A soil salinity inversion method, system and device based on multi-modal remote sensing data fusion and a storage medium
By fusing multimodal remote sensing data, utilizing Sentinel-1 and Sentinel-2 data, and combining random forest algorithm and recursive feature elimination method, the feature combination was optimized, solving the problems of accuracy and efficiency in soil salinity inversion, and realizing accurate soil salinity monitoring in complex environments.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Patents(China)
- Current Assignee / Owner
- Filing Date
- 2025-04-22
- Publication Date
- 2026-03-20
AI Technical Summary
Existing technologies suffer from low efficiency and poor accuracy when monitoring large-scale soil salinization, especially under complex environmental conditions and human influences, making it difficult to achieve accurate soil salinity inversion.
A multimodal remote sensing data fusion method was adopted, combining radar remote sensing data and optical remote sensing data. The feature combination was optimized by random forest algorithm and recursive feature elimination method. Soil electrical conductivity was predicted using Sentinel-1 and Sentinel-2 data, and a soil salinity inversion model was constructed.
It improves the accuracy of soil salinity inversion under complex environments and human influences, and provides a more efficient and reliable method for soil salinity monitoring, which is applicable to areas such as the Hetao Irrigation District.
Smart Images
Figure CN120427870B_ABST
Abstract
Description
TECHNICAL FIELD
[0001] The present application relates to the technical field of soil quality monitoring, and particularly relates to a soil salinity inversion method, system and device based on multi-modal remote sensing data fusion and a storage medium. BACKGROUND
[0002] Soil salinization is a serious environmental challenge, particularly in arid and semi-arid regions, which severely impacts agricultural productivity, ecosystem health, and water resource management. Salinization occurs when excessive salt accumulates in the soil, impairing plant growth and soil fertility, and this process is often exacerbated by improper irrigation practices, poor drainage, and high evaporation rates, among other natural factors. As agriculture remains the backbone of many economies, monitoring and mitigating soil salinization is crucial to ensuring long-term food security and sustainable land use.
[0003] Traditional methods of soil salinization monitoring, such as ground sampling and laboratory analysis, while accurate, are labor-intensive, time-consuming, and limited in spatial coverage. Therefore, there is an urgent need for more efficient, scalable, and accurate methods to monitor soil salinization over large areas. Remote sensing technology has emerged as a promising tool for soil salinization monitoring, offering a non-invasive, cost-effective, and scalable approach to assess soil conditions over large geographical areas. In particular, satellite-based remote sensing provides valuable data that can be used to detect and monitor changes in soil and vegetation cover caused by salinization. Various studies have utilized optical and radar satellite data to monitor soil salinity, with varying degrees of success depending on the type of data used, spatial resolution, and analysis techniques. Optical data is often used to derive vegetation indices, such as the Normalized Difference Vegetation Index (NDVI) and Salinity Index (SI), which indirectly indicate soil salt through detected vegetation stress. Radar data can provide information on soil moisture and surface roughness, which are also related to salinity.
[0004] However, despite these advances, accurately inverting soil salinity using remote sensing data remains challenging. The combination of multiple remote sensing data sources, such as optical and radar imagery, has the potential to improve the accuracy of salinity monitoring by leveraging the strengths of each data type. For example, optical data can capture spectral reflectance related to vegetation stress and soil properties, while radar data can penetrate clouds and provide consistent measurements of soil surface conditions independent of weather conditions.
[0005] Therefore, it is an urgent problem for those skilled in the art to provide a soil salinity inversion method, system and device based on multi-modal remote sensing data fusion, which can realize more accurate inversion of soil salinity under the interaction of complex environments and human factors. SUMMARY
[0006] In view of this, the present application provides a soil salinity inversion method, system, device and storage medium based on multi-modal remote sensing data fusion.
[0007] In order to achieve the above purpose, the present application adopts the following technical solutions:
[0008] A soil salinity inversion method based on multi-modal remote sensing data fusion comprises:
[0009] Step 1: Obtain radar remote sensing data and optical remote sensing data of the soil salinity area to be inverted, and determine the soil conductivity as an inversion index of soil salinity.
[0010] Step 2: Obtain multi-modal remote sensing features based on radar remote sensing data and optical remote sensing data fusion, and combine radar remote sensing data and optical remote sensing data, based on the random forest algorithm, use recursive feature elimination method to optimize the remote sensing feature combination, use cross-validation to evaluate the model performance of the optimized remote sensing feature combination, and obtain the optimal remote sensing feature combination.
[0011] Step 3: Based on the optimal remote sensing feature combination, construct and train a soil conductivity prediction model based on the random forest algorithm, input the data to be measured into the soil conductivity prediction model after adjusting the hyperparameters, obtain the soil conductivity, and complete the soil salinity inversion.
