A method for monitoring salinization based on multi-source remote sensing and interpretable machine learning

Abstract

The present application relates to the field of geographic information science and remote sensing monitoring technology, and particularly relates to a kind of salinization monitoring method based on multi-source remote sensing and explainable machine learning.The method obtains multi-source remote sensing data of target area, and carries out space-time benchmark unification processing, obtains optical microwave collaborative feature, combines meteorological, hydrological, human activity and land use data to construct multi-dimensional feature vector, matches multi-dimensional feature vector with ground measured salt data to construct training sample set, calculates cost-sensitive weight according to ground measured salt and land use type, trains explainable machine learning model according to multi-dimensional feature vector and cost-sensitive weight, obtains salinization inversion model, applies the model to predict target area, generates soil salinization spatial distribution map, and identifies salinization significant change area after trend and significance test.The present application effectively improves salinization inversion precision and model generalization ability, and significantly improves the accuracy of salinized land monitoring results.

Description

Technical Field

[0001] This invention relates to the fields of geographic information science and remote sensing monitoring technology, specifically to a method for monitoring salinization based on multi-source remote sensing and interpretable machine learning. Background Technology

[0002] Soil salinization is a key environmental problem that restricts the sustainable development of agriculture and ecological security in arid and semi-arid regions. Traditional ground monitoring methods are costly and have limited coverage, making it difficult to reflect the spatiotemporal dynamics of large-scale salinization in a timely manner. Therefore, using remote sensing technology to achieve efficient, accurate, and dynamic salinization monitoring is of great practical significance for regional land resource management, irrigation optimization, and ecological governance.

[0003] However, existing remote sensing-based methods for monitoring salinization still have significant limitations: on the one hand, existing technologies mainly rely on optical bands and lack effective means of detecting soil dielectric constant, which leads to a significant decrease in the inversion accuracy of a single optical index during irrigation season or after rainfall, resulting in a large number of misjudgments. Moreover, in irrigated farmland or seasonal wetlands in arid areas, soil moisture has a significant masking effect on spectral features, and moist non-saline soil appears dark in optical images, which is easily confused with dry soil with high salt content. In the early stages of salinization, the high reflectivity of moist soil is often masked by water absorption features, resulting in a significant decrease in the inversion accuracy of relying solely on optical indices during irrigation season.

[0004] On the other hand, although machine learning techniques have been used for multi-source remote sensing data fusion, existing multi-source remote sensing data fusion processes often suffer from "spatiotemporal mismatch" problems. The spatiotemporal resolution and projection of multi-source remote sensing data are not uniform in scale, leading to large regional contrast biases. Ground salinization sampling points are usually discrete data from specific years, while model training often directly splices together multi-year average climate factors, ignoring interannual fluctuations in meteorological conditions and reducing the model's training reliability. Existing methods often mix cultivated and non-cultivated land in modeling, lacking scene differentiation and differential processing, making it difficult for the model to capture irrigation-induced secondary salinization features. Furthermore, existing machine learning modeling often uses random partitioning of training and test sets, resulting in test set samples often being adjacent to training set samples. This leads to artificially high accuracy on the validation set and poor generalization ability when applied to unsampled areas, often resulting in poor performance in low-salinity areas with large sample sizes. Overfitting occurs, particularly with small sample sizes of highly saline soils. During training, existing machine learning inversion methods often employ a "black box" approach, typically randomly partitioning the training set. Due to the strong spatial autocorrelation of geographical data, random partitioning leads the model to "memorize" geographical locations rather than learn physical laws, resulting in inflated validation set accuracy but poor generalization ability in practical applications. Traditional equal-weight training causes the model to tend to fit large areas of non-saline backgrounds, while failing to adequately learn about key, scattered secondary salinization areas. Traditional regression analysis is mostly global statistics, unable to answer spatial differentiation questions at the pixel scale, i.e., unable to distinguish whether salinization in a particular area is primarily caused by climate warming or improper irrigation. Existing technologies struggle to quantify and separate the relative contributions of natural and anthropogenic factors, making it difficult for the final monitoring results to directly support differentiated and precise governance decisions. Summary of the Invention

