Soil organic matter content estimation method based on deep learning and hyperspectrum
Through the soil organic matter content estimation method based on deep learning and hyperspectral, the interference network model is used for reverse optimization, which solves the problem that learning models are difficult to optimize in the existing technology, and achieves the accuracy and stability of soil organic matter content estimation.
Patent Information
- Application Number
- CN202411640836.8
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2024-11-18
- Publication Date
- 2025-02-14
AI Technical Summary
The existing soil organic matter content estimation methods are difficult to optimize the learning model when facing a large amount of data and data refinement process, which makes it difficult to guarantee the estimation accuracy and cannot meet the usage needs of existing personnel.
The soil organic matter content estimation method based on deep learning and hyperspectral is adopted to train and tune the deep learning model by obtaining hyperspectral data, and set up the interference network model for reverse optimization to ensure the authenticity and stability of the model's estimation results. .
The continuous optimization and stability of the deep learning model is achieved, the accuracy and data stability of soil organic matter content estimation are ensured, and the use needs of existing personnel are met.
Smart Images
Figure CN119446322A_ABST
Abstract
Description
Technical Field
[0001] The present invention relates to the field of data processing technology, and in particular to a soil organic matter content estimation method based on deep learning and hyperspectral. Background Art
[0002] Hyperspectral technology can provide soil reflection or emission information in continuous and narrow spectral bands; this information includes detailed spectral characteristics of soil composition, structure, moisture and other characteristics; by measuring and analyzing these spectral characteristics, the soil organic matter content can be indirectly estimated, and by cooperating with the learning model, the soil organic matter content can be quickly estimated. However, although the existing soil organic matter content estimation is combined with the learning model for autonomous estimation, it is difficult to optimize the learning model in use in the face of a large amount of data and the process of data refinement. It is impossible to ensure the estimation accuracy of the learning model in a timely manner without affecting the learning model, and it is difficult to meet the usage needs of existing personnel. Summary of the invention
[0003] The purpose of the present invention is to provide a soil organic matter content estimation method based on deep learning and hyperspectral to address the shortcomings of the background technology.
[0004] In order to achieve the above object, the present invention provides the following technical solution: a soil organic matter content estimation method based on deep learning and hyperspectral, comprising the following steps: Acquire spectral characteristic data of target soil samples based on hyperspectral; The deep learning model is trained and tuned based on historical data to obtain a target model, wherein the historical data includes historical spectral feature data and corresponding soil organic matter content; An interference network model is set corresponding to the target model, and reverse optimization is performed on the target model through the interference network model; Input the spectral characteristic data of the target soil sample into the target model and output the soil organic matter estimation result; The soil model of soil organic matter content is obtained based on the visualization of soil organic matter results.
[0005] In a preferred embodiment, the step of training and tuning the deep learning model based on historical data to obtain a target model includes: Obtain historical spectral feature data and corresponding soil organic matter content as historical data, clean the historical data, and divide the cleaned historical data into a training set and a test set; The training set is input into the learning model for training and the learning model is tested through the test set, and the learning model that meets the test conditions is used as the target model.
[0006] In a preferred embodiment, the step of setting an interference network model corresponding to the target model and reversely optimizing the target model through the interference network model includes: Set up a preview space, copy the target model as the preview target model and put it into the preview space; Put the target model into the storage for storage; The storage is connected to the preview space and interference conditions are set. When the preview target model in the preview space meets the interference conditions, the preview target model is reversely optimized to the target model.
[0007] In a preferred embodiment, the step of setting up a preview space, copying the target model as the preview target model and putting it into the preview space includes: Setting a rehearsal space, wherein the rehearsal space includes a plurality of independent spaces and a main space; The target model is copied as a preview target model, and the preview target model is cloned at fixed points to obtain multiple sub-target models; The preview target model is put into the main space, and the main space is connected with multiple independent spaces; Put multiple sub-goal models one by one into independent spaces, and connect multiple independent spaces with big data; Obtain soil spectral characteristic data under big data to divide the plates, and obtain soil spectral characteristic data of multiple plates; The number of plate soil spectral feature data is the same as the number of independent spaces. The plate soil spectral feature data are put into independent spaces to continuously optimize the sub-target model.