[0012] Optionally, in step 1, the radar remote sensing data comprises VV band and VH band.
[0013] Optionally, in step 1, the optical remote sensing data comprises blue band B2, green band B3, red band B4, first red edge band B5, second red edge band B6, third red edge band B7, near-infrared band B8, fourth red edge band B8A, first short-wave infrared band B11 and second short-wave infrared band B12.
[0014] Optionally, in step 2, the multi-modal remote sensing features based on radar remote sensing data and optical remote sensing data fusion include: normalized difference vegetation index NDVI, enhanced vegetation index EVI, soil-adjusted vegetation index SAVI, modified soil-adjusted vegetation index MSAVI, slope-adjusted vegetation index TSAVI, adjusted normalized difference vegetation index OSAVI, adaptive soil-adjusted vegetation index ATSAVI, vegetation index PVI, near-infrared vegetation index NIRv, chlorophyll index MTCI, moisture-sensitive vegetation index SSAI, chroma index Clr, vegetation health index CIg, first normalized difference index NDRE1, second normalized difference index NDRE2, first normalized difference vegetation index NDVIr1, second normalized difference vegetation index NDVIr2, third normalized difference vegetation index NDVIr3, radar vegetation index VV / VH Ratio, radar vegetation index RVI, dual-polarization SAR vegetation index DPSVI, normalized difference polarization index NDPI, a spectral transformation method Brightness, greenness index Greenness, and wetness index Wetness.
[0015] Optionally, in step 2, based on the random forest algorithm, the recursive feature elimination method is used to optimize the combination of remote sensing features, specifically:
[0016] In each iteration, the importance score of the feature is calculated by using the index, and the feature with the smallest contribution to the model is removed;
[0017] The index is an index that measures impurity. For a node , The index is as follows:
[0018]
[0019] wherein, is the total number of categories; is the proportion of samples belonging to category to the total number of samples in the node;
[0020] The smaller the index, the purer the samples in the node, representing that most samples belong to the same category;
[0021] When a certain feature is used to split the node, the data is divided into left and right child nodes, The gain measures the reduction in the index after splitting, as follows:
[0022]
[0023] wherein, is the index of the pre-split node ; and are the indices of the left and right child nodes after splitting, respectively ; and are the number of samples in the left and right child nodes after splitting, respectively is the total number of samples in the current node
[0024] The greater the gain, the greater the contribution of the feature to classification
[0025] In a random forest, the gain of each split node of each tree is calculated based on a certain feature The importance score of a feature is the cumulative average of the gain of all trees in the forest, as follows:
[0026]
[0027] where is the total number of trees in the random forest is the gain of feature in a split in the th tree ;
[0028] The random forest evaluates the importance of each feature by the above formula and performs recursive feature elimination.
[0029] Optionally, in step 2, cross-validation is used to evaluate the model performance of the optimized remote sensing feature combination, specifically:
[0030] Using 5-fold cross-validation, the sample data is randomly divided into 5 mutually exclusive subsets, and each time 4 of them are used as the training set and the remaining 1 is used as the test set, for a total of 5 iterations
[0031] In each iteration, the random forest model is used to train the training set, and the trained model is used to predict the test set to obtain the predicted value and the true value, and the average determination coefficient and the root mean square error of the 5 iterations are used to evaluate the model performance of different remote sensing feature combinations.
[0032] Optionally, in step 3, hyperparameter adjustment, including: the number of trees in the random forest ntrees, the maximum number of nodes per tree maxnodes, the minimum number of samples in the terminal node nodesize, and the number of samples extracted from the training set for each tree sampsize.
[0033] The application also provides a soil salinity inversion system based on multi-modal remote sensing data fusion, which utilizes a soil salinity inversion method based on multi-modal remote sensing data fusion, and comprises the following modules.
[0034] A data acquisition module is configured to acquire radar remote sensing data and optical remote sensing data of a soil salinity inversion area, and to determine the conductivity of the soil as an inversion index of the soil salinity.
[0035] A feature screening module is configured to acquire multi-modal remote sensing features based on fusion of the radar remote sensing data and the optical remote sensing data, to combine the radar remote sensing data and the optical remote sensing data, to optimize a remote sensing feature combination by using a recursive feature elimination method based on a random forest algorithm, to perform model performance evaluation on the optimized remote sensing feature combination by using cross-validation, and to obtain an optimal remote sensing feature combination.
