A soil salinity inversion method for multi-source remote sensing image cross-modal fusion

Abstract

The application discloses a soil salt content inversion method based on multi-source remote sensing image cross-modal fusion, and relates to the technical field of soil salt content inversion. The specific implementation scheme is as follows: the starting point of a remote sensing observation time sequence is determined through sampling and meteorological data, spectral features are extracted from remote sensing images to form a time sequence spectral feature set, a dynamic response feature parameter is constructed based on a salt content and a water content index, curve parameters are extracted through curve fitting and fused into a comprehensive feature set, the contribution of each feature in the comprehensive feature set to the salt content response is analyzed to construct a feature subset, the feature subset is input into a salt content inversion model, and soil salt content spatial distribution data are synchronously output. Through the application, the surface spectrum saturation limitation is effectively broken, and integrated remote sensing diagnosis of high-precision inversion of total salt content is realized.

Description

Technical Field

[0001] This application relates to the field of soil salinity inversion technology, specifically to a method for soil salinity inversion through cross-modal fusion of multi-source remote sensing images. Background Technology

[0002] In the field of remote sensing monitoring of soil salinization, timely and accurate understanding of the total amount of soil salinity and its dynamic changes is crucial for guiding irrigation management and saline-alkali land improvement. Remote sensing inversion methods can provide macro-scale salinity information, effectively compensating for the limitations of ground sampling. Currently, most common soil salinity remote sensing inversion methods rely on single-temporal or a few-temporal optical remote sensing images. By extracting spectral features such as salinity index, vegetation index, and moisture index, statistical regression or machine learning models are established with the surface soil salinity data from sampling points to achieve spatial distribution mapping of salinity. Some studies have attempted to introduce multi-temporal images, superimposing single indices from different times or performing simple change analysis, but their analytical framework is essentially still static or quasi-static modeling, failing to utilize the dynamic process of salinity content changing with external conditions as an independent feature dimension.

[0003] The existing methods described above have significant shortcomings. When surface soil salts accumulate to a high concentration through evaporation and form salt crusts or patches, the spectral reflectance in remote sensing images approaches saturation. Spectral indices cannot effectively characterize the difference between surface and deep soil salts, making it difficult to deduce the true total salt content. More importantly, rainfall infiltration and post-rain evaporation processes cause strong vertical transport of soil salts. This dynamic evolution contains crucial information for distinguishing between surface and deep salts, but current technologies lack quantitative descriptions of the dynamic process of water-salt coupling. They cannot extract and filter effective features reflecting the direction, sensitivity, and decoupling degree of salt response to rainfall from time-series remote sensing information. Consequently, the inversion model not only has limited accuracy in estimating total salt content but also completely lacks the ability to diagnose and map the dynamic response process of salts. Summary of the Invention

[0004] In view of this, the purpose of this application is to provide a soil salinity inversion method based on cross-modal fusion of multi-source remote sensing images, which can overcome the limitation of surface spectral saturation and achieve high-precision inversion of total salinity through integrated remote sensing diagnosis.

[0005] According to a first aspect of this application, a method for soil salinity inversion through cross-modal fusion of multi-source remote sensing images is provided, comprising: Sampling points are set for the soil in the area to be inverted. Sampling and analysis of the soil at the sampling points are performed to obtain sampling data. Meteorological data of the area to be inverted is obtained. Based on the meteorological data, the remote sensing time point for acquiring remote sensing images is determined. Remote sensing images of the inverted area are acquired based on remote sensing time points and preprocessed. A set of temporal spectral feature parameters that can characterize the soil state are extracted pixel by pixel from the preprocessed remote sensing images. Based on the salinity index and moisture index in the set of spectral feature parameters at each remote sensing time point, the dynamic response feature parameters of salinity change over time are analyzed, and the change curve of salinity is fitted. Based on the change curve, the curve parameters are extracted, and the spectral feature parameters, dynamic response feature parameters, and curve parameters are constructed into a comprehensive feature set. The comprehensive feature set of each sampling point is paired with the corresponding total salt content. For each feature in the comprehensive feature set, the univariate regression determination coefficient with the total salt content, the process response degree corresponding to the change before and after rainfall, and the average synergistic gain value after the feature is combined with other features are calculated. The determination coefficient, process response degree and average synergistic gain value are normalized and weighted to obtain the contribution ability of the feature. The features are sorted and filtered according to their contribution ability to construct a feature subset. A salinity inversion model was constructed and trained. A subset of features was used as input to the trained model, and the spatial distribution data of soil salinity content was output in real time.

[0006] In some embodiments, the sampling data includes: surface salt content and total salt content; The specific process for obtaining the sampled data is as follows: The soil to be inverted is divided into uniform grids, and the center of each grid is used as a sampling point. Several random sampling plots are set up at each sampling point. The surface salt content and total salt content of the soil in each sampling plot are collected. The average of the surface salt content and total salt content of all sampling plots under the same sampling point is calculated and used as the surface salt content and total salt content of the soil at that sampling point.

[0007] In some embodiments, meteorological data includes rainfall amount and rainfall duration; The specific method for determining the remote sensing time point for acquiring remote sensing images based on meteorological data is as follows: Based on the relationship between soil surface salinity in the sampling data and the changes in surface salinity after historical rainfall, the dissolved rainfall threshold was determined. Within the observation period, the moment when the cumulative rainfall first reaches the dissolved rainfall threshold is taken as the remote sensing time starting point. The remote sensing time length is set and divided into equal parts, with each division point being a remote sensing time point.

[0008] In some embodiments, radiometric calibration and atmospheric correction are performed sequentially on multi-temporal remote sensing images to eliminate atmospheric effects, restore the true reflectivity of the land surface, and then orthorectification is performed. Pixel-level spatial alignment is then performed, and the orthorectified remote sensing images are masked to obtain remote sensing images with unified radiometric and geometric references.

[0009] In some embodiments, spectral characteristic parameters include; Normalized salinity index, brightness index, vegetation category index, water category index, temperature and humidity index; The method for extracting spectral feature parameters is as follows: Based on the preprocessed remote sensing images acquired at remote sensing time points, a set of surface reflectance values ​​for different bands corresponding to each pixel are obtained. According to the type of spectral feature parameter, and the corresponding band combination operation rules for each type of spectral feature parameter, pixel-level mathematical operations are performed on all pixels of all remote sensing images at all remote sensing time points. The original, single-band surface reflectance information is transformed into a series of spectral feature parameters that have clear indicative capabilities for soil salinity, vegetation status, surface moisture, and temperature and humidity elements. After calculating all spectral feature parameters for a single remote sensing time point, all spectral feature images for that time point are stacked according to the feature dimension to form a set of spectral feature parameters for that single remote sensing time point. The above calculation process is repeated for all selected remote sensing time points. Finally, the sets of spectral feature parameters for each remote sensing time point are aligned and stitched together along the time dimension to form a set of spectral feature parameters for time-series features.

[0010] In some embodiments, dynamic response characteristic parameters include the relative salinity response index and the water-salt decoupling index; The process of obtaining the relative salinity response index is as follows: Using the remote sensing time points corresponding to the salinity and moisture indices before rainfall as the baseline time, and based on the salinity and moisture indices corresponding to adjacent remote sensing time points after rainfall, the change values ​​of salinity and moisture indices are obtained according to the time series. The remote sensing time point corresponding to the maximum value of the change values ​​of salinity and moisture indices is taken as the peak time. The difference between the salinity index at the peak time and the salinity index at the baseline time is taken as the net change in salinity. The absolute value of the difference between the moisture index at the peak time and the moisture index at the baseline time is taken as the change in moisture. A constant is introduced, and the ratio obtained by dividing the net change in salinity by the sum of the change in moisture and the constant is the relative salinity response index. The process for obtaining the water-salt decoupling index is as follows: Based on the salinity index and moisture index in the spectral feature parameter set, they are sorted according to time series to form salinity index series and moisture index series, respectively. They are then normalized and scaled to the same interval. A dynamic time warping algorithm is used to find the alignment path with the minimum cumulative distance between the two normalized sequences by bending, stretching or compressing the time axis, and this shortest cumulative distance is recorded. The ratio obtained by dividing the shortest cumulative distance by the actual step length of the warped path is the water-salt decoupling index.

[0011] In some embodiments, the curve parameters include the salinity change range, the maximum rate of change, and the area under the curve. The method for fitting the curve is as follows: Based on the salinity index in the set of spectral feature parameters, the parameters are sorted according to time series and fitted using cubic spline interpolation to obtain a salinity change curve that is continuous and smooth in time, which is the change curve. The method for extracting the range of salinity changes is as follows: Based on the change curve, find the highest and lowest points on the entire change curve, and subtract the salinity index of the lowest point from the salinity index of the highest point. The difference is the salinity change range. The method for extracting the maximum rate of change is as follows: Based on the change curve, calculate the first derivative of the change curve, and the maximum absolute value of the first derivative is the maximum rate of change. The method for extracting the area of ​​a curve is as follows: Using the salinity index before rainfall as a baseline, the area enclosed by the entire change curve and this baseline is the area of ​​the curve.

[0012] In some embodiments, the analysis process for contribution capability is as follows: Based on the measured total salt content in the soil at all sampling points, the comprehensive feature set of each sampling point is paired with the corresponding total salt content one by one. The univariate regression determination coefficient between each feature in the comprehensive feature set and the total salt content is calculated to obtain the determination coefficient between the feature and the total salt content. The average change of this characteristic at all sampling points before and after rainfall, as well as the average change of the total salt content at the corresponding sampling points, were analyzed. The relative deviation between the two changes was obtained, and the process response was obtained by subtracting the relative deviation from 1. Based on any two features in the comprehensive feature set, perform univariate regression on the total salt content to obtain the explanatory power of the two individual features. Perform binary regression on the measured total salt content using the two features together to obtain the joint explanatory power. Subtract the sum of the explanatory powers of the two individual features from the joint explanatory power to obtain the collaborative gain value of the features. The contribution capability of this feature is obtained by normalizing the coefficient of determination, process responsiveness, and synergistic gain value and then fusing them according to preset weights.

[0013] In some embodiments, the steps for prioritizing each feature based on its contribution capability and constructing a feature subset are as follows: based on the contribution capability value, the features are sorted in descending order from high to low to form an ordered sequence. A curve is plotted in a plane coordinate system with the sorted sequence number as the horizontal axis and the contribution capability value corresponding to each feature as the vertical axis. The inflection point of the transition region between the steep and flat sections of the curve is found. All features with contribution capabilities greater than the contribution capability value corresponding to this inflection point are selected to construct the feature subset.

[0014] In some embodiments, the salt inversion model is a deep learning model built based on temporal attention mechanism and multi-task learning. The specific construction steps are as follows: The model construction includes: Input layer: Organizes the feature subset into a standardized data format that the model can process; Temporal coding layer: It consists of two long short-term memory network units in opposite directions, which are used to automatically learn the dependencies and evolution patterns of features at different time points along the time axis; Perceptual attention layer: It contains two linear transformations and a non-linear activation function, which is used to give the model the ability to autonomously judge the importance of information at different time points; Multi-task learning output layer: includes a salinity regression head and a spatial mapping head, wherein the salinity regression head consists of a 3-layer fully connected network and is used to output the soil salinity content corresponding to each pixel; The spatial mapping head reconstructs the salt content of each pixel into a spatial distribution map with the same resolution as the input remote sensing image based on the pixel coordinates and output results.

[0015] One embodiment of the above application has the following advantages or beneficial effects: This application uses future rainfall events that generate effective water infiltration and salt migration as triggering conditions to dynamically set remote sensing observation time series, acquire multi-temporal images, and extract multiple types of spectral features representing soil state pixel by pixel. These features are then stacked along the time dimension to form a spectral feature set. This approach systematically integrates remote sensing spectral information from different time phases and types, transforming previously isolated and static spectral indices into expressions with clear temporal evolution patterns. This allows for real-time depiction of the complete process of salt infiltration and evaporation-induced salt return. Furthermore, this scheme constructs dynamic response characteristic parameters and extracts dynamic curve parameters by fitting continuously changing curves, tightly coupling remote sensing spectral information with the dynamic evolution of salt driven by rainfall. These dynamic features can quantitatively describe the salt response behavior to water from multiple dimensions, including direction, sensitivity, synchronicity, and cumulative deviation. Even when the surface salt concentration has reached spectral saturation, the overall migration of deep salt can still be effectively perceived and calibrated through these process features, thus solving the problem that traditional static remote sensing can only reflect the surface and cannot detect the total salt content of the profile.

