Soil salinity remote sensing inversion method and system based on mixed feature optimization
By fusing multi-source remote sensing data and integrating learning models, the problems of insufficient generalization ability of single data sources and models in UAV remote sensing soil salinity estimation are solved, achieving high-precision and stable soil salinity estimation and supporting salinization monitoring and management in precision agriculture.
Patent Information
- Application Number
- CN202511750355.7
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- Filing Date
- 2025-11-26
- Publication Date
- 2026-03-06
AI Technical Summary
Existing UAV remote sensing technology for soil salinity estimation relies on a single data source, lacks near-infrared and red-edge bands, has low spatial resolution, is easily affected by light and vegetation transpiration, has weak generalization ability of a single model, and has uncertain selection of feature parameters, resulting in insufficient accuracy.
By employing multi-source remote sensing data fusion, recursive feature elimination, and ensemble learning methods, the optimal feature combination is selected, an ensemble learning model is constructed, and combined with a multi-index evaluation system, soil salinity estimation is optimized.
It achieves high-precision and stable soil salinity estimation under complex vegetation cover, improves the model's generalization ability and estimation accuracy, and provides a powerful monitoring tool for precision agriculture.
Smart Images

Figure CN121616989A_ABST
Abstract
Description
Technical Field
[0001] This invention relates to the field of agricultural water and soil engineering technology, specifically to a remote sensing inversion method and system for soil salinity based on mixed feature optimization. Background Technology
[0002] Soil salinization severely weakens the water absorption capacity of crop roots and gradually damages soil structure. This not only leads to farmland degradation and reduced crop yields, but also poses a significant threat to crop growth and regional biodiversity. To support prevention and control and land management, and to provide rapid, macroscopic, and time-series monitoring methods, it is necessary to quantitatively assess salt content through remote sensing inversion. Remote sensing inversion is the process of using spectral, thermal infrared, and radar data observed by satellites or drones to invert surface salinity or related indicators through algorithms.
[0003] Existing UAV-based soil salinity estimation technologies largely rely on single data sources. Visible light imagery (RGB) data only covers the visible light band, lacking salinity-sensitive bands such as near-infrared and red-edge images. Multispectral imagery (MS) data has low spatial resolution, and thermal infrared imagery (TIR) data is easily affected by light, wind speed, and vegetation transpiration, resulting in limited inversion accuracy when used alone. Furthermore, existing technologies for selecting remote sensing feature variables often focus on single methods, primarily assessing the contribution of feature parameters while neglecting the selection of the number of parameters. Although thresholds can be set to filter parameter variables, their selection involves many uncertainties, further impacting the model's soil salinity estimation accuracy. Simultaneously, machine learning models for soil salinity estimation built based on UAV remote sensing mostly focus on using single machine learning algorithms. Single models exhibit significant spatiotemporal heterogeneity, resulting in weak generalization ability, overfitting problems, and a failure to fully exploit the complementarity of multi-source remote sensing data.
[0004] Therefore, there is an urgent need for a new method to achieve remote sensing inversion in complex vegetation cover scenarios. Summary of the Invention
[0005] Purpose of the invention:
[0006] The present invention aims to provide a soil salinity estimation model method and system that can achieve deep fusion of multi-source remote sensing data, perform stable and efficient feature screening, and have strong generalization ability.
[0007] Technical solution:
[0008] This invention proposes a remote sensing inversion method for soil salinity based on hybrid feature optimization, comprising:
[0009] Remote sensing images from different data sources within the target area are acquired, soil samples at different depths are collected, and their salinity is measured. The remote sensing images are preprocessed to extract the original band values of the remote sensing images, and a set of feature variables is constructed for each data source. The data sources include visible light images, multispectral images, and thermal infrared images.
[0010] Using the salinity content of soil samples at different depths as the target variable, a recursive feature elimination method was used to screen out the optimal feature subsets of each feature variable set; all optimal feature subsets were randomly permuted and combined to obtain several feature combination schemes, each of which includes at least one optimal feature subset;
[0011] According to the feature combination scheme, the feature variables are combined using the salinity content of soil samples at different depths as an index to obtain several modeling datasets; for each modeling dataset, a corresponding ensemble learning model is constructed through fusion training.
