A single-model multi-task joint inversion method for multiple soil indicators
By introducing a shared and task-specific dual-layer band gating mechanism within a single model framework, the problems of repetitive training and redundant bands in the inversion of multiple soil indicators are solved, thereby improving the stability and accuracy of multi-indicator soil inversion. This approach is suitable for unified modeling and batch inference of multiple soil indicators.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- CHINA UNIV OF GEOSCIENCES (WUHAN)
- Filing Date
- 2026-03-13
- Publication Date
- 2026-07-31
AI Technical Summary
Existing technologies suffer from repetitive training, deployment, and parameter tuning when inverting multiple soil indicators. Equal-weighted input across the entire hyperspectral band leads to redundant bands and noise interference. Furthermore, it is difficult to balance the commonalities and individual characteristics of tasks in multi-task shared modeling, which affects the stability and accuracy of joint inversion of multiple indicators.
A single-model, multi-task joint inversion method is adopted. Global band noise reduction is achieved through shared gating, and a dual-layer band gating mechanism with shared and task-specific gating is introduced in the spectral branch. Multi-index synchronous regression is achieved by combining multi-task output heads, and end-to-end training is performed using a multi-task joint loss function.
Under a single-model framework, the joint inversion of multiple soil indices was achieved, which improved the stability and accuracy of the inversion, reduced the computation and maintenance costs, and enhanced the applicability and generalization ability of multi-index inversion.
Smart Images
Figure CN122490973A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of intelligent soil information inversion and remote sensing data processing technology, and more specifically, to a single-model multi-task joint inversion method for multiple soil indicators. Background Technology
[0002] With the improvement of hyperspectral remote sensing data acquisition capabilities, using spectral information to invert soil properties such as organic matter content, salinity, nutrient content, and pH has become an important technical means in soil resource surveys, agricultural production management, and ecological environment monitoring. Existing technologies typically achieve quantitative prediction of target soil indicators by constructing regression models between spectral characteristics and soil indices.
[0003] For example, CN110376139A discloses a method for soil organic matter inversion based on hyperspectral data. This method estimates the soil organic matter content by extracting hyperspectral characteristic bands and constructing a regression model.
[0004] In addition, CN111508569B discloses a method for constructing a spatial distribution map of soil properties based on multi-source soil data and environmental variables. By constructing a soil property prediction model, the spatial distribution of soil indicators can be estimated.
[0005] Furthermore, CN120427870A discloses a soil salinity inversion method based on remote sensing data fusion and machine learning model, which uses remote sensing data features as input to a regression model to achieve soil salinity prediction.
[0006] However, the aforementioned existing technologies still have the following shortcomings: First, existing methods typically build independent prediction models for a single or a small number of soil indicators. When multiple soil indicators need to be inverted, multiple models need to be trained separately, resulting in repeated training, repeated calls and repeated deployments, which increases computational and maintenance costs.
[0007] Second, existing methods often directly use the entire hyperspectral band as model input, lacking a mechanism for dynamically adjusting the contribution of bands within the model structure. This can easily lead to redundant and noisy bands affecting the model's learning performance, thereby reducing model stability and prediction accuracy.
[0008] Third, although some methods adopt a multi-task prediction structure, they usually only make predictions through shared features and lack a structured adjustment mechanism for the differences in spectral sensitivity of different soil indicators. This can easily lead to insufficient expression of some indicators by shared features, thereby affecting the performance of multi-indicator joint inversion.
[0009] Fourth, existing methods, when applied to engineering applications, struggle to simultaneously achieve unified modeling of multiple indicators, suppression of hyperspectral redundancy, and representation of task differences. There is a lack of technical solutions that can simultaneously achieve global band adjustment and task difference compensation within a single model framework.
[0010] Therefore, there is an urgent need for a technical solution that can achieve joint inversion of multiple soil indices under a single model framework, and can simultaneously achieve redundant band suppression and task difference compensation through a structured band adjustment mechanism, so as to improve the stability and engineering applicability of multi-index soil inversion. Summary of the Invention
[0011] The purpose of this invention is to provide a single-model, multi-task joint inversion method for multiple soil indices, which can realize joint inversion of multiple soil indices, suppression of redundant bands, and compensation for task differences within a single-model framework, thereby improving the stability and engineering applicability of multi-indicator soil inversion.
[0012] This invention provides a single-model, multi-task joint inversion method for various soil indicators, comprising the following steps: S1: Perform data standardization processing on hyperspectral data, auxiliary features, and target values to obtain sample input features and a multi-task target set; S2: Construct a joint inversion model; S3: Train the joint inversion model using the multi-task joint loss function to obtain the trained joint inversion model; S4: Use the trained joint inversion model to predict the sample to be predicted, and obtain a multi-task prediction output set.
