampling points set, and all underlying surface types are covered as much as possible. Since the land cover types in the study area are relatively complex (including some sampling - operation - difficult areas where it is difficult for people and vehicles to reach), when actually carrying out the sampling work, the road accessibility and other situations need to be comprehensively considered. Referring to the Technical Specification for Soil Environmental Monitoring (HJ / T 166 - 2004), each soil sample is obtained by the five - point mixed sampling method in a plum - blossom shape, and 1.5 kg of samples are taken after quartering. The GPS is used to record the sample point coordinates, and at the same time, the land cover type, landform, vegetation cover and other information at the sample point location are recorded. After the soil samples are taken back to the laboratory, small stones and plant residues are removed, and after natural air - drying, all are ground through a 2 - mm sieve for the determination of soil organic carbon in the laboratory. The determination of organic carbon uses the dichromate volumetric method - external heating method.

[0049] In some embodiments, the collection of remote - sensing data and analysis include:

[0050] Collect Landsat series remote - sensing images covering the entire region in the corresponding period;

[0051] Pre - process the Landsat series remote - sensing images, and classify the utilization methods of the images in combination with the field sampling records and visual interpretation results;

[0052] Adopt the comprehensive index model of land use degree, classify the land according to different utilization methods, and assign classification indexes to obtain a quantitative representation of the land use degree.

[0053] Specifically, the Landsat series (TM / ETM+ / OLI) remote sensing images covering the entire territory of Guangdong Province during the corresponding period all have a spatial resolution of 30 m. The image date is consistent with the sampling time. If the cloud cover of the current image is large (>20%), images of similar months or similar years during the same period are downloaded for replacement. In addition, the difference in the spectral response functions of multi-source remote sensing sensors is a problem that needs to be solved in data conversion. Considering the high correlation between the bands of remote sensing sensors, the present invention uses a statistical regression method to cross-calibrate (normalize) multi-source remote sensing reflectance data, reduce potential differences, and ensure the consistency of continuous long-term remote sensing data.

[0054] Use ENVI 5.3 software to perform preprocessing on Landsat image data, such as geometric correction, registration, radiometric correction, mosaicking, and cropping. Combining field sampling records and visual interpretation results, use the object-oriented support vector machine classification method to classify the images, and the categories are unused land, cultivated land, forest land, grassland, water body, and urban settlement land. Adopt the comprehensive index model of land use degree, divide the land into 4 levels according to its different utilization methods and assign grading indexes, so as to obtain a quantitative representation of the land use degree.

[0055] In some embodiments, the environmental factors include but are not limited to soil properties, climate, hydrology, biology, topography, and human activity factors.

[0056] Specifically, the soil-landscape model theory (Scorpan spatial prediction function) believes that soil properties are a comprehensive mapping of climate, biology, geomorphology, parent material, time, and spatial information. For soil organic carbon, climate, topography, parent material, hydrology, biology, and human activities are the reasons for the change of soil organic carbon. In addition, since the homogeneity of the soil parent material within the study area is relatively strong and the difference is small, the parent material factor is not within the scope of consideration of the present invention. Correspondingly, the environmental factors selected in the present invention mainly include 6 major soil-forming factors: soil properties, climate, hydrology, biology, topography, and human activity factors. Since these factors come from different data sources, they need to be resampled in ArcGIS 10.1 and uniformly converted into raster data with a resolution of 90 m.

[0057] In some embodiments, a geographical detector model is used to screen for optimal environmental variable information, including: A geographical detector is a statistical method for detecting spatial heterogeneity, which can reveal the driving factors affecting spatial heterogeneity. Its core idea is based on the assumption of the similarity principle, that is, if the dependent variable is significantly affected by the independent variable, then the spatial distributions of the independent variable and the dependent variable should be similar. Currently, geographical detectors have been widely used in geographical space similarity and heterogeneity research. Compared with machine learning algorithms, it has stronger generalization ability and interpretability, and is also more intuitive and faster than traditional linear statistical analysis methods such as principal component analysis and multiple regression. More importantly, geographical detectors are not sensitive to the number of samples, that is, reliable analysis results can be obtained with limited input samples. These advantages can all help improve the reliability of variable screening for digital mapping of soil organic carbon density. The present invention uses two geographical detectors, namely, differentiation and factor detection, and interaction detection, to explore the influencing factors of regional soil organic carbon density and their interactions to fully understand the soil-environment relationship; and on this basis, identify the optimal environmental variable information.

