Method for constructing soil fertility index comprehensive evaluation model based on multiple soil indexes

Abstract

The application provides a method for constructing a soil fertility index comprehensive evaluation model based on multiple soil indexes, which comprises the following steps: (1) collecting soil sample image data in different regions and preprocessing the same to obtain panoramic soil profile hyperspectral images; (2) extracting soil sample data and standard reflectance plate data; (3) performing radiation correction and preprocessing on the soil sample data and the standard reflectance plate data; (4) extracting characteristic wavelengths from the soil sample spectral reflectance data at different depths; (5) measuring the parameter data of the soil samples at different depths respectively; (6) constructing a multi-parameter inversion model at different depths; (7) establishing a soil fertility comprehensive evaluation model; and (8) obtaining different soil parameter and soil fertility index color distribution maps. The application can obtain soil profile spectral data at different depths through low-loss and rapid acquisition, thereby inverting soil index data and greatly improving the soil fertility evaluation efficiency.

Description

Technical Field

[0001] This invention relates to the field of soil fertility evaluation technology, specifically to a method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators. Background Technology

[0002] Soil health refers to the soil's ability, within the scope of ecosystems and land use, to sustain itself as a living system, maintain biological productivity, preserve air and water quality, and promote the health of plants, animals, and humans. Soil fertility refers to the soil's ability to consistently and appropriately supply and coordinate the air, temperature, nutrients, and non-toxic substances required for plant growth. Soil fertility is a comprehensive reflection of various physical and chemical properties of the soil; it is the primary function and essential attribute of the soil. Soil fertility is the result of the combined effects of the soil's inherent materials, structure, and physical and chemical properties with external environmental conditions.

[0003] Traditional soil analysis relies on laboratory chemical methods, which are destructive to samples and time-consuming. Conventional soil profile spectral observations have long depended on laboratory sampling and analysis. A few studies have conducted in-situ soil observations by excavating large square profiles (e.g., 2.2m long × 0.8m wide × 1.2m deep) for manual spectral measurements. However, this method suffers from high profile excavation costs, low efficiency, significant damage to soil structure, and the inability to achieve continuous dynamic monitoring.

[0004] Currently, soil parameter inversion based on spectral data is mostly single-index inversion, and the accuracy of parameter inversion is not high. Moreover, the spectral characteristics of soil at different depths will shift due to differences in physicochemical properties. Due to the lack of deep profile data, there are few studies on the changes in soil parameter content at different depths. Summary of the Invention

[0005] To address the technical problems existing in the background art, this invention proposes a method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators. The method is reasonable and can obtain spectral data of soil profiles at different depths with low loss and speed, thereby retrieving soil indicator data such as total nitrogen, organic matter, calcium carbonate, and water-soluble salts. Combined with the proposed soil fertility index (SFI) calculation, the efficiency of soil fertility evaluation is greatly improved.

[0006] To address the aforementioned technical problems, this invention provides a method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators, which specifically includes the following steps:

[0007] (1) Collect soil sample image data from different regions, and then preprocess the image data to obtain a panoramic view of the soil profile hyperspectral image;

[0008] (2) For the obtained soil profile hyperspectral image, different regions are divided according to soil depth and manual screening and sampling are carried out to screen out abnormal regions and standard reflectance plate regions. After removing the data of abnormal regions, soil sample data within each depth range is extracted. Then, the spectral data of the standard reflectance plate region is homogenized and the standard reflectance plate data is extracted.

[0009] (3) Perform radiometric correction on the extracted soil sample data and standard reflectance plate data, convert them into radiance data, and then convert the soil sample radiance data into soil sample spectral reflectance data based on the standard reflectance plate data. Finally, preprocess the soil sample spectral reflectance data.

[0010] (4) Extract characteristic wavelengths from the spectral reflectance data of soil samples at different depths and use them as inputs for the multi-parameter inversion model;

[0011] (5) Use chemical methods to measure the parametric data of soil samples at different depths, and use them as the output of the multi-parameter inversion model training;

[0012] (6) Design a deep learning network structure and train it to obtain a multi-parameter inversion model at different depths. This multi-parameter inversion model is used for multi-parameter spectral quantitative inversion of soil.