[0013] The present invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described single-model multi-task joint inversion method for multiple soil indicators.
[0014] Implementing the single-model multi-task joint inversion method for multiple soil indices provided by this invention has the following beneficial effects: This invention addresses the problems of repetitive training, deployment, and parameter tuning caused by single-indicator single-model in existing soil index inversion processes, as well as the problems of redundant bands and noise interference caused by equal-weighted hyperspectral full-band input, which affect the stability and accuracy of multi-objective joint learning, and the difficulty in balancing task commonality and task individuality in multi-task shared modeling. Targeting hyperspectral band sequences and auxiliary feature data, this invention achieves global band noise reduction through shared gating and introduces a two-layer band gating mechanism in the spectral branch, using shared gating and task-specific gating residuals. The task-specific gating residuals compensate for index differences, and the multi-task output head enables synchronous regression of multiple indices, achieving synchronous output of more than ten soil indices in a single inference, thus improving the stability and applicability of multi-indicator joint inversion.
[0015] Specifically, this invention constructs soil properties as a multi-objective set and builds input features composed of hyperspectral band sequences and auxiliary features. A shared band gating vector is set before the spectral branch to perform global adaptive weighting of the hyperspectral input to achieve redundancy suppression and noise reduction. In the task branching stage, task-specific band gating residuals are further introduced to perform fine-grained reweighting of the spectral features of each task, enhancing the ability of different soil indicators to represent differences in sensitive bands and avoiding false suppression of effective bands for specific indicators by a single gating mechanism. Subsequently, a common feature representation is extracted through a shared backbone, and multiple task output heads are set to regress the corresponding soil indicators. During the training phase, a unified optimization strategy is used to complete end-to-end joint training. During the inference phase, a single forward computation synchronously outputs the rasterized inversion results of more than ten soil indicators, thereby reducing overall training and deployment costs while improving the stability, accuracy, and generalization ability of multi-indicator inversion. Attached Figure Description
[0016] The present invention will be further described below with reference to the accompanying drawings and embodiments. In the accompanying drawings: Figure 1 This is a flowchart of the single-model multi-task joint inversion method for multiple soil indicators provided by the present invention; Figure 2 This is a schematic diagram of the overall process of the single-model multi-task joint inversion method provided by the present invention; Figure 3 This is a diagram of the dual-band gating structure provided by the present invention; Figure 4 This is a diagram of the multi-task joint inversion network structure provided by the present invention; Figure 5 This is a diagram showing the experimental results of dual-band gated ablation provided by the present invention. Detailed Implementation
[0017] To provide a clearer understanding of the technical features, objectives, and effects of the present invention, specific embodiments of the present invention will now be described in detail with reference to the accompanying drawings.
[0018] Figure 1 This diagram illustrates a single-model multi-task joint inversion method for multiple soil indices according to this embodiment. In this embodiment, the single-model multi-task joint inversion method for multiple soil indices includes the following steps: S1: Perform data standardization processing on hyperspectral data, auxiliary features, and target values to obtain sample input features and a multi-task target set.
[0019] In one exemplary embodiment, the hyperspectral data includes spectral data of multiple samples across multiple bands.
[0020] The auxiliary features include at least one of the following: topographic factors, meteorological factors, multispectral statistical features, and soil type coding.
[0021] The target value is the actual value of the soil index measured in the laboratory.
[0022] The soil indicators include at least organic matter content or pH value.
[0023] S2: Construct a joint inversion model.
[0024] In one exemplary embodiment, the joint inversion model is configured as follows: The hyperspectral input is globally weighted and modulated using shared band gating to obtain the modulated spectral input; The modulated spectral input and the auxiliary features are encoded and fused to obtain a shared spectral feature representation, an auxiliary feature representation, and a joint feature representation; The shared backbone network is used to extract the multi-task general features of the joint feature representation to obtain the shared general feature representation; The mission-specific gating weights are calculated based on the mission-specific band gating parameters. The shared spectral feature representation is then reweighted at the mission level using the mission-specific gating weights to obtain the mission-specific spectral features. Based on the task-specific spectral features and the shared general feature representation, the predicted values of each indicator are generated using the task adapter and the multi-task output head, resulting in a multi-task prediction output set.
[0025] In one exemplary embodiment, the formula for globally weighted modulation of the hyperspectral input using shared band gating is:
[0026]
[0027] in, The weight vector for shared band gating; Represents the Sigmoid mapping function; Shared band gating parameters set before spectral branching; Indicates the number of hyperspectral bands; This represents the spectral input after global weighted modulation via shared band gating; Indicates hyperspectral input; This indicates element-wise multiplication by band.