[0036] A salinity inversion module is configured to construct and train a soil conductivity prediction model based on a random forest algorithm based on the optimal remote sensing feature combination, to input test data into the soil conductivity prediction model after hyperparameter adjustment, to obtain soil conductivity, and to complete soil salinity inversion.
[0037] The application also provides an electronic device, which comprises:
[0038] A memory is configured to store a computer program.
[0039] A processor is configured to implement steps of a soil salinity inversion method based on multi-modal remote sensing data fusion when the computer program is executed.
[0040] The application also provides a computer readable storage medium, which stores a computer program, and the computer program is configured to implement steps of a soil salinity inversion method based on multi-modal remote sensing data fusion when the computer program is executed by a processor.
[0041] Compared with the prior art, the application provides a soil salinity inversion method, system, device and storage medium based on multi-modal remote sensing data fusion. The application extracts multi-modal remote sensing features based on radar remote sensing data and optical remote sensing data fusion by integrating two complementary Sentinel-1 radar data (providing insights into soil moisture and surface characteristics) and Sentinel-2 optical data (capturing vegetation and soil spectral properties related to salinity), and performs feature combination optimization and performance evaluation through recursive feature elimination and cross-validation method, and obtains the optimal feature combination (Greenness, Wetness, VV, B11 and NDVI) suitable for the most typical Hetao Irrigation District. Finally, the model adjusted by the hyperparameters is applied to the soil salinity inversion of the Hetao Irrigation District in 2021, proving the accuracy of the application. In summary, the application realizes more accurate inversion of soil salinity under the interaction of complex environment and human factors. BRIEF DESCRIPTION OF DRAWINGS
[0042] In order to more clearly illustrate the technical solutions in the embodiments of the application or the prior art, the following will briefly introduce the drawings needed to be used in the embodiments or the prior art description. Obviously, the drawings in the following description are only embodiments of the application, and for those skilled in the art, other drawings can be obtained without creative labor on the basis of the provided drawings.
[0043] Figure 1 The method flowchart of the application.
[0044] Figure 2 The soil sample point distribution diagram of the Hetao Irrigation District in 2021 of the application.
[0045] Figure 3 The relationship between the variable number and the cross-validation RMSE based on the random forest regression model of the application.
[0046] Figure 4 The feature variable importance ranking diagram of the application.
[0047] Figure 5 The relationship between the EC actual value and the predicted value in 2021 of the application.
[0048] Figure 6 The Hetao Irrigation District farmland salinization degree classification diagram in 2021 of the application. DETAILED DESCRIPTION
[0049] With reference to the accompanying drawings, the technical solutions in the embodiments of the present application will be described clearly and completely. Obviously, the described embodiments are only a part of the embodiments of the present application, rather than all the embodiments of the present application. Based on the embodiments in the present application, all the other embodiments obtained by a person of ordinary skill in the art without creative effort are within the protection scope of the present application.
[0050] Embodiment 1
[0051] Embodiment 1 of the present application discloses a soil salinity inversion method based on multi-modal remote sensing data fusion, as shown in Figure 1 , which comprises:
[0052] Step 1: Obtain radar remote sensing data and optical remote sensing data of the soil salinity area to be inverted, and determine the soil conductivity as an inversion index of soil salinity.
[0053] The radar remote sensing data includes VV band and VH band.
[0054] The radar echo signal of VV band (vertical-vertical polarization) is sensitive to soil humidity and surface roughness, and can provide soil moisture information of salinization area. For salinization area, VV band can effectively reveal the wetness degree of saline soil, and then reflect the degree of salinization.
[0055] VH band (vertical-horizontal polarization) helps to identify the special surface features of salinized soil by reflecting the roughness and structure of soil surface. In the area with serious salinization, the soil surface usually becomes hard and rough, and VH band can capture these changes, thereby improving the identification accuracy of salinized soil.
[0056] The optical remote sensing data includes blue band B2 (Blue), green band B3 (Green), red band B4 (Red), first red edge band B5 (Red Edge1), second red edge band B6 (Red Edge2), third red edge band B7 (Red Edge3), near-infrared band B8 (Near-infrared (NIR), fourth red edge band B8A (Red Edge4), first short-wave infrared band B11 (SWIR1), and second short-wave infrared band B12 (SWIR2).
[0057] Step 2: Obtain multi-modal remote sensing features based on the fusion of radar remote sensing data and optical remote sensing data, and combine the radar remote sensing data and the optical remote sensing data. Based on the random forest algorithm, use the Recursive Feature Elimination method to optimize the combination of remote sensing features. Use Cross-Validation to evaluate the performance of the optimized remote sensing feature combination model, and obtain the optimal remote sensing feature combination.