[0008] In a preferred embodiment, the step of copying the target model as a preview target model and performing fixed-point clones of the preview target model to obtain multiple sub-target models includes: Copying the target model as a preview target model; The preview target model is fixed as a main fixed point and information is bound. Based on the main fixed point, multiple dispersed fixed points are dispersed to obtain a communication relationship between the main fixed point and the multiple dispersed fixed points. Based on the main fixed point, the preview target model is virtually differentiated to obtain multiple sub-target models, and the multiple sub-target models are information-bound with the dispersed fixed points; The preview target model is connected with multiple sub-target models according to the communication relationship between the main fixed point and the distributed fixed points.
[0009] In a preferred embodiment, the step of connecting the storage to the preview space and setting interference conditions, and when the preview target model in the preview space satisfies the interference conditions, performing reverse optimization of the preview target model to the target model comprises: Set interference conditions corresponding to the preset space; Obtain the optimization index of the plate soil spectral characteristic data for the sub-target model in the preview space, where the formula for the optimization index is: ,in, To optimize the index, is the number of sub-goal models, is the soil organic matter content obtained by the nth sub-target model trained with plate soil spectral feature data, is the reliability coefficient of the plate soil spectral characteristic data, The soil organic matter content obtained for the target model; When the optimization index corresponding to the preview target model in the preview space meets the interference condition, the preview target model is modified to complete the reverse optimization.
[0010] In a preferred embodiment, the step of inputting the spectral characteristic data of the target soil sample into the target model and outputting the soil organic matter estimation result comprises: Determine the spectral characteristic data of the target soil sample and perform data cleaning; The spectral characteristic data of the target soil sample after data cleaning is input into the target model, and the soil organic matter estimation result is output.
[0011] In a preferred embodiment, the step of visually displaying the soil organic matter content soil model based on the soil organic matter result includes: Obtain soil organic matter estimation results and corresponding soil locations as visualization information; The soil organic matter content soil model is obtained by constructing soil three-dimensionally using multiple continuous visualization information.
[0012] In the above technical solution, the technical effects and advantages provided by the present invention are: The present invention sets interference conditions corresponding to the preview space. When the preview target model meets the interference conditions, the preview target model can be used to optimize the target model to ensure the authenticity of the target model estimation result. The target model is operated in isolation in the storage to ensure the stability of the model operation. At the same time, the target model can be reversely optimized through the preview target model, which can ensure continuous optimization of the target model and stability of the target model under normal use. It has good data stability and ensures the accuracy of subsequent soil organic matter content estimation. BRIEF DESCRIPTION OF THE DRAWINGS
[0013] In order to more clearly illustrate the embodiments of the present application or the technical solutions in the prior art, the drawings required for use in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments recorded in the present invention. For ordinary technicians in this field, other drawings can also be obtained based on these drawings.
[0014] Figure 1 The present invention is a flow chart of the method. DETAILED DESCRIPTION
[0015] In order to make the purpose, technical solution and advantages of the embodiments of the present invention clearer, the technical solution in the embodiments of the present invention will be clearly and completely described below in conjunction with the drawings in the embodiments of the present invention. Obviously, the described embodiments are part of the embodiments of the present invention, not all of the embodiments. Based on the embodiments of the present invention, all other embodiments obtained by ordinary technicians in this field without creative work are within the scope of protection of the present invention.