[0013] (7) Establish a comprehensive evaluation model for soil fertility based on the data of various parameters at different soil depths in different regions;

[0014] (8) By establishing the inversion algorithm model and the comprehensive evaluation model of soil fertility, multi-parameter inversion and soil fertility index calculation are performed on each pixel of the hyperspectral image to obtain color distribution maps of different soil parameters and soil fertility index.

[0015] The method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators, wherein the specific process of step (1) is as follows: using the constructed spectral sensing module and fiber optic image bundle to synthesize imaging spectral equipment, sample spectral data are collected for soil in different regions; then, for the selected soil sampling area, a pit is dug out, i.e., the location for soil sample collection, a standard reflectance plate is placed at the pit opening, and the imaging spectral equipment is placed at the pit opening and moved downwards to obtain image data; wavelet threshold denoising algorithm is used to remove honeycomb artifacts from the image data in sequence, and then the image edge distortion is eliminated based on the Brown-Conrady model, and the image is horizontally and vertically stitched using cylindrical coordinate system projection to obtain a panoramic view of the soil profile hyperspectral image.

[0016] The method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators includes: in step (3), the extracted soil sample data and standard reflectance plate data are radiantly corrected by using calibration data measured in the laboratory to convert the original digital number, i.e., the DN value, into a radiance value, so as to more accurately reflect the radiation characteristics of the soil.

[0017] The method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators includes: in step (3), the preprocessing of soil sample spectral reflectance data specifically involves Savitzky-Golay filtering and multivariate scattering correction; the Savitzky-Golay filtering smooths the data by performing local polynomial fitting on the soil sample spectral reflectance data to remove noise; the multivariate scattering correction is used to reduce changes caused by spectral scattering and eliminate spectral changes caused by differences in the physical properties of the samples.

[0018] The method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators, wherein the specific process of step (4) is as follows: the competitive adaptive reweighted sampling algorithm is used to extract the characteristic wavelengths of the spectral reflectance data of soil samples at different depths, the Monte Carlo sampling number is initialized to 1000, 80% are randomly selected as training set samples, and the remaining 20% ​​are validation set samples. The wavelength importance weight is calculated by the PLSR algorithm, and iterative sampling is continuously performed. The wavelength retention frequency is statistically analyzed, and the top 10% of wavelengths with the highest frequency are selected as the input of the multi-parameter inversion model.

[0019] The method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators, wherein step (5) involves using chemical methods to measure soil parameters such as organic matter, total nitrogen, salt content and calcium carbonate of soil samples in each region, which are used as outputs for training the multi-parameter inversion model.

[0020] The method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators, wherein the specific process of step (6) is as follows: establish an inversion algorithm model for four soil parameters, namely organic matter, total nitrogen, salinity and calcium carbonate, under different depth conditions, design a deep learning network and use the deep learning network to automatically learn features from hyperspectral data and perform soil parameter inversion.

[0021] The deep learning network structure includes an input layer, a feature extraction layer, multiple fully connected layers, and an output layer. The input layer takes in the spectral reflectance data after feature extraction. The feature extraction layer extracts important features through a residual module and a channel attention module. The multiple fully connected layers fuse features step by step through each fully connected layer, and finally output four soil parameters—organic matter, total nitrogen, salinity, and calcium carbonate—through a linear transformation. The output layer outputs the predicted values ​​of the four soil parameters.

[0022] During training, the deep learning network requires manual parameter setting: the convolutional layer activation function is ReLU, the optimizer is Adam, the loss function is mean squared error, the batch size is 64, and the learning rate is 0.0001. For soil samples at different depths, 90% of the samples in each region are set as the training set and 10% as the test set. The sample reflectance data is input into the deep learning network to train the inversion model and obtain the multi-parameter inversion network parameters for different depth ranges.