[0058] In some embodiments, the geographical detector model includes differentiation and factor detection, and interaction detection;

[0059] The differentiation and factor detection includes detecting the spatial heterogeneity of the dependent variable soil organic carbon density and the degree of explanation of the spatial heterogeneity of each independent variable x factor on the dependent variable, measured by the q value, and its expression is shown in Equations 2 and 3:

[0060]

[0061] where h is the stratification or partition of the independent variable x; N h and N are the number of units in layer h and the whole region, respectively; σ 2 h and σ 2 are the variances of the soil organic carbon density values in layer h and the whole region, respectively; X SSW and X SST are the within-layer variance SSW generated by the independent variable x and the total variance of the whole region, respectively; the q value indicates that the independent variable x can explain q×100% of the organic carbon density, and the range threshold of the q value is [0,1]; among them, the larger the q value, the stronger the explanatory power of a certain independent variable x on the dependent variable, and vice versa, the weaker the explanatory power. However, in extreme cases, if the q value is 1, it means that the independent variable x completely controls the spatial distribution of the dependent variable soil organic carbon density value.

[0062] The interaction detection includes identifying different independent variables x nFor the interaction between soil organic carbon density values, that is, to determine whether the effects of different driving factors on the dependent variable are independent of each other, and at the same time, the enhancement or weakening effect of the combined action of the two factors on the explanatory power of the dependent variable can be further evaluated. Specifically, it includes the following steps:

[0063] Calculate the q-values of the two independent variable factors x 1 and x 2 on the dependent variable separately: q(x 1 ) and q(x 2 );

[0064] Calculate the q-value of the interaction of different independent variables: q(x 1 ∩x 2 );

[0065] By comparing the magnitudes of q(x 1 ), q(x 2 ), and q(x 1 ∩x 2 ), classify the type of interaction.

[0066] Specifically, the types of interactions between the independent variables and the dependent variable of the geographical detector model are shown in Table 1.

[0067] Table 1 Types of interactions between the independent variables and the dependent variable of the geographical detector model

[0068] Basis for judgment Interaction <![CDATA[q(x 1 ∩x 2 )>q(x 1 )+q(x 2 )]]> Nonlinear enhancement <![CDATA[q(x 1 ∩x 2 )=q(x 1 )+q(x 2 )]]> Mutually independent <![CDATA[q(x 1 ∩x 2 )>Max(q(x 1 ),q(x 2 ))]]> Two-factor enhancement <![CDATA[Max(q(x 1 ),q(x 2 ))>q(x 1 ∩x 2 )>Min(q(x 1 ),q(x 2 ))]]> Single-factor nonlinear weakening <![CDATA[q(x 1 ∩x 2 )<Min(q(x 1 ),q(x 2 ))]]> Nonlinear weakening

[0069] In this embodiment, based on the geographical detector model, with soil organic carbon density as the dependent variable and environmental factors such as soil properties, climate, hydrology, biology, topography, and human activity factors as independent variables, the explanatory power of regional soil organic carbon density is analyzed to fully understand the soil-environment relationship, and on this basis, the identification of the optimal environmental variable information is realized.

[0070] In this embodiment, a limited number of soil samples are augmented in accordance with the law of geographical similarity. The core idea of the law of geographical similarity is that "the more similar the geographical environment, the more similar the geographical features (processes)", that is, comprehensively considering the similarity between the geographical element combinations (geographical environment) at one point (or one domain) and those at another point (or domain) to measure the similarity of geographical phenomenon features and processes. Therefore, the principle of geographical similarity mainly refers to the comprehensive similarity degree of two geographical spatial positions in the geographical environment (including spatial and non-spatial elements). What the law of geographical similarity summarizes is not only the similarity characteristics of geographical phenomena at different positions, but more importantly, it concisely expounds the mechanism of action between geographical phenomena and their geographical environment, that is, closely linking geographical phenomena with their geographical environment. The conceptual framework for the quantitative implementation of the law of geographical similarity is as Figure 2 shown.