[0005] To address the problems existing in current technologies, this invention provides a method for monitoring salinization based on multi-source remote sensing and interpretable machine learning. This method acquires multi-source remote sensing data of the target area, performs spatiotemporal benchmark unification processing, obtains optical-microwave co-features, and constructs a multi-dimensional feature vector by combining meteorological, hydrological, human activity, and land use data. The multi-dimensional feature vector is then matched with ground-measured salinity data to construct a training sample set. Cost-sensitive weights are calculated based on ground-measured salinity and land use type. An interpretable machine learning model is trained based on the multi-dimensional feature vector and cost-sensitive weights to obtain a salinization inversion model. This model is then applied to predict the target area, generating a spatial distribution map of soil salinization. After trend and significance tests, areas with significant salinization changes are identified. This invention effectively improves the accuracy of salinization inversion and the model's generalization ability, significantly enhancing the accuracy of saline-alkali land monitoring results.

[0006] This invention adopts the following technical solution: a method for monitoring salinization based on multi-source remote sensing and interpretable machine learning, comprising: Acquire multi-source remote sensing data of the target area and perform spatiotemporal benchmark unification processing; the multi-source remote sensing data includes at least optical remote sensing images, synthetic aperture radar images, meteorological data, hydro-topographic data, human activity data, and land use data; Optical-microwave coordinated features are obtained from the multi-source remote sensing data, and a multi-dimensional feature vector is constructed based on the optical-microwave coordinated features and the multi-source remote sensing data. The multidimensional feature vectors are spatiotemporally matched with preset ground-measured salinity data to construct a training sample set; Based on the measured salinity data of each training sample in the training sample set and the land use type corresponding to that training sample, obtain the cost-sensitive weight of each training sample. Based on the multidimensional feature vector and cost-sensitive weights of each training sample, an interpretable machine learning model is trained using a spatial constraint strategy to obtain a trained saltification inversion model. Based on the trained salinization inversion model, the target area is predicted, and a spatial distribution map of soil salinization in the target area is generated. Based on the spatial distribution map of soil salinization, trend and significance tests were performed to identify areas of significant salinization changes.

[0007] Furthermore, acquiring optical microwave cooperative features based on the multi-source remote sensing data includes: Extract the optical salinity index from the optical remote sensing image; Microwave moisture index is obtained from the backscattering coefficient of synthetic aperture radar imagery. Optical-microwave synergistic characteristics are obtained based on the optical salinity index and the microwave moisture index.

[0008] Furthermore, a multi-dimensional feature vector is constructed based on the optical-microwave coordinated features and the multi-source remote sensing data, specifically as follows: The vegetation index and red edge index are obtained from the optical remote sensing image and combined with the optical microwave collaborative features to obtain a subset of optical features. The backscattering coefficients of the synthetic aperture radar image are extracted, and a microwave feature subset is constructed based on the temporal statistics of the backscattering coefficients within a preset time window. Annual-scale climate variables are extracted from the meteorological data as a subset of climate features; Based on the derived variables representing hydrological processes and topographic conditions in the hydrological and topographic data, a subset of hydrological and topographic features is constructed; Based on the human activity data, variables representing the intensity of human interference are extracted to construct a subset of human activity features; Based on the land use data, soil physicochemical property variables and topographic factors are extracted to construct a static background feature subset. A multidimensional feature vector is obtained by concatenating the optical feature subset, microwave feature subset, climate feature subset, hydrological feature subset, human activity feature subset, and static background feature subset.