[0023] The method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators, wherein the specific process of establishing the inversion algorithm model is as follows:

[0024] (6.1.1) Data preparation

[0025] Input data, namely the hyperspectral reflectance data of the soil sample, is used as the input to the inversion algorithm model; output data, namely the experimental measurements of the four soil parameters corresponding to the soil sample, namely organic matter, total nitrogen, salinity and calcium carbonate, are used as the output of the inversion algorithm model.

[0026] (6.1.2) Selecting an inversion algorithm

[0027] The PLSR inversion algorithm is selected to map the input data to a low-dimensional space by dimensionality reduction and to predict soil parameters through regression analysis.

[0028] (6.1.3) Training and Optimization

[0029] Using k-fold cross-validation, the data is divided into k subsets. Each time the model is trained, k-1 subsets are used for training, and the remaining subset is used for validation. This process is repeated k times. The model performance is improved by adjusting the model parameters. The optimal number of features is selected, and the prediction error is calculated to evaluate the model's accuracy.

[0030] The method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators, wherein the comprehensive evaluation model of soil fertility established in step (7) is as follows:

[0031] ;

[0032] In the above formula, These represent the evaluation scores for four soil parameters: total nitrogen, organic matter, calcium carbonate, and salinity, respectively. These represent the proportions of the four soil parameters—total nitrogen, organic matter, calcium carbonate, and salt content—in the comprehensive evaluation model of soil fertility.

[0033] The method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators, wherein the specific process of step (8) is as follows: based on the multi-parameter inversion model and the comprehensive evaluation model of soil fertility index under different depth ranges, quantitative inversion is performed on each pixel of the pre-processed panoramic hyperspectral image to obtain the inversion grayscale images of the four soil parameters of total nitrogen, organic matter, calcium carbonate and salt content and soil fertility index respectively. The grayscale images are processed by pseudo-color to obtain a visual color distribution map of organic matter, total nitrogen, salt content, calcium carbonate content and soil fertility index on the panoramic soil profile.

[0034] By adopting the above technical solution, the present invention has the following beneficial effects:

[0035] This invention can acquire spectral data of soil profiles at different depths with low loss and speed, thereby retrieving soil index data such as total nitrogen, organic matter, calcium carbonate, and water-soluble salts. Combined with the proposed Soil Fertility Index (SFI) calculation, it can significantly improve the efficiency of soil fertility evaluation.

[0036] The comprehensive evaluation model for soil fertility index of this invention, through low-loss and rapid spectral data acquisition and multi-parameter inversion, not only significantly improves the efficiency of soil fertility evaluation, but also has the following beneficial effects and advantages:

[0037] (1) High efficiency and low loss acquisition

[0038] This invention employs an imaging spectral device that combines a spectral sensing module with an optical fiber image bundle, enabling rapid acquisition of soil profile spectral data at different depths without damaging the soil sample. Compared to traditional laboratory sampling and profile excavation methods, this significantly reduces the complexity of soil damage and sampling processes, achieving efficient and low-loss monitoring of large-scale soil areas.

[0039] (2) Multi-parameter integrated inversion improves accuracy

[0040] This invention is not limited to inverting a single soil parameter (such as organic matter, total nitrogen, calcium carbonate, etc.), but combines multiple soil indicators to comprehensively evaluate soil fertility. By using deep learning networks to invert multiple soil indicators, the accuracy of soil fertility evaluation is greatly improved, and the health status of the soil can be reflected more comprehensively and in detail.

[0041] (3) Dynamic monitoring and continuous evaluation

[0042] Traditional methods typically rely on static laboratory analysis, which cannot provide continuous and dynamic soil monitoring data. However, this invention, through non-destructive sampling and inversion techniques based on spectral data, enables long-term and continuous monitoring of soil fertility, providing real-time data support for farmland management.

[0043] (4) It is highly adaptable and suitable for different regions and soil types.

[0044] The method of this invention can be flexibly adapted to different soil types in different regions; by combining the specific properties of the soil with hyperspectral data for deep learning training, the model can be widely applied to different soil environments and has strong universality.

[0045] (5) Reduce manual intervention and costs

[0046] Compared to traditional manual sampling and chemical experimental analysis methods, the method of this invention reduces human intervention, lowers labor costs and laboratory analysis expenses. The automated data acquisition and analysis system not only improves efficiency but also reduces human error.