[0058] In the present application, the multi-modal remote sensing features based on the fusion of radar remote sensing data and optical remote sensing data are selected based on their ability to effectively capture key information related to soil salinization, such as vegetation conditions, surface water distribution, brightness characteristics, and humidity. The specific selection criteria are as follows:
[0059] (1) Vegetation index: Quantifies vegetation growth conditions, coverage, and health, indirectly reflecting the impact of soil salinization on vegetation. Salinized soil usually inhibits vegetation growth, resulting in a significant decrease in vegetation index, thus providing a reliable indirect indicator for soil salinization inversion.
[0060] (2) Water index: Soil salinization significantly affects the water retention capacity of soil, and these indices can well depict the surface and soil water characteristics, providing key support for the extraction of water-related variables in inversion, thereby improving the inversion accuracy.
[0061] (3) Spectral index: By selecting sensitive spectral bands, it can capture the significant changes in spectral characteristics of salinized soil and vegetation, especially the spectral reflectance changes caused by salinization, providing rich spectral information for quantitative analysis of soil salinization degree.
[0062] (4) Radar features: Radar images have high sensitivity to water and surface structure, especially in arid and saline areas, which can supplement the soil moisture and surface roughness information that optical images cannot obtain, enhancing the monitoring capability of saline soil.
[0063] (5) Tower transformation features: Extract brightness, greenness, and wetness features based on the tower transformation method of multispectral images, which comprehensively reflect the surface features. These features effectively distinguish the spectral differences between saline soil and other soil types, improving the reliability and applicability of the inversion.
[0064] In summary, these features cover multi-dimensional information of optical and radar images, which helps to comprehensively evaluate the degree of soil salinization and provide accurate inversion results, thereby providing scientific basis for regional soil salinization management and improvement strategy formulation.
[0065] Therefore, based on the multi-modal remote sensing features fused by radar remote sensing data and optical remote sensing data, as shown in Table 1, include: normalized vegetation index NDVI, enhanced vegetation index EVI, soil-adjusted vegetation index SAVI, modified soil-adjusted vegetation index MSAVI, slope-adjusted vegetation index TSAVI, adjusted normalized difference vegetation index OSAVI, adaptive soil-adjusted vegetation index ATSAVI, vegetation index PVI, near-infrared vegetation index NIRv, chlorophyll index MTCI, moisture-sensitive vegetation index SSAI, chroma index Clr, vegetation health index CIg, first normalized difference index NDRE1, second normalized difference index NDRE2, first normalized difference vegetation index NDVIr1, second normalized difference vegetation index NDVIr2, third normalized difference vegetation index NDVIr3, radar vegetation index VV / VH Ratio, radar vegetation index RVI, dual-polarization SAR vegetation index DPSVI, normalized difference polarization index NDPI, a spectral transformation method Brightness, greenness index Greenness, and wetness index Wetness.
[0066] Table 1 Multi-modal remote sensing features fused by radar remote sensing data and optical remote sensing data
[0067]
[0068]
[0069]
[0070] Based on the random forest algorithm, the recursive feature elimination method is used to optimize the combination of remote sensing features, specifically:
[0071] In each iteration, the importance score of each feature is calculated by using exponential function, and the features with the smallest contribution to the model are removed; for example, the importance scores of the slope direction and soil classification features are 0.01 and 0.03 respectively in the 5th iteration, and they are considered to have limited explanatory power for the target variable, and are finally removed.
[0072] The index is a measure of impurity, and for a node , The index is as follows:
[0073]
[0074] wherein, is the total number of classes; is the proportion of samples belonging to class to the total number of samples in the node;
[0075] The smaller the index, the purer the samples in the node, representing that most samples belong to the same class;
[0076] When a feature is used to split a node, the data is divided into left and right child nodes, Gain measures the improvement of the split The reduction of the index, as follows:
[0077]
[0078] where, is the index of the node before splitting; and are the indexes of the left and right child nodes after splitting, respectively; and are the sample sizes of the left and right child nodes after splitting, respectively; is the total sample size of the current node;
[0079] The greater the gain, the greater the contribution of the feature to classification;
[0080] In a random forest, the gain of each split node of each tree is calculated based on a certain feature The importance score of a feature is the cumulative average of the gain of all trees in the forest, as follows:
[0081]
[0082] where, is the total number of trees in the random forest; is the gain of the th tree in a split of the th feature;
[0083] Random forest evaluates the importance of each feature through the above formula and performs recursive feature elimination.
[0084] Cross-validation is used to evaluate the performance of the optimized remote sensing feature combination, specifically:
[0085] 5-fold cross-validation is used, and the sample data is randomly divided into 5 mutually exclusive subsets. Each time, 4 of them are used as the training set, and the remaining 1 is used as the test set, for a total of 5 iterations.
