GNSS-R soil salinity inversion method and system based on cross-modal fusion

Abstract

The present application belongs to the technical field of soil salinity inversion, and specifically provides a GNSS-R soil salinity inversion method and system based on cross-modal fusion, comprising: performing data fusion and quality control on the Tianmu-1 satellite observation data set and the SMAP satellite auxiliary data set to obtain DDM images and multi-dimensional auxiliary features; based on the DDM images and the multi-dimensional auxiliary features, combining a T-GateFusionNet module, a HybridFormer module and a second dynamic fusion module to construct a cross-modal dynamic fusion model; and based on the cross-modal dynamic fusion model, performing inversion on the soil salinity of a preset area to obtain a soil salinity inversion result. The present application uses multiple types of features to perform soil salinity inversion, thereby improving the accuracy of the soil salinity inversion result.

Description

Technical Field

[0001] This invention belongs to the field of soil salinity inversion technology, specifically relating to a GNSS-R soil salinity inversion method and system based on cross-modal fusion. Background Technology

[0002] Soil salinity, or soil salt content, is a physical quantity representing the amount of soluble salts in soil and is an important indicator of the degree of soil salinization. Soil salinization is a major cause of soil desertification and degradation, especially in irrigated farmland and grasslands and coastal wetlands in arid and semi-arid regions, where the phenomenon is more severe. Therefore, monitoring soil salinity is essential.

[0003] Soil salinity can be obtained through in-situ observation and remote sensing methods. In-situ observation can obtain high-precision soil salinity values, but it is time-consuming and difficult to meet the needs of large-scale salinization monitoring. Optical remote sensing missions (such as MODIS and Landsat) can capture soil salinity with high spatiotemporal resolution, but the spectrum cannot directly perceive the characteristics of the surface soil, and its penetration ability is insufficient, making it significantly affected by weather conditions. Compared with optical remote sensing, microwave remote sensing has the advantage of all-weather observation due to its ability to penetrate vegetation and its immunity to weather conditions, and is considered an effective method for monitoring soil salinity. Existing research mainly focuses on active microwave remote sensing and has confirmed that synthetic aperture radar (SAR) imagery can serve as an effective tool for salinization monitoring. However, active microwave remote sensing technology is limited by image geometric distortion and complex noise, and its temporal resolution is not ideal, making it difficult to achieve high spatiotemporal resolution and high-precision soil salinity inversion. To address the limitations of the above monitoring technologies, Global Navigation Satellite System Reflectance Measurement (GNSS-R) technology has shown unique advantages, opening up new avenues for soil salinization monitoring. GNSS-R technology, with its advantages of all-weather observation, high spatiotemporal coverage, and low cost, has been widely applied in various remote sensing fields such as soil moisture, ocean altimetry, and sea surface wind speed. In recent years, with the development of multi-constellation navigation systems, the potential of GNSS-R in soil parameter inversion has become increasingly prominent, making it a research hotspot in the field of remote sensing. Summary of the Invention

[0004] This invention provides a GNSS-R soil salinity inversion method and system based on cross-modal fusion, which utilizes multiple features to perform soil salinity inversion and obtain more accurate soil salinity inversion results.

[0005] To achieve the above objectives, the present invention provides the following solution: GNSS-R soil salinity inversion methods based on cross-modal fusion include: Data fusion and quality control were performed on the Tianmu-1 satellite observation dataset and the SMAP satellite auxiliary dataset to obtain DDM images and multidimensional auxiliary features. Based on DDM images and multidimensional auxiliary features, a cross-modal dynamic fusion model is constructed by combining the T-GateFusionNet module, the HybridFormer module, and the second dynamic fusion module. Based on the cross-modal dynamic fusion model, the soil salinity of the preset area is inverted to obtain the soil salinity inversion results.