[0009] Furthermore, based on the measured ground salinity data of each training sample in the training sample set and the corresponding land use type, the cost-sensitive weight of each training sample is obtained, specifically as follows: Based on the ground-measured salinity data of each training sample in the training sample set, they are divided into preset salinity levels; The salinity level is combined with the land use type corresponding to the training sample to form multiple composite categories of training samples; Based on the number of training samples for each composite category and the spatial heterogeneity index of its region, the cost-sensitive weight of the training sample is determined according to a preset weight calculation formula.

[0010] Furthermore, based on the multidimensional feature vectors and cost-sensitive weights of each training sample, a space-constrained strategy is employed to train an interpretable machine learning model, including: Based on the region where the training samples are located in the training sample set, the target region is divided into multiple spatial blocks; Based on the divided spatial blocks, the training sample set is divided into a model training set and a validation set using the grouped cross-validation method; Obtain the geographical boundaries of the spatial blocks corresponding to the model training set, and remove data from each training sample whose spatial location exceeds the set range of the geographical boundary to obtain the final training set; Using the training samples corresponding to the final training set as input, and constructing the loss function of the interpretable machine learning model based on its cost-sensitive weights, and using the measured salinity data corresponding to the training samples as output, the interpretable machine learning model is trained to obtain the trained salinization inversion model.

[0011] Furthermore, after obtaining the trained saltification inversion model, the following steps are also included: Using the XGBoost-SHAP interpretability framework, we obtain the contribution value of each input feature in the multidimensional feature vector to the prediction result of each pixel. According to preset rules, the input features are divided into multiple driving factor groups, including: climate factor group, hydrological factor group, human activity factor group and static background factor group. The sum of the absolute values ​​of the contributions of all input features within each driving factor group is taken as the contribution of that driving factor group. The driving factor group whose contribution ratio exceeds the set threshold is taken as the dominant driving factor of the corresponding pixel; A governance priority image is generated based on the dominant driving factors of all pixels.

[0012] The beneficial effects of this invention are as follows: By integrating heterogeneous data from multiple sources, including optical, microwave, meteorological, hydrological, land use, and soil properties, and unifying their spatiotemporal references, this invention overcomes the problems of single data sources and scale mismatches in traditional methods. It provides a multidimensional and consistent data foundation for comprehensively characterizing the causes of salinization, improving the completeness of monitoring factors from the source. Furthermore, by utilizing the sensitivity of microwave remote sensing to soil moisture to correct optical signals, it physically weakens the interference of moisture on salinity spectra in "heterogeneous matter with the same spectrum" at the feature level, improving the quality of features. The constructed multidimensional feature vector integrates spectral, environmental, human, and background information, enabling the model to learn more complex salinization driving relationships. During model training, a sample library is constructed by spatiotemporally matching remote sensing inversion features with ground-measured data, ensuring the accuracy of the model's learning objectives and providing reliable supervision information for establishing a high-precision inversion model. Simultaneously, weights are allocated based on the salinity level and land use type of the samples, enabling the model to focus more on secondary salinization areas with sparse samples but significant remediation value during training. This effectively alleviates the model prediction bias caused by class imbalance and improves the ability to identify salinization hotspots. Then, through a validation strategy of spatial segmentation and setting up isolation zones, the inherent spatial autocorrelation in geographical data is broken, preventing the model from overfitting due to "memorizing" the locations of neighboring samples. This significantly improves the model's generalization ability in new areas and the robustness of the inversion results. Finally, the trained interpretable machine learning model is used for regional prediction, achieving rapid, large-scale, and high-precision mapping of salinization levels. At the same time, statistical trend analysis of the salinization distribution map can identify areas of significant and continuously deteriorating changes from interannual fluctuations, realizing a leap from static assessment to dynamic early warning and providing a scientific basis for accurately identifying remediation target areas. Attached Figure Description

[0013] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0014] Figure 1 This is a schematic diagram of a method for monitoring salinization based on multi-source remote sensing and interpretable machine learning, according to an embodiment of the present invention. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. 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.