[0071] In some embodiments, the geographical similarity includes a measure of the similarity of the geographical environment and the construction of a similarity threshold index;

[0072] To comprehensively consider the similarity of each environmental variable, it is necessary to infer the importance weight of each environmental covariate in the overall environmental variables of Guangdong Province; in this inference method, it is necessary to weight the similarity of each environmental variable to consider the importance weight of this environmental variable, so as to calculate the environmental similarity between the point to be inferred and the existing sample points.

[0073] Specifically, the measure of the geographical environment similarity includes the steps of:

[0074] Calculate the environmental similarity between the point to be inferred and the existing sample points, and the calculation method is shown in Equation 4:

[0075]

[0076] where S i,j is the environmental similarity between the point to be inferred i and the sample point j calculated by the weighted average method, and the value range is [0, 1]. is the importance weight value of the m-th environmental variable in characterizing the soil-environment relationship of the point to be inferred i, is the similarity of a single environmental variable and respectively represent the N-th covariate values of the unvisited location i and the sampling location j.

[0077] In some embodiments, the similarity of a single environmental variable can be calculated in two ways according to the nature of the environmental variable. When the environmental variable is a continuous environmental variable, the output value is determined by the relationship between the specific values of the environmental variable at the two points and the value range range(e n ) of the corresponding environmental variable in the study area, as shown in Equation 5:

[0078]

[0079] where |e i,n -e j,n | is the absolute difference between the point to be inferred i and the sample point j based on the n-th environmental variable, and range(e n ) is the value range of the n-th environmental variable in the study area.

[0080] By considering the importance weight values of the similarity of each environmental variable, the above method can reduce the distortion of the speculation result caused by the excessive influence of a certain environmental variable. Through the weighted average method, the similarity of each environmental variable can be more comprehensively considered, making the calculated environmental similarity more accurate and reliable. This method calculates the environmental similarity between the point to be speculated and each sample point, and uses the importance weight of each environmental variable as a measure of its contribution degree, so as to comprehensively consider the influence of each environmental variable on the dependent variable. Based on this, the value of the prediction point can be more accurately speculated according to the environmental similarity between the two points and the soil property value of the sample point, as shown in Equation 6:

[0081]

[0082] Among them, T is the number of sample points, V j is the soil property value of the jth sample point, S threshold is the environmental similarity threshold set by the user. The function iif(S i,j ≥S threshold ,S i,j ,0) means that if S i,j ≥S threshold , the return value of the function is S i,j , otherwise the return value of the function is 0;

[0083] In this embodiment, the soil samples under different S threshoid values are traversed at a step size of 0.01 and an interval of 0.5 - 0.9. A similarity threshold index (STI) is constructed to evaluate the model prediction results under different similarity thresholds, as shown in Equation 7:

[0084]

[0085] STI is the similarity threshold index, n 1 is the preferred sample under different similarity thresholds, n is the total sample in the corresponding period, RMSE 1 is the root mean square error of the model, and RMSE max is the maximum root mean square error.

[0086] Specifically, the similarity threshold index considers three aspects of factors. First, the proportion of the preferred sample (n 1 ) under different similarity thresholds in the total sample (n) of this period; second, the model prediction error, which is calculated by dividing RMSE 1 under different similarity thresholds by RMSE max in the same period; finally, the accuracy of the model. In this embodiment, 1 - R 2Representation. In this embodiment, a weight of 1 / 3 is assigned to each of the three parameters. The change range of STI is between 0 and 1, and the closer the value is to 1, the more accurate the effect of the model is under the similarity threshold.

[0087] In some embodiments, the prediction of the temporal-spatial distribution of soil organic carbon density mainly includes: the design of a temporal modeling scheme, the construction of a spatio-temporal prediction model and uncertainty analysis, and the accuracy verification of mapping.

[0088] Specifically, on the basis of considering the spatio-temporal distribution of sample points, 70% of all samples are selected as the modeling set, and the other 30% are used as an independent verification set (not participating in modeling). Based on the Scorpan digital soil mapping modeling idea, the random forest algorithm is selected to construct a soil-landscape model for spatial prediction mapping of soil organic carbon density in Guangdong Province.

[0089] In some embodiments, in order to realize the temporal mapping of regional soil organic carbon density, the temporal soil mapping idea includes one-time modeling and multiple uses (using one model multiple times) and multiple modeling and separate uses (using each model once); then, accordingly, 4 different modeling schemes are designed based on different sample inputs (whether to introduce similarity samples).