[0016] Example 1, please refer to Figure 1 As shown, the soil organic matter content estimation method based on deep learning and hyperspectral in this embodiment includes the following steps: S1. Obtain spectral characteristic data of target soil samples based on hyperspectral; S2. Training and optimizing the deep learning model based on historical data to obtain a target model, wherein the historical data includes historical spectral feature data and corresponding soil organic matter content; S3, setting an interference network model corresponding to the target model, and reversely optimizing the target model through the interference network model; S4, inputting the spectral characteristic data of the target soil sample into the target model, and outputting the soil organic matter estimation result; S5. Visualize and display the soil organic matter results to obtain a soil model of soil organic matter content; As described in the above steps S1-S4, hyperspectral technology can provide reflection or emission information of soil in a continuous and narrow spectral band; this information contains detailed spectral features of soil composition, structure, moisture and other characteristics; by measuring and analyzing these spectral features, the soil organic matter content can be indirectly estimated, and by cooperating with the learning model, the estimation result of the soil organic matter content can be quickly obtained. However, although the existing soil organic matter content estimation combines the learning model for autonomous estimation, it is difficult to optimize the learning model in use in the face of a large amount of data and the process of data refinement, and it is impossible to ensure the estimation accuracy of the learning model in time without affecting the learning model, and it is difficult to meet the use needs of existing personnel; and the present application sets interference conditions for the preview space. When the preview target model meets the interference conditions, the preview target model can be optimized to the target model to ensure the authenticity of the target model estimation result, and the target model can be isolated in the storage to ensure the stability of the model operation. At the same time, the target model can be reversely optimized through the preview target model, which can ensure the continuous optimization of the target model and the stability of the target model under normal use, and has good data stability and ensures the accuracy of subsequent soil organic matter content estimation; In one embodiment, the step S2 of training and tuning the deep learning model based on historical data to obtain a target model includes: S21, obtaining historical spectral feature data and corresponding soil organic matter content as historical data, cleaning the historical data, and dividing the cleaned historical data into a training set and a test set; S22, input the training set into the learning model for training and test the learning model through the test set, and use the learning model that meets the test conditions as the target model; As described in the above steps S21-S22, in order to better understand the estimation results of soil organic matter content, it is necessary to analyze and summarize the characteristics of historical data through a learning model. Here, a multivariate stepwise regression model is used for model training to obtain a model suitable for soil organic matter content estimation. First, it is necessary to obtain historical spectral feature data, wherein the historical spectral feature data includes spectral features in the visible light region (in the visible light band (about 400-1000 nm), organic matter molecules in the soil will absorb or emit light of specific wavelengths, resulting in characteristic changes in the spectral curve. These features are used to distinguish different types of soil and determine the relative abundance of organic matter content), spectral features in the near-infrared region (in the near-infrared band (about 1000-2000 nm), soil organic matter will undergo a series of chemical bonding reactions, resulting in characteristic changes in the spectral curve. These features can be used to evaluate the absolute amount of organic matter content), spectral slope (reflects the sensitivity of organic matter in the soil to changes in spectral wavelength. Generally, the larger the spectral slope, the higher the organic matter content in the soil), spectral intercept (reflects the depth of the light absorption area on the soil spectral curve. Generally, the deeper the spectral intercept, the higher the soil organic matter content) and spectral noise (reflects the measurement error of the spectral curve, which will affect the analysis results of the spectral characteristics and other spectral characteristic data), clean the historical data, and divide the cleaned historical data into training sets and test sets. The cleaning of historical data includes removing noise, outliers, etc. to ensure the accuracy and reliability of the data. After that, the training set is input into the learning model for training and the learning model is tested through the test set. The learning model that meets the test conditions is used as the target model. In order to ensure the accuracy of the learning model, it is necessary to optimize the learning model according to the loss function to obtain the target model T = a0' + a1*v1' + a2'*v2' + ... + am'*vm' + c', where a0' , a1, a2, ... am are the new regression coefficients obtained after stepwise regression, which are the dynamic adjustments made during the training process. v1', v2' ... vm' are the soil spectral characteristic data retained in the model after the training learning model, which can obtain the corresponding soil organic matter content estimation and have good estimation ability.