[0006] Preferably, the DDM image includes the original count DDM image, the analog power DDM image, the bistatic radar cross section DDM image, and the effective scattering area DDM image; the multidimensional auxiliary features include features related to DDM and features related to the land surface; the features related to DDM include: corrected surface reflectivity, incident angle, DDM mean, DDM peak value, signal-to-noise ratio, and the row and column of the DDM peak value; the features related to the land surface include: time, specular reflection point longitude, specular reflection point latitude, vegetation optical thickness, surface roughness, surface temperature, vegetation moisture content, soil bulk density, and the proportion of clay and sand in the soil.

[0007] Preferably, in the cross-modal dynamic fusion model, the T-GateFusionNet module is used to extract physical prior features from the multi-dimensional auxiliary features, and to fuse the physical prior features, the core tensor statistics, and the singular value distribution of the factor matrix of the DDM image using Tucker tensor decomposition and parsing, combined with a gated attention mechanism, to obtain a one-dimensional fused feature; the HybridFormer module is used to perform feature projection and spatial fusion on the DDM image and the multi-dimensional auxiliary features to obtain a two-dimensional fused image, and to jointly model the local texture features and global structural distribution features of the two-dimensional fused image using a parallel dual-path Transformer structure to obtain the spatial structural features in the two-dimensional fused image; the second dynamic fusion module is used to perform adaptive fusion of the one-dimensional fused feature and the spatial structural feature across modalities using a dynamic attention fusion mechanism to obtain the fused feature.

[0008] Preferably, the T-GateFusionNet module includes: a residual module for effectively representing and enhancing the multidimensional auxiliary features to obtain the physical prior features; a Tucker low-rank decomposition module for performing Tucker decomposition on the DDM image to obtain the core tensor statistics and the singular value distribution of the factor matrix, and concatenating the core tensor statistics and the singular value distribution of the factor matrix to obtain a one-dimensional feature vector; and a gated fusion module for adaptively fusing the physical prior features and the one-dimensional feature vector using a dual-path collaborative mechanism to obtain the one-dimensional fused features.

[0009] Preferably, the HybridFormer module includes: a conditional encoding module, used to perform layer normalization processing on the multidimensional auxiliary features, and project them onto a spatial dimension consistent with the DDM image through a linear layer to obtain a spatial attention map, and to concatenate the spatial attention map with the original DDM image in the channel dimension, and to complete feature fusion through 3×3 convolution to obtain the two-dimensional fused image; The multi-scale residual enhancement module is used to model the two-dimensional fused image in parallel under different receptive fields using a multi-branch structure, and to stitch together the features extracted by each branch in the channel dimension to obtain multi-scale spatial information. A downsampling layer is used to downsample the multi-scale spatial information using a differentiated downsampling strategy; the differentiated downsampling strategy includes a conservative downsampling strategy and an aggressive downsampling strategy; The local attention module is used to downsample the multi-scale spatial information using the conservative downsampling strategy to obtain local texture features; A global attention module is used to aggregate the global context information of the multi-scale spatial information using the aggressive downsampling strategy to obtain the global structural distribution features; The first dynamic fusion module is used to adaptively fuse the local texture features with the global structural distribution features to obtain the spatial structural features in the two-dimensional fused image.

[0010] Preferably, the second dynamic fusion module includes three parts: differential feature extraction, dynamic convolution optimization, and adaptive attention fusion.

[0011] This invention relates to a cross-modal fusion GNSS-R soil salinity inversion system for implementing the method, comprising: The data processing module is used to perform data fusion and quality control on the Tianmu-1 satellite observation dataset and the SMAP satellite auxiliary dataset to obtain DDM images and multidimensional auxiliary features. The model building module is used to construct a cross-modal dynamic fusion model based on DDM images and multi-dimensional auxiliary features, combined with the T-GateFusionNet module, HybridFormer module and second dynamic fusion module. The salinity inversion module is used to invert the soil salinity of a preset area based on the cross-modal dynamic fusion model to obtain the soil salinity inversion results.