[0016] A schematic diagram of a method for monitoring salinization based on multi-source remote sensing and interpretable machine learning according to an embodiment of the present invention is shown below. Figure 1 As shown, it includes: Acquire multi-source remote sensing data of the target area and perform spatiotemporal benchmark unification processing; In this embodiment of the invention, the multi-source remote sensing data includes at least optical remote sensing images, synthetic aperture radar (SAR) images, meteorological data, hydro-topographic data, human activity data, and land use data. Specifically, due to the inconsistency in spatiotemporal scales among multi-source remote sensing data, a unified data cube needs to be constructed first. In this embodiment, the optical remote sensing images are obtained from the L2A level surface reflectance product of the European Space Agency's Sentinel-2 multispectral imager; the SAR images are obtained from the C-band SAR interferometric wide swath (IW) mode ground detection distance product of the Sentinel-1 satellite, including VV and VH dual polarization data; the meteorological data is obtained from the ERA5-Land monthly reanalysis dataset released by the European Centre for Medium-Range Weather Forecasts (ECMWF), which includes 2-meter temperature, total precipitation, net solar radiation, and potential evapotranspiration; the hydro-topographic data can be obtained from the MERIT dataset released by the University of Tokyo. The Hydro hydro-topography dataset was used to extract derived variables such as elevation, flow direction, and catchment area. Land use data was obtained by acquiring the 1000-meter resolution cultivated land distribution product from the Global Food Security Support Analysis Data (GFSAD) to distinguish between cultivated and non-cultivated land. Soil property data was obtained from the OpenLandMap soil property dataset released by the International Soil Reference and Information Centre (ISRIC), including 0-5cm surface layer data such as soil pH and sand content.

[0017] When performing spatiotemporal benchmark unification, the spatial reference system of all data is first converted to the WGS 84 geographic coordinate system. Then, a resampling method is used to unify the spatial resolution of all data to 1 kilometer. For continuous variables such as surface reflectance and meteorological variables, bilinear interpolation is used for resampling. For categorical variables such as land use types, the majority aggregation method is used for resampling. Finally, all raster data are cropped according to the geographic boundaries of the target area to complete the spatial benchmark unification. For time-series data of optical remote sensing images and synthetic aperture radar images, processing is performed on an annual scale. For example, for optical remote sensing images, the median composite method is used to generate annual representative images for the growing season window from April to October each year to eliminate the influence of clouds, snow and short-term phenological fluctuations. For monthly-scale data such as meteorological data, monthly-scale variables are accumulated or averaged over the years to generate annual totals or averages. For static background data, it is treated as a background field that does not change over time, and spatial matching is performed with the annual time-series data to complete the temporal benchmark unification. When completing the spatiotemporal reference unification process, further quality control is performed on optical remote sensing imagery and synthetic aperture radar imagery data. Specific quality control methods can refer to any of the existing technologies.

[0018] Optical-microwave coordinated features are obtained from multi-source remote sensing data, and multi-dimensional feature vectors are constructed based on optical-microwave coordinated features and multi-source remote sensing data. In this embodiment of the invention, the method for obtaining optical-microwave coordinated features is as follows: extracting the optical salinity index from the optical remote sensing image; obtaining the microwave moisture index based on the backscattering coefficient of the synthetic aperture radar image; and obtaining the optical-microwave coordinated features based on the optical salinity index and the microwave moisture index.

[0019] In a specific embodiment of the present invention, pixel-level optical salinity index and brightness index are extracted from optical remote sensing images. The extraction method can be referred to in the prior art. To address the problem of optical indices being affected by soil moisture, the sensitivity of the microwave backscattering coefficient to the real part of the soil dielectric constant dominated by moisture is further utilized to construct a moisture suppression probability function, which physically corrects the optical indices. Its expression is as follows: In the formula, The normalized microwave moisture index, This represents the backscattering coefficient after radiation correction in a vertically polarized transmit-vertically polarized receive configuration. This represents the backscattering coefficient after radiation correction in a vertically polarized transmit-horizontally polarized receive configuration.