[0090] Specifically, the modeling schemes for constructing the spatio-temporal prediction model include at least one of the following: constructing different-year prediction models and mapping based on local samples in four periods respectively; integrating local samples in four periods and constructing a model and mapping centrally; constructing different-year prediction models and mapping based on similarity samples in four periods respectively; integrating similarity samples in four periods and constructing a model and mapping centrally. By systematically comparing the effects of different models and selecting the optimal scheme; finally, long-term temporal-spatial prediction mapping of soil organic carbon density is carried out based on the optimal scheme and the random forest algorithm.

[0091] In some of these embodiments, as Figure 3 shown, the random forest RF is composed of many single classification and regression trees (CART, Classification And Regression Tree) as meta-classification and regression estimators, and the finally predicted variable is the average value of the output values of multiple single tree models. The Bagging method is used to form a different training sample set, and the optimal division of each node of the tree model is selected from randomly selected subsets of predictors.

[0092] In this embodiment, the RF model is established through the "randomForest" toolkit (version 4.6-14) in the R language platform software. During the establishment of the RF model, it is necessary to set three main parameters in the model: the number of decision trees (ntree), the minimum number of samples per node (mtry), and the number of variables tried per node (node).

[0093] In some embodiments, to characterize the reliability of the mapping results, the temporal-spatial prediction mapping quantifies the uncertainty of the model by using the non-parametric simulation Bootstrap method. By performing spatial modeling on each Bootstrap sample, the probability distribution of the modeling results is obtained. Among them, the uncertainty of the modeling is quantified by the calculated confidence interval, and the final temporal-spatial distribution prediction result is the average value of N times of modeling. The calculation method is shown in Equation 8:

[0094]

[0095] where N is the number of random modeling times, and SOCD k is the finally predicted soil organic carbon of the kth grid, and SOCD(i) k is the soil organic carbon predicted by the ith model. For each grid, the upper and lower limits of its 95% confidence interval are calculated based on the values calculated from multiple random modelings.

[0096] The calculation method of the uncertainty is shown in Equation 9:

[0097] Uncertainty k =(SOCD upper,k -SOCD lower,k ) (Equation 9)

[0098] where Uncertainty k is the uncertainty of the kth grid; SOCD upper,k and SOCD lower,k correspond to the upper and lower limits of the confidence interval of the predicted value of the soil organic carbon density of the kth grid, respectively.

[0099] In some embodiments, in step S40, the model construction dataset (i.e., the sample dataset) is used to train the model to obtain a spatio-temporal prediction model, and 10-fold cross-validation is used to evaluate the training result of the model. Then, an independent validation dataset is used to evaluate the prediction result. The evaluation indicators include the mean absolute error (MAE, Mean Absolute Error), the root mean square error (RMSE, Root Mean of Squared Error), and the coefficient of determination (R 2,Coefficient of Determination) and Lin's Concordance Correlation Coefficient (LCCC), and their expressions are shown in Equations 10, 11, 12, and 13 respectively:

[0100]

[0101] Among them, n is the number of modeling samples; y and are the measured value and predicted value of the model respectively; and are the average values of the measured value and predicted value respectively; σ y and are the variances of the measured value and predicted value respectively; r is the Pearson correlation coefficient. Among them: MAE is used to evaluate the deviation degree of the predicted value. The closer it is to 0, the better the prediction result. RMSE is used to evaluate the overall accuracy of the prediction result. The smaller its value, the higher the prediction accuracy of the model. R 2 is used to evaluate the goodness of fit of the model. The closer its value is to 1, the higher the reference value of the model. LCCC is used to measure the degree of distribution of the predicted value and the measured value on the 1:1 line. The closer it is to 1, the higher the degree of coincidence between the predicted value and the observed value, indicating the stronger the prediction ability of the model.