[0086] The training set is trained using a random forest model in each iteration, the prediction value and the true value are obtained by predicting the test set using the trained model, and the average determination coefficient and root mean square error of 5 iterations are used to evaluate the model performance of different remote sensing feature combinations.
[0087] Step 3: Based on the optimal remote sensing feature combination, a soil conductivity prediction model based on the random forest algorithm is constructed and trained, the measured data is input into the soil conductivity prediction model after the hyperparameter adjustment, the soil conductivity is obtained, and the soil salinity inversion is completed.
[0088] Hyperparameter adjustment includes: the number of trees ntrees in the random forest, the maximum number of nodes maxnodes of each tree, the minimum number of samples nodesize of the terminal node, and the number of samples sampsize extracted from the training set for each tree.
[0089] Embodiment 2
[0090] Embodiment 2 of the present application discloses a specific application of a soil salinity inversion method based on multi-modal remote sensing data fusion, as follows:
[0091] The radar remote sensing data and the optical remote sensing data are taken from Sentinel-1 and Sentinel-2 data respectively.
[0092] The synthetic aperture radar (SAR) data of Sentinel-1 provides important insights into the surface dynamics of the Hetao Irrigation District, especially the direct impact of soil moisture and roughness on salinization. The C-band dual-polarization capability (VV and VH) of Sentinel-1 has unparalleled advantages in tracking changes in soil texture and moisture, even in the absence of visible light or cloudy weather conditions. Unlike optical data, Sentinel-1 can penetrate vegetation and capture surface features that are crucial for salinization assessment, making it suitable for salinization monitoring in both cultivated and uncultivated areas. The 12-day temporal resolution ensures frequent updates, ensuring that seasonal changes in soil moisture and salt migration can be monitored on a detailed temporal scale. This high revisit frequency allows radar data to provide a consistent record of soil conditions over time, almost like a “pulse”. When combined with environmental variables such as irrigation practices and precipitation, the data archive of Sentinel-1 from 2017 to 2023 provides users with a time series perspective, helping them understand how human activities and natural factors influence the salinization patterns in the region. The 10-meter spatial resolution can capture detailed heterogeneity, especially in transition zones where salt levels increase or decrease due to changes in land management or natural drainage.
[0093] As a complement to the Sentinel-1 radar capabilities, Sentinel-2 provides detailed spectral analysis of the surface of the Hetao Irrigation District through its 13 multispectral bands. Sentinel-2 is able to identify small changes in vegetation health and soil reflectance, which is crucial to understand the manifestation of salinization in the landscape. Sentinel-2 data are particularly valuable in areas where salinization impacts crop growth, as its spectral bands, including visible, near-infrared (NIR), and shortwave infrared (SWIR), are able to capture vegetation stress, water changes, and the formation of saline crusts. By taking advantage of the 5-day revisit cycle of Sentinel-2, the present invention is able to ensure temporal trends of vegetation and bare soil, accurately linking remote sensing indices (such as NDVI, salinity index (SI)) to in-situ measured electrical conductivity (EC). This frequent monitoring ensures that rapid changes in crop health during critical growing seasons can be observed, identifying areas where salinization can impede agricultural productivity.
[0094] By combining the radar data of Sentinel-1 and the spectral capabilities of Sentinel-2, the present invention is able to not only monitor the static state of salinization but also track its evolution as a dynamic process. From this perspective, it is possible to track the evolution of soil salinization over time, detect early signs of degradation, and provide actionable insights for land management in the Hetao Irrigation District.
[0095] In the present invention, Sentinel-1 / 2 synergic data in 2021 were used to retrieve the salinized soil of cultivated land. Among them, the spectral data obtained from Sentinel-2 are shown in Table 2.
[0096] Table 2 Optical remote sensing data in Sentinel-2
[0097]
[0098] In the present invention, field sampling operations were carried out on the soil of the Hetao Irrigation District in 2021. The specific sampling methods and quantities are as follows:
[0099] Sampling point layout: The sampling range selected land plots greater than 100 mu in the study area to ensure the spatial representativeness of the plots and cover the main types of soil salinization in the area. Within each plot, sampling points were laid out with a sampling interval greater than 10 meters to avoid spatial autocorrelation between sampling points and ensure the independence of the sampling points. Soil samples were taken from the 0-10 cm surface layer at each sampling point. To reduce sampling errors, a mixed sampling method was used. Three positions were selected around each sampling point (usually in an equilateral triangle or ring distribution), and soil samples were collected at each position. The soil samples from the three positions were mixed evenly to form a mixed sample.