[0028] In one exemplary embodiment, the formula for encoding and fusion is: , , , in, Representation of shared spectral features; For spectral branch coding functions; For auxiliary feature representation; As an auxiliary feature; Z is the auxiliary branch encoding function; Z is the joint feature representation. This is the feature fusion function.
[0029] In one exemplary embodiment, the feature fusion function is any one of splicing, addition, and / or attention fusion.
[0030] In one exemplary embodiment, the formula for calculating the task-specific gating weight is:
[0031] in, Task-specific gating weights; For task-specific gated residual amplitude control coefficients; It is the hyperbolic tangent mapping function; To set mission-specific band gating parameters; This is the sequence number of the soil indicator task.
[0032] In one exemplary embodiment, the calculation formula for the task-specific spectral features is as follows:
[0033] in, For the first Task-specific spectral features corresponding to each task.
[0034] In one exemplary embodiment, the formula for generating the predicted values of each indicator is:
[0035]
[0036] in, For the task The predicted value of the corresponding indicator; Indicates a multi-tasking output header; To share common feature representations; Indicates a task adapter; This represents the set of multi-task prediction outputs.
[0037] S3: Train the joint inversion model using the multi-task joint loss function to obtain the trained joint inversion model.
[0038] In one exemplary embodiment, the calculation formula for the multi-task joint loss function is as follows: , , , in, For multi-task joint loss function; For the first Each task's loss weight; Indicates the first Predicted loss for each task, Indicates the first The true values of each soil indicator; Indicates the constraint loss term; Indicates a gating regularization term; and These are the weighting coefficients for the corresponding loss terms; Balance coefficients for shared gating and task-specific gating regularization terms; This represents the average value; This indicates the quantity of soil indicators.
[0039] S4: Use the trained joint inversion model to predict the sample to be predicted, and obtain a multi-task prediction output set.
[0040] In some embodiments, the above-described single-model multi-task joint inversion method for multiple soil indices can also be implemented in the following ways.
[0041] like Figure 2 The diagram shows the overall process of the single-model multi-task joint inversion method. In this embodiment, the single-model multi-task joint inversion method for multiple soil indicators includes the following steps: Step 1: Constructing the input feature and multi-task objective set Constructing sample input features Including hyperspectral spectral features with auxiliary features And construct a multi-task target set. Its definition is as follows: (1) (2) in, Indicates the number of samples. Indicates the number of hyperspectral bands. This indicates the number of soil indicators to be jointly inverted; Indicates the first The true value of each soil indicator.
[0042] Preferably, the auxiliary feature This may include topographic factors, meteorological factors, multispectral statistical characteristics, soil type coding, etc.
[0043] To improve training stability, the input features are standardized, which can be represented as: (3) in, Indicates the first 3D input features, and These represent the mean and standard deviation of the feature on the training set, respectively. To prevent constants with a denominator of zero.
[0044] Step 2: Construct a shared band gating system and globally weight the hyperspectral input. Set shared band gating parameters before spectral branching The shared gating weight vector is obtained through bounded mapping. The hyperspectral input is then weighted band by band, as shown in the following formula: (4) (5) in, This represents the Sigmoid mapping function. This indicates element-wise multiplication by band. This represents the spectrum input after shared-gated modulation.
[0045] This step enables the model to adaptively adjust band contributions based on the overall multi-task objective during joint training, suppressing the interference of redundant and noisy bands on subsequent shared feature learning.
[0046] Step 3: Shared Spectral Branch Coding and Auxiliary Feature Coding Input the spectrum after shared-gated modulation Input the shared spectrum encoding module to obtain the shared spectrum feature representation. ; auxiliary features Inputting into the auxiliary encoding module yields auxiliary feature representations. Its expression can be written as: (6) (7) in, For spectral branch coding functions, This is an auxiliary branch encoding function.
[0047] Furthermore, the features from the two branches are fused to form a joint feature representation (Z): (8) in, The feature fusion function is preferably any one of splicing, summation, or attention fusion.
[0048] Step 4: Extracting multi-task common features from the shared backbone network Joint feature representation Input a shared backbone network and extract general feature representations for multiple soil indicator tasks. Its expression is: (9) in, To share the backbone network mapping function, it is used to learn the common correlation information between multiple tasks.
[0049] This step enables shared modeling of multiple indicator tasks within a single model framework, reducing the repetitive modeling process for a single indicator.