[0016] The salinity inversion method based on contribution capability proposed in this application can accurately identify feature combinations that, although weak individually, significantly enhance the explanatory power of salinity dynamics when combined, ensuring that the selected feature subset has high sensitivity and interactive explanatory power for salinity evolution.

[0017] The selected feature subset is input into the salinity inversion model. The model can autonomously learn the temporal dependence before and after rainfall and focus on key response moments. At the same time, it outputs a spatial distribution map of soil salinity content. It truly realizes the integrated inversion of multi-temporal remote sensing spectral fusion → rainfall response interactive feature selection → salinity dynamic evolution. It greatly improves the ability to break through the surface spectral saturation limit and realize high-precision integrated remote sensing of total salinity inversion.

[0018] Other effects of the above-mentioned alternative methods will be described below in conjunction with specific embodiments. Attached Figure Description

[0019] The accompanying drawings are provided for a better understanding of this solution and do not constitute a limitation of this application. Wherein: Figure 1 This is a flowchart illustrating a method for soil salinity inversion based on cross-modal fusion of multi-source remote sensing images provided in an embodiment of this application. Figure 2 (a) A time-series variation graph of salt and moisture index provided in an embodiment of this application; Figure 2 (b) Relative salinity response index for different rainfall events provided in the embodiments of this application; Figure 2(c) A Dynamic Time Warping (DTW) alignment analysis diagram provided for an embodiment of this application; Figure 2 (d) is a comparison chart of water-salt decoupling indices for different soil types provided in the embodiments of this application; Figure 3 (a) is a schematic diagram of the fitting of the difference of cubic splines of the salt index provided in the embodiments of this application; Figure 3 (b) A graph showing the rate of salt change provided in an embodiment of this application; Figure 3 (c) A schematic diagram of curve area calculation provided in an embodiment of this application; Figure 3 (d) is a comparison chart of curve parameters for different rainfall events provided in the embodiments of this application; Figure 4 (a) A predicted distribution map of this embodiment provided for an example of this application; Figure 4 (b) A measured salt distribution map provided in an embodiment of this application; Figure 4 (c) Distribution map of CNN method predictions provided in the embodiments of this application; Figure 4 (d) is a distribution map predicted by the random forest method provided in the embodiments of this application; Figure 5 The predicted residual distribution diagram provided for the embodiments of this application. Detailed Implementation

[0020] In the following description, specific details such as particular system architectures and techniques are set forth for illustrative purposes and not for limitation, in order to provide a thorough understanding of the embodiments of this application. However, those skilled in the art will understand that this application may also be implemented in other embodiments without these specific details. In other instances, detailed descriptions of well-known systems, apparatuses, circuits, and methods have been omitted so as not to obscure the description of this application with unnecessary detail.

[0021] To facilitate understanding of this application, the embodiments of this application will be briefly described below: This application provides a method for soil salinity inversion based on cross-modal fusion of multi-source remote sensing images, primarily applied to the precise monitoring and dynamic diagnosis of soil salinization in semi-arid irrigation areas. In this scenario, it is necessary to perform remote sensing inversion of the total soil salinity at the irrigation area scale and its transport process after irrigation or rainfall to guide scientific water and salt regulation and saline-alkali land improvement.

[0022] See Figure 1 This is a flowchart illustrating a method for soil salinity inversion based on cross-modal fusion of multi-source remote sensing images provided in this application embodiment. Figure 1The execution subject of the method shown can be a combination of software and / or hardware, specifically, it can be one or more of various types of terminals, hardware systems, cloud computing, etc.

[0023] Figure 1 The method for soil salinity inversion by cross-modal fusion of multi-source remote sensing images, as shown, includes steps S101 to S105, as detailed below: S101: Set up sampling points for the soil in the area to be inverted, sample and analyze the soil at the sampling points to obtain sampling data, obtain meteorological data for the area to be inverted, and determine the remote sensing time point for acquiring remote sensing images based on the meteorological data.

[0024] This step first obtains surface salt content through sampling and then uses the concept of dissolved rainfall to determine the critical rainfall amount required for effective water infiltration, i.e., dissolved rainfall. Subsequently, only when the predicted future rainfall exceeds this threshold is it considered the starting point of an effective rainfall event, and the event is divided into a series of remote sensing time points according to a set duration. This design ensures that each remote sensing observation corresponds to a real, rainfall-induced soil salinity redistribution process.

[0025] This step, based on physical mechanisms, transforms the process from passively receiving remote sensing images to actively designing observation sequences, ensuring that the acquired time-series data has a clear process orientation. This directly solves the existing problem that, due to the lack of effective dynamic observation strategies, remote sensing cannot infer the total salinity in the deeper layers when surface salinity is saturated. The accurate time-series data acquired in this step provides an indispensable spectral evolution record that is strictly synchronized with water migration for constructing the dynamic characteristics of rainfall response in subsequent steps. This makes it possible to capture the direction and sensitivity of salt accumulation to the surface, as well as the degree of water-salt decoupling, thus overcoming the limitations of traditional static spectroscopy.

[0026] From the perspective of its contribution to the overall technical solution, S101 is the initial driving force that connects the entire technical concept. It acts like a "process-aware clock" for the entire inversion method system, with all subsequent steps revolving around the time-series data generated by this clock. Specifically, it provides a unified time anchor for the stacking of multi-temporal spectral feature parameters in S102, is the sole data source for constructing the relative salinity response index, water-salt decoupling index, and various curve parameters in S103, directly determines the effectiveness of the feature's responsiveness to the rainfall process and the synergistic gain value in S104, and ultimately inputs standardized time-series data containing the true dynamic evolution law of salinity into the deep learning model of S105. It can be said that it is precisely because S101 guarantees the event-driven nature of the time-series data from the source that this solution can closely link spectral fusion, interactive feature selection and fusion with the dynamic evolution of salinity, ultimately achieving the progressive effect of integrating high-precision total salinity inversion with salinity dynamic process diagnosis.

[0027] In this embodiment, the sampling data includes: surface salt content and total salt content; The specific process for obtaining the sampled data is as follows: The soil to be inverted is divided into uniform grids, and the center of each grid is used as a sampling point. Several random sampling plots are set up at each sampling point. The surface salt content and total salt content of the soil in each sampling plot are collected. The average of the surface salt content and total salt content of all sampling plots under the same sampling point is calculated and used as the surface salt content and total salt content of the soil at that sampling point.

[0028] The sampling process was designed following the principles of remote sensing pixel scale matching and soil stratification. First, the baseline size of the ground sampling grid was determined based on the spatial resolution of the remote sensing image. Typically, the grid side length was set to an integer multiple of the side length of a single pixel in the remote sensing image, ensuring that each sampling grid unit could completely cover several pixels. This ensured a one-to-one spatial correspondence between the measured ground data and the remote sensing image pixels. During grid division, a regular fishing net-like grid was generated within the vector boundary of the area to be inverted, using the determined grid side length. The physical center of the grid was the preset sampling point. For field boundaries or abrupt topographic changes, the sampling point position could be appropriately offset to fall near the center, ensuring the representativeness of the sample points. Four sampling plots were radially arranged outwards from the sampling point at each grid intersection, maintaining a distance of more than one meter between sampling plots to avoid interference from local heterogeneity. Simultaneously, all sampling plots were located within the same grid unit to ensure they corresponded to the same set of remote sensing pixels.

[0029] More specifically, the sampling depth is determined by comprehensively considering the thickness of the active layer for vertical salt transport in the study area and the distribution range of crop roots. The overall sampling depth can be set to 100 centimeters, and the surface salt content is the content from the soil surface depth of 0.1 meters to the surface. During sampling, the soil at the top 0.1 meters is first collected, bagged, and its location and depth are marked. Then, the soil at the remaining depth is sampled and its location and depth are marked.

[0030] The method for obtaining surface salinity is as follows: surface soil samples are extracted from four sampling plots at each sampling point, soil extracts are prepared, and the conductivity of the extracts is measured using a conductivity meter. Then, the salinity of each sample is calculated using a pre-established regional conductivity-salinity conversion formula. The average surface salinity of all sampling plots at that sampling point is used as the surface salinity of that point. This is used to determine the effective rainfall threshold that can initiate salt transport, serving as the triggering basis for the entire dynamic remote sensing observation strategy.

[0031] The method for obtaining the total salt content is as follows: for soil samples from the remaining depth of the four sampling plots at each sampling point, the salt content is obtained using the same method described above, and the average salt content of the remaining depth of all sampling plots at each sampling point is calculated as the salt content of the remaining depth. The total salt content is obtained by adding the salt content of the remaining depth at the same sampling point to the surface salt content.

[0032] In this embodiment, the meteorological data includes rainfall amount and rainfall time; The specific method for determining the remote sensing time point for acquiring remote sensing images based on meteorological data is as follows: Based on the relationship between soil surface salinity in the sampling data and the changes in surface salinity after historical rainfall, the dissolved rainfall threshold was determined. Within the observation period, the moment when the cumulative rainfall first reaches the dissolved rainfall threshold is taken as the remote sensing time starting point. The remote sensing time length is set and divided into equal parts, with each division point being a remote sensing time point.

[0033] More specifically, meteorological data can be obtained through the short-term quantitative precipitation forecast grid products issued by the China Meteorological Administration, whose timeliness can meet the needs of advance planning for remote sensing image acquisition.

[0034] Dissolved rainfall can be quantitatively determined through outdoor sprinkler experiments. The specific steps are as follows: First, select several sampling points representing different surface salinity levels in the area to be remotely sensed. Set up small closed or semi-closed simulated rainfall devices on these points. Apply water treatments with different water volume gradients to the soil surface using artificial sprinklers or portable rainfall simulators. The water volume for each treatment increases sequentially, such as setting multiple levels like 5 mm, 10 mm, 15 mm, and 20 mm. After each sprinkler, observe and sample at different times to monitor the changes in the downward depth of the water content increase front in the soil profile. That is, if the water content increase front has moved down to below the surface, it indicates that there is a significant net migration of salt from the surface to the lower layers. Record the minimum experimental sprinkler volume at which significant water infiltration is first observed as the dissolved rainfall of that sampling point. Statistically analyze the measured values ​​of multiple sampling points and take the average value or establish an empirical relationship curve between surface salinity and dissolved rainfall as a unified standard for determining dissolved rainfall in the entire study area.

[0035] Based on real-time rainfall forecast information transmitted from meteorological stations, the rainfall amounts for all expected rainfall events in the future period are screened out. The forecast cumulative rainfall for each rainfall event is compared with the pre-measured dissolved rainfall to identify rainfall events where the forecast rainfall first exceeds the dissolved rainfall threshold. The dissolved rainfall threshold represents the minimum amount of rainfall required for effective dissolution, infiltration, and migration of surface soil salts. Within the observation period, only when the cumulative rainfall first reaches this threshold does the soil water and salt state undergo a fundamental shift from the relatively stable background before rainfall, and salts begin to be truly driven. If the second or later threshold-reaching moment is chosen as the starting point, it means that the soil has already experienced at least one effective water and salt disturbance before this point, and surface salts may have been partially leached or redistributed. Soil moisture content is also not in its initial state. At this point, setting a reference time can no longer represent the true background conditions before rainfall. Subsequent calculations of net salt change, moisture change, and relative salt response index will be distorted due to the residual effects of the previous rainfall. Therefore, the actual date of the rainfall or the predicted start time is used as the starting point for the entire time-series remote sensing observation. Pre-rainfall remote sensing images are acquired before the rainfall occurs, and then the data is analyzed based on rainfall intensity and local conditions. A reasonable remote sensing time length is determined by factors such as soil drainage conditions and crop root depth. This time length should be sufficient to fully cover the entire dynamic process of salt leaching downwards under rainfall and then gradually returning upwards with evaporation, typically between 7 days and one month. Finally, the remote sensing time length is divided equally, for example, the total duration is evenly divided into several observation intervals, and the end point of each interval is set as a remote sensing time point. The density of the division depends on the revisit cycle of available satellites and the temporal resolution requirements of the study. This generates a time point sequence evenly distributed from the start point to the end point around the effective rainfall event, which is used to guide the acquisition of multi-temporal remote sensing images.