[0102] In addition, the present invention also provides a prediction system for the temporal and spatial distribution prediction method of soil organic carbon density under limited sample conditions, including:

[0103] A data acquisition module, configured to acquire the publicly available historical data of soil organic carbon density and the supplemented soil organic carbon density data, and perform statistical analysis on the soil organic carbon density and the corresponding environmental variables;

[0104] An analysis module, configured to analyze the explanatory power of the environmental variables for the regional soil organic carbon density based on the geographical detector model to obtain the soil-environment relationship;

[0105] A sample expansion module, configured to expand the limited samples based on the soil-environment relationship and geographical similarity to obtain a sample data set;

[0106] A model construction module, configured to use the sample data set to construct a spatio-temporal prediction model based on a machine learning algorithm and the idea of temporal soil mapping;

[0107] A prediction module, configured to perform temporal and spatial prediction mapping of soil organic carbon density based on the spatio-temporal prediction model to achieve the temporal and spatial distribution prediction of regional soil organic carbon density.

[0108] In addition, the present invention further provides a terminal, which includes a memory, a processor, and a program stored on the memory and executable on the processor. When the program is executed by the processor, the steps of the method for predicting the temporal-spatial distribution of soil organic carbon density under limited sample conditions are implemented.

[0109] In some embodiments, the memory may be an internal storage unit of the terminal, such as the hard disk or memory of the terminal. In other embodiments, the memory may also be an external storage device of the terminal, such as a plug-in hard disk equipped on the terminal, a Smart Media Card (SMC), a Secure Digital (SD) card, a Flash Card, etc. Further, the memory may also include both the internal storage unit and the external storage device of the terminal. The memory is used to store application software installed on the terminal and various types of data, such as the program code for installing the terminal. The memory may also be used to temporarily store data that has been output or will be output. In one embodiment, a program for predicting the temporal-spatial distribution of soil organic carbon density under limited sample conditions is stored on the memory, and this program for predicting the temporal-spatial distribution of soil organic carbon density under limited sample conditions can be executed by the processor, thereby implementing the method for predicting the temporal-spatial distribution of soil organic carbon density under limited sample conditions of the present invention.

[0110] In some embodiments, the processor may be a Central Processing Unit (CPU), a microprocessor, or other data processing chips, and is used to run the program code stored in the memory or process data, such as implementing the method for predicting the temporal-spatial distribution of soil organic carbon density under limited sample conditions.

[0111] In addition, the present invention further provides a computer-readable storage medium, on which a computer program is stored. The computer-readable storage medium stores a program for predicting the temporal-spatial distribution of soil organic carbon density under limited sample conditions. When the program for predicting the temporal-spatial distribution of soil organic carbon density under limited sample conditions is executed by the processor, the steps of the method for predicting the temporal-spatial distribution of soil organic carbon density under limited sample conditions are implemented.

[0112] In summary, a method, system, and terminal for predicting the temporal and spatial distribution of soil organic carbon density under limited sample conditions provided by the present invention. The method for predicting the temporal and spatial distribution of soil organic carbon density under limited sample conditions includes the steps of: obtaining the publicly available historical data of soil organic carbon density and the supplemented soil organic carbon density data, and conducting statistical analysis of the soil organic carbon density and the corresponding environmental variables; analyzing the explanatory power of the environmental variables for the regional soil organic carbon density based on the geographical detector model to obtain the soil-environment relationship; expanding the limited samples based on the soil-environment relationship and geographical similarity to obtain a sample dataset; using the sample dataset, based on machine learning algorithms and the idea of temporal soil mapping, to construct a spatio-temporal prediction model; and carrying out temporal and spatial prediction mapping of soil organic carbon density based on the spatio-temporal prediction model to achieve the prediction of the temporal and spatial distribution of regional soil organic carbon density. The present invention completes the screening of environmental variables based on the obtained historical data and the supplemented soil organic carbon density data, analyzes the explanatory power of each variable for the regional soil organic carbon density based on the geographical detector model, fully understands the soil-environment relationship, and realizes the identification of the optimal environmental variable information; develops a method for expanding soil samples considering geographical similarity, selects the optimal samples suitable for modeling and expands the limited samples; finally, by systematically comparing different modeling schemes, relying on machine learning algorithms to construct different spatio-temporal prediction models, systematically comparing the model effects, selecting the optimal scheme and accordingly accurately exploring the spatio-temporal pattern of soil organic carbon. This prediction method has the advantages of short measurement time, low cost, and wide applicability; moreover, this prediction method is conducive to promoting the digital soil mapping technology to the challenging task of limited sample conditions, solving the three problems of "variable screening, sample expansion, and modeling scheme" in digital soil mapping research, and can provide scientific support for enhancing the soil carbon sequestration potential of regional terrestrial ecosystems.