[0012] A multi-index evaluation system was established to select the optimal soil salinity estimation model from all integrated learning models. The optimal soil salinity estimation model was then deployed to the equipment to quantitatively invert the soil salinity of the target area.
[0013] Furthermore, the remote sensing images from different data sources include visible light images, multispectral images, and thermal infrared images; the preprocessing includes mosaicking, radiometric calibration, and orthorectification.
[0014] Furthermore, the original values of the bands include the brightness values of the R, G, and B bands of the visible light image, the reflectance of the blue, green, red edges, and near-infrared bands of the multispectral image, and the brightness temperature values of the thermal infrared image.
[0015] Furthermore, the initial set of feature variables includes:
[0016] Spectral indices, including canopy coverage, color index, vegetation index, salinity index, canopy temperature value, and canopy temperature-derived indices;
[0017] Texture features are extracted using the gray-level co-occurrence matrix method. Feature parameters include mean, variance, entropy, homogeneity, contrast, correlation, dissimilarity, and second moment.
[0018] Furthermore, the method for recursive feature elimination includes:
[0019] An initial dataset is formed by any set of feature variables and a set of soil sample salinity. Random forest is used as the base model. Starting with all features in the current initial dataset, ten-fold cross-validation is used. At each step, the feature with the lowest importance score is removed. The iteration stops when the root mean square error of cross-validation is minimized.
[0020] The initial datasets from all iterations are combined according to the data source, and common features from soil samples at different depths are retained to obtain the optimal feature subset corresponding to each data source.
[0021] Furthermore, for each modeling dataset, constructing a corresponding ensemble learning model through fusion training includes:
[0022] Using the aforementioned modeling dataset, training is performed on several pre-defined base models, including Gaussian process regression, support vector machine, BP-MLP, and extreme learning machine.
[0023] Automatic hyperparameter optimization is performed by using nested cross-validation with an outer 5-fold layer and an inner 10-fold layer, combined with Bayesian optimization, to obtain the predicted values of the modeling dataset in the base model. Each predicted value is then used as a new input value for fusion training to obtain the corresponding ensemble learning model.
[0024] Furthermore, the fusion training includes the following methods, each of which generates an ensemble learning model:
[0025] (1) Ridge regression is used as the meta-model for secondary training on the training set to obtain the stacked ensemble model;
[0026] (2) Assign weight values to each base model based on the magnitude of the inverse RMSE of each base model on the training set to obtain a weighted average ensemble model;
[0027] (3) Average the predicted values of each base model to obtain the arithmetic mean ensemble model.
[0028] Furthermore, the multi-indicator evaluation system includes:
[0029] Calculate the evaluation metrics for each ensemble learning model, including: coefficient of determination, root mean square error, mean absolute error, performance-to-interquartile range ratio, and performance-to-standard deviation ratio.
[0030] The comprehensive score of each ensemble learning model is obtained according to the comprehensive scoring formula, and the ensemble learning model with the highest score is the optimal soil salinity estimation model; the comprehensive scoring formula is expressed as follows:
[0031]
[0032] In the formula, F i Let R represent the score of the i-th model, where α, β, μ, δ, and θ represent the weight values. 2 min RMSE is the minimum value of the coefficient of determination. max The root mean square error (MAE) is the maximum value of the root mean square error. max RPIQ is the maximum value of the mean absolute error. minThe minimum ratio of performance to interquartile range, RPD min It is the minimum value of the ratio of performance to standard deviation.
[0033] Furthermore, the deployment of the optimal soil salinity estimation model to the device includes:
[0034] The optimal soil salinity estimation model is persistently stored to form a model file that can be called independently. When it is necessary to perform inversion on the target area, the model file is loaded on the device, new remote sensing data after preprocessing and feature extraction is input, forward inference is performed, and a quantitative estimate of soil salinity is output.