[0017] In one embodiment, the step S3 of setting an interference network model corresponding to the target model and reversely optimizing the target model through the interference network model includes: S31, setting a preview space, copying the target model as a preview target model and putting it into the preview space; S32, putting the target model into a storage device for storage; S33, connecting the storage to the preview space and setting interference conditions, when the preview target model in the preview space meets the interference conditions, reversely optimizing the preview target model to the target model; As described in the above steps S31-S33, the rehearsal space here is a storage space that carries the operation of the rehearsal target model. The rehearsal target model is the same as the target model in the initial state. The rehearsal target model is copied from the target model. In order to facilitate the subsequent reverse optimization of the target model through the rehearsal target model, the target model is put into the storage for storage and operation. The rehearsal space here can be used as an optimization library of the target model. After that, the storage is connected to the rehearsal space, and interference conditions are set for the rehearsal space. When the rehearsal target model meets the interference conditions, the rehearsal target model can be used to optimize the target model to ensure the authenticity of the target model estimation result. The target model can be run in isolation in the storage to ensure the stability of the model operation. At the same time, the target model can be reversely optimized through the rehearsal target model, which can effectively ensure the accuracy and stability of the model estimation and ensure the stable operation of the system. In one embodiment, the step S31 of setting up a preview space, copying the target model as a preview target model and putting it into the preview space includes: S311, setting a rehearsal space, wherein the rehearsal space includes a plurality of independent spaces and a main space; S312, copying the target model as a preview target model, and performing fixed-point clones on the preview target model to obtain multiple sub-target models; S313, putting the preview target model into the main space, and connecting the main space with the multiple independent spaces; S314, putting multiple sub-goal models into independent spaces one by one, and connecting the multiple independent spaces with the big data; S315, obtaining soil spectral characteristic data under big data to perform plate division, and obtaining soil spectral characteristic data of multiple plates; S316, the number of plate soil spectral feature data is the same as the number of independent spaces, and the plate soil spectral feature data is put into the independent space to continuously optimize the sub-target model; As described in the above steps S311-S316, the rehearsal space is a space that is connected to the external big data. The rehearsal space is used as the optimization database of the target model. In order to ensure the comprehensiveness and referenceability of the data, multiple independent spaces and a main space are divided in the rehearsal space. The main space is connected to multiple independent spaces, and the multiple independent spaces are independent of each other. Then, the target model is copied as the rehearsal target model. The copied rehearsal target model is put into operation in the main space. The rehearsal target model is cloned at a fixed point to obtain multiple sub-target models. The rehearsal target model is information-connected between the multiple sub-target models. Then, multiple sub-target models are put into independent spaces one by one, and multiple independent spaces are connected with big data. The soil spectral feature data under the big data is obtained to divide the plates to obtain multiple plate soil spectral feature data. One plate soil spectral feature data corresponds to the continuous training of a sub-target model in an independent space. The preview target model in the main space does not participate in the training. The preview target model in the main space is used to summarize the training results of multiple independent spaces, which can ensure the continuous optimization of the target model and the stability of the target model under normal use, have good data stability and ensure the accuracy of subsequent soil organic matter content estimation; In one embodiment, the step S312 of copying the target model as a preview target model and performing fixed-point clones of the preview target model to obtain multiple sub-target models includes: S3121, copying the target model as a preview target model; S3122, a fixed point of the preview target model is used as a main fixed point and information is bound, and multiple dispersed fixed points are obtained by dispersing based on the main fixed point, and a communication relationship is established between the main fixed point and the multiple dispersed fixed points; S3123, performing virtual differentiation on the preview target model based on the main fixed point to obtain multiple sub-target models, and binding the multiple sub-target models with the dispersed fixed points; S3124, connecting the preview target model with the plurality of sub-target models according to the communication relationship between the main fixed point and the dispersed fixed points; As described in the above steps S3121-S3124, the target model is an external model for estimating soil organic matter, the preview target model is a model for reverse optimization of the target model, the preview target model is run in the preview space, the preview target model is fixed as the main fixed point and information binding is performed, and multiple dispersed fixed points are obtained based on the main fixed point. There is a communication relationship between the main fixed point and the dispersed fixed points. Here, the main fixed point and the dispersed fixed points are both servers, and the dispersed fixed point server is a part of the dispersed association of the main fixed point server. The dispersed fixed point server belongs to the main fixed point server and is served by the main fixed point server. The server for running the sub-target model divided from the server is used to virtually differentiate the preview target model based on the main fixed point to obtain multiple sub-target models. The multiple sub-target models are information-bound with the dispersed fixed points. The virtual differentiation here is because the multiple sub-target models differentiated from the preview target model are still running in the dispersed fixed points belonging to the main fixed point server. Therefore, the computing power occupied by the multiple sub-target models is still the same as that of the preview target model. In this way, the preview target model can synchronize the training data of the sub-target model, and at the same time, it can also realize the dispersed data of the sub-target model for model training, so that the training result is more accurate. In one embodiment, the step S33 of connecting the storage to the preview space and setting the interference condition, and performing reverse optimization of the preview target model to the target model when the preview target model in the preview space meets the interference condition, comprises: S331, setting interference conditions corresponding to the preset space; S332, obtaining the optimization index of the plate soil spectral characteristic data in the preview space for the sub-target model, wherein the formula of the optimization index is: ,in, To optimize the index, is the number of sub-goal models, is the soil organic matter content obtained by the nth sub-target model trained with plate soil spectral feature data, is the reliability coefficient of the plate soil spectral characteristic data (a reliability evaluation value of the channel for obtaining plate soil spectral characteristic data), The soil organic matter content obtained by the target model. The larger the optimization index is, the less accurate the target model is. The optimization index is calculated by summarizing the preview target model in the main space. S333, when the optimization index corresponding to the preview target model in the preview space meets the interference condition, the preview target model is modified to complete the reverse optimization; As described in the above steps S331-S333, an interference condition is set for the corresponding preset space, and the interference condition is the threshold value of the corresponding optimization index exceeding the reasonable range. When the optimization index corresponding to the preview target model in the preview space meets the interference condition (that is, the optimization index exceeds the reasonable range), the preview target model is modified to complete the reverse optimization. The optimization here is to modify the target model through the preview target model after training and adjustment, which is reverse optimization. It can ensure the normal use of the target model in daily life, and can also optimize the target model in time through the preview target model in the background to ensure the accuracy of the estimation of soil organic matter content.