[0020] This invention further designs a mapping operator based on the Sigmoid function to convert the microwave moisture index into a moisture suppression probability, with a value range of [0, 1]. The optical-microwave cooperative feature is finally calculated using the following formula: In the formula, It is an optical-microwave coordinated feature. For pixels, The optical salinity index. The adaptive coupling coefficient can be dynamically calculated using the optical-microwave correlation within a local window. The normalized microwave moisture index, It is a nonlinear mapping operator based on the Sigmoid function, used to map microwave signals to moisture suppression probabilities. This optical-microwave cooperative feature eliminates false low-salt signals in high-humidity areas from a physical mechanism perspective.

[0021] In this embodiment of the invention, the specific method for constructing a multidimensional feature vector based on optical microwave co-location features and multi-source remote sensing data is as follows: Based on the red edge band in the optical remote sensing image, red edge indices are extracted, specifically including the red edge normalized vegetation index and the red edge chlorophyll index, which are combined with optical salinity and vegetation indices. Here, the vegetation indices include the normalized vegetation index (NDVI), enhanced vegetation index 2 (EVI2), and soil-modified vegetation index (SAVI), etc. The specific methods for obtaining the above vegetation indices can refer to existing technologies to obtain a subset of optical features; the backscattering coefficient of the synthetic aperture radar image is extracted, and based on this backscattering coefficient, a preset time... A microwave feature subset is constructed using time-series statistics within the time window; annual-scale climate variables are extracted from meteorological data as a climate feature subset; a hydro-topographic feature subset is constructed based on derived variables characterizing hydrological processes and topographic conditions from hydro-topographic data; a human activity feature subset is constructed by extracting variables characterizing the intensity of human disturbance from human activity data; a static background feature subset is constructed by extracting soil physicochemical properties and topographic factors from land use data; and a multidimensional feature vector is obtained by concatenating the optical feature subset, microwave feature subset, climate feature subset, hydro-topographic feature subset, human activity feature subset, and static background feature subset.

[0022] In a specific embodiment of the present invention, the optical feature subset mainly includes core optical features, vegetation stress features, and salinity auxiliary features. The core optical features are the optical salinity index and vegetation index; the vegetation stress feature is represented by the red-edge chlorophyll index in the red-edge index; and the salinity auxiliary feature is the brightness index. The microwave feature subset provides independent microwave information, which can typically be obtained by calculating three time-series statistics of the polarization backscattering coefficient of the synthetic aperture radar image for each pixel within an annual time window: median, variance, and 90th quantile. The climate feature subset can specifically be the annual cumulative precipitation and annual average 2-meter temperature extracted from ERA5-Land annual data. The hydro-topographic feature subset can specifically be from MERIT... Elevation and topographic moisture index are extracted from the Hydro data; the human activity feature subset can be obtained from the integrated binary mask data of cultivated and non-cultivated land; the static background feature subset is obtained by extracting soil pH and sand content from the OpenLandMap soil attribute dataset published by the International Soil Reference and Information Center (ISRIC); finally, all the above feature subsets are concatenated at the pixel scale to form a comprehensive multidimensional feature vector, which is used to characterize the comprehensive state of each pixel in a specific year, represented as: In the formula, Represents a pixel In the year Multidimensional feature vectors at time A subset of optical features A subset of microwave features A subset of climate characteristics It is a subset of hydrogeomorphic features. A subset of human activity characteristics This is a subset of static background features.