[0047] (6) Precise deep soil profile analysis

[0048] By employing spectral inversion technology, this invention can accurately analyze soil composition at different depths, particularly overcoming the difficulty of obtaining deep soil data using traditional techniques. It can not only assess topsoil but also provide precise fertility evaluations for deep soils, offering more comprehensive soil health data support for agricultural production.

[0049] (7) Environmentally friendly and meets the requirements of sustainable development.

[0050] The soil assessment method of this invention uses non-destructive spectral imaging technology, which avoids the environmental damage caused by traditional soil sampling, meets the needs of sustainable agriculture and ecological environmental protection, and helps to improve the green development of agricultural production.

[0051] (8) Improve the scientific nature of soil fertility management and agricultural decision-making.

[0052] This invention helps agricultural producers, researchers, and policymakers manage and utilize soil resources more scientifically by providing more accurate and comprehensive soil fertility assessment data. This not only helps improve the precision of soil management but also effectively reduces fertilizer application and lowers the environmental burden on agricultural production.

[0053] (9) Provide technical support for precision agriculture

[0054] This invention provides technical support for soil management in precision agriculture, enabling precise guidance for soil improvement and fertilization plans, improving the efficiency of agricultural resource utilization, optimizing agricultural production processes, and reducing resource waste and environmental pollution.

[0055] (10) Potential for large-scale application

[0056] The soil assessment model of this invention can process data from large-scale areas, is suitable for large-scale soil monitoring and assessment, has broad application prospects, and can meet the needs of modern agriculture for large-scale soil monitoring and precision management. Attached Figure Description

[0057] To more clearly illustrate the specific embodiments of the present invention or the technical solutions in the prior art, the drawings used in the description of the specific embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are some embodiments of the present invention. For those skilled in the art, other drawings can be obtained from these drawings without creative effort.

[0058] Figure 1 The flowchart illustrates the method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators, as described in this invention. Detailed Implementation

[0059] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0060] The present invention will be further explained below with reference to specific embodiments.

[0061] like Figure 1 As shown in the figure, the method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators provided in this embodiment specifically includes the following steps:

[0062] S100. Soil sample image data from different regions are acquired using a hyperspectral imaging device, and then the image data is preprocessed to obtain a panoramic hyperspectral image of the soil profile; the specific process is as follows:

[0063] Using a constructed spectral sensing module and fiber optic image bundle synthesized imaging spectrometer, spectral data of soil samples from different regions were collected. A cylindrical pit with a diameter of 0.5m and a height of 2m was excavated. A standard reflectance plate was placed at the pit opening, and the imaging spectrometer was positioned at the pit opening and moved downwards to acquire a batch of image data. Wavelet thresholding denoising algorithms were applied to the image data to remove honeycomb artifacts, and the Brown-Conrady model was used to eliminate image edge distortion. Lateral and vertical image stitching was performed using cylindrical coordinate system projection to obtain a panoramic hyperspectral image of the soil profile with a spectral range of 400-1000nm and a spectral resolution of 5nm.

[0064] S200. Divide the soil into different regions according to depth, manually screen for abnormal areas, extract 1000 spectral data points from each depth range, and homogenize the spectral data of the standard reflectance plate region, extracting one spectral data point from it; the specific process is as follows:

[0065] For the acquired hyperspectral image of the soil profile, the vertical axis represents the soil depth, which is 2m in reality. The hyperspectral image is divided vertically into 20 regions at a depth of 10cm, along with the standard reflectance plate region at the pit opening. Manual sampling is performed at different depths in the hyperspectral image of the soil profile. After removing abnormal data, 1000 soil sample data are extracted from each region. The standard reflectance plate region in the hyperspectral image of the soil profile is then selected by framing. The data in the selected region are averaged to obtain one standard reflectance plate data.