[0036] It is worth noting that after initially selecting the remote sensing time starting point based on the forecast rainfall exceeding the dissolved rainfall, the stability of the weather window needs to be verified. Before rainfall, it is necessary to verify that no other rainfall events have occurred at least 10 days prior to the forecast rainfall to ensure that the soil water and salt profile at the reference time is in a relatively stable state, unaffected by recent hydraulic disturbances. Only in this way can the measured salinity and moisture spectral indices represent the true background values ​​before rainfall.

[0037] After rainfall, it is necessary to review the subsequent weather conditions within the entire set remote sensing time period (e.g., 7 to 30 days) in the forecast. The core is to ensure that there are no further significant rainfall events within this monitoring window. This is because the core purpose of this scheme is to observe the complete salt transport process driven by a single effective rainfall. If there is rainfall again within the window, the salt change curve will be disturbed or even interrupted in the middle of the decline. The extracted change amplitude and maximum rate will no longer reflect the effect of a single rainfall.

[0038] This method fundamentally changes the traditional model of acquiring remote sensing images, directly linking the start time of remote sensing observations to the physical initiation conditions of salt movement. Only rainfall events that can actually dissolve and move salt will trigger a complete set of subsequent remote sensing observation plans, thus completely eliminating invalid observation events that, despite rainfall, fail to produce effective infiltration. This significantly improves the targeting of remote sensing data acquisition and the efficiency of information utilization. Furthermore, because the remote sensing time start and observation window are uniformly set based on the same set of dissolved rainfall thresholds, the time series of all sample points within the entire study area are strictly comparable and synchronous, providing a unified time benchmark for subsequent construction of water-salt coupling dynamic characteristics, extraction of time-series change curves, and cross-sample point collaborative analysis. Furthermore, by dividing the time sequence into equal parts, the sampling uniformity of the salinity change curve on the time axis is ensured, making the curve fitted by cubic spline interpolation smoother and more reliable. The curve parameters such as the salinity change amplitude, maximum change rate and curve area extracted subsequently can truly reflect the complete dynamic process of salinity migration. Thus, the entire method chain from spectral fusion to dynamic feature screening and deep learning modeling can accurately capture the evolution law of salinity content with water disturbance.

[0039] S102: Based on remote sensing time points, acquire remote sensing images of the inverted area and preprocess them. Extract a set of temporal spectral feature parameters that can characterize the soil state from each pixel of the preprocessed remote sensing image.

[0040] This step effectively transforms the discrete remote sensing images driven by rainfall events, identified in the previous step, into a highly organized and information-rich temporal spectral feature cube. For each geometrically registered remote sensing time point, this step calculates spectral feature parameters covering multiple indicators such as salinity, vegetation, moisture, brightness, and temperature and humidity pixel-by-pixel. Then, the multi-feature images from single time phases are stacked along the feature dimension, and the stacked feature volumes from all remote sensing time points are aligned and stitched together along the time dimension to form a three-dimensional data set of pixel-feature type-time point. This significantly overcomes the dimensional limitations of traditional methods, constructing a spatiotemporally continuous spectral feature space that can characterize soil salinity and simultaneously reflect vegetation response and changes in surface water and heat conditions, providing a rich information foundation for analyzing the dynamic evolution of salinity.

[0041] The moisture index and vegetation index extracted in this step exhibit dramatic responses during rainfall, indirectly carrying traces of vertical salt transport. Furthermore, the stacking of multi-temporal characteristics transforms spectral changes from isolated values ​​into a family of curves evolving over time. This provides direct material for quantitatively capturing the dynamic behavior of salt driven by moisture, advancing remote sensing of soil salinity from simple static content estimation to a dynamic process description. This allows the connection between surface spectra and deep salinity to be indirectly revealed through the evolutionary patterns of the process.

[0042] This step is the core of spectral fusion, responsible for maximizing the extraction and structuring of spectral information at each remote sensing time point. The output time-series spectral feature parameter set serves as the raw material for all subsequent dynamic feature construction. It also provides a broad feature pool for assessing the contribution of rainfall response to S104 and screening interactive features, enabling the screening algorithm to identify the key combinations most helpful for retrieving total salinity from numerous spectral features and their time-series evolution patterns.

[0043] In this embodiment, the preprocessing method is as follows: Radiometric calibration and atmospheric correction are performed sequentially on multi-temporal remote sensing images to eliminate atmospheric effects and restore the true reflectivity of the Earth's surface. Then, orthorectification is performed, followed by pixel-level spatial alignment. The orthorectified remote sensing images are then masked to obtain remote sensing images with unified radiometric and geometric references.

[0044] More specifically, the preprocessing procedure for multi-temporal remote sensing images is as follows: First, radiometric calibration is performed. Using calibration coefficients provided by the sensor manufacturer, the raw digital quantization values ​​of the image are converted into physically meaningful radiance values ​​or apparent reflectance. This unifies the radiometric scale of images acquired on different dates. It can uniformly convert the raw digital quantization values ​​of images acquired on different dates and by different sensors into physically meaningful radiance or apparent reflectance, eliminating the problem of inconsistent dimensions caused by differences in sensor gain settings and quantization levels. This allows the same ground feature to have a comparable radiometric benchmark in images from different time phases, which is the metric premise for the entire time series analysis to be valid.

[0045] Next, atmospheric correction is performed, using mature algorithms such as FLAASH, 6S radiative transfer model or dark target method to eliminate the interference of atmospheric molecular scattering, aerosol extinction and water vapor absorption on the surface reflection signal pixel by pixel, and further restore the apparent reflectance to the true surface reflectance, so as to ensure that the same ground object has comparable spectral response under different times and different atmospheric conditions. Atmospheric correction can eliminate the interference of time-varying factors such as atmospheric molecular scattering, aerosol extinction, and water vapor absorption on the surface reflection signal on a pixel-by-pixel basis, restoring the apparent reflectance to the true surface reflectance. Because atmospheric humidity and aerosol content often differ significantly before and after rainfall, without correction, the temporal changes of moisture index and salinity index will be severely polluted by atmospheric noise. Subsequently, the relative salinity response index and water-salt decoupling index constructed by S103 based on these indices will not be able to truly reflect the soil's own water and salt dynamics, but will instead be mixed with false signals brought about by changes in atmospheric conditions.

[0046] Then, orthorectification is performed. Using a high-precision digital elevation model of the study area and rational polynomial coefficients or a rigorous sensor model of the imagery, pixel geometric displacement and perspective distortion caused by terrain undulations and sensor tilt observations are eliminated. This ensures the entire image possesses the planar geometric accuracy of orthorectified projection, achieving precise pixel-level spatial alignment between images from different time phases and between the imagery and actual ground locations. This scheme requires extracting a complete spectral feature time series at each fixed pixel location. If spatial alignment is not strictly enforced, a pixel may correspond to different points on the ground at different time phases. Consequently, the salinity index variation curve pieced together will no longer reflect the salinity transport process but rather a random walk of surface spatial heterogeneity. Subsequent curve fitting, amplitude, and rate extraction will all lose their physical meaning.

[0047] Finally, the orthorectified remote sensing image is masked. Using the image's built-in band quality label file or through manual thresholding and deep learning cloud detection algorithms, invalid pixel areas such as clouds, cloud shadows, snow cover, and water bodies are identified and removed. Only clear surface pixels are retained for subsequent calculations, ultimately resulting in a clean remote sensing image dataset with consistent radiometrics across multiple time phases and unified spatial geometric references. Masking removes invalid pixels such as clouds, cloud shadows, and snow cover, preventing the spectral values ​​of these abnormal pixels from being mixed into the calculation of spectral feature parameters.

[0048] In this embodiment, the spectral characteristic parameters include: Normalized salinity index, brightness index, vegetation category index, water category index, temperature and humidity index; The method for extracting the spectral feature parameters is as follows: Based on the preprocessed remote sensing images acquired at remote sensing time points, a set of surface reflectance values ​​for different bands corresponding to each pixel are obtained. According to the type of spectral feature parameter, and the corresponding band combination operation rules for each type of spectral feature parameter, pixel-level mathematical operations are performed on all pixels of all remote sensing images at all remote sensing time points. The original, single-band surface reflectance information is transformed into a series of spectral feature parameters that have clear indicative capabilities for soil salinity, vegetation status, surface moisture, and temperature and humidity elements.

[0049] The normalized salinity index is used because it utilizes a specific combination of visible light and near-infrared or short-wave infrared bands to sensitively capture changes in spectral reflectance on the soil surface caused by salt crystallization or salt crust formation. It is a core parameter that directly characterizes the soil salinity state. In the entire technical solution, it is not only a basic variable reflecting the spatiotemporal distribution of salinity, but also the main data source upon which subsequent dynamic characteristics such as salinity dynamic change curves, extraction of relative salinity response index, salinity change amplitude, and maximum rate are relied upon. Its evolution trajectory over time directly depicts the leaching and surface accumulation process of salt under the action of rainfall.

[0050] More specifically, the normalized salinity index (NSI) is calculated as follows: for each pixel, the surface reflectance values ​​in the red and near-infrared bands are extracted separately, and the ratio of the difference between the two values ​​to their sum is calculated. The result is the NSI. This index utilizes the spectral characteristics of saline soils, which have relatively high reflectance in the visible red band and relatively low reflectance in the near-infrared band. By normalizing the ratio, it enhances the salinity signal and suppresses interference from the soil brightness background.

[0051] Since saline-alkali soils often have white salt spots or salt crusts on their surface, their overall reflectivity is significantly higher than that of normal soils. The brightness index can quantify the overall reflectivity of the surface by weighting multiple visible light bands. When the salinity index no longer changes significantly due to spectral saturation caused by high concentrations on the surface, the brightness index can still provide auxiliary information on the degree of salt accumulation on the soil surface, complementing the salinity index and enhancing the ability of the spectral feature set to identify high salinity ranges.

[0052] The brightness index is calculated as follows: for each pixel, the surface reflectance values ​​in the three visible light bands (blue, green, and red) are extracted, and the square root of the sum of the squares of the reflectance values ​​in the three bands is calculated. The result is the brightness index. This index quantifies the overall brightness of the Earth's surface by integrating the total reflected energy in the visible light bands. Salt accumulation, forming salt crusts or salt patches, will significantly increase this value.

[0053] In arid and semi-arid regions, the growth status of crops and natural vegetation is highly correlated with the degree of root soil salinity stress. Salinity inhibits vegetation growth through osmotic stress and ion toxicity, leading to changes in canopy greenness, coverage, and physiological state. Vegetation indexes can sensitively capture these vegetation response signals through comparative calculations of red and near-infrared bands. When the surface salinity spectrum is saturated and cannot directly infer the total amount of deep salinity, the temporal changes in vegetation growth can indirectly convey the dynamic information of root salinity stress levels, providing biological clues for models to infer the deep soil salinity status.

[0054] The vegetation index is calculated as follows: for each pixel, the surface reflectance values ​​in the near-infrared and red bands are extracted separately. The difference between the near-infrared and red band values ​​is calculated and divided by the sum of the two values; the result is the vegetation index. This index utilizes the spectral contrast characteristics of healthy green vegetation—strong reflectance in the near-infrared band and strong absorption in the red band—to enhance the vegetation signal through a normalized ratio. Its value directly reflects vegetation cover and growth vitality, and indirectly indicates the degree of salt stress in the root zone soil.