[0035] On the other hand, the present invention also proposes a soil salinity remote sensing inversion system based on mixed feature optimization, including a memory, a processor and a computer program stored in the memory, wherein the processor executes the computer program to implement the steps of any of the aforementioned methods.
[0036] Beneficial effects: Compared with the prior art, the technical solution provided by the present invention has the following significant beneficial effects:
[0037] (1) Advantages of data fusion: By fusing visible light, multispectral and thermal infrared multi-source remote sensing data, information complementarity is achieved at the feature level, overcoming the shortcomings of one-sided information from a single data source, and providing a rich data foundation for high-precision inversion.
[0038] (2) Advantages of feature selection: By using a hybrid strategy of random forest importance assessment and recursive feature elimination, the optimal number and combination of features can be determined automatically and stably, avoiding the subjectivity and uncertainty of manually setting thresholds and improving the robustness and reliability of feature selection.
[0039] (3) Model performance advantages: The integrated learning framework combines the prediction capabilities of multiple heterogeneous machine learning models, effectively reducing the risk of overfitting and estimation bias of a single model, and significantly improving the generalization ability and estimation accuracy of the model under different spatiotemporal conditions.
[0040] (4) Application value advantages: The integrated solution finally formed by the present invention can efficiently and accurately estimate the salt content of soil layers at different depths under crop cover, providing a powerful technical tool for monitoring and managing saline soil in precision agriculture, and has high promotion value and economic and social benefits. Attached Figure Description
[0041] Figure 1 This is a flowchart of the method of the present invention;
[0042] Figure 2 shows the screening results of the recursive feature elimination method in this embodiment. Specifically, Figure 2(a) corresponds to the initial dataset of RGB 0~10cm, Figure 2(b) corresponds to the initial dataset of RGB 10~40cm, Figure 2(c) corresponds to the initial dataset of MS 0~10cm, Figure 2(d) corresponds to the initial dataset of MS 10~40cm, Figure 2(e) corresponds to the initial dataset of TIR 0~10cm, and Figure 2(f) corresponds to the initial dataset of TIR 10~40cm.
[0043] Figure 3 The results of soil salinity estimation and evaluation indices from 0 to 10 cm in the embodiments of this invention;
[0044] Figure 4 The results of soil salinity estimation and evaluation indices for the model 10~40cm in this embodiment of the invention;
[0045] Figure 5 This is the soil salinity estimation verification result of the optimal model in the embodiments of the present invention. Detailed Implementation
[0046] The present invention will be further explained below with reference to the accompanying drawings and specific embodiments. The present invention provides a method for remote sensing inversion of soil salinity based on mixed feature optimization, the flowchart of which is shown below. Figure 1 As shown, it includes the following steps:
[0047] Step 1: Construction of multi-source remote sensing feature dataset and initial feature variable set.
[0048] Remote sensing images of the target area are acquired using visible light, multispectral, and thermal infrared sensors mounted on the UAV. After preprocessing the images, the original band values, spectral indices, and texture features are extracted from each data source to construct three initial feature variable sets.
[0049] Spectral indices include canopy coverage, color index, vegetation index, salinity index, canopy temperature, and canopy temperature-derived indices. Texture features are extracted using the gray-level co-occurrence matrix method, and their feature parameters include mean, variance, entropy, homogeneity, contrast, correlation, dissimilarity, and second moment.
[0050] This embodiment uses the Yangzhou University Eco-Hydrology Experimental Field in Yangzhou City, Jiangsu Province as the study area. Throughout the entire barley growth period, a DJI Inspire 2 drone equipped with a MicaSense Altum multispectral-thermal infrared camera and a Zenmuse X5S RGB camera was used for aerial flights under clear, light-wind conditions (11:00–14:00 local time) to acquire high-resolution remote sensing images of the study area. During the drone flight, the electrical conductivity of soil profiles at depths of 0-10 cm and 10-40 cm was measured using an EC450 conductivity meter, and soil samples were collected. After air-drying, grinding, and sieving, the total soluble salt content of the samples was determined. Finally, the conductivity values were converted into soil salinity using the empirical equation SSC = 0.0013EC + 0.0008 (g / kg). A total of spectral data and corresponding soil salinity values were collected from 180 sample points throughout the entire growth period.