[0012] Preferably, in the model building module, the cross-modal dynamic fusion model includes: The T-GateFusionNet module is used to extract physical prior features from the multidimensional auxiliary features, and to use Tucker tensor decomposition and parsing of the core tensor statistics and the singular value distribution of the factor matrix of the DDM image, combined with a gated attention mechanism to fuse the physical prior features, the core tensor statistics and the singular value distribution of the factor matrix to obtain a one-dimensional fused feature. The HybridFormer module is used to perform feature projection and spatial fusion on the DDM image and multi-dimensional auxiliary features to obtain a two-dimensional fused image. It also uses a parallel dual-path Transformer structure to jointly model the local texture features and global structural distribution features of the two-dimensional fused image to obtain the spatial structural features in the two-dimensional fused image. The second dynamic fusion module is used to adaptively fuse the one-dimensional fusion feature and the spatial structure feature across modal features using a dynamic attention fusion mechanism to obtain the fusion feature.

[0013] Compared with the prior art, the beneficial effects of the present invention are as follows: 1. To fully leverage the complementary advantages of DDM images and auxiliary features, this invention employs a two-branch cross-modal input strategy. The two-dimensional DDM image and one-dimensional auxiliary features are used as two-branch cross-modal inputs, enabling interactive modeling of cross-modal information at both the feature domain and spatial domain levels. This strategy preserves the time-delay-Doppler structural information while strengthening the statistical correlation of physical features, thereby improving feature representation capabilities and model generalization performance in both structural and semantic dimensions.

[0014] 2. Tucker tensor decomposition is introduced to construct regularized constraint branches, extracting the statistics of the core tensor of the DDM image and the singular value features of the factor matrix from the feature space, so as to capture the coupling effect of the time delay-Doppler domain and extract the noise-resistant DDM main scattering mode, effectively integrate auxiliary parameter information, and improve the targeting of feature selection.

[0015] 3. To achieve effective complementarity of multi-scale features and efficient fusion of multi-source information, this invention constructs a Transformer architecture with parallel local and global attention and a dynamic feature fusion module. Specifically, the local branch uses conservative downsampling to preserve detailed information, while the global branch uses aggressive downsampling to expand the receptive field, thus taking into account both local details and global semantics in multi-scale modeling. At the same time, the dynamic feature fusion module introduces an adaptive attention fusion mechanism and a dynamic convolution weighting method to achieve adaptive aggregation and effective collaboration of different features, effectively improving feature representation capabilities and the overall performance of the model. Attached Figure Description

[0016] To more clearly illustrate the technical solution of the present invention, the drawings used in the embodiments are briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0017] Figure 1 This is a flowchart of the GNSS-R soil salinity inversion method based on cross-modal fusion, as described in an embodiment of the present invention. Figure 2 This is a schematic diagram of the cross-modal dynamic fusion model structure according to an embodiment of the present invention; Figure 3 This is a schematic diagram of the dual-path residual channel attention module structure according to an embodiment of the present invention; Figure 4 This is a schematic diagram of the gated fusion module structure according to an embodiment of the present invention; Figure 5 This is a schematic diagram of the structure of the condition coding module and the gated fusion module in an embodiment of the present invention; Figure 6 This is a schematic diagram of the multi-scale residual enhancement module structure according to an embodiment of the present invention; Figure 7 This is a schematic diagram of the conservative downsampling layer structure according to an embodiment of the present invention; Figure 8 This is a schematic diagram of the radical downsampling layer structure according to an embodiment of the present invention; Figure 9 This is a schematic diagram of the local attention module structure according to an embodiment of the present invention; Figure 10 This is a schematic diagram of the global attention module structure according to an embodiment of the present invention; Figure 11 This is the first dynamic fusion module in this embodiment of the invention; Figure 12 This is a schematic diagram of the dynamic convolution module structure according to an embodiment of the present invention; Figure 13 This is a scatter plot of the inversion results from an embodiment of the present invention; Figure 14 This is a line graph showing the inversion results of an embodiment of the present invention. Detailed Implementation

[0018] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0019] To make the above-mentioned objects, features and advantages of the present invention more apparent and understandable, the present invention will be further described in detail below with reference to the accompanying drawings and specific embodiments.