[0055] Since soil moisture is the direct carrier of salt dissolution and transport, soil moisture content changes drastically before and after rainfall. The water classification index can effectively detect changes in surface and shallow soil moisture content by utilizing the absorption characteristics of near-infrared and short-wave infrared bands. Its time series directly reflects the infiltration depth, duration, and receding rate of water. The subsequent calculation of the water-salt decoupling index and the relative salt response index requires the water index series as the other half of the input for coupling analysis. At the same time, the water index also provides a basis for distinguishing between changes in the salt spectral response caused by salt itself and interference caused by changes in water.

[0056] The water classification index is calculated as follows: for each pixel, the surface reflectance values ​​in the near-infrared and short-wave infrared bands are extracted separately, and the ratio of the difference between the two to their sum is calculated. The result is the water classification index. This index utilizes the spectral difference that liquid water has a strong absorption valley in the short-wave infrared band and relatively weak absorption in the near-infrared band. Through normalized ratio calculation, it sensitively captures changes in the moisture content of the surface and shallow soil.

[0057] Because surface temperature is closely related to soil moisture content and evaporation rate, moist soil cools down through evaporation, while dry saline-alkali surfaces heat up rapidly due to weak evaporation and low heat capacity. The temperature and humidity index, through the thermal infrared band or a combination of near-infrared and thermal infrared bands, can invert surface temperature or thermal inertia information. From the perspective of energy balance, it can help determine the change in soil dryness and wetness and the change in evaporation intensity after rainfall, and supplement the constraints of the invisible thermal infrared dimension for a complete analysis of the dynamic evolution of salinity driven by water.

[0058] The calculation rule for the temperature and humidity index is as follows: for each pixel, the land surface temperature obtained by inversion after radiometric calibration and atmospheric correction of the thermal infrared band is used as the temperature and humidity index. This index reflects the spatiotemporal changes of soil evaporation intensity and dryness / wetness through surface thermal conditions or canopy moisture information.

[0059] After calculating all spectral feature parameters for a single remote sensing time point, all spectral feature images for that time point are stacked according to the feature dimension to form a set of spectral feature parameters for that single remote sensing time point. The above calculation process is repeated for all selected remote sensing time points. Finally, the sets of spectral feature parameters for each remote sensing time point are aligned and stitched together along the time dimension to form a set of spectral feature parameters for time-series features.

[0060] More specifically, using a remote sensing image size of 500 rows × 500 columns, four remote sensing time points were selected, denoted in chronological order as T0, T1, T2, and T3. Each time phase has been preprocessed to obtain an image of the true surface reflectance.

[0061] Taking time T0 as an example, the spectral feature parameters of the preprocessed image at that time phase are calculated pixel by pixel. For any pixel position in the image... The surface reflectance values ​​for each band are obtained. Then, according to the band combination operation rules, five spectral feature parameters for each pixel are calculated: normalized salinity index, brightness index, vegetation type index, water classification index, and temperature and humidity index. This operation is performed on all 500×500 pixels of the entire image, generating a single-band feature image of 500 rows × 500 columns for each spectral feature parameter. A total of five feature images are generated from the five spectral feature parameters. These five feature images at time T0 are stacked along the feature dimension to form a three-dimensional array of dimension [5, 500, 500]. This is the set of spectral feature parameters at time T0. The same band combination operation rules are used for the preprocessed images at times T1, T2, and T3, and the above calculation process is repeated to obtain an independent set of [5, 500, 500] spectral feature parameters for each time. The sets of spectral feature parameters from four remote sensing time points are aligned and stitched together along the time dimension in the order of T0→T1→T2→T3. Each time point set is a three-dimensional array [5, 500, 500]. After the four time points are superimposed in sequence, a four-dimensional array with dimensions [4, 5, 500, 500] is formed, which is the set of spectral feature parameters with temporal characteristics.

[0062] S103: Based on the salinity index and moisture index in the set of spectral characteristic parameters at each remote sensing time point, analyze the dynamic response characteristic parameters of salinity over time, fit the change curve of salinity, extract the curve parameters based on the change curve, and construct a comprehensive feature set from the spectral characteristic parameters, dynamic response characteristic parameters, and curve parameters.

[0063] This step goes beyond simply using multi-temporal remote sensing information as isolated static snapshots. Instead, it abstracts and quantifies a set of dynamic characteristic parameters and curve parameters from the temporal set of spectral feature parameters to specifically describe how salinity responds to changes in moisture. This step constructs a relative salinity response index by defining a baseline and peak time, using the ratio of net salinity change to moisture change to quantitatively answer the direction and sensitivity of salinity's response to moisture disturbances. Secondly, it innovatively proposes a water-salt decoupling index by introducing a dynamic time warping algorithm. By measuring the alignment of the salinity and moisture sequences on the time axis, it quantitatively answers the extent to which salinity dynamics are decoupled from purely hydraulic drive. Furthermore, it fits discrete salinity index points into a continuous, smooth curve using cubic spline interpolation, extracting three curve parameters: salinity change amplitude, maximum rate of change, and area under the curve. These parameters comprehensively characterize the entire process of salinity transport in a single rainfall event from the perspectives of total displacement scale, instantaneous intensity, and cumulative deviation.

[0064] The comprehensive feature set constructed in this step expands the information type from "what" to "how it changes." Even when the surface salinity index tends to saturate at high concentrations and no longer fluctuates significantly numerically, its temporal response pattern relative to water changes—for example, whether it responds with a lag or changes immediately and synchronously with increasing water, and whether the change curve is steep and then gradually declines or gradually rises and then steeply declines—still carries unique information about deep salinity reserves and soil matrix characteristics. The relative salinity response index directly indicates the direction of salinity movement, the water-salt decoupling index reveals whether salinity changes are influenced by other non-hydraulic factors, and the curve parameters supplement the scale and rhythm of the changes. This series of dynamic features together constitutes an incremental dimension of information, enabling the indirect inference of the total amount of salt in the deep soil from the dynamic behavior pattern of salinity even under conditions of saturated surface spectral signals.

[0065] From the perspective of its contribution to the overall technical solution, this step plays a crucial pivotal role in elevating the results of "spectral fusion" to the capability of "process analysis." The time-series spectral feature parameter set output by S102 is given a physical process meaning in this step. The feature values, originally arranged only by time, are transformed into dynamic response indicators with clear physical significance. This provides a truly discriminative pool of candidate features for feature selection in S104—the assessment of rainfall response contribution and the calculation of synergistic gain are based on the interpretative power of these dynamic features for total salinity, rather than remaining at the level of the original spectral indices.

[0066] In this embodiment, the dynamic response characteristic parameters include the relative salinity response index and the water-salt decoupling index; The process of obtaining the relative salinity response index is as follows: Using the remote sensing time points corresponding to the salinity and moisture indices before rainfall as the baseline time, and based on the salinity and moisture indices corresponding to adjacent remote sensing time points after rainfall, the change values ​​of salinity and moisture indices are obtained according to the time series. The remote sensing time point corresponding to the maximum value of the change values ​​of salinity and moisture indices is taken as the peak time. The difference between the salinity index at the peak time and the salinity index at the baseline time is taken as the net change in salinity. The absolute value of the difference between the moisture index at the peak time and the moisture index at the baseline time is taken as the change in moisture. A constant is introduced, and the ratio obtained by dividing the net change in salinity by the sum of the change in moisture and the constant is the relative salinity response index. The relative salinity response index does not simply compare changes in salinity itself, but rather standardizes salinity changes within a reference frame of moisture changes, thereby extracting the pure response sensitivity of salinity to moisture disturbances. By using a ratio form, it eliminates the dimensional influence caused by differences in soil background salinity at different sampling points, allowing for horizontal comparison of response intensity under different salinity backgrounds. The directional information of salinity changes is represented by the relative salinity response index, enabling a single index to simultaneously answer two key questions: "In which direction does salinity move?" and "How sensitive is salinity to moisture?" The introduction of a constant term solves the problem of ratio instability common in arid regions when the denominator is too small or even zero, ensuring the index's numerical stability and computational feasibility under any moisture conditions.

[0067] The reference time is selected as the salinity index and moisture index at the last remote sensing time point before the rainfall occurs. Its significance lies in establishing the background state before the disturbance. This is the reference origin for measuring all subsequent changes. Only when the soil water and salt profile at the reference time is in a relatively stable state can the subsequent changes be attributed to the rainfall event.

[0068] The net change in salinity refers to the difference between the salinity index at the peak time and the baseline time. It quantifies the net change in the salinity spectral signal caused by the rainfall at the maximum response level. A positive value indicates that the salinity index increases, that is, the upward accumulation of salt is enhanced, while a negative value indicates that the salinity index decreases, that is, the salt is leached downward.

[0069] The moisture change value refers to the absolute value of the difference between the peak moisture index and the baseline moisture index. It quantifies the magnitude of change in soil moisture conditions from the background state to the peak of the salinity response. An absolute value is used because the scale of change, whether moisture increases or decreases, constitutes the driving force of the salinity response. The purpose of introducing a constant is to prevent the ratio from approaching infinity and losing computational stability when rainfall is too low or infiltration is weak, causing the moisture change value to approach zero. This constant is usually a very small positive number, and its specific value can be determined based on the typical range of the moisture index. For example, when the normalized moisture index fluctuates between 0 and 1, the constant can be set between 0.001 and 0.01, making it much smaller than the normal moisture change value but not zero.

[0070] The ratio obtained by dividing the net change in salinity by the sum of the change in moisture and the constant is the relative salinity response index. The essence of this division operation is to normalize the salinity response amplitude with respect to the moisture driving intensity, so as to obtain the amount of salinity change caused by a unit change in moisture, thereby allowing the salinity response under different rainfall and soil types to be compared on the same scale.

[0071] When the relative salinity response index is positive, it indicates that the net change in salinity is positive, meaning that the salinity index increases. This means that salt accumulates on the surface after rainfall, which usually occurs when the rainfall is small and only wets the surface layer, and strong evaporation after the rain pulls deep salt to the surface. When the index is negative, it indicates that the net change in salinity is negative, meaning that the salinity index decreases. This means that salt is leached downwards, which usually occurs when the rainfall is sufficient and infiltration water carries salt from the surface to the deeper layers.

[0072] The magnitude of the relative salinity response index reflects the sensitivity of soil salinity to water disturbance. The larger the value, the more intense the salinity spectral response caused by a unit change in water, and the more easily the soil salinity is migrated by hydraulic drive. The smaller the value and the closer it is to zero, the more stable the salinity spectral signal is even if there is a significant change in water, which usually means that the soil salinity is firmly fixed in the soil matrix or has reached a dynamic equilibrium.

[0073] The process for obtaining the water-salt decoupling index is as follows: Based on the salinity index and moisture index in the spectral feature parameter set, they are sorted according to time series to form salinity index series and moisture index series, respectively. They are then normalized and scaled to the same interval. A dynamic time warping algorithm is used to find the alignment path with the minimum cumulative distance between the two normalized sequences by bending, stretching or compressing the time axis, and this shortest cumulative distance is recorded. The ratio obtained by dividing the shortest cumulative distance by the actual step length of the warped path is the water-salt decoupling index.

[0074] The dynamic time warping algorithm, rather than simple Pearson correlation coefficients or Euclidean distance, is used to obtain the water-salt decoupling index because the response of salinity changes to moisture changes often exhibits a time lag or misalignment. After rainfall infiltration, the moisture index typically responds immediately, while the dissolution, migration, and subsequent surface accumulation of salt require time. This results in the peaks and troughs of the two curves not being strictly aligned on the time axis. Traditional correlation analysis requires point-to-point correspondence, which underestimates the true water-salt coupling strength due to this time lag. The advantage of the dynamic time warping algorithm lies in its ability to allow for local bending and stretching of the time axis, enabling intelligent matching of a lagging response point on the salinity curve with an earlier driving point on the moisture curve, thereby capturing coupling relationships missed by traditional methods. This allows the water-salt decoupling index to more accurately reflect the intrinsic correlation between the two in the dynamic process.