[0100] Sampling sample processing: the weight of each mixed sample is about 1 kg, to ensure sufficient sample quantity for subsequent determination; after the soil sample is collected, it is packed into a clean plastic bag, and the sampling number, coordinates (latitude and longitude), plot number, sampling time, etc. of each sample are recorded in detail.
[0101] Laboratory determination: after invalid soil samples are removed, a total of 487 soil samples are determined, as shown in Figure 2 The determination method is as follows: a fixed proportion (such as soil-water ratio 1:5) of each soil sample is mixed with deionized water, and the supernatant is extracted after standing. The conductivity of the soil supernatant is measured using professional conductivity instruments. Only EC is measured as an important indicator for analyzing soil salinity. After the determination results are returned, the subsequent analysis and model verification are combined with the location information of the sampling points.
[0102] Data preprocessing: cloud detection and mask: use Sentinel-2 quality assurance band (QA60) to remove clouds and cirrus; multi-temporal image synthesis: average value synthesis of images within the selected time range to reduce data noise; radar data processing: apply a 3x3 Boxcar filter to radar data to smooth noise.
[0103] After obtaining the multi-modal remote sensing features based on radar remote sensing data and optical remote sensing data fusion in Hetao Irrigation District, combining radar remote sensing data and optical remote sensing data, based on the random forest algorithm, using recursive feature elimination method to optimize the combination of remote sensing features, using cross-validation to evaluate the model performance of the optimized remote sensing feature combination, the optimal remote sensing feature combination is obtained, which is:
[0104] Recursive elimination and cross-validation of features by REF algorithm, as shown in Figure 3 It shows the changes of the main precision indicators RMSE under different feature number combinations. As can be seen, in the case of setting 5, 10, 15, 20 and all feature combinations, RMSE is only 0.938 when 5 feature combinations, reaching the highest prediction level, that is, the REF algorithm extracts the best feature combination for subsequent modeling based on the random forest algorithm. The optimal feature combination after recursive verification is Greenness, Wetness, VV, B11 and NDVI, as shown in Figure 4 .
[0105] Based on the optimal feature set (Greenness, Wetness, VV, B11 and NDVI), regression prediction was made on the validation dataset (20% of samples) based on the random forest algorithm. Figure 4 The results of feature variable importance analysis are shown, which is usually used to evaluate the relative importance of each input variable in the random forest model. IncMSE represents the relative impact of each variable on the model prediction error, usually measured by removing the variable and observing the change in mean square error (MSE). The horizontal axis is IncMSE, i.e. the percentage increase in model prediction error after removing the variable. The larger the value, the greater the impact of the variable on the prediction effect of the model, i.e. the higher its importance. From the figure, we can see that Wetness has the greatest impact on the model prediction error and is the most important variable. VV is second. NDVI and Greenness have similar importance. B11 has the smallest impact on the model prediction error. IncNodePurity represents the degree of improvement of each variable on node purity, which is usually used to measure the importance of variables in the decision tree splitting process. The higher the node purity, the more effectively the variable can distinguish data. The horizontal axis is IncNodePurity, and the larger the value, the greater the contribution of the variable to the model construction. From the figure, we can see that Greenness has the greatest improvement on node purity and is the most important variable. VV and Wetness are second. NDVI and B11 have relatively low importance.
[0106] Overall, Wetness and Greenness are the two most important variables in the model, dominating in error impact IncMSE and node purity improvement IncNodePurity, respectively. VV radar vertical polarization intensity) is also an important variable, ranking second. NDVI and B11 have lower importance but still contribute to model construction. In addition, the validation results of random forest regression prediction on 20% of samples are shown in Figure 5 .
[0107] Through hyperparameter adjustment, the optimal random forest regression model configuration is finally obtained, as shown in Table 3.
[0108] Table 3 Optimal random forest regression model hyperparameters
[0109]
[0110] Based on the random forest regression model under the hyperparameter adjustment in Table 3, the degree of soil salinization of cultivated land in Hetao Irrigation District in 2021 was predicted. Specifically as follows:
[0111] In the present application, by comparing and analyzing different wave bands and features, the optimal combination containing Greenness, Wetness, VV, B11, NDVI and other features is finally selected. This combination takes into account the spectral response characteristics of crop growth and salinization and the sensitivity of vegetation index, especially in the complex background of salinization dynamics in arid and semiarid areas, which helps to improve the precision prediction ability of the model for soil salinization.