[0020] The following explanations are provided for the relevant technical terms appearing in the embodiments: GNSS-R: Global Navigation Satellite System-Reflectometry is a remote sensing technology that uses direct signals from global navigation satellite systems and surface reflection signals to invert surface parameters. GNSS-R technology adopts a dual (multi)-base radar mode, with the globally shared GNSS constellation serving as the source of multi-source microwave signals, and the receiving system deployed on an independent platform separate from the source, such as a shore-based, airborne, or spaceborne platform.

[0021] Tianmu-1 is China's first commercial low-Earth orbit meteorological satellite system, supporting four major navigation systems: GPS, BeiDou (BDS), Galileo (GAL), and GLONASS (GLO). It is the world's first satellite system to simultaneously support four navigation systems, significantly improving global atmospheric data acquisition capabilities and enabling the acquisition of characteristic parameters such as global distribution and all-weather atmospheric, land surface, and sea surface data. Tianmu-1 began operation on January 9, 2023, providing GNSS occultation and reflection data.

[0022] DDM (Digital Power Distribution Map) is a two-dimensional power distribution map acquired by a GNSS-R receiving system. It represents the received signal strength under different propagation delays and Doppler shifts. For the Tianmu constellation, the DDM consists of 61 delay rows and 20 Doppler columns. The DDM comprehensively reflects the surface scattering characteristics and observation geometry, and is a key data format for inverting parameters such as soil salinity and soil moisture.

[0023] Tucker decomposition: Tucker decomposition is a high-order tensor decomposition method that can decompose the original tensor into a product of a core tensor and factor matrices of various orders, thereby achieving data compression and feature extraction.

[0024] Example 1: like Figure 1 As shown, the GNSS-R soil salinity inversion method based on cross-modal fusion includes: S1: Perform data fusion and quality control on the Tianmu-1 satellite observation dataset and the SMAP satellite auxiliary dataset to obtain DDM images and multidimensional auxiliary features.

[0025] In this embodiment, the Tianmu-1 data product used is L1-level reflectance data from the occultation reflector. The Tianmu-1 dataset provides DDM images and single features. The SMAP data product used is enhanced L3-level data with a grid resolution of 9km × 9km and HDF5 format. The SMAP dataset provides data on vegetation optical thickness, surface roughness coefficient, surface temperature, vegetation water content, soil bulk density, clay content in the soil, and soil moisture. Soil moisture data is used as reference data, and the other data are also used as single features.

[0026] The Tianmu-1 observation data contains many low-quality sampling points. To ensure the reliability of the inversion results, this paper performs standardized quality control on the raw data to remove observation samples that do not meet the requirements. The specific screening criteria are as follows: (1) Incident Angle Screening: When the incident angle is greater than 65°, the path of the GNSS-R reflected signal is too long and is easily affected by factors such as surface roughness and atmospheric refraction, resulting in signal attenuation and enhanced scattering, thereby reducing the inversion accuracy. Therefore, this invention only retains observation samples with an incident angle of less than 65°.

[0027] (2) Antenna gain screening: When the receiver antenna gain is less than 0 dB, the signal receiving sensitivity is insufficient, which can easily lead to a low signal-to-noise ratio or even signal loss. To ensure the availability of data, all samples with a gain less than 0 dB are removed.

[0028] (3) SNR range screening: A low SNR (less than 0) means that the signal is almost completely submerged by noise and has no physical interpretability; while an excessively high SNR (greater than 14) is usually caused by abnormal reflection or hardware error and does not conform to the physical laws of soil scattering characteristics. Existing GNSS-R data processing experience also shows that the SNR range of 0–14 can better reflect the surface scattering characteristics of soil and vegetation.

[0029] A further implementation involves using a DDM image that includes a raw count DDM image, an analog power DDM image, a bistatic radar cross section DDM image, and an effective scattering area DDM image. The DDM is obtained by analyzing the delay and Doppler shift information of the reflected signal, characterizing its time delay and frequency shift properties. Therefore, the DDM changes as the salinity of the soil changes. This invention uses an analog power DDM image as input to the model.