[0113] It should be understood that the application of the present invention is not limited to the above examples. For those of ordinary skill in the art, improvements or transformations can be made according to the above description, and all such improvements and transformations should fall within the protection scope of the appended claims of the present invention.

Claims

1. A method for predicting the temporal spatial distribution of soil organic carbon density under finite sample conditions, characterized in that: Includes steps: Obtain publicly available historical data on soil organic carbon density and supplementary data on soil organic carbon density, and conduct statistical analysis of soil organic carbon density and corresponding environmental variables; The explanatory power of the environmental variables on regional soil organic carbon density was analyzed based on the geographic detector model to obtain the soil-environment relationship; Based on the soil-environment relationship and geographical similarity, the limited sample is expanded to obtain a sample data set; Using the sample data set, a spatiotemporal prediction model is constructed based on machine learning algorithms and time series soil mapping ideas; Based on the spatiotemporal prediction model, temporal spatial prediction mapping of soil organic carbon density is carried out to achieve temporal spatial distribution prediction of regional soil organic carbon density.

2. The method for predicting the temporal spatial distribution of soil organic carbon density under limited sample conditions according to claim 1 is characterized in that: Obtain publicly available historical data on soil organic carbon density and supplementary data on soil organic carbon density, and conduct statistical analysis of soil organic carbon density and corresponding environmental variables, including: collecting soil sample data sets and conducting physical and chemical analysis, collecting remote sensing data and analyzing them, and selecting environmental factors required for soil organic carbon prediction and mapping.

3. The method for predicting the temporal spatial distribution of soil organic carbon density under limited sample conditions according to claim 2, characterized in that: The soil sample dataset includes soil organic carbon density information of the topsoil layer and profile layer of the ecosystem at different periods, and the soil sample dataset needs to remove abnormal and missing values ​​of soil organic carbon density information; In the physical and chemical analysis, the calculation method of soil organic carbon density at different periods is shown in Formula 1: Among them, SOCD soil organic carbon density in a certain layer (kg Cm -2 ), H is the soil layer thickness (cm), SOC S is the soil organic carbon content (g kg -1 ), BD represents soil bulk density (kg m -3 ), Gr represents the gravel percentage.

4. The method for predicting the temporal spatial distribution of soil organic carbon density under limited sample conditions according to claim 2, characterized in that: The remote sensing data collection and analysis includes: Collect Landsat series remote sensing images covering the entire region during the corresponding period; Preprocess the Landsat series remote sensing images, combine field sampling records and visual interpretation results, and classify the image utilization methods; The land use degree comprehensive index model is adopted to classify the land according to different utilization modes and assign classification indexes to obtain a quantitative representation of the land use degree. The environmental factors include soil properties, climate, hydrology, biology, topography, and human activities.

5. The method for predicting the temporal spatial distribution of soil organic carbon density under limited sample conditions according to claim 1, characterized in that: The geographic detector model includes differentiation and factor detection, and interaction detection; The differentiation and factor detection includes detecting the spatial differentiation of the dependent variable soil organic carbon density and the degree of explanation of the spatial differentiation of the dependent variable by each variable x factor, which is measured by the q value, and its expression is shown in Formula 2 and Formula 3: Where h is the stratification or partition of the independent variable x; N h and N are the number of units in layer h and the whole area respectively; σ 2 h and σ 2 are the variances of soil organic carbon density values ​​in layer h and the whole area; X SSW and X SST They are the intra-layer variance SSW and the total variance of the whole area generated by the independent variable x; the q value means that the independent variable x can explain q×100% of the organic carbon density, and the range threshold of the q value is [0,1]; The interaction detection involves identifying different independent variables x n The interaction between soil organic carbon density values ​​includes the following steps: Calculate the q values ​​of the two independent variable factors x1 and x2 for the dependent variable separately: q(x1) and q(x2); Calculate the q value of the interaction between different independent variables: q(x1∩x2); By comparing the size relationship among q(x1), q(x2) and q(x1∩x2), the interaction types are classified.