[0075] The water-salt decoupling index represents the degree to which salt dynamics are decoupled from purely hydraulic driving forces. Physically, it represents the minimum morphological difference remaining between the time-varying curves of the salt index and the water index after eliminating time lag effects. A lower value indicates stronger synchronicity and tighter coupling between the time-varying curves of salt and water, meaning that the dynamic changes in salt are almost entirely driven by water changes. A higher value indicates a greater difference in the trends of the two curves; even considering time lag effects, the change patterns of salt and water are still significantly different. This implies that other non-hydraulic factors, such as strong adsorption by the soil matrix, chemical precipitation-dissolution equilibrium, or salt absorption by vegetation, are dominating salt dynamics, and the degree to which salt is decoupled from purely hydraulic driving forces is higher.

[0076] The Dynamic Time Warping (RTW) algorithm is used to measure the similarity of two time series of potentially unequal lengths. It finds an optimal alignment path that minimizes the cumulative difference between the two series along this path. Specifically, the process involves constructing a distance matrix with the normalized salinity index and moisture index sequences as two dimensions, representing the number of rows and columns of the two sequences, respectively. Each element in the matrix is ​​the absolute value of the difference between the values ​​at corresponding time points in the two sequences, or the Euclidean distance. Then, starting from the bottom left corner of the matrix, dynamic programming is used to progressively calculate the minimum cumulative distance to each grid point. Each move allows for three directions: right, up, or diagonally to the upper right, corresponding to stretching the moisture sequence, stretching the salinity sequence, or synchronizing the time of both sequences, respectively. The minimum cumulative distance to the top right corner of the matrix is ​​the shortest cumulative distance between the two sequences.

[0077] The shortest cumulative distance represents the sum of morphological differences between the salinity index and moisture index sequences throughout the entire process, after allowing for optimal scaling and alignment of the time axis. It is not simply the sum of distances between time points, but rather the globally minimum morphological difference obtained after finding the best matching path. Dividing this shortest cumulative distance by the actual step length of the normalized path normalizes the path length, preventing longer sequences from naturally having larger cumulative distances due to containing more time points. The resulting ratio is the water-salt decoupling index, whose value represents the average morphological difference between the two sequences per unit step length; the smaller the difference, the tighter the coupling, and the larger the difference, the higher the degree of decoupling.

[0078] Please see Figure 2 (a)- Figure 2 (d) are respectively the time series variation diagram of salinity and moisture index provided in the embodiments of this application, the relative salinity response index of different rainfall events provided in the embodiments of this application, the dynamic time warping (DTW) alignment analysis diagram provided in the embodiments of this application, and the water-salt decoupling index comparison diagram of different soil types provided in the embodiments of this application; Figure 2 (a) shows the temporal changes in salinity index (NSI) and moisture index (WI) after a rainfall event. The following can be observed from the figure: After rainfall (T0→T2), the moisture index rose rapidly from 0.30 to a peak of 0.60, with a response delay of about 1-2 time steps. The salinity index responded subsequently, rising from 0.35 to a peak of 0.55, with a response delay of about 2-3 time steps, showing a significant hysteresis effect. After the peak (T2→T6), both the moisture and salinity indices gradually recovered to pre-rainfall levels, but the salinity recovered more slowly than the moisture. This time-delay response pattern is the physical basis for the dynamic response characteristics proposed in this scheme—the response of salinity changes to moisture changes has a time lag or misalignment, and traditional static analysis methods cannot capture this dynamic process.

[0079] Figure 2 (b) Shows the Relative Salinity Response Index (RSRI) values ​​for different rainfall events. The RSRI, by standardizing salinity changes against a reference frame of moisture changes, extracts the pure response sensitivity of salinity to moisture disturbances. As can be seen from the figure: When RSRI is positive (blue bar), it indicates that salts accumulate upwards to the surface after rainfall, which usually occurs when rainfall is small and evaporation is strong after rain. When RSRI is negative (orange bar), it indicates that salts are leached downwards, which usually occurs when rainfall is sufficient and infiltration water carries salts from the surface to deeper layers. The value of RSRI reflects the sensitivity of soil salinity to water disturbance. The larger the value, the more intense the spectral response of salts caused by a unit change in water. RSRI can characterize the direction of salt movement and the sensitivity of salts to water disturbance, and can provide a basis for judging the degree of soil salinization risk and assessing irrigation leaching efficiency.

[0080] Figure 2 (c) shows the alignment analysis of the salinity and moisture sequences by the Dynamic Time Warping (DTW) algorithm. The physical meaning of the water-salt decoupling index (WSI) is the minimum morphological difference that remains between the time variation curves of the salinity index and the time variation curves of the moisture index after the time lag effect has been eliminated. Figure 2 (c) The gray line represents the optimal alignment path found by the DTW algorithm. By bending, stretching or compressing the time axis, the lagging response point on the salinity curve is intelligently matched with the earlier driving point on the moisture curve. When WSI=0.23, it means that there is still a certain degree of morphological difference between the two curves after optimal alignment, indicating that salinity dynamics are not entirely driven by moisture, but are also affected by non-hydraulic factors such as soil matrix adsorption and chemical precipitation dissolution equilibrium.

[0081] Figure 2 (d) shows a comparison of the water-salt decoupling index (WSI) for different soil types, and compares the WSI values ​​of four soil types. Sandy soil has the lowest WSI (0.15), indicating that its water-salt coupling is the tightest and its salinity dynamics are almost entirely driven by water. Saline-alkali soil has the highest WSI (0.48), indicating that its salinity dynamics are affected by multiple non-hydraulic factors and its decoupling degree is the highest. This difference provides a basis for selecting salinity inversion strategies for different soil types.

[0082] In this embodiment, the curve parameters include the salinity change range, the maximum rate of change, and the curve area; The method for fitting the curve is as follows: Based on the salinity index in the set of spectral feature parameters, the parameters are sorted according to time series and fitted using cubic spline interpolation to obtain a salinity change curve that is continuous and smooth in time, which is the change curve. The method for extracting the range of salinity changes is as follows: Based on the change curve, find the highest and lowest points on the entire change curve, and subtract the salinity index of the lowest point from the salinity index of the highest point. The difference is the salinity change range. The method for extracting the maximum rate of change is as follows: Based on the change curve, calculate the first derivative of the change curve, and the maximum absolute value of the first derivative is the maximum rate of change. The method for extracting the area of ​​a curve is as follows: Using the salinity index before rainfall as a baseline, the area of ​​the region enclosed between the entire change curve and this baseline is the area of ​​the curve. Curve parameters are obtained through curve fitting, rather than directly using the raw salinity index values ​​at discrete time points. This is because discrete observation points can only reflect the state of salinity at a few specific moments, failing to depict the complete trajectory and dynamic rhythm of salinity's continuous evolution throughout the entire rainfall event. The advantage of cubic spline interpolation is that it can generate a smooth curve with continuous first and second derivatives using piecewise low-order polynomials, ensuring the curve passes through all known observation points. This avoids the space oscillations that may occur with high-order polynomial interpolation and can reconstruct the gradual change process of salinity between two adjacent remote sensing time points with high mathematical accuracy. This continuous and smooth curve transforms the discrete "snapshot" sequence into a time-differentiable function, making it possible to further extract rate and cumulative parameters. This provides much richer dynamic process information for subsequent feature selection and model learning than the original discrete points.

[0083] More specifically, cubic spline interpolation is a numerical analysis method used to construct a smooth and continuous function curve that passes through all known points, given a set of discrete data points. Its core idea is to fit the intervals between every two adjacent data points using piecewise cubic polynomials. These polynomials must satisfy the condition of continuity of the first and second derivatives at the connection points of adjacent intervals, thus ensuring that the entire curve is visually and mathematically smooth without abrupt changes. Compared to simple linear interpolation, cubic splines do not produce broken lines or sharp corners. Compared to overall interpolation using higher-order polynomials, it effectively avoids the Runge phenomenon—the violent oscillations of the interpolation polynomial at the interval edges. Therefore, it is particularly suitable for reconstructing the continuous and gradual change in salinity between several discrete remote sensing observation points.

[0084] The specific process of fitting using cubic spline interpolation is as follows: First, the existing remote sensing time points are arranged in chronological order, with time as the independent variable and the corresponding normalized salinity index as the dependent variable, resulting in a set of ordered discrete data points. Then, a cubic polynomial is constructed between every two adjacent time points. Each polynomial is a cubic function of the time variable within that interval, containing four undetermined coefficients. Since there are multiple intervals, the entire curve consists of multiple piecewise cubic polynomials, and the total number of coefficients for all intervals needs to be determined. The solution for these coefficients relies on three types of constraints: the first is the interpolation condition, which requires that the function value of each piecewise polynomial at the endpoint of the interval must be equal to the known salt index observation value corresponding to that endpoint, ensuring that the curve passes through all known points; the second is the continuity condition, which requires that at the connection point of adjacent piecewise polynomials, the first derivative and the second derivative of the preceding and following segments are equal, ensuring a smooth connection of the entire curve without sharp corners; the third is the boundary condition, which usually adopts a natural boundary condition, that is, making the second derivative of the entire curve zero at the start and end points, which is equivalent to allowing the ends of the curve to relax naturally without additional bending. By simulating all the above constraints into a system of linear equations, solving for all the coefficients of each piecewise polynomial, and substituting them back into the piecewise expressions, the complete analytical form of the entire changing curve can be obtained. Based on this, the interpolated estimate of the salt index at any time can be calculated, and further differentiation or integration can be performed to extract curve parameters such as the maximum rate of change and the area of ​​the curve.

[0085] Choosing the magnitude of salinity change as a curve parameter directly reflects the overall scale of salinity change caused by rainfall. It represents the difference between the highest and lowest salinity index points on the entire curve. A larger value indicates a greater overall displacement of soil salinity driven by water during this rainfall event. Whether salinity is leached downwards through infiltration or accumulates upwards through evaporation, the total magnitude of the displacement is summarized in this single value. This parameter provides the model with a one-dimensional summary index of the dynamic "magnitude" of salinity, enabling the model to quickly determine whether the soil represented by that pixel has undergone a significant salinity redistribution process.

[0086] The maximum rate of change captures the most dramatic instantaneous velocity during salinity changes. It represents the maximum absolute value of the first derivative of the change curve, physically corresponding to the peak leaching velocity of infiltration water rapidly carrying salt downwards immediately after rainfall ends, or the peak velocity of salt rapidly accumulating on the surface with capillary water during the strong evaporation phase after rain. This parameter helps identify sudden and dramatic events in salt transport. A high maximum rate of change often indicates the presence of a large amount of easily soluble and migratable salts in the soil, along with good soil hydraulic conductivity. This is a significant indicator for assessing the degree of soil salinization risk and irrigation leaching efficiency.

[0087] The area under the curve measures the cumulative deviation of salinity from its pre-rainfall baseline state over the entire monitoring period. It means that the integral area enclosed by the salinity index before rainfall is used as a baseline. This parameter comprehensively reflects the total duration and total deviation of salinity from its initial equilibrium state within the entire observation window after a rainfall event. A larger area indicates a longer time and more complex path for the salinity system to recover to its original state after being disturbed by rainfall. This may suggest the existence of a buffering mechanism in the soil salinity or a stepwise dissolution and precipitation process of multiple forms of salt, providing a comprehensive indicator for distinguishing the salinity dynamics of different soils.

[0088] Please see Figure 3 (a)- Figure 3 (d) are respectively a schematic diagram of the fitting of the difference of cubic splines of salinity index provided in the embodiments of this application, an analysis diagram of the rate of change of salinity provided in the embodiments of this application, a schematic diagram of the calculation of curve area provided in the embodiments of this application, and a comparison diagram of curve parameters of different rainfall events provided in the embodiments of this application.