[0030] Multidimensional auxiliary features include features related to DDM and features related to the land surface; features related to DDM include: corrected land surface reflectance, incident angle, DDM mean, DDM peak value, signal-to-noise ratio, and the row and column of DDM peak value; features related to the land surface include: time, longitude of specular reflection point, latitude of specular reflection point, vegetation optical thickness, land surface roughness, land surface temperature, vegetation moisture content, soil bulk density, and the proportion of clay and sand in the soil.

[0031] Studies have shown a strong correlation between surface reflectance and soil salinity. Since most land surfaces are relatively smooth, GNSS-R signals are mainly coherent reflection components. Based on this assumption, surface reflectance can be calculated using formula (1).

[0032] (1) In the formula, This represents the power at the peak point of the DDM image. This indicates the power of the transmitter's transmitted signal. This indicates the antenna gain of the transmitter. This indicates the distance between the transmitter and the point of reflection on the mirror. This indicates the distance between the specular reflection point and the receiver. This indicates the antenna gain of the receiver. This indicates the wavelength (0.19m) of the GNSS navigation satellite signal. Indicates at the angle of incidence The surface reflectance. After converting all items to dB, the surface reflectance... The following formula (2) can be used to solve the problem: (2) The DDM mean refers to the average intensity of reflected signals in a DDM image. It characterizes the overall energy level or average distribution of scattering intensity across the entire DDM image. A higher DDM mean indicates stronger electromagnetic reflection capability of the Earth's surface; conversely, a lower DDM mean indicates weaker GNSS-R signal scattering capability. DDM peak value, incident angle, latitude and longitude of specular reflection points, and signal-to-noise ratio data are stored in a NetCDF file, which can be obtained from Tianmu-1 data. Other auxiliary data include surface roughness, vegetation cover, vegetation moisture content, surface temperature, soil bulk density, clay content in the soil, and sand content in the soil; these can all be obtained from SMAP satellite data and HWSD2.0 data.

[0033] S2: Based on DDM images and multi-dimensional auxiliary features, a cross-modal dynamic fusion model (CDF-Net) is constructed by combining the T-GateFusionNet module, the HybridFormer module, and the second dynamic fusion module. For example... Figure 2 As shown.

[0034] A further implementation involves the T-GateFusionNet module in the cross-modal dynamic fusion model, which extracts physical prior features from multi-dimensional auxiliary features and uses Tucker tensor decomposition and parsing of the core tensor statistics and factor matrix singular value distribution of the DDM image. Combined with a gated attention mechanism, the physical prior features, core tensor statistics, and factor matrix singular value distribution are fused to obtain one-dimensional fused features.

[0035] A further implementation method is that the T-GateFusionNet module includes: The residual module is used to effectively characterize and enhance multidimensional auxiliary features to obtain physical prior features. The 17 multidimensional auxiliary features are time, longitude of the specular reflection point, latitude of the specular reflection point, corrected surface reflectance, DDM mean, DDM peak value, row of DDM peak value, column of DDM peak value, signal-to-noise ratio, incident angle, surface roughness coefficient, vegetation optical thickness, vegetation water content, surface temperature, soil bulk density, clay ratio in soil, and sand ratio in soil.

[0036] Specifically, the residual module, centered on the channel attention mechanism, consists of linear layers, layer normalization, nonlinear activation, deactivation layers, and a dual-residual channel attention module (DRCA), possessing strong feature transformation and regularization capabilities. The DRCA module is an improved channel attention mechanism that independently models the channel features of average pooling and max pooling, introduces a dual-path fully connected network, and uses learnable fusion weights α for adaptive weight fusion, thereby achieving a more flexible and robust channel attention mechanism. Figure 3 As shown. Specifically: The DRCA module uses a dual-path independent pooling structure, employing average pooling and max pooling to extract features from different perspectives. Average pooling focuses on the overall trend, while max pooling emphasizes prominent local features. The complementarity of these two methods helps to more comprehensively understand the importance between channels. In addition, a learnable fusion weight α is added to adaptively weight and fuse the two attention results, rather than simply averaging or fusion with a fixed ratio, as shown in formula (3). This allows the model to automatically adjust its focus based on actual data, further enhancing the flexibility and expressiveness of the attention mechanism.