[0023] A training sample set is constructed by spatiotemporally matching the multidimensional feature vector with the preset ground-measured salinity data. In this embodiment of the invention, the ground-measured salinity data is the measured data of historical soil samples collected within the target area. The data source can be a publicly available scientific research dataset or field sampling data, which includes the geographic coordinates of the sampling point pixels, sampling time, salinity observation value, and land use type of the sampling point pixels, etc. When performing spatiotemporal matching of the multidimensional feature vector with the preset ground-measured salinity data, the same operation as the spatiotemporal benchmark unification processing of multi-source remote sensing data is performed, thereby forming a complete training sample for each valid sampling point pixel in the target area and its corresponding year. The training samples corresponding to all sampling point pixels are collected to construct a training sample set for training interpretable machine learning models.

[0024] Based on the measured salinity data of each training sample in the training sample set and the land use type corresponding to that training sample, obtain the cost-sensitive weight of each training sample. In this embodiment of the invention, to address the problem of insufficient identification of secondary salinization of arable land by traditional models, when training the interpretable machine learning model, a simple random sampling method is no longer used. Instead, cost-sensitive weights for the training samples are constructed and input into the model's loss function to improve the model's ability to identify spatially marginal areas or areas with drastic variations. Specifically, the cost-sensitive weights are determined as follows: based on the measured salinity data of each training sample in the training sample set, the samples are divided into preset salinity levels; the salinity levels are combined with the land use type corresponding to the training sample to form multiple composite categories of training samples; and the cost-sensitive weight of each training sample is determined according to the number of training samples in each composite category and the spatial heterogeneity index of its region, based on a preset weight calculation formula.

[0025] In a specific embodiment of the present invention, the spatial heterogeneity index of the region where the training samples of each composite category are located is as follows: a window is set with the pixel corresponding to each training sample as the center. Within this window, the coefficient of variation of the measured salinity data of all training samples under the same land use type is calculated as the local spatial heterogeneity index of the training sample. The local spatial heterogeneity index is then converted into a cost-sensitive weight using the following formula: In the formula, For the first Cost-sensitive weights for each training sample. For the first Each training sample corresponds to the number of training samples in the composite category. To adjust the parameters, It is a local spatial heterogeneity index. For the number of composite categories, This represents the total number of training samples.

[0026] Based on the multidimensional feature vector and cost-sensitive weights of each training sample, an interpretable machine learning model is trained using a spatial constraint strategy to obtain a trained saltification inversion model. In this embodiment of the invention, the process of training an interpretable machine learning model using a spatial constraint strategy is as follows: the target region is divided into multiple spatial blocks according to the region where the training samples are located in the training sample set; based on the multiple spatial blocks after division, the training sample set is divided into a model training set and a validation set using a grouped cross-validation method; the geographical boundaries of the spatial blocks corresponding to the model training set are obtained, and data in each training sample whose spatial location exceeds the set range of the geographical boundary are removed to obtain the final training set; the training samples corresponding to the final training set are used as input, and the loss function of the interpretable machine learning model is constructed according to its cost-sensitive weights; the measured salinity data corresponding to the training samples are used as output to train the interpretable machine learning model to obtain the trained salinization inversion model.

[0027] In one specific embodiment of the present invention, the interpretable machine learning model uses a nonlinear gradient boosting tree model as the base learner and the Huber loss function as the optimization objective of the model. During group cross-validation, the Bayesian optimization method is used to optimize the key hyperparameters of the model, with the optimization objective being to minimize the root mean square error and mean absolute error on the validation set.

[0028] Based on the trained salinization inversion model, the target area is predicted, and a spatial distribution map of soil salinization in the target area is generated. In this embodiment of the invention, the feature vectors of all pixels in the target area are assembled into a matrix. Using a block or batch prediction method, the matrix is ​​divided into multiple batches and sequentially input into the trained salinization inversion model for prediction. The prediction result is a continuous salinization index prediction value for each pixel. In this embodiment of the invention, this value is labeled as a relative salinization degree index between 0 and 1. Then, the salinization index prediction values ​​of all pixels are reorganized according to their original geographic coordinates and filled into a blank raster template with the same spatial range, projection, and resolution as the target area. This raster data is output as a standard GeoTIFF format file, and its pixel value is the predicted soil salinization index, thereby obtaining a spatial distribution map of soil salinization in the target area.