[0112] In the model construction stage (2021), the model was strictly trained and verified using training samples, and finally achieved an excellent performance of an average R2 of 0.985 and an average RMSE of 0.39 dS / m. This indicates that the model can well fit the relationship between soil salinization degree and feature variables, and has high prediction accuracy and reliability. The inversion results of the soil salinization degree of the Hetao Irrigation District in 2021 are shown in FIG. 1. Figure 6
[0113] Embodiment 3
[0114] Embodiment 3 of the present application discloses a soil salinity inversion system based on multi-modal remote sensing data fusion using a soil salinity inversion method based on multi-modal remote sensing data fusion, comprising:
[0115] The data acquisition module is used to acquire radar remote sensing data and optical remote sensing data of the soil salinity area to be inverted, and to determine the conductivity of the soil as an inversion index of soil salinity.
[0116] The feature selection module is used to acquire multi-modal remote sensing features based on the fusion of radar remote sensing data and optical remote sensing data, and to combine the radar remote sensing data and the optical remote sensing data, based on the random forest algorithm, to optimize the remote sensing feature combination using the recursive feature elimination method, to evaluate the model performance of the optimized remote sensing feature combination using cross-validation, and to obtain the optimal remote sensing feature combination.
[0117] The salinity inversion module is used to construct and train a soil conductivity prediction model based on the random forest algorithm based on the optimal remote sensing feature combination, to input the data to be measured into the soil conductivity prediction model with adjusted hyperparameters, to obtain the soil conductivity, and to complete the soil salinity inversion.
[0118] Embodiment 4
[0119] Embodiment 4 of the present application discloses an electronic device, comprising:
[0120] The memory is used to store a computer program.
[0121] The processor is used to implement the steps of a soil salinity inversion method based on multi-modal remote sensing data fusion when executing the computer program.
[0122] Embodiment 5:
[0123] Embodiment 5 of the present application discloses a computer readable storage medium, and the computer readable storage medium stores a computer program.
[0124] The present application discloses a soil salinity inversion method and system based on multi-modal remote sensing data fusion, and a device and a storage medium. The present application extracts multi-modal remote sensing features based on radar remote sensing data and optical remote sensing data fusion by integrating two complementary Sentinel-1 radar data (providing insights into soil moisture and surface features) and Sentinel-2 optical data (capturing vegetation and soil spectral properties related to salinity), and performs feature combination optimization and performance evaluation through recursive feature elimination and cross-validation, obtaining the optimal feature combination (Greenness, Wetness, VV, B11 and NDVI) suitable for the most typical Hetao Irrigation District, and finally applying the model adjusted by hyperparameters to the soil salinity inversion of the Hetao Irrigation District in 2021, proving the accuracy of the present application.
[0125] The embodiments in the specification are described in a progressive manner, and each embodiment focuses on the differences from other embodiments, and the same or similar parts between the embodiments can be referred to each other.
[0126] The above description of the disclosed embodiments enables a person skilled in the art to implement or use the present application. Various modifications to the embodiments will be apparent to those skilled in the art, and the general principles defined herein can be implemented in other embodiments without departing from the spirit or scope of the present application. Therefore, the present application will not be limited to the embodiments shown herein, but will conform to the widest scope consistent with the principles and novel features disclosed herein.
Claims
1. A method for soil salinity inversion based on multimodal remote sensing data fusion, characterized in that, include: Step 1: Obtain radar and optical remote sensing data of the area where soil salinity needs to be inverted, and measure the electrical conductivity of the soil as an indicator for soil salinity inversion. Step 2: Obtain multimodal remote sensing features based on the fusion of radar and optical remote sensing data. Combine the radar and optical remote sensing data, optimize the remote sensing feature combination using the random forest algorithm and the recursive feature elimination method, and evaluate the model performance of the optimized remote sensing feature combination using cross-validation to obtain the optimal remote sensing feature combination. Step 3: Based on the optimal combination of remote sensing features, construct and train a soil electrical conductivity prediction model based on the random forest algorithm, input the data to be measured into the soil electrical conductivity prediction model after hyperparameter adjustment, obtain the soil electrical conductivity, and complete the soil salinity inversion. In step 2, based on the random forest algorithm, the recursive feature elimination method is used to optimize the combination of remote sensing features, specifically as follows: In each iteration, by utilizing The importance score of features is calculated using an index, and features that contribute the least to the model are removed. The index is a measure of impurity for a given node. , The index is as follows: in, Total number of categories; To belong to category The proportion of samples in the total number of nodes; The smaller the index, the purer the samples in the node, indicating that most samples belong to the same category; When a certain feature is used to split a node, the data is divided into two child nodes, left and right. Gain measures the result after splitting The decrease in the index is as follows: in, For the node before splitting index; and These are the left and right child nodes after the split. index; and These represent the number of samples in the left and right child nodes after the split; This represents the total number of samples in the current node. The greater the gain, the greater the contribution of the feature to the classification. In a random forest, each split node of each tree is calculated based on a certain feature. Gain, the importance score of features is calculated for all trees in the forest. The result of accumulating and averaging the gains is as follows: in, The total number of trees in the random forest; For the first Features in a granular tree In a split Gain; Random forests evaluate the importance of each feature using the formula above and perform recursive feature elimination.