[0066] S300. Based on the laboratory calibration coefficients, the extracted spectral data are radiometrically corrected and converted into radiance data. Then, based on the standard reflectance radiance data, the soil sample radiance data is converted into reflectance and preprocessed. The specific process is as follows:

[0067] The soil sample data and standard reflectance plate data were radiometrically corrected using laboratory-measured calibration data to convert DN values ​​into radiance values. Based on the standard reflectance plate data, the soil sample radiance data were converted into soil sample spectral reflectance data. Then, the soil sample reflectance data were preprocessed by Savitzky-Golay filtering and multivariate scattering correction to reduce noise interference and baseline shifts and amplitude variations caused by soil particle size and roughness.

[0068] S400: The Competitive Adaptive Reweighted Sampling (CARS) algorithm is used to extract feature wavelengths from the spectral reflectance data of soil samples at different depths. The Monte Carlo sampling number is initialized to 1000. 80% of the samples are randomly selected as training set samples and the remaining 20% ​​are used as validation set samples. The wavelength importance weights are calculated using the PLSR algorithm. Iterative sampling is performed continuously, and the wavelength retention frequency is statistically analyzed. The top 10% of wavelengths with the highest frequency are selected as inputs to the multi-parameter inversion model.

[0069] S500 uses chemical methods to measure the organic matter, total nitrogen, salt content, and calcium carbonate of soil samples in each region, which are then used as outputs for training the multi-parameter inversion model.

[0070] S600. A deep learning network structure was designed and trained to obtain multi-parameter inversion models at different depths. This multi-parameter inversion model is used for multi-parameter spectral quantitative inversion of soil. Specifically, inversion algorithm models for four soil parameters (organic matter, total nitrogen, salinity, and calcium carbonate) under different depth conditions were established, and a deep learning network was designed.

[0071] The specific process of establishing the inversion algorithm model is as follows:

[0072] The goal of the inversion algorithm model is to extract soil parameters such as organic matter, total nitrogen, salinity, and calcium carbonate from spectral data. The inversion algorithm model typically uses methods such as regression analysis, least squares, or partial least squares regression (PLSR). The specific steps are as follows:

[0073] S610, Data Preparation

[0074] Input data: Prepare hyperspectral reflectance data of soil samples (e.g., wavelength features extracted using the CARS algorithm), which will be used as input to the inversion model;

[0075] Output data: Prepare experimental measurements of the four soil parameters (organic matter, total nitrogen, salinity and calcium carbonate) from the soil samples as the output of the inversion model.

[0076] S620, Selecting the Inversion Algorithm

[0077] Choosing the right inversion algorithm is crucial. Common inversion algorithms include multiple linear regression, partial least squares regression (PLSR), and support vector machine regression (SVR). Among these, PLSR is the most common and suitable for high-dimensional data and multi-parameter inversion.

[0078] The PLSR inversion algorithm is Partial Least Squares Regression (PLSR), a method combining Principal Component Analysis (PCA) and least squares regression. The goal of PLSR is to map input data to a lower-dimensional space through dimensionality reduction and predict soil parameters through regression analysis; the specific process is as follows:

[0079] S621. Standardize input data: Standardize the spectral data so that its mean is zero and its standard deviation is one, in order to eliminate dimensional differences.

[0080] S622, Dimensionality Reduction (PCA): Principal Component Analysis (PCA) is performed on the standardized spectral data to extract the most important principal components; usually, the first few principal components are selected, as they can explain most of the data variability.

[0081] S623. Regression Analysis: Regression analysis is performed using the selected principal components and target soil parameters to establish an inversion model.

[0082] S630, Training and Optimization

[0083] Cross-validation: To avoid overfitting and ensure the model's generalization ability, cross-validation is used to evaluate the model's performance. Typically, k-fold cross-validation is used, dividing the data into k subsets. Each time the model is trained, k-1 subsets are used for training, and the remaining subset is used for validation, repeating this process k times.

[0084] Model optimization: Improve model performance by adjusting model parameters (such as the number of principal components in PLSR). Select the optimal number of features and evaluate model accuracy by calculating prediction errors (e.g., root mean square error RMSE).

[0085] The specific process of designing the deep learning network described above is as follows:

[0086] The goal of designing deep learning networks is to automatically learn features from hyperspectral data and invert soil parameters. Common deep learning network architectures include multilayer perceptrons (MLPs), convolutional neural networks (CNNs), and recurrent neural networks (RNNs), but since the task is mainly a regression problem, fully connected neural networks (MLPs) are usually used.