[0051] Pix4Dmapper software was used to stitch together, radiometrically calibrate, and orthorectify the acquired high-resolution remote sensing images (including visible light, multispectral, and thermal infrared images) to generate reflectance images. The thermal infrared brightness values were inverted to the true temperature using Altum blackbody calibration data and GeoTIFF was generated. ArcGIS was used to complete geometric correction and surface temperature image production to obtain the preprocessed images.
[0052] From the preprocessed images, the R, G, and B band brightness values of the visible light image, the blue, green, red edge, and near-infrared band reflectance of the multispectral image, and the brightness temperature values of the thermal infrared image were extracted as the raw band values. Based on these bands, a total of 17 vegetation indices, 14 salinity indices, 29 color indices, and 2 canopy temperature-derived drought indices were calculated. Using the gray-level co-occurrence matrix tool of ENVI 5.3, eight texture features were calculated for each band of the visible light, multispectral, and thermal infrared images, including: mean, variance, homogeneity, contrast, dissimilarity, entropy, second moment of angle, and correlation. Finally, three initial feature variable sets were constructed: Feature Variable Set 1 (corresponding to visible light data, with a total of 57 features), Feature Variable Set 2 (corresponding to multispectral data, with a total of 76 features), and Feature Variable Set 3 (corresponding to thermal infrared data, with a total of 11 features).
[0053] Step 2: Hybrid feature optimization selection.
[0054] A recursive feature elimination method is adopted to perform feature screening for each initial feature variable set. The recursive feature elimination uses random forest as the base model, and eliminates the features with the lowest importance through iteration. The optimal feature subset under each data source is determined by minimizing the root mean square error of cross-validation.
[0055] This step is implemented using code written in Matlab R2024b. The feature variable sets from step 1 are combined with soil salinity sample data from 0~10cm and 10~40cm depths, respectively. Using the spectral parameters of the feature variable sets as feature variables and the measured soil salinity content as the target variable, six initial datasets are constructed, named sequentially by lowercase letters in this embodiment. Taking the initial dataset (a) combining feature variable set 1 with the 0~10cm soil salinity content as an example, the feature selection steps specifically include: using a random forest as the base model, setting the number of trees to 100, and leaving the other parameters as default values; employing a recursive feature elimination method, starting with all features, using 10-fold cross-validation, removing the feature with the lowest importance score at each step, and recording the root mean square error of the cross-validation for each feature subset. In this embodiment, during the screening of the initial dataset (a), the RMSE reached its minimum value of 0.895 when the feature subset contained 21 features. Therefore, these 21 features were considered as the optimal feature subset selected from the visible light data when the soil salinity content of 0-10cm was the target variable. Similarly, the above operation was performed on the initial datasets (b), (c), (d), (e), and (f), selecting 9, 16, 17, 5, and 5 optimal features, respectively. Figure 2 shows the result of feature screening in this embodiment.
[0056] To maintain consistency in the spectral characteristic parameters of soil layers at depths of 0–10 cm and 10–40 cm, only the characteristic variables selected in both soil layers are retained, and these shared features are used as the final selection result of the feature set. That is, the two optimal feature subsets from the same feature set are combined by selecting the common characteristic variables, which is then used as the final result obtained after feature optimization. For example, the common characteristic variables obtained from the two initial datasets (a) and (b) in Figure 2 are used as the final result of feature set 1; similarly, the common characteristic variables in (c) and (d) are used as the final result of feature set 2; and the common characteristic variables in (e) and (f) are used as the final result of feature set 3. The final number of features selected for feature sets 1, 2, and 3 are 9, 14, and 4, respectively, which are used as the optimal feature subsets obtained from the final selection.