[0029] Based on the spatial distribution map of soil salinization, trend and significance tests were conducted to identify areas of significant changes in salinization.

[0030] In this embodiment of the invention, the soil salinization index corresponding to each pixel in the spatial distribution map of soil salinization is extracted. A corresponding time series is constructed based on the soil salinization index of multiple years. Then, Sen's slope trend estimation is performed on the time series of each pixel, that is, its Sen slope is calculated as an unbiased estimate of the trend strength. The specific calculation method can refer to the existing technology. Then, the Mann-Kendall nonparametric trend test is performed on the time series to evaluate whether the trend is statistically significant. Specifically, the Mann-Kendall statistic and the standardized test statistic of the time series are calculated. The p-value of the two-tailed test is calculated based on the standardized test statistic. The smaller the p-value, the more significant the trend. In specific implementation, a significance level threshold can be set according to the actual situation. When the p-value is less than the set significance level threshold, the salinization change trend of the pixel is considered to be statistically significant.

[0031] In another specific embodiment of the present invention, the XGBoost-SHAP interpretability framework is further utilized to obtain the contribution value of each input feature in the multidimensional feature vector to the prediction result of each pixel; according to preset rules, the input features are divided into multiple driving factor groups, including: climate factor group, hydrological factor group, human activity factor group, and static background factor group; the sum of the absolute values ​​of the contribution values ​​of all input features in each driving factor group is obtained as the contribution degree of the driving factor group, expressed as: In the formula, Contribution to each driving factor group The contribution of input features in each driving factor group to the pixel prediction result.

[0032] The driving factor group whose contribution ratio exceeds a set threshold is regarded as the dominant driving factor of the corresponding pixel. The determination criteria are as follows: In the formula, As the dominant driving factor type, To set a threshold, The above formula indicates that if the contribution percentage is greater than the set threshold, the salinization of the pixel is determined to be dominated by this factor type; otherwise, it is marked as a mixed or uncertain type.

[0033] A governance priority image is generated based on the dominant driving factors of all pixels, as follows: In this embodiment of the invention, in order to directly serve governance decisions with monitoring and attribution results, a governance priority index is constructed to rank pixels. This index integrates information on the severity of salinization changes, trend strength, and dominant driving factors, and is expressed as follows: in, For pixels Governance priority index This is a hotspot indicator function. The contribution percentage of the leading group To normalize the trend strength, , , For weights.

[0034] The governance priority index value is calculated for all pixels in the target area. The calculation result is spatialized to generate a governance priority image. In specific implementation, the priority values ​​can be classified and basic geographical elements such as administrative boundaries and water systems can be superimposed to form a thematic map that can be directly used for consultation and engineering deployment.

[0035] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring salinization based on multi-source remote sensing and interpretable machine learning, characterized in that, include: Acquire multi-source remote sensing data of the target area and perform spatiotemporal benchmark unification processing; The multi-source remote sensing data includes at least optical remote sensing images, synthetic aperture radar images, meteorological data, hydro-topographic data, human activity data, and land use data. Optical-microwave coordinated features are obtained from the multi-source remote sensing data, and a multi-dimensional feature vector is constructed based on the optical-microwave coordinated features and the multi-source remote sensing data. The multidimensional feature vectors are spatiotemporally matched with preset ground-measured salinity data to construct a training sample set; Based on the measured salinity data of each training sample in the training sample set and the land use type corresponding to that training sample, obtain the cost-sensitive weight of each training sample. Based on the multidimensional feature vector and cost-sensitive weights of each training sample, an interpretable machine learning model is trained using a spatial constraint strategy to obtain a trained saltification inversion model. Based on the trained salinization inversion model, the target area is predicted, and a spatial distribution map of soil salinization in the target area is generated. Based on the spatial distribution map of soil salinization, trend and significance tests were performed to identify areas of significant salinization changes.