[0018] In one embodiment, the step S4 of inputting the spectral characteristic data of the target soil sample into the target model and outputting the soil organic matter estimation result includes: S41, determining the spectral characteristic data of the target soil sample and performing data cleaning; S42, inputting the spectral characteristic data of the target soil sample after data cleaning into the target model, and outputting the soil organic matter estimation result; As described in the above steps S41 and S42, the spectral characteristic data of the target soil sample is determined and data cleaning is performed. The data cleaning method is: identify and delete outliers. Outliers are usually inconsistent with most of the data. Outliers can be identified by statistical methods (such as the 3σ principle, that is, data points that deviate from the mean value by more than three times the standard deviation are considered outliers), visualization methods (such as scatter plots, box plots), and intuitive observation. For the identified outliers, strategies such as deletion, setting as missing values, filling (such as using the mean, median, mode, etc.) or not processing can be adopted. Processing missing values: For missing values, filling (such as using the mean, median, mode, interpolation, etc.), deletion or interpolation can be adopted. When processing missing values, it is necessary to keep the overall distribution and characteristics of the data unchanged; data standardization is a method of scaling data according to a certain rule to eliminate the dimensional differences between different data. It can make data easier to compare and analyze; commonly used data standardization methods include Z-score standardization (i.e. converting data into a distribution with a mean of 0 and a standard deviation of 1) and Min-Max standardization (i.e. scaling the data to between a specified minimum and maximum value); the spectral feature data of the target soil sample after data cleaning is then input into the target model, and the soil organic matter estimation result is output, where the soil organic matter estimation result is the estimated soil organic matter content.
[0019] In one embodiment, the step S5 of visually displaying the soil organic matter content soil model based on the soil organic matter result includes: S51, obtaining soil organic matter estimation results and corresponding soil positions as visualization information; S52, constructing soil three-dimensionally using the continuous multiple visualization information to obtain a soil model of soil organic matter content; As described in the above steps S51 and S52, the soil organic matter estimation result and the corresponding soil position are obtained as visualization information. The soil organic matter estimation result here is processed in blocks, so the organic matter estimation results at different positions are given separately. Therefore, a soil model of soil organic matter content is obtained by constructing a continuous plurality of visualization information in three dimensions, and a soil model analysis of soil organic matter content can be obtained, which is convenient for the staff to analyze the soil organic matter content. The above is only a specific implementation of the present application, but the protection scope of the present application is not limited thereto. Any person skilled in the art who is familiar with the present technical field can easily think of changes or substitutions within the technical scope disclosed in the present application, which should be included in the protection scope of the present application. Therefore, the protection scope of the present application should be based on the protection scope of the claims.
Claims
1. A soil organic matter content estimation method based on deep learning and hyperspectral, characterized in that: The following steps are involved: Acquire spectral characteristic data of target soil samples based on hyperspectral; The deep learning model is trained and tuned based on historical data to obtain a target model, wherein the historical data includes historical spectral feature data and corresponding soil organic matter content; An interference network model is set corresponding to the target model, and reverse optimization is performed on the target model through the interference network model; Input the spectral characteristic data of the target soil sample into the target model and output the soil organic matter estimation result; The soil model of soil organic matter content is obtained based on the visualization of soil organic matter results.