[0089] Figure 3 (a) shows the fitting effect of cubic spline interpolation on discrete salinity index observation points. As can be seen from the figure: The fitted curve (blue solid line) passes through all observation points (red dots), satisfies the interpolation conditions, and the curve is smooth and continuous as a whole, without obvious oscillations or abrupt changes. This verifies the superiority of cubic spline interpolation in reconstructing the continuous gradual change process of salinity. The highest and lowest points are accurately identified, providing a reliable basis for calculating the magnitude of salinity changes. Compared with simple linear interpolation, cubic splines do not produce broken line corners. Compared with high-order polynomial global interpolation, it can effectively avoid Runge phenomenon and is particularly suitable for reconstructing the continuous gradual change process of salinity between several discrete remote sensing observation points.

[0090] Figure 3 (b) shows the first derivative of the salinity change curve, i.e., the rate of salinity change. It can be observed from the figure that: The rate of change is positive at the beginning of rainfall, indicating that the salinity index is increasing (salt accumulates upwards). The maximum rate of change occurs during the peak of the rainfall response, corresponding to the moment when salt migration is most intense. The rate of change then turns negative, indicating that salt begins to be leached and migrate downwards. The maximum rate of change parameter can identify sudden and intense events in the salt migration process. A very high maximum rate of change often means that there is a large amount of easily soluble and migratable salt in the soil, and that the soil has good hydraulic conductivity.

[0091] Figure 3(c) This demonstrates the method for calculating the area under the curve—using the pre-rainfall salinity index as a baseline, the area enclosed by the curve and the baseline is calculated. Positive deviation (red area) indicates salt accumulation, while negative deviation (blue area) indicates salt leaching. The area under the curve = 0.243 comprehensively reflects the cumulative deviation of salinity from its pre-rainfall background state over the entire monitoring period. A larger area indicates a longer recovery time and a more complex path for the salinity system after rainfall disturbance.

[0092] Figure 3 (d) The curve parameters under four different rainfall intensities were compared. The graph shows that: As rainfall intensity increased from light rain to heavy rain, the magnitude of salinity change, the maximum rate of change, and the area under the curve all showed a monotonically increasing trend. Under heavy rain events, the magnitude of salinity change reached 0.35, the maximum rate was 0.32, and the area under the curve was 0.58, all of which were significantly higher than those under light rain events. This verifies that the curve parameters can effectively distinguish the degree of influence of different rainfall intensities on salinity transport, providing a powerful tool for the quantitative characterization of salinity dynamics.

[0093] S104: Pair the comprehensive feature set of each sampling point with the corresponding total salt content one by one. For each feature in the comprehensive feature set, calculate its univariate regression determination coefficient with the total salt content, the process response degree corresponding to the change before and after rainfall, and the average cooperative gain value after the feature is combined with other features. Normalize and weight the determination coefficient, process response degree and average cooperative gain value to obtain the contribution ability of the feature. Sort and filter each feature according to its contribution ability to construct a feature subset.

[0094] This step elevates the screening criteria from traditional static statistical significance to a dual evaluation level of dynamic process contribution and synergistic gain between features. The innovation of this step lies in constructing two entirely new evaluation dimensions: first, the contribution of rainfall response, which is obtained by weighting the feature's overall ability to characterize the salinity-rainfall response by combining the static coefficient of determination of each feature with total salinity and the process response of that feature to changes before and after rainfall and changes in measured salinity; second, the synergistic gain value, which precisely quantifies the additional information increment brought about by the interaction of two features by comparing the explanatory power of two features individually in a univariate regression of total salinity with the explanatory power of their combined binary regression. Combining these two factors and assigning weights forms the final ranking of contribution capabilities. This design ensures that the selected features are not only spatially correlated with total salinity but also sensitively follow salinity dynamics over time during rainfall events, and that the synergistic interaction between features further amplifies their explanatory power for salinity evolution.

[0095] The comprehensive feature set constructed in S103 includes various types of high-dimensional features such as spectral feature parameters, dynamic response feature parameters, and curve parameters. Directly inputting all of these into the model would not only cause the curse of dimensionality and the risk of overfitting, but more importantly, many of these features may only carry noise or have no practical indicative meaning in the dynamic process. Traditional static screening methods tend to ignore key variables that, when individually weakly linearly related to total salinity, exhibit unique dynamic response patterns during rainfall and whose explanatory power is multiplied when combined with other features. The two indicators specifically designed in this step—rainfall response contribution and synergistic gain value—can precisely uncover these hidden, effective features. This allows the final feature subset to retain or even enhance the comprehensive expressive power of total salinity and dynamic processes with a smaller number of variables, providing feature-level assurance for the accuracy and generalization ability of the subsequent model.

[0096] From the perspective of its contribution to the overall technical solution, S104 plays a core role in feature space optimization and information purification. It inherits the comprehensive feature set output by S103, filtering out the features that truly explain the dynamic evolution of salinity, eliminating redundant and noisy features, and providing a refined and efficient feature subset for the deep learning model of S105. The input of this feature subset allows the temporal attention mechanism and multi-task learning model to focus on the most valuable dynamic information channels, avoiding distractions on irrelevant features, thereby improving the model's prediction accuracy and training efficiency for salinity content and dynamic response features.

[0097] In this embodiment, the analysis process for the contribution capability is as follows: Based on the measured total salt content in the soil at all sampling points, the comprehensive feature set of each sampling point is paired with the corresponding total salt content one by one. The univariate regression determination coefficient between each feature in the comprehensive feature set and the total salt content is calculated to obtain the determination coefficient between the feature and the total salt content. The average change of this characteristic at all sampling points before and after rainfall, as well as the average change of the total salt content at the corresponding sampling points, were analyzed. The relative deviation between the two changes was obtained, and the process response was obtained by subtracting the relative deviation from 1. Based on any two features in the comprehensive feature set, perform univariate regression on the total salt content to obtain the explanatory power of the two individual features. Perform binary regression on the measured total salt content using the two features together to obtain the joint explanatory power. Subtract the sum of the explanatory powers of the two individual features from the joint explanatory power to obtain the collaborative gain value of the features. The contribution capability of this feature is obtained by normalizing the coefficient of determination, process responsiveness, and synergistic gain value and then fusing them according to preset weights.

[0098] More specifically, the specific implementation steps are as follows: Using the total salt content measured in the soil at all sampling points, a univariate linear regression was performed on each feature in the comprehensive feature set to obtain the coefficient of determination between the feature and the total salt content. The higher the value of the coefficient of determination, the stronger the static statistical correlation between the feature and the salt content. That is, for the first in the comprehensive feature set Features , with all The characteristic value of each sampling point is the independent variable, with the total salinity of the soil at the corresponding sampling point being the measured value. Using univariate linear regression as the dependent variable, a regression model is obtained:

[0099] This characteristic is related to the coefficient of determination of total salt content. The calculation formula is:

[0100] in, Features The predicted value, Features The intercept, Features The regression coefficients, Features The coefficient of determination, For the first The measured total salt content at each sampling point For this sampling point based on features The predicted value, This represents the average total salt content across all sampling points.

[0101] The coefficient of determination (COD) represents the static explanatory power of a feature in the spatial dimension. It measures the extent to which a single feature value directly corresponds to the level of total soil salinity. As a measure of spatial correlation, a higher COD indicates a stronger ability of the feature to map the spatial distribution of salinity, and a stronger static statistical correlation between the feature and total soil salinity. A lower COD indicates a weaker correlation, ensuring that the selected features have a solid statistical basis and are the spatial components contributing to the overall capability.

[0102] Calculate the average change of this feature at all sampling points before and after rainfall, and the average change of the total salt content at the corresponding sampling points. Then calculate the relative deviation between the two changes. Subtract the relative deviation from 1 to obtain the process response. Calculate the number of sampling points before and after rainfall. The average change of each feature :

[0103] in, For the first The sampling point The values ​​of each feature before rainfall occurs. For the first The sampling point The values ​​of each feature after rainfall occurs. It is a positive integer. .

[0104] Secondly, calculate the average change in total salt content corresponding to the sampling points. :

[0105] in, The first Total salt content measured at each sampling point before and after rainfall; Then, calculate the relative deviation between the two changes. ;

[0106] in, It is a very small positive number. To prevent the denominator from being zero, the smaller the relative deviation value, the more consistent the change in the feature is with the measured change in salinity, and the stronger the feature's ability to track the true changes in salinity driven by rainfall.

[0107] Finally, subtract this relative deviation from 1 to obtain the process response degree. :

[0108] in, For the first The process responsiveness of each feature The closer the value is to 1, the more synchronous the dynamic changes of this feature are with the measured changes in salinity during the rainfall process. If it is negative, it means that the direction of the feature change is opposite to or seriously deviates from the direction of the measured changes in salinity, and the feature lacks indicative ability in the process dimension.

[0109] The process response represents the dynamic following ability of a feature in the time dimension. By comparing the consistency between the feature changes before and after rainfall and the measured changes in salinity, it measures whether the feature can keenly capture the dynamic migration process of salinity. As a benchmark for dynamic processes, the larger the value, the more accurate the feature's characterization of salinity migration. It makes up for the deficiency of the coefficient of determination, which only considers static conditions, and ensures that the features selected later can reflect how salinity changes. It is a dynamic component that contributes to the ability.

[0110] Based on any two features in the comprehensive feature set, perform univariate regression on the total salt content to obtain the explanatory power of the two individual features. Perform binary regression on the measured total salt content using the two features together to obtain the joint explanatory power. Subtract the sum of the explanatory powers of the two individual features from the joint explanatory power to obtain the paired synergistic gain value of the features. The mean of the paired synergistic gain values ​​is the synergistic gain value. The contribution capability is obtained by weighted fusion of the coefficient of determination, process responsiveness, and synergistic gain value. For any two distinct features in the comprehensive feature set and ( The following regression analyses were performed respectively: Performing univariate linear regressions on the total salt content separately yields the determination coefficients of the two single features. and These two values ​​represent the single feature explanation rate.

[0111] Combine the two features and perform a binary linear regression on the total salt content:

[0112] in, Features and Calculate the determination coefficient of the binary regression using the joint predicted values. This refers to the joint explanatory power, which also ranges from [0,1]. Calculate the cooperative gain value :

[0113] when When the two features are used together, it indicates a positive synergistic effect, and the interaction between them can explain more of the total salt content variation than the sum of their individual explanations; when When the two features are independent, their combined use yields no additional gain; when... If this occurs, it indicates that the two features are redundant or even interfere with each other, and using them together is not as effective as using them alone.

[0114] For each feature The synergistic gain value of this feature is averaged by calculating the synergistic gain values ​​obtained by pairwise pairing it with all other features. :

[0115] in, The total number of features in the comprehensive feature set. The larger the value, the stronger the synergistic effect of the feature with other features; Synergy gain represents the information complementarity of a feature in the interaction dimension. It measures whether the feature, when combined with other features, can generate an additional information increment that is "greater than the sum of its parts." As a benchmark for interaction value, the larger the value, the better the feature is at complementing other features, and the greater the accuracy of the inversion when combined. It is used to discover "golden partners" that are not strong individually but have outstanding combined effects, and is an interaction component that contributes to the overall capability.

[0116] The contribution capability is obtained by weighting and integrating the coefficient of determination, process responsiveness, and synergistic gain value. ;

[0117] in, For the first The contribution capacity of each feature The weighting coefficients used to determine the coefficients can be adjusted according to the actual application scenario; they can be increased when spatial distribution inversion accuracy is emphasized. , This is the weighting coefficient for process responsiveness. If more emphasis is placed on the ability to analyze the dynamic process of salinity, it can be increased. , The weighting coefficients for the collaborative gain value.

[0118] Contribution capacity is the weighted sum of the above three factors, representing a comprehensive evaluation of the characteristics in salt inversion, and serving as the sole ranking criterion for final selection. This is achieved by adjusting the weighting coefficients ( , , The screening criteria can be flexibly customized according to the application scenario. For example, if the main goal is to obtain a high-precision salinity distribution map, the weight of the coefficient of determination can be increased. If the focus is on diagnosing and analyzing the dynamic changes in salinity, then the weight of process responsiveness can be increased. Analysis revealed that many features, when viewed individually, have low coefficients of determination, but when combined, they have extremely strong explanatory power. Alternatively, when the soil salinity in the study area is complex and a single indicator is insufficient to fully characterize it, increasing the coefficient of determination may be beneficial. This total score ensures that the final selected feature subset is a group of elite feature combinations that understand spatial distribution, dynamic evolution, and can perfectly characterize the entire salinization process through interactive cooperation.