[0087] ① Network structure design

[0088] Fully connected network (MLP) architecture is suitable for regression tasks, offering high flexibility and expressiveness.

[0089] Input layer: The number of nodes in the input layer is equal to the number of features in the reflectance data (i.e., the number of selected spectral wavelengths). For example, if 100 feature wavelengths are selected, then the input layer will contain 100 nodes.

[0090] Hidden layers: Generally, two to three hidden layers are set. The number of neurons in each layer can be adjusted according to the complexity of the data. Typically, you can start with 64 to 256 neurons and gradually adjust them.

[0091] Activation function: The hidden layer uses the ReLU activation function, which can improve the non-linear expressiveness of the model and avoid the gradient vanishing problem.

[0092] Output layer: Since this is a regression task, the number of nodes in the output layer is equal to the number of soil parameters that need to be predicted (e.g., four parameters: organic matter, total nitrogen, salinity, and calcium carbonate). The output layer typically uses a linear activation function (i.e., no activation function).

[0093] ②Network training

[0094] Loss function: Since this is a regression task, the loss function is usually chosen as the mean squared error (MSE), and its formula is as follows:

[0095] Optimizer: The Adam optimizer is used to optimize network parameters. The Adam optimizer combines momentum and an adaptive learning rate, enabling fast convergence in most problems.

[0096] Batch size and learning rate: Batch gradient descent or mini-batch gradient descent is typically used. Set an appropriate batch size (e.g., 64 or 128) and adjust the learning rate (e.g., 0.001) to ensure good convergence.

[0097] ③ Training process

[0098] Data splitting: The dataset is divided into a training set (usually 80%) and a test set (20%) for training and evaluating the model's performance.

[0099] Training: Train the network using the training set data, gradually update the network parameters, and minimize the loss function.

[0100] Evaluation: Evaluate the model's performance on the test set, calculate the root mean square error (RMSE) or other regression evaluation metrics, and ensure the model's generalization ability.

[0101] The final structure of the aforementioned deep learning network includes an input layer, a feature extraction layer, multiple fully connected layers, and an output layer. The input layer takes the extracted spectral reflectance data as input; the feature extraction layer extracts key features using residual modules and channel attention modules; each of the multiple fully connected layers progressively fuses the features, and finally, through linear transformation, outputs four soil parameters (organic matter, total nitrogen, salinity, and calcium carbonate); the output layer outputs the predicted values ​​of these four soil parameters. This design enables the network not only to capture spectral features related to soil fertility but also to automatically learn and optimize the model, improving inversion accuracy.

[0102] The input layer of the deep learning network is the reflectance data of the feature wavelengths selected in step S400. A residual module and a channel attention module are added to the feature extraction layer of the deep learning network, and finally, four fully connected layers are connected to obtain four types of soil parameter inversion data. The network hyperparameters (here, network hyperparameters refer to parameters that need to be manually set during deep learning model training. These parameters are not automatically optimized by the model through learning data, but need to be adjusted according to the actual situation. These hyperparameters are crucial to the model training process, affecting the model's convergence speed, learning effect, and final performance) are as follows: the convolutional layer activation function is ReLU, the optimizer is Adam, the loss function is mean squared error (MSE), the batch size is 64, and the learning rate is 0.0001. For 20 soil samples at different depths, 90% of the samples in each region are set as the training set, and 10% of the samples are set as the test set. The sample reflectance data is input into the deep network for inversion model training, obtaining multi-parameter inversion network parameters for 20 different depth ranges.

[0103] S700 establishes a comprehensive soil fertility evaluation model (SFI) for four types of parameters (organic matter, total nitrogen, salinity, and calcium carbonate) at different soil depths in different regions:

[0104] ;

[0105] In the above formula, These represent the evaluation scores for each soil parameter, i.e., scoring each type of parameter based on the data retrieved from the model and Table 1. These represent the proportion of each type of parameter in the evaluation model. Experts can assign values ​​to the proportions of each parameter in the evaluation model based on the actual land type to obtain the comprehensive fertility evaluation score (out of 100) for the soil sample.