[0057] Step 3: Generate multi-scheme datasets.
[0058] The optimal feature subset obtained in step 2 is combined according to the data source type to form multiple sets of modeling datasets, including single data source schemes, multi-data source combination schemes, and full data source fusion schemes.
[0059] The three optimal feature subsets obtained in step 2 are combined, and RGB represents visible light data, MS represents multispectral data, and TIR represents thermal infrared data, to generate the following seven spectral data feature combination schemes:
[0060] Option 1: RGB+MS+TIR (27 features);
[0061] Option 2: RGB+MS (23 features);
[0062] Option 3: RGB+TIR (13 features);
[0063] Option 4: MS+TIR (18 features);
[0064] Option 5: RGB (9 features);
[0065] Option 6: MS (14 features);
[0066] Option 7: TIR (4 features).
[0067] Step 4: Building and optimizing integrated machine learning models.
[0068] Based on each set of modeling datasets, multiple base models of different types are trained respectively; the prediction results of the above base models are fused using stacked ensemble method, weighted average method or arithmetic average method to construct ensemble learning model; a multi-index evaluation system including coefficient of determination, root mean square error and mean absolute error is used to validate all ensemble models, and the optimal soil salinity estimation model is selected based on the preset comprehensive scoring formula.
[0069] The stacked ensemble method uses the prediction results of the base machine learning model as input features and a linear regression model as the meta-model for secondary training. The weighted average method assigns weight values based on the performance of the base machine learning model; the better the model's performance, the larger the weight percentage.
[0070] Evaluation indicators include the coefficient of determination (R²). 2 The evaluation metrics include: root mean square error (RMSE), mean absolute error (MAE), performance-to-interquartile range (RPIQ), and performance-to-standard deviation (RPD). The comprehensive scoring formula integrates these multiple evaluation metrics to calculate a comprehensive score for each model, and selects the best model based on its score. The formula is as follows:
[0071]
[0072] In the formula, F i Let Ri represent the score of the i-th model, and let α, β, μ, δ, and θ represent Ri, ... 2 The weights of RMSE, MAE, RPIQ, and RPD, R 2 min The minimum value of the coefficient of determination, RMSE max The root mean square error (MAE) is the maximum value of the error. maxRPIQ is represented by the maximum value of the mean absolute error. min The minimum value of the ratio of performance to interquartile range, RPD min This represents the minimum value of the ratio of performance to standard deviation.
[0073] In this embodiment, the seven spectral feature combination schemes designed in step 2 are combined with the soil salinity content at 0~10 cm and 10~40 cm respectively to construct 14 modeling datasets, which are numbered sequentially. Modeling datasets 1~7 are obtained by combining the soil salinity content at 0~10 cm according to feature combination schemes 1~7, and modeling datasets 8~14 are obtained by combining the soil salinity content at 10~40 cm according to feature combination schemes 1~7.
[0074] Let each X represent a feature variable, and Y represent soil salinity. Each X has 180 values, corresponding to the values of that feature variable at 180 sample points collected during the crop's growth period. Y corresponds to the soil salinity value at each of those 180 sample points. Using a set {X1, X2, X3...X...} n} and Y construct a modeling dataset.
[0075] 180 soil salinity data points were collected for soil layers 0–10 cm and 10–40 cm deep. Each modeling dataset was randomly divided into a training set (126 samples) and a validation set (54 samples) in a 7:3 ratio. For each of the 14 modeling datasets, the following four base machine learning models were trained using their training set data: Gaussian Process Regression, Support Vector Machine, Backpropagation-MLP, and Extreme Learning Machine. Automatic hyperparameter optimization was performed using nested cross-validation with an outer 5-fold and an inner 10-fold fold, combined with Bayesian optimization, to obtain the predicted values for each modeling dataset using the four base models.