[0119] In this embodiment, the steps for prioritizing each feature based on its contribution capability and constructing a feature subset are as follows: Based on the contribution energy values, the features are sorted in descending order to form an ordered sequence. Using the sorted sequence number as the horizontal axis and the contribution energy value corresponding to each feature as the vertical axis, a curve is plotted in a planar coordinate system where the contribution energy decreases as the number of features increases. The inflection point of the transition region between the steep and flat sections of the curve is found, and all features whose contribution energy value is greater than the contribution energy value corresponding to this inflection point are selected to construct a feature subset.

[0120] More specifically, the elbow rule is used to construct feature subsets because it provides an objective, intuitive, and efficient means to automatically identify the optimal boundary between effective and redundant features based on the natural distribution of feature contribution capabilities.

[0121] The contribution capability value calculated by S104 is a comprehensive index that integrates static correlation, process response, and synergistic gain. After all features are sorted by this index from high to low, their values ​​usually exhibit a clear non-uniform distribution: a few features with high contribution capabilities and significant differences between them form the steep drop section of the curve; while a large number of features with low contribution capabilities and similar values ​​form the gentle section of the curve. In the transition region between these two sections, there is an elbow-like inflection point. This inflection point naturally divides the feature set into two categories: features before the inflection point have high contribution and information content, belonging to core effective features; features after the inflection point have weak and stable contribution capabilities, belonging to redundant or noisy features that can be eliminated. When there are multiple elbow-like inflection points, the point with the most drastic bend is selected as the rejection boundary. This point represents that before this point, the feature contribution capability drops sharply with the increase of the sorting number, while once this point is crossed, the downward trend of the feature contribution capability becomes very smooth and tends to stabilize. This cliff-like abrupt change in slope clearly separates high-contribution features from redundant / noisy features into two independent intervals. The elbow rule determines the screening threshold by capturing this natural breakpoint, avoiding the arbitrariness of setting fixed values ​​or quantities by humans, and making the construction process of feature subsets data-driven and repeatable.

[0122] The elbow rule for constructing feature subsets maximizes the retention of core features that substantially contribute to the inversion of total salinity and the analysis of dynamic processes, while eliminating redundant and noisy features to the greatest extent possible. This effectively compresses the size of the feature subset, which is particularly crucial for the high-dimensional small sample problem with a limited number of field sampling points in this scheme. It can significantly reduce the risk of overfitting of the S105 deep learning model. Since the inflection point is determined by observing the morphological changes of the overall contribution capability curve, the screening results have intuitive physical interpretability. Each retained or eliminated feature can be clearly located on the sorting curve. Thirdly, this method does not rely on preset threshold parameters and has the ability to adapt to the overall distribution differences of contribution capability in different study areas and under different data conditions, enhancing the robustness and universality of the entire method system.

[0123] S105: Construct a salinity inversion model and train it. Use a subset of features as input to the trained model and output the spatial distribution data of soil salinity in real time.

[0124] This step designed a salinity inversion model specifically adapted to the dynamic response process of salinity precipitation. By organically integrating a bidirectional long short-term memory network with a perceptual attention mechanism and a multi-task learning framework, the model can not only automatically learn the evolution pattern of salinity over time from the optimized feature subset, but also autonomously identify which key moments in the entire time series are most important for the final inversion, and simultaneously complete the two tasks of total salinity prediction and dynamic response feature output. Traditional deep learning applications in salinity remote sensing mainly rely on convolutional neural networks or shallow fully connected networks. Their inputs are usually simple stacks of spectral indices from single or multiple time phases. These models treat features from different times as isolated static slices, failing to capture the dependencies and evolution rhythm of salinity changes on the time axis, and even more so failing to distinguish the information weight differences between key time points and ordinary time points. As a result, the model's ability to infer deep salinity is severely limited by the saturation effect of surface spectral signals.

[0125] Compared to existing technologies, bidirectional long short-term memory networks simultaneously encode feature sequences along the forward and reverse time axes, enabling the model to understand the complete causal chain of salinity from its pre-rainfall background state to its post-rainfall leaching and re-aggregation. This temporal modeling capability is not possessed by traditional convolutional neural networks or ordinary fully connected networks. The perceptual attention layer, through two layers of linear transformation and nonlinear activation functions, endows the model with the ability to autonomously judge the importance of information at different remote sensing time points. This allows the model to automatically focus attention weights on key moments of most dramatic salinity changes, such as peak moments, thereby suppressing responses to noise signals during stable periods. This adaptive temporal focusing mechanism significantly improves the efficiency of extracting key information on the dynamic evolution of salinity. The multi-task learning output layer enables the model to simultaneously output a salinity dynamic response feature map while predicting the total salinity. The two tasks share the underlying temporal encoding and attention representation, which not only improves the generalization accuracy of each task but also gives the model the ability to directly diagnose the salinity change process. The output result is no longer a single total salinity value or distribution map, but a complete product containing both total amount information and process information.

[0126] S105 is the final output and integration verification stage of the entire methodology. S101 provides the model with high-quality event-driven time-series observation data; S102 and S103 construct a comprehensive feature space containing information on the dynamic evolution of salinity; S104 selects the most refined and efficient feature subset for the model; and S105 fully mines and utilizes the results of these previous works through its unique model architecture. The temporal encoding capability of the bidirectional long short-term memory network is precisely for analyzing the evolution of the dynamic response index and curve parameters constructed in S103 on the time axis. The adaptive focusing capability of the attention mechanism and the selection logic of rainfall response contribution and synergistic gain value in S104 form an internally and externally complementary information purification mechanism. The design of multi-task learning makes the dynamic response features extracted by a lot of computation in S103 no longer just intermediate variables, but become the final product that the model can directly output and use directly by users. Therefore, S105 advances the entire method from "feature extraction and screening" to the final stage of "process-aware modeling and integrated output", enabling the realization of a complete technical closed loop of "spectral fusion - interactive feature screening - salinity dynamic evolution inversion", which is the key transformation link from theoretical design to practical application of this scheme.

[0127] In this embodiment, the salt inversion model is a deep learning model built based on temporal attention mechanism and multi-task learning. The specific construction steps are as follows: The model construction includes: Input layer: Organizes the feature subset into a standardized data format that the model can process; Temporal coding layer: It consists of two long short-term memory network units in opposite directions, which are used to automatically learn the dependencies and evolution patterns of features at different time points along the time axis; Perceptual attention layer: It contains two linear transformations and a non-linear activation function, which is used to give the model the ability to autonomously judge the importance of information at different time points; Multi-task learning output layer: includes a salinity regression head and a spatial mapping head, wherein the salinity regression head consists of a 3-layer fully connected network and is used to output the soil salinity content corresponding to each pixel; The spatial mapping head reconstructs the salt content of each pixel into a spatial distribution map with the same resolution as the input remote sensing image based on the pixel coordinates and output results.

[0128] More specifically, in the input layer, the original data of the feature subset needs to be standardized. The Z-score standardization method is usually used, which is to subtract the mean of the feature at all time points of all samples for each feature dimension, and then divide it by its standard deviation, so that the mean of the data of each feature dimension is 0 and the standard deviation is 1, thereby eliminating the difference in the units of measurement between different features and making the model training process more stable.

[0129] The temporal coding layer consists of two long short-term memory network units in opposite directions: a forward long short-term memory network and a reverse long short-term memory network. The two work in parallel and together form a bidirectional long short-term memory network structure.

[0130] The positive long short-term memory network follows the natural order of time, from arrive Process the input sequence step by step. At each time step... This unit receives the input vector at the current time. That is, the pixel in the th A vector consisting of the values ​​of all F features in the feature subset at each time point, and simultaneously receiving values ​​from the previous time step. Hidden states of Long Short-Term Memory Networks and cell state The hidden state at the current time step is calculated through an internal gating mechanism: the forget gate determines which old information to discard, the input gate determines which new information to store, and the output gate determines which information to output. and the updated cell state This positive processing enables the network to learn the causal evolution of salinity from before to after rainfall.

[0131] The inverse long short-term memory network follows the reverse order of time points, from arrive Reverse the same input sequence and compute the reverse hidden state at each time step. This reverse processing enables the network to learn the inverse dependencies between subsequent time steps and earlier time steps, capturing the implicit feedback of salinity changes on its previous state.

[0132] After processing all time steps, the forward and backward hidden states of each time step are concatenated along the feature dimension to form the bidirectional hidden representation of that time step. The dimension of the output matrix is ​​the sum of the number of forward and backward hidden units. All bidirectional hidden representations at n time steps are stacked together to form the output matrix of the temporal coding layer, with dimensions [T, H], where H is the total number of bidirectional hidden units, typically set to 64 or 128. This output matrix transforms discrete, potentially time-delayed spectral feature sequences into deep temporal feature representations containing contextual dependencies, providing an informational basis for subsequent attention layers to focus on key time points.

[0133] The design goal of the perception attention layer is to give the model the ability to autonomously judge the importance of information at different remote sensing time points, so that the model can automatically focus its attention weight on the critical moment when salinity changes most drastically, rather than treating all time points equally.

[0134] The specific structure of this layer includes two linear transformations and a nonlinear activation function.

[0135] Its working process is as follows: First, the bidirectional hidden representation of each time step output by the timing coding layer is... This serves as the input vector for the attention layer. The input vector first undergoes a first-level linear transformation, mapping the dimension from H to a smaller intermediate dimension, typically set to H / 2 or a fixed value such as 32 dimensions. This linear transformation learns a weight matrix. and bias vector To achieve this, that is ,in, The intermediate representation vector is then used. A nonlinear activation function, typically the hyperbolic tangent function tanh, is applied to the result of the linear transformation to compress the numerical values ​​to the (-1, 1) interval, introducing nonlinear expressive power. ,in, To activate the output vector, the activated output vector is then fed into a second linear transformation, through a weight vector. Map it to a scalar fraction, i.e. This scalar score represents the original importance score for that time step.

[0136] In obtaining all Energy value at each time step Then, the softmax function is used to normalize these energy values ​​to obtain the attention weights at each time step. = Attention weights The sum of the weights for all time steps is 1. Each weight value is between 0 and 1. The larger the value, the more important the model considers that time step to be for the final inversion task.

[0137] Finally, the calculated attention weights are used to perform a weighted summation of the bidirectional hidden representations at all time steps to generate a fixed-length context vector: This context vector, with a dimension of H, encapsulates the most crucial information from the entire time series, highlighting the characteristic contributions of key time points such as peak salinity changes, while suppressing interference from noise signals during stationary periods. The context vector will serve as the shared input to the multi-task learning output layer.

[0138] It includes a salinity regression head and a spatial mapping head. The salinity regression head consists of a 3-layer fully connected network and is used to output the soil salinity content corresponding to each pixel. The spatial mapping head reconstructs the salt content of each pixel into a spatial distribution map with the same resolution as the input remote sensing image based on the pixel coordinates and output results.

[0139] The three-layer fully connected network takes the context vector c output by the perception attention layer as input, and gradually extracts high-order feature representations through layer-by-layer nonlinear transformation, and finally completes the prediction output of soil salinity content.

[0140] The first fully connected layer maps the dimension of the context vector from H to a hidden dimension, typically set as the first hidden layer of H, such as 64 or 128 neurons. The output of this hidden layer is... ,in, This is the weight matrix for this layer. Let be the set vector of this layer, and ReLU is the modified linear unit activation function, which introduces nonlinearity and alleviates the gradient vanishing problem.

[0141] The second fully connected layer further compresses and transforms the hidden representation of the first layer into a second hidden layer, typically set to half the number of neurons in the first layer, such as 32 or 64 neurons. The output of this hidden layer is... ,in, This is the weight matrix for this layer. This is the set vector for this layer. The third fully connected layer is the final task output layer. For the task of predicting soil salinity, the third layer maps the hidden representation of the second layer to the output dimension, namely the spatial distribution map of soil salinity.