2. The method for monitoring salinization based on multi-source remote sensing and interpretable machine learning according to claim 1, characterized in that: Obtaining optical microwave coordinated features based on the multi-source remote sensing data includes: Extract the optical salinity index from the optical remote sensing image; Microwave moisture index is obtained from the backscattering coefficient of synthetic aperture radar imagery. Optical-microwave synergistic characteristics are obtained based on the optical salinity index and the microwave moisture index.

3. The method for monitoring salinization based on multi-source remote sensing and interpretable machine learning according to claim 1, characterized in that: A multi-dimensional feature vector is constructed based on the optical-microwave coordinated features and the multi-source remote sensing data, specifically as follows: The vegetation index and red edge index are obtained from the optical remote sensing image, and a subset of optical features is obtained by combining them with the optical microwave collaborative features. The backscattering coefficients of the synthetic aperture radar image are extracted, and a microwave feature subset is constructed based on the temporal statistics of the backscattering coefficients within a preset time window. Annual-scale climate variables are extracted from the meteorological data as a subset of climate features; Based on the derived variables representing hydrological processes and topographic conditions in the hydrological and topographic data, a subset of hydrological and topographic features is constructed; Based on the human activity data, variables representing the intensity of human interference are extracted to construct a subset of human activity features; Based on the land use data, soil physicochemical property variables and topographic factors are extracted to construct a static background feature subset. A multidimensional feature vector is obtained by concatenating the optical feature subset, microwave feature subset, climate feature subset, hydro-topographic feature subset, human activity feature subset, and static background feature subset.

4. The method for monitoring salinization based on multi-source remote sensing and interpretable machine learning according to claim 1, characterized in that: Based on the measured salinity data of each training sample in the training sample set and the corresponding land use type, the cost-sensitive weight of each training sample is obtained, specifically as follows: Based on the ground-measured salinity data of each training sample in the training sample set, they are divided into preset salinity levels; The salinity level is combined with the land use type corresponding to the training sample to form multiple composite categories of training samples; Based on the number of training samples for each composite category and the spatial heterogeneity index of its region, the cost-sensitive weight of the training sample is determined according to a preset weight calculation formula.

5. The method for monitoring salinization based on multi-source remote sensing and interpretable machine learning according to claim 1, characterized in that: Based on the multidimensional feature vectors and cost-sensitive weights of each training sample, an interpretable machine learning model is trained using a space-constrained strategy, including: Based on the region where the training samples are located in the training sample set, the target region is divided into multiple spatial blocks; Based on the divided spatial blocks, the training sample set is divided into a model training set and a validation set using the grouped cross-validation method; Obtain the geographical boundaries of the spatial blocks corresponding to the model training set, and remove data from each training sample whose spatial location exceeds the set range of the geographical boundary to obtain the final training set; Using the training samples corresponding to the final training set as input, and constructing the loss function of the interpretable machine learning model based on its cost-sensitive weights, and using the measured salinity data corresponding to the training samples as output, the interpretable machine learning model is trained to obtain the trained salinization inversion model.

6. The method for monitoring salinization based on multi-source remote sensing and interpretable machine learning according to claim 1, characterized in that: After obtaining the trained saltification inversion model, the following steps are also included: Using the XGBoost-SHAP interpretability framework, we obtain the contribution value of each input feature in the multidimensional feature vector to the prediction result of each pixel. According to preset rules, the input features are divided into multiple driving factor groups, including: climate factor group, hydrological factor group, human activity factor group and static background factor group. The sum of the absolute values ​​of the contributions of all input features within each driving factor group is taken as the contribution of that driving factor group. The driving factor group whose contribution ratio exceeds the set threshold is taken as the dominant driving factor of the corresponding pixel; A governance priority image is generated based on the dominant driving factors of all pixels.

Images