[0050] Step 5: Construct mission-specific band gating residuals and compensate for index differences. For each soil indicator task Set task-specific band gating parameters Task-specific gating weights are generated using bounded residuals. The formula is as follows: (10) in, This is the task-specific gated residual amplitude control coefficient. It is the hyperbolic tangent mapping function.
[0051] Based on the task-specific gating weights, the shared spectrum encoding features are... Perform task-level reweighting to obtain task-specific spectral features. : (11) in, Indicates the first Spectral features corresponding to each task.
[0052] This step provides differentiated compensation for mission-sensitive bands based on shared gating, avoiding false suppression of effective bands for certain indicators by a single shared gating system.
[0053] like Figure 3 The diagram shown is of a dual-band gating structure.
[0054] Step Six: Construct the task adapter and multi-task output header to generate predicted values for each metric. For each task Build a task adapter With task output head and utilize shared common features Task-specific spectrum features Generate predicted values Its expression is: (12) in, This is used to map task-specific spectral features into a task-compensated representation with the same dimension as the shared features. Used to output the regression prediction value of the (k)th soil index.
[0055] This yields the multi-task prediction output set: (13) This step enables a single model to simultaneously predict multiple soil indicators through regression.
[0056] Step 7: Construct the joint loss function and perform end-to-end training Constructing a multi-task joint loss function The losses from each task are weighted and combined, and then the constraint loss and gating regularization term are added together, as shown in the following formula: (14) in, Indicates the first Predicted loss for each task, For the first Each task's loss weight; Indicates the constraint loss term; Indicates a gating regularization term; and These are the weighting coefficients for the corresponding loss terms.
[0057] Preferably, the gating regularization term simultaneously constrains shared gating and task-specific gating, and its expression is: (15) in, This represents the balance coefficient between shared gating and task-specific gating regularization terms. By minimizing this joint loss function, the shared gating parameters, task-specific gating parameters, shared backbone parameters, task adapter parameters, and task output header parameters are jointly updated.
[0058] like Figure 4 The diagram shows the structure of the multi-task joint inversion network.
[0059] Step 8: Model Inference and Simultaneous Output of Multiple Indicators During the inference phase, the raster data of the sample or region to be predicted is input into the trained model, a forward propagation is performed, and a multi-task prediction output set is obtained. : (16) in, This represents the joint inversion function of the single model after training. This represents the set of model parameters.
[0060] Furthermore, the first Predicted values for each task Mapped to corresponding soil index raster results : (17) in, This represents the rasterized output function. This step enables the synchronous output and unified delivery of more than ten soil indicators in a single inference process.
[0061] Compared with existing technologies, this invention does not simply apply multi-task learning methods to soil index inversion scenarios. Instead, it introduces a dual-layer band gating mechanism within a single-model, multi-task joint inversion framework. This allows for the coordinated completion of global redundancy suppression of hyperspectral input and compensation for differences in sensitive bands of each index during unified training. This ensures simultaneous output of multiple indices while accommodating both common task modeling and individual task expression. By incorporating multiple soil indices into the same model for joint training and inference, it reduces the problems of model expansion, computational resource consumption, and increased maintenance complexity caused by repetitive training, parameter tuning, and deployment in traditional "single-index, single-model" schemes, thus improving the overall implementation efficiency of multi-index inversion tasks. Furthermore, by using shared band gating... The hyperspectral input undergoes global adaptive weighting, which dynamically adjusts the contribution of each band according to the joint optimization objective, reducing the interference of low-contribution bands and noisy bands on the learning of shared features and improving the effectiveness of input feature utilization. By using task-specific band-gated residuals to differentiate the spectral features of each task on the basis of shared gating, the responsiveness of different soil indicators to sensitive bands can be enhanced, reducing the risk of false suppression of some effective bands of indicators by a single shared gating. At the same time, the synergistic optimization of dual-layer band gating and multi-task joint loss also gives the invention good training stability, structural scalability and engineering deployment applicability in multi-indicator joint inversion scenarios, making it suitable for unified modeling, batch inference and rasterized output of various soil indicators.
[0062] To verify the technical effectiveness of the dual-band gating mechanism of this invention for the joint inversion of multiple indicators, ablation experiments were conducted on 12 soil indicators, including pH, OC, TN, TP, AP, AK, AN, TK, CEC, BD, POROSITY, and CLAY. Figure 5As shown in the figure. Three models were set up in the experiment: no gating, shared gating only, and two-layer gating, and compared under the same training and validation sample conditions. The results show that the average R of the two-layer gating model is higher. 2 The average RMSE and average MAE are both superior to the comparative model. This indicates that the dual-band gating mechanism proposed in this invention can further enhance the differentiated response capability of different soil indicators to key sensitive bands on the basis of shared band screening, reduce redundant band interference, and improve the joint inversion performance of multiple indicators.