6. The method for predicting the temporal spatial distribution of soil organic carbon density under limited sample conditions according to claim 1, characterized in that: The geographical similarity includes the measurement of geographical environment similarity and the construction of a similarity threshold index; The measurement of the geographical environment similarity comprises the steps of: Calculate the environmental similarity between the point to be inferred and the existing sample points. The calculation method is shown in Formula 4: Among them, S i,j is the environmental similarity between the inferred point i and the sample point j calculated by the weighted average method, and its value range is [0,1]. is the importance weight of the mth environmental variable in describing the soil-environment relationship of the inferred point i, is the similarity of a single environmental variable; and denote the Nth covariate value of the unvisited position i and the sampled position j, respectively; When the environmental variable is a continuous environmental variable, the output value is determined by the specific value of the environmental variable at two points and the value range of the corresponding environmental variable in the study area (e n ), as shown in Formula 5: Among them, |e i,n -e j,n | is the absolute difference between the inferred point i and the sample point j based on the nth environmental variable, range(e n ) is the value range of the nth environmental variable in the study area; The value of the prediction point is inferred based on the environmental similarity of the two points and the soil property value of the sample point, as shown in Formula 6: Where T is the number of samples, V j is the soil property value of the jth sample point, S threshold is the environment similarity threshold set by the user, function iif(S i,j ≥S threshold ,S i,j ,0) means if S i,j ≥S threshold , the return value of the function is S i,j , otherwise the return value of the function is 0; The similarity threshold index (STI) is constructed to evaluate the model prediction results under different similarity thresholds, as shown in Formula 7: STI is the similarity threshold index, n1 is the preferred sample under different similarity thresholds, n is the total samples in the corresponding period, RMSE1 is the root mean square error of the model, RMSE max is the maximum root mean square error.

7. The method for predicting the temporal spatial distribution of soil organic carbon density under limited sample conditions according to claim 1, characterized in that: The time series soil mapping idea includes one-time modeling and multiple use and multiple modeling and separate use; The modeling scheme for constructing the spatiotemporal prediction model includes at least one of the following: based on local samples of four periods, construct prediction models for different ages and make maps respectively; integrating local samples of four periods, centrally constructing models and making maps; based on similarity samples of four periods, constructing prediction models for different ages and making maps respectively; integrating similarity samples of four periods, centrally constructing models and making maps.

8. The method for predicting the temporal spatial distribution of soil organic carbon density under limited sample conditions according to claim 1, characterized in that: The time series spatial prediction mapping quantifies the uncertainty of the model by adopting the parameter-free simulation Bootstrap method, and obtains the probability distribution of the modeling results by performing spatial modeling on each Bootstrap sample; wherein the uncertainty of the modeling is quantified by the calculated confidence interval, and the final time series spatial distribution prediction result is the average value of N modelings, and the calculation method is shown in Formula 8: Where N is the number of random modeling, SOCD k is the final predicted soil organic carbon of the kth grid, SOCD(i) k is the soil organic carbon predicted by the ith model; The calculation method of uncertainty is shown in formula 9: Uncertainty k =(SOCD upper,k -SOCD lower,k )(Equation 9) Among them, Uncertainty k is the uncertainty of the kth grid; SOCD upper,k With SOCD lower,k They correspond to the upper and lower limits of the confidence interval of the predicted value of soil organic carbon density in the kth grid, respectively.

9. A prediction system for the soil organic carbon density temporal spatial distribution prediction method under limited sample conditions as described in any one of claims 1 to 8, characterized in that: include: The data acquisition module is used to obtain the publicly available historical data on soil organic carbon density and supplementary data on soil organic carbon density, and to conduct statistical analysis of soil organic carbon density and corresponding environmental variables; An analysis module, for analyzing the explanatory power of the environmental variables for regional soil organic carbon density based on a geographic detector model, and obtaining a soil-environment relationship; A sample expansion module, used to expand the limited samples based on the soil-environment relationship and geographical similarity to obtain a sample data set; A model building module, used to use the sample data set to build a spatiotemporal prediction model based on a machine learning algorithm and a time series soil mapping idea; The prediction module is used to carry out temporal spatial prediction mapping of soil organic carbon density based on the spatiotemporal prediction model, so as to realize the temporal spatial distribution prediction of regional soil organic carbon density.

10. A terminal, characterized in that: The terminal includes a memory, a processor, and a program stored in the memory and executable on the processor. When the program is executed by the processor, the steps of the method for predicting the temporal spatial distribution of soil organic carbon density under limited sample conditions as described in any one of claims 1 to 8 are implemented.

Images