2. The soil organic matter content estimation method based on deep learning and hyperspectral according to claim 1 is characterized in that: The step of training and tuning the deep learning model based on historical data to obtain a target model includes: Obtain historical spectral feature data and corresponding soil organic matter content as historical data, clean the historical data, and divide the cleaned historical data into a training set and a test set; The training set is input into the learning model for training and the learning model is tested through the test set, and the learning model that meets the test conditions is used as the target model.
3. The soil organic matter content estimation method based on deep learning and hyperspectral according to claim 1 is characterized in that: The step of setting an interference network model corresponding to the target model and reversely optimizing the target model through the interference network model includes: Set up a preview space, copy the target model as the preview target model and put it into the preview space; Put the target model into the storage for storage; The storage is connected to the preview space and interference conditions are set. When the preview target model in the preview space meets the interference conditions, the preview target model is reversely optimized to the target model.
4. The soil organic matter content estimation method based on deep learning and hyperspectral according to claim 3 is characterized in that: The step of setting up a preview space, copying the target model as the preview target model and putting it into the preview space includes: Setting a rehearsal space, wherein the rehearsal space includes a plurality of independent spaces and a main space; The target model is copied as a preview target model, and the preview target model is cloned at fixed points to obtain multiple sub-target models; The preview target model is put into the main space, and the main space is connected with multiple independent spaces; Put multiple sub-goal models one by one into independent spaces, and connect multiple independent spaces with big data; Obtain soil spectral characteristic data under big data to divide the plates, and obtain soil spectral characteristic data of multiple plates; The number of plate soil spectral feature data is the same as the number of independent spaces. The plate soil spectral feature data are put into independent spaces to continuously optimize the sub-target model.
5. The soil organic matter content estimation method based on deep learning and hyperspectral according to claim 4 is characterized in that: The step of duplicating the target model as a preview target model and performing fixed-point clones of the preview target model to obtain a plurality of sub-target models comprises: Copying the target model as a preview target model; The preview target model is fixed as a main fixed point and information is bound. Based on the main fixed point, multiple dispersed fixed points are dispersed to obtain a communication relationship between the main fixed point and the multiple dispersed fixed points. Based on the main fixed point, the preview target model is virtually differentiated to obtain multiple sub-target models, and the multiple sub-target models are information-bound with the dispersed fixed points; The preview target model is connected with multiple sub-target models according to the communication relationship between the main fixed point and the distributed fixed points.
6. The soil organic matter content estimation method based on deep learning and hyperspectral according to claim 5 is characterized in that: The step of connecting the storage to the preview space and setting interference conditions, and when the preview target model in the preview space satisfies the interference conditions, performing reverse optimization of the preview target model to the target model comprises: Set interference conditions corresponding to the preset space; Obtain the optimization index of the plate soil spectral characteristic data for the sub-target model in the preview space, where the formula for the optimization index is: ,in, To optimize the index, is the number of sub-goal models, is the soil organic matter content obtained by the nth sub-target model trained with plate soil spectral feature data, is the reliability coefficient of the plate soil spectral characteristic data, The soil organic matter content obtained for the target model; When the optimization index corresponding to the preview target model in the preview space meets the interference condition, the preview target model is modified to complete the reverse optimization.
7. The soil organic matter content estimation method based on deep learning and hyperspectral according to claim 1 is characterized in that: The step of inputting the spectral characteristic data of the target soil sample into the target model and outputting the soil organic matter estimation result comprises: Determine the spectral characteristic data of the target soil sample and perform data cleaning; The spectral characteristic data of the target soil sample after data cleaning is input into the target model, and the soil organic matter estimation result is output.
8. The soil organic matter content estimation method based on deep learning and hyperspectral according to claim 1, characterized in that: The step of visually displaying the soil organic matter results to obtain a soil model of soil organic matter content comprises: Obtain soil organic matter estimation results and corresponding soil locations as visualization information; The soil organic matter content soil model is obtained by constructing soil three-dimensionally using multiple continuous visualization information.