[0106] S800. Using an inversion algorithm model and an evaluation model, multi-parameter inversion and soil fertility index calculation are performed on each pixel of the hyperspectral image to obtain a color distribution map of different soil parameters and soil fertility indices; the specific process is as follows:

[0107] Based on the multi-parameter inversion model and the comprehensive evaluation model of soil fertility index at different depth ranges, quantitative inversion is performed on each pixel of the pre-processed panoramic hyperspectral image to obtain inversion grayscale images of four soil parameters and soil fertility index. The grayscale images are then processed with pseudo-color to obtain a visualized color distribution map of organic matter, total nitrogen, salt content, calcium carbonate content and soil fertility index on the soil panoramic profile.

[0108] Table 1. Scores of Soil Fertility Parameters:

[0109]

[0110] This invention can acquire spectral data of soil profiles at different depths with low loss and speed, thereby retrieving soil index data such as total nitrogen, organic matter, calcium carbonate, and water-soluble salts. Combined with the proposed Soil Fertility Index (SFI) calculation, it can significantly improve the efficiency of soil fertility evaluation.

[0111] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention, and not to limit them; although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some or all of the technical features; and these modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators, characterized in that, Specifically, the following steps are included: (1) Collect soil sample image data from different regions, and then preprocess the image data to obtain a panoramic view of the soil profile hyperspectral image; (2) For the acquired soil profile hyperspectral image, different regions are divided according to soil depth and manual screening and sampling are carried out to screen out abnormal regions and standard reflectance plate regions. After removing the data of abnormal regions, soil sample data within each depth range is extracted. Then, the spectral data of the standard reflectance plate region is homogenized and the standard reflectance plate data is extracted. The specific process is as follows: the constructed spectral sensing module and fiber optic image bundle are used to synthesize imaging spectral equipment to collect sample spectral data for soil in different regions. Then, for the selected soil sampling area, the pit is dug out, i.e. the location of soil sample collection. The standard reflectance plate is placed at the pit opening, and the imaging spectral equipment is placed at the pit opening and moved downward to sweep and obtain image data. The image data is then subjected to wavelet threshold denoising algorithm to remove honeycomb artifacts. Then, the image edge distortion is eliminated based on the Brown-Conrady model. The image is then stitched horizontally and vertically using cylindrical coordinate system projection to obtain a panoramic view of the soil profile hyperspectral image. (3) Perform radiometric correction on the extracted soil sample data and standard reflectance plate data, convert them into radiance data, and then convert the soil sample radiance data into soil sample spectral reflectance data based on the standard reflectance plate data. Finally, preprocess the soil sample spectral reflectance data. (4) Extract characteristic wavelengths from the spectral reflectance data of soil samples at different depths and use them as inputs for the multi-parameter inversion model; (5) Use chemical methods to measure the parametric data of soil samples at different depths, and use them as the output of the multi-parameter inversion model training; Specifically, the soil parameters of organic matter, total nitrogen, salt content and calcium carbonate in each region are measured by chemical methods and used as the output of the multi-parameter inversion model for training. (6) Design a deep learning network structure and train it to obtain a multi-parameter inversion model at different depths. This multi-parameter inversion model is used for multi-parameter spectral quantitative inversion of soil. The specific process is as follows: establish an inversion algorithm model for four soil parameters, namely organic matter, total nitrogen, salinity and calcium carbonate, under different depth conditions, design a deep learning network and use the deep learning network to automatically learn features from hyperspectral data and perform soil parameter inversion. The deep learning network structure includes an input layer, a feature extraction layer, multiple fully connected layers, and an output layer. The input layer takes in the spectral reflectance data after feature extraction. The feature extraction layer extracts important features through a residual module and a channel attention module. The multiple fully connected layers fuse features step by step through each fully connected layer, and finally output four soil parameters—organic matter, total nitrogen, salinity, and calcium carbonate—through a linear transformation. The output layer outputs the predicted values ​​of the four soil parameters. The deep learning network requires manual parameter setting during training: the convolutional layer activation function is ReLU, the optimizer is Adam, the loss function is mean squared error, the batch size is 64, and the learning rate is 0.0001. For soil samples at different depths, 90% of the samples in each region are set as the training set and 10% as the test set. The sample reflectance data is input into the deep learning network to train the inversion model and obtain the multi-parameter inversion network parameters for different depth ranges. (7) Establish a comprehensive soil fertility evaluation model for various parameters at different soil depths in different regions; the established comprehensive soil fertility evaluation model is as follows: ； In the above formula, These represent the evaluation scores for four soil parameters: total nitrogen, organic matter, calcium carbonate, and salinity, respectively. These represent the proportions of the four soil parameters—total nitrogen, organic matter, calcium carbonate, and salinity—in the comprehensive soil fertility evaluation model, respectively. (8) By establishing the inversion algorithm model and the comprehensive evaluation model of soil fertility, multi-parameter inversion and soil fertility index calculation are performed on each pixel of the hyperspectral image to obtain color distribution maps of different soil parameters and soil fertility index.