2. The soil salinity inversion method based on multimodal remote sensing data fusion according to claim 1, characterized in that, In step 1, the radar remote sensing data includes VV band and VH band.
3. The soil salinity inversion method based on multimodal remote sensing data fusion according to claim 1, characterized in that, In step 1, the optical remote sensing data includes: blue band B2, green band B3, red band B4, first red edge band B5, second red edge band B6, third red edge band B7, near-infrared band B8, fourth red edge band B8A, first shortwave infrared band B11, and second shortwave infrared band B12.
4. The soil salinity inversion method based on multimodal remote sensing data fusion according to claim 1, characterized in that, In step 2, multimodal remote sensing features based on the fusion of radar and optical remote sensing data are used, including: Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI), Soil-Adjusted Vegetation Index (SAVI), Modified Soil-Adjusted Vegetation Index (MSAVI), Slope-Adjusted Vegetation Index (TSAVI), Adjusted Normalized Difference Vegetation Index (OSAVI), Adaptive Soil-Adjusted Vegetation Index (ATSAVI), Vegetation Index (PVI), Near-Infrared Vegetation Index (NIRv), Chlorophyll Index (MTCI), Water-Sensitive Vegetation Index (SSAI), Color Index (Clr), Vegetation Health Index (CIg), First Normalized Difference Index (NDRE1), Second Normalized Difference Index (NDRE2), First Normalized Difference Vegetation Index (NDVIr1), Second Normalized Difference Vegetation Index (NDVIr2), Third Normalized Difference Vegetation Index (NDVIr3), and Radar Vegetation Index (VV / VH). Ratio, Radar Vegetation Index (RVI), Dual Polarization SAR Vegetation Index (DPSVI), Normalized Differential Polarization Index (NDPI), Brightness (a spectral transformation method), Greenness, and Wetness.
5. The soil salinity inversion method based on multimodal remote sensing data fusion according to claim 1, characterized in that, In step 2, cross-validation is used to evaluate the model performance of the optimized remote sensing feature combination, specifically as follows: The five-fold cross-validation method is adopted, and the sample data is randomly divided into five mutually exclusive subsets. Each time, four subsets are used as the training set and the remaining one subset is used as the test set. A total of five iterations are performed. In each iteration, a random forest model is used to train the training set. The trained model is then used to predict the test set to obtain the predicted and true values. The average coefficient of determination and root mean square error of the five iterations are used to evaluate the model performance of different combinations of remote sensing features.
6. The soil salinity inversion method based on multimodal remote sensing data fusion according to claim 1, characterized in that, In step 3, the hyperparameter adjustment includes: the number of trees in the random forest (ntrees), the maximum number of nodes in each tree (maxnodes), the minimum number of samples for the terminal node (nodesize), and the number of samples drawn from the training set for each tree (sampsize).
7. A soil salinity inversion system based on multimodal remote sensing data fusion, utilizing the soil salinity inversion method based on multimodal remote sensing data fusion as described in any one of claims 1-6, characterized in that, include: Data acquisition module: used to acquire radar remote sensing data and optical remote sensing data of the area where soil salinity needs to be inverted, and to measure the electrical conductivity of the soil as an indicator for soil salinity inversion; Feature selection module: used to acquire multimodal remote sensing features based on the fusion of radar remote sensing data and optical remote sensing data, and combine radar remote sensing data and optical remote sensing data, optimize the remote sensing feature combination based on the random forest algorithm and the recursive feature elimination method, and use cross-validation to evaluate the model performance of the optimized remote sensing feature combination to obtain the optimal remote sensing feature combination. Salinity inversion module: Based on the optimal combination of remote sensing features, it constructs and trains a soil electrical conductivity prediction model based on the random forest algorithm, inputs the data to be measured into the soil electrical conductivity prediction model after hyperparameter adjustment, obtains the soil electrical conductivity, and completes the soil salinity inversion.
8. An electronic device, characterized in that, include: Memory, used to store computer programs; A processor, configured to execute the computer program to implement the steps of a soil salinity inversion method based on multimodal remote sensing data fusion as described in any one of claims 1-6.
9. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the steps of a soil salinity inversion method based on multimodal remote sensing data fusion as described in any one of claims 1-6.