[0063] This embodiment provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described single-model multi-task joint inversion method for multiple soil indicators.
[0064] The embodiments of the present invention have been described above with reference to the accompanying drawings. However, the present invention is not limited to the specific embodiments described above. The specific embodiments described above are merely illustrative and not restrictive. Those skilled in the art can make many other forms under the guidance of the present invention without departing from the spirit and scope of the claims. All of these forms are within the protection scope of the present invention.
Claims
1. A single-model, multi-task joint inversion method for multiple soil indices, characterized in that, Includes the following steps: S1: Perform data standardization processing on hyperspectral data, auxiliary features, and target values to obtain sample input features and a multi-task target set; S2: Construct a joint inversion model; S3: Train the joint inversion model using the multi-task joint loss function to obtain the trained joint inversion model; S4: Use the trained joint inversion model to predict the sample to be predicted, and obtain a multi-task prediction output set.
2. The single-model multi-task joint inversion method for multiple soil indices according to claim 1, characterized in that, The joint inversion model is configured as follows: The hyperspectral input is globally weighted and modulated using shared band gating to obtain the modulated spectral input; The modulated spectral input and the auxiliary features are encoded and fused to obtain a shared spectral feature representation, an auxiliary feature representation, and a joint feature representation; The shared backbone network is used to extract the multi-task general features of the joint feature representation to obtain the shared general feature representation; The mission-specific gating weights are calculated based on the mission-specific band gating parameters. The shared spectral feature representation is then reweighted at the mission level using the mission-specific gating weights to obtain the mission-specific spectral features. Based on the task-specific spectral features and the shared general feature representation, the predicted values of each indicator are generated using the task adapter and the multi-task output head, resulting in a multi-task prediction output set.
3. The single-model multi-task joint inversion method for multiple soil indices according to claim 1, characterized in that, The hyperspectral data includes spectral data of multiple samples in multiple bands; the auxiliary features include at least one of topographic factors, meteorological factors, multispectral statistical features, and soil type codes; the target value is the true value of soil indicators measured in the laboratory; the soil indicators include at least organic matter content or pH value.
4. The single-model multi-task joint inversion method for multiple soil indices according to claim 2, characterized in that, The formula for globally weighted modulation of the hyperspectral input using shared band gating is as follows: ， ， in, The weight vector for shared band gating; Represents the Sigmoid mapping function; Shared band gating parameters set before spectral branching; Indicates the number of hyperspectral bands; This represents the spectral input after global weighted modulation via shared band gating; Indicates hyperspectral input; This indicates element-wise multiplication by band.
5. The single-model multi-task joint inversion method for multiple soil indices according to claim 2, characterized in that, The formulas for encoding and fusion are as follows: ， ， ， in, Representation of shared spectral features; For spectral branch coding functions; For auxiliary feature representation; As an auxiliary feature; Z is the auxiliary branch encoding function; Z is the joint feature representation. This is the feature fusion function.
6. The single-model multi-task joint inversion method for multiple soil indices according to claim 2, characterized in that, The formula for calculating the task-specific gating weight is as follows: ， in, Task-specific gating weights; For task-specific gated residual amplitude control coefficients; It is the hyperbolic tangent mapping function; To set mission-specific band gating parameters; This is the sequence number of the soil indicator task.
7. The single-model multi-task joint inversion method for multiple soil indices according to claim 2, characterized in that, The formula for calculating the task-specific spectral features is as follows: ， in, For the first Task-specific spectral features corresponding to each task.
8. The single-model multi-task joint inversion method for multiple soil indices according to claim 2, characterized in that, The formula for generating the predicted values of each indicator is as follows: ， ， in, For the task The predicted value of the corresponding indicator; Indicates a multi-tasking output header; To share common feature representations; Indicates a task adapter; This represents the set of multi-task prediction outputs.
9. The single-model multi-task joint inversion method for multiple soil indices according to claim 2, characterized in that, The formula for calculating the multi-task joint loss function is as follows: ， ， ， in, For multi-task joint loss function; For the first Each task's loss weight; Indicates the first Predicted loss for each task, Indicates the first The true values of each soil indicator; Indicates the constraint loss term; Indicates a gating regularization term; and These are the weighting coefficients for the corresponding loss terms; Balance coefficients for shared gating and task-specific gating regularization terms; This represents the average value; This indicates the quantity of soil indicators.
10. A computer program product, comprising a computer program, characterized in that, When executed by a processor, the computer program implements the steps of the single-model multi-task joint inversion method for multiple soil indices as described in any one of claims 1-9.