2. The method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators as described in claim 1, characterized in that: In step (3), the extracted soil sample data and standard reflectance plate data are subjected to radiation correction. This involves using calibration data measured in the laboratory to convert the original digital number, i.e., the DN value, into a radiance value, so as to more accurately reflect the radiation characteristics of the soil.

3. The method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators as described in claim 1, characterized in that: In step (3), the preprocessing of the soil sample spectral reflectance data specifically involves Savitzky-Golay filtering and multivariate scattering correction. Savitzky-Golay filtering smooths the data by performing local polynomial fitting on the soil sample spectral reflectance data to remove noise. Multivariate scattering correction is used to reduce changes caused by spectral scattering and eliminate spectral changes caused by differences in the physical properties of the samples.

4. The method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators as described in claim 1, characterized in that, The specific process of step (4) is as follows: the competitive adaptive reweighted sampling algorithm is used to extract the characteristic wavelengths of the spectral reflectance data of soil samples at different depths. The Monte Carlo sampling number is initialized to 1000. 80% are randomly selected as training set samples and the remaining 20% ​​are validation set samples. The wavelength importance weight is calculated by the PLSR algorithm. Iterative sampling is continuously performed, the wavelength retention frequency is statistically analyzed, and the top 10% of wavelengths with the highest frequency are selected as the input of the multi-parameter inversion model.

5. The method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators as described in claim 1, characterized in that, The specific process for establishing the inversion algorithm model is as follows: (6.1.1) Data preparation Input data, namely the hyperspectral reflectance data of the soil sample, is used as the input to the inversion algorithm model; output data, namely the experimental measurements of the four soil parameters corresponding to the soil sample, namely organic matter, total nitrogen, salinity and calcium carbonate, are used as the output of the inversion algorithm model. (6.1.2) Selecting an inversion algorithm The PLSR inversion algorithm is selected to map the input data to a low-dimensional space by dimensionality reduction and to predict soil parameters through regression analysis. (6.1.3) Training and Optimization Using k-fold cross-validation, the data is divided into k subsets. Each time the model is trained, k-1 subsets are used for training, and the remaining subset is used for validation. This process is repeated k times. The model performance is improved by adjusting the model parameters. The optimal number of features is selected, and the prediction error is calculated to evaluate the model's accuracy.

6. The method for constructing a comprehensive evaluation model of soil fertility index based on multiple soil indicators as described in claim 1, characterized in that, The specific process of step (8) is as follows: Based on the multi-parameter inversion model and the comprehensive evaluation model of soil fertility index under different depth ranges, quantitative inversion is performed on each pixel of the pre-processed panoramic hyperspectral image to obtain the inversion grayscale images of the four soil parameters of total nitrogen, organic matter, calcium carbonate and salt content and soil fertility index respectively. The grayscale images are processed by pseudo-color to obtain the visualized color distribution map of organic matter, total nitrogen, salt content, calcium carbonate content and soil fertility index on the panoramic soil profile.

Images