[0142] During model training, the feature data corresponding to all sampling points in the study area are used as training samples. The input of each sample is a time-series feature matrix extracted from the feature subset constructed from S104. Its dimension is the number of time points multiplied by the number of features retained after filtering. Each row of the matrix corresponds to a remote sensing time point, and each column corresponds to a feature. The value is the feature value of the sampling point at the corresponding time point. The ground truth label of each sample is the measured total soil salinity obtained by the sampling point through field sampling in S101 and indoor analysis.

[0143] At the start of training, all samples are randomly divided into training and validation sets in a 7:3 ratio, and the weights of each layer of the model are randomly initialized. The predicted values ​​are compared with the actual measured total salt content, and the mean squared error is used as the loss function to calculate the prediction error. The error gradient is then propagated backward along the network using the backpropagation algorithm. The Adam optimizer updates the weights and biases of each layer of the model based on the gradient, completing one parameter iteration. This process is repeated multiple times on the entire training set. After each training round, the model performance is evaluated on the validation set. Training stops when the loss function value on the validation set no longer decreases or reaches the preset early stopping condition, and the optimal model parameters on the validation set are saved.

[0144] After training, the temporal feature matrices of all effective pixels in the study area are input into the trained model one by one for inference. The model outputs a predicted total salt content value for each pixel. These predicted values ​​are then restored into a two-dimensional raster image according to the spatial row and column positions of the pixels, thus generating a spatial distribution map of soil salinity in the entire study area.

[0145] Please see Figure 4 (a)- Figure 4 (d) and Figure 5 These are, respectively, the predicted distribution map of this application embodiment, the measured salinity distribution map of this application embodiment, the predicted distribution map of the CNN method of this application embodiment, the predicted distribution map of the random forest method of this application embodiment, and the predicted residual distribution map of this application embodiment.

[0146] Figure 4 (a) The predicted spatial distribution of salt in this embodiment and Figure 4 (b) The measured distributions are highly consistent, with a uniform spatial pattern. The location, shape, and extent of high- and low-salinity areas have been accurately reconstructed. Figure 5 The predicted residual distribution map provided in this application shows that the residuals are mainly concentrated in the range of ±1 g / kg, with no obvious spatial clustering bias. Figure 4 (c) The prediction results of the CNN method deviate to some extent from the measured distribution, especially in areas with large salinity gradients. Figure 4 (d) The prediction results of the random forest method are relatively smooth, but it is not good at characterizing the spatial heterogeneity of salinity.

[0147] The above embodiments can be implemented, in whole or in part, by software, hardware, firmware, or any other combination thereof. When implemented in software, the above embodiments can be implemented, in whole or in part, as a computer program product. Those skilled in the art will recognize that the units and algorithm steps of the various examples described in conjunction with the embodiments disclosed herein can be implemented by electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are implemented in hardware or software depends on the specific application and design constraints of the technical solution.

[0148] The units described as separate components may or may not be physically separate. The components shown as units may or may not be physical units; they may be located in one place or distributed across multiple network units. Some or all of the units can be selected to achieve the purpose of this embodiment, depending on actual needs.

[0149] The above description is merely a specific embodiment of this application, but the scope of protection of this application is not limited thereto. Any changes or substitutions that can be easily conceived by those skilled in the art within the scope of the technology disclosed in this application should be included within the scope of protection of this application.

Claims

1. A method for soil salinity inversion through cross-modal fusion of multi-source remote sensing images, characterized in that, include: Sampling points are set for the soil in the area to be inverted. Sampling and analysis of the soil at the sampling points are performed to obtain sampling data. Meteorological data of the area to be inverted is obtained. Based on the meteorological data, the remote sensing time point for acquiring remote sensing images is determined. Remote sensing images of the inverted area are acquired based on remote sensing time points and preprocessed. A set of temporal spectral feature parameters that can characterize the soil state are extracted pixel by pixel from the preprocessed remote sensing images. Based on the salinity index and moisture index in the set of spectral feature parameters at each remote sensing time point, the dynamic response feature parameters of salinity change over time are analyzed, and the change curve of salinity is fitted. Based on the change curve, the curve parameters are extracted, and the spectral feature parameters, dynamic response feature parameters, and curve parameters are constructed into a comprehensive feature set. The comprehensive feature set of each sampling point is paired with the corresponding total salt content. For each feature in the comprehensive feature set, the univariate regression determination coefficient with the total salt content, the process response degree corresponding to the change before and after rainfall, and the average synergistic gain value after the feature is combined with other features are calculated. The determination coefficient, process response degree and average synergistic gain value are normalized and weighted to obtain the contribution ability of the feature. The features are sorted and filtered according to their contribution ability to construct a feature subset. A salinity inversion model was constructed and trained. A subset of features was used as input to the trained model, and the spatial distribution data of soil salinity content was output in real time.

2. The method according to claim 1, characterized in that, The sampling data includes: surface salt content and total salt content; The specific process for obtaining the sampling data is as follows: The soil to be inverted is divided into uniform grids, and the center of each grid is used as a sampling point. Several random sampling plots are set up at each sampling point. The surface salt content and total salt content of the soil in each sampling plot are collected. The average of the surface salt content and total salt content of all sampling plots under the same sampling point is calculated and used as the surface salt content and total salt content of the soil at that sampling point.

3. The method according to claim 1, characterized in that, Meteorological data includes rainfall amount and rainfall duration; The specific method for determining the remote sensing time point for acquiring remote sensing images based on meteorological data is as follows: Based on the relationship between soil surface salinity in the sampling data and the changes in surface salinity after historical rainfall, the dissolved rainfall threshold was determined. Within the observation period, the moment when the cumulative rainfall first reaches the dissolved rainfall threshold is taken as the remote sensing time starting point. The remote sensing time length is set and divided into equal parts, with each division point being a remote sensing time point.

4. The method according to claim 1, characterized in that, The preprocessing method is as follows: radiometric calibration and atmospheric correction are performed on the multi-temporal remote sensing images in sequence to eliminate atmospheric effects and restore the true reflectivity of the surface. Then, orthorectification is performed, pixel-level spatial alignment is performed, and the orthorectified remote sensing images are masked to obtain remote sensing images with unified radiometric and geometric references.

5. The method according to claim 1, characterized in that, Spectral characteristic parameters include; Normalized salinity index, brightness index, vegetation category index, water category index, temperature and humidity index; The method for extracting spectral feature parameters is as follows: Based on the preprocessed remote sensing images acquired at remote sensing time points, a set of surface reflectance values ​​for different bands corresponding to each pixel are obtained. According to the type of spectral feature parameter, and the corresponding band combination operation rules for each type of spectral feature parameter, pixel-level mathematical operations are performed on all pixels of all remote sensing images at all remote sensing time points. The original, single-band surface reflectance information is transformed into a series of spectral feature parameters that have clear indicative capabilities for soil salinity, vegetation status, surface moisture, and temperature and humidity elements. After calculating all spectral feature parameters for a single remote sensing time point, all spectral feature images for that time point are stacked according to the feature dimension to form a set of spectral feature parameters for that single remote sensing time point. The above calculation process is repeated for all selected remote sensing time points. Finally, the sets of spectral feature parameters for each remote sensing time point are aligned and stitched together along the time dimension to form a set of spectral feature parameters for time-series features.

6. The method according to claim 1, characterized in that, Dynamic response characteristic parameters include the relative salinity response index and the water-salt decoupling index; The process of obtaining the relative salinity response index is as follows: Using the remote sensing time points corresponding to the salinity and moisture indices before rainfall as the baseline time, and based on the salinity and moisture indices corresponding to adjacent remote sensing time points after rainfall, the change values ​​of salinity and moisture indices are obtained according to the time series. The remote sensing time point corresponding to the maximum value of the change values ​​of salinity and moisture indices is taken as the peak time. The difference between the salinity index at the peak time and the salinity index at the baseline time is taken as the net change in salinity. The absolute value of the difference between the moisture index at the peak time and the moisture index at the baseline time is taken as the change in moisture. A constant is introduced, and the ratio obtained by dividing the net change in salinity by the sum of the change in moisture and the constant is the relative salinity response index. The process for obtaining the water-salt decoupling index is as follows: Based on the salinity index and moisture index in the spectral feature parameter set, they are sorted according to time series to form salinity index series and moisture index series, respectively. They are then normalized and scaled to the same interval. A dynamic time warping algorithm is used to find the alignment path with the minimum cumulative distance between the two normalized sequences by bending, stretching or compressing the time axis, and this shortest cumulative distance is recorded. The ratio obtained by dividing the shortest cumulative distance by the actual step length of the warped path is the water-salt decoupling index.

7. The method according to claim 1, characterized in that, The curve parameters include the magnitude of salinity change, the maximum rate of change, and the area under the curve. The method for fitting the curve is as follows: Based on the salinity index in the set of spectral feature parameters, the parameters are sorted according to time series and fitted using cubic spline interpolation to obtain a time-continuous and smooth reaction salinity change curve, which is the change curve. The method for extracting the range of salinity changes is as follows: Based on the change curve, find the highest and lowest points on the entire change curve, and subtract the salinity index of the lowest point from the salinity index of the highest point. The difference is the salinity change range. The method for extracting the maximum rate of change is as follows: Based on the change curve, calculate the first derivative of the change curve, and the maximum absolute value of the first derivative is the maximum rate of change. The method for extracting the area of ​​a curve is as follows: Using the salinity index before rainfall as a baseline, the area enclosed by the entire change curve and this baseline is the area of ​​the curve.

8. The method according to claim 1, characterized in that, The analysis process for contribution capability is as follows: Based on the measured total salt content in the soil at all sampling points, the comprehensive feature set of each sampling point is paired with the corresponding total salt content one by one. The univariate regression determination coefficient between each feature in the comprehensive feature set and the total salt content is calculated to obtain the determination coefficient between the feature and the total salt content. The average change of this characteristic at all sampling points before and after rainfall, as well as the average change of the total salt content at the corresponding sampling points, were analyzed. The relative deviation between the two changes was obtained, and the process response was obtained by subtracting the relative deviation from 1. Based on any two features in the comprehensive feature set, perform univariate regression on the total salt content to obtain the explanatory power of the two individual features. Perform binary regression on the measured total salt content using the two features together to obtain the joint explanatory power. Subtract the sum of the explanatory powers of the two individual features from the joint explanatory power to obtain the collaborative gain value of the features. The contribution capability of this feature is obtained by normalizing the coefficient of determination, process responsiveness, and synergistic gain value and then fusing them according to preset weights.

9. The method according to claim 1, characterized in that, The steps for prioritizing and constructing a feature subset based on contribution ability are as follows: Based on the contribution ability value, sort the features in descending order to form an ordered sequence. Plot a curve in a plane coordinate system with the sorted sequence number as the horizontal axis and the contribution ability value corresponding to each feature as the vertical axis. Find the inflection point in the transition region between the steep and flat sections of the curve. Select all features whose contribution ability is greater than the contribution ability value corresponding to this inflection point to construct the feature subset.

10. The method according to claim 1, characterized in that, The salt inversion model is a deep learning model built on temporal attention mechanism and multi-task learning. The specific construction steps are as follows: The model construction includes: Input layer: Organizes the feature subset into a standardized data format that the model can process; Temporal coding layer: It consists of two long short-term memory network units in opposite directions, which are used to automatically learn the dependencies and evolution patterns of features at different time points along the time axis; Perceptual attention layer: It contains two linear transformations and a non-linear activation function, which is used to give the model the ability to autonomously judge the importance of information at different time points; Multi-task learning output layer: includes a salinity regression head and a spatial mapping head, wherein the salinity regression head consists of a 3-layer fully connected network and is used to output the soil salinity content corresponding to each pixel; The spatial mapping head reconstructs the salt content of each pixel into a spatial distribution map with the same resolution as the input remote sensing image based on the pixel coordinates and output results.

Images