[0076] The predicted values of four base models corresponding to a modeling dataset are used as new input features and fused using three different fusion methods, specifically:
[0077] (1) Ridge regression is used as the meta-model for secondary training on the training set to obtain the stacked ensemble model;
[0078] (2) Weighted average ensemble assigns weight values based on the magnitude of the inverse RMSE of each base model on the training set;
[0079] (3) Arithmetic mean ensemble is the average prediction result of each base model.
[0080] The three fusion methods need to be calculated separately, and the fusion results obtained from each method are then compared. Using validation set data, the model's predictive performance on unseen numerical values is validated, and the model's generalization ability is evaluated. The R², RMSE, MAE, RPIQ, and RPD values for each ensemble model are calculated using the prediction data from the validation set, as shown in the following formulas:
[0081]
[0082]
[0083]
[0084]
[0085]
[0086] In the formula, y i Indicates the observed value, ŷ i The predicted value is represented by IQR (interquartile range of observations = Q3 - Q1), where Q3 and Q1 are the 75th and 25th percentiles, respectively. SD represents the standard deviation of y, and n represents the sample size. The index values of each modeling dataset after the above processing are as follows: Figure 3 , Figure 4 As shown.
[0087] Apply the preset comprehensive scoring formula F i For the final screening, the comprehensive scoring formula is set with α=0.4, β=0.3, μ=0.1, δ=0.1, and θ=0.1, as follows:
[0088]
[0089] The calculated scores are shown in Table 1. As can be seen from the figure, considering both soil depth and feature combination schemes, "Feature Combination Scheme 2" in this embodiment has the highest model score F constructed by the "Stacked Integration Method" and is selected as the optimal soil salinity estimation model. It should be further noted that the modeling datasets 7 and 14 in Table 1, which only use thermal infrared data, have the worst model performance due to insufficient remote sensing feature information, and their overall score is zero.
[0090] Table 1. Scores (F) for each model
[0091]
[0092] Step 5: Model Deployment and Application.
[0093] The optimal soil salinity estimation model is deployed on a computing device to quantitatively invert the soil salinity in the target area.
[0094] The optimal soil salinity estimation model is persistently stored to form an independently callable model file. When it is necessary to perform inversion on the target area, the model file is loaded on the computing device, and new remote sensing data after preprocessing and feature extraction is input. Forward inference is then performed to output a quantitative estimate of soil salinity. The estimated value can be further processed through a geographic information system or visualization library to generate a spatial distribution map of soil salinity for analysis.
[0095] Compared to existing technologies that rely on a single data source and manually set feature thresholds, this embodiment constructs multiple ensemble learning models by fusing visible light, multispectral, and thermal infrared data and employing a hybrid screening strategy that combines random forests and recursive feature elimination. Figure 5 The linear fitting relationship between predicted and measured salinity of the optimal stacked ensemble model in surface soil and rhizosphere soil is shown respectively. Figure 5 In the first half, the training and validation set accuracies of the soil salinity prediction model in the 0–10 cm depth were R²=0.76 and RMSE=0.69%, respectively. Figure 5 In the second half, the accuracy of the soil salinity prediction model for the 10–40 cm soil layer on the training and test sets were R²=0.78 and RMSE=0.77%, respectively, and R²=0.77 and RMSE=0.99%. The consistency between the training and validation set results indicates that the model has good stability and generalization ability. This model significantly outperforms any single data source dataset, effectively demonstrating the technical advantages of this invention in improving estimation accuracy and model generalization ability.
Claims
1. A soil salinity remote sensing inversion method based on hybrid feature optimization, characterized in that, The application relates to a method for estimating soil salt content, comprising the following steps: acquiring remote sensing images of different data sources in a target area, collecting soil samples of different depths and measuring the salt content of the soil samples; preprocessing the remote sensing images, extracting the band original values of the remote sensing images, and constructing a feature variable set for the remote sensing images of each data source; the data sources include visible light images, multispectral images and thermal infrared images; taking the salt content of the soil samples of different depths as a target variable, using a recursive feature elimination method to screen an optimal feature subset of each feature variable set; randomly arranging and combining all the optimal feature subsets to obtain a plurality of feature combination schemes, and each feature combination scheme includes at least one optimal feature subset; combining the optimal feature subsets according to the feature combination schemes and taking the salt content of the soil samples of different depths as an index to obtain a plurality of modeling data sets; for each modeling data set, an integrated learning model is constructed through fusion training; a multi-index evaluation system is established, and an optimal soil salt content estimation model is screened from all the integrated learning models; the optimal soil salt content estimation model is deployed to a device to quantitatively invert the soil salt content of the target area.
2. The method according to claim 1, wherein, The remote sensing images of different data sources include visible light images, multispectral images and thermal infrared images; the preprocessing includes splicing, radiation calibration and orthographic correction.
3. The method according to claim 2, wherein, The band original values include R, G and B band brightness values of the visible light images, blue, green, red edge and near-infrared band reflectivity of the multispectral images and brightness temperature values of the thermal infrared images.
4. The method according to claim 3, wherein, The initial feature variable set includes: spectrum indexes including canopy coverage, color index, vegetation index, salt index, canopy temperature value and canopy temperature derived index; texture features extracted by a gray level co-occurrence matrix method, and feature parameters including mean value, variance, entropy, homogeneity, contrast, correlation, difference and second moment.
5. The method according to claim 4, wherein, The recursive feature elimination method comprises the following steps: any one feature variable set and a group of soil sample salt contents constitute an initial data set, a random forest is used as a base model, all features in the current initial data set are taken as a starting point, ten-fold cross validation is used, one feature with the lowest importance score is removed at each step, and the iteration is stopped when the root mean square error of cross validation is the smallest; all the initial data sets with completed iteration are combined according to data sources, common features in soil samples of different depths are retained, and an optimal feature subset corresponding to each data source is obtained.
6. The method of claim 5, wherein, For each modeling data set, an integrated learning model is constructed through fusion training, comprising the following steps: the modeling data set is used to train a plurality of preset base models, and the base models include Gaussian process regression, support vector machine, BP-MLP and extreme learning machine; outer 5-fold and inner 10-fold nested cross validation are combined with Bayesian optimization to automatically optimize hyperparameters, the prediction values of the modeling data set in the base models are obtained, the prediction values are taken as new input values for fusion training, and a corresponding integrated learning model is obtained.
7. The method of claim 6, wherein, The fusion training comprises the following methods, and each method generates an integrated learning model: (1) using ridge regression as a meta-model to perform secondary training on the training set to obtain a stacked ensemble model; (2) According to the weight value of the RMSE inverse of each base model on the training set, a weighted average integrated model is obtained; (3) The prediction values of each base model are averaged to obtain an arithmetic average integrated model.
8. The method of claim 7, wherein, The multi-index evaluation system comprises: The evaluation indexes of each integrated learning model are calculated, including: determination coefficient, root mean square error, mean absolute error, ratio of performance to quartile range, and ratio of performance to standard deviation; According to the comprehensive score formula, the comprehensive score of each integrated learning model is obtained, and the integrated learning model with the highest score is the optimal soil salt estimation model; the comprehensive score formula is represented as: In the formula, F i is a score value of the i-th model, and α, β, μ, δ, θ are weight values, R 2 min is a minimum value of a determination coefficient, RMSE max is a maximum value of a root mean square error, MAE max is a maximum value of a mean absolute error, RPIQ min is a minimum value of a ratio of performance to quartile distance, RPD min is a minimum value of a ratio of performance to standard deviation.
9. The method of claim 8, wherein, The optimal soil salt estimation model is deployed to the equipment, comprising: The optimal soil salt estimation model is persistently stored to form a model file that can be independently called; when it is necessary to perform inversion on a target area, the model file is loaded on the equipment, the new remote sensing data after feature extraction is input, forward reasoning is performed, and the quantitative estimation value of soil salt is output. 10.A soil salinity remote sensing inversion system based on hybrid feature optimization, comprising a memory, a processor and a computer program stored on the memory, characterized in that, The processor executes the computer program to realize the steps of any method in claims 1-9.