[0037] (3) In the formula, Represents adaptive fusion features. The representation is the feature extracted by the average pooling (AVGPool) branch. This represents the feature representation extracted through the max pooling (MAXPool) branch.

[0038] The introduction of the DRCA module significantly enhances the model's ability to focus on key channels, while its residual-enhanced output design ensures that the discriminative information of the original features is preserved during attention learning. Batch normalization is also introduced throughout the network to accelerate convergence and improve training stability.

[0039] The Tucker low-rank decomposition module is used to perform Tucker decomposition on the DDM image, obtain the core tensor statistics and the singular value distribution of the factor matrix, and concatenate the core tensor statistics and the singular value distribution of the factor matrix to obtain a one-dimensional feature vector. Specifically, the Tucker low-rank decomposition module decomposes the DDM image using formula (4). Decomposed into a core tensor and two factor matrices , Then, the core tensor statistics, including mean, standard deviation, L2 norm, maximum, and minimum, are calculated to capture the core distribution characteristics. The statistics of the factor matrix are calculated, including the mean, standard deviation, and L2 norm along the columns, to analyze the feature changes of the DDM in different dimensions. Furthermore, the singular value decomposition (SVD) method is introduced to extract principal component information, which helps to identify the most important patterns in the data and improve the representativeness of the features. Finally, the features are concatenated into a one-dimensional feature vector, which is used as the input of the fully connected layer.

[0040] (4) in, This represents the product of the core tensor and the factor matrix along the row dimension. This represents the product of the core tensor and the factor matrix along the column dimension.

[0041] The Gate Fusion Block employs a dual-path collaborative mechanism to adaptively fuse physical prior features with a one-dimensional feature vector to obtain a one-dimensional fused feature, which is then reduced in dimensionality by a linear layer. It's worth noting that because the reconstruction loss itself reflects the degree of fit of the low-rank representation to the original DDM data, it helps to enhance the identification of boundary samples and outliers; therefore, this invention also includes the reconstruction loss as part of the final loss function. The specific structure is as follows... Figure 4As shown, the core of this module lies in the collaborative design of a feature filter and an attention enhancer: the feature filter transforms low-rank features into a dynamic gating mask that matches the dimensions of auxiliary features through linear transformation and the sigmoid activation function, achieving selective filtering of low-rank features; the attention enhancer adopts a bottleneck structure design, extracting key attention weights of auxiliary features through dimensionality reduction and dimensionality enhancement operations. The synergistic effect of these two mechanisms enables the module to intelligently filter valuable low-rank feature information and adaptively enhance important dimensions in auxiliary features. Compared to traditional feature fusion methods, this module has three significant advantages: First, the dynamic gating mechanism achieves intelligent filtering of low-rank features, effectively suppressing noise interference; second, the attention mechanism automatically focuses on key feature dimensions, improving the discriminative power of feature representation; and finally, the unique feature enhancement method achieves non-linear enhancement while preserving the integrity of the original features.

[0042] The HybridFormer module is used to perform feature projection and spatial fusion of DDM images and multi-dimensional auxiliary features to obtain a two-dimensional fused image. It then employs a parallel dual-path Transformer structure to jointly model the local texture features and global structural distribution features of the fused image, obtaining the spatial structural features within the fused image. The HybridFormer module is an innovative parallel dual-branch Transformer architecture that achieves efficient multi-scale feature modeling by fusing local and global attention mechanisms. Its core advantage lies in its heterogeneous dual-branch design: one branch uses local window attention combined with a conservative downsampling strategy to preserve detailed features, while the other branch combines global attention and aggressive downsampling to capture global information. It should be noted that a Dynamic Fusion Module (DFM) and a DRCA module are embedded in this module; the DFM adaptively integrates dual-path features, while the DRCA module uses dual-path pooling to enhance feature representation capabilities. The HybridFormer module designed in this invention consists of a conditional coding module, a multi-scale residual enhancement module, a downsampling layer, a local attention module, a global attention module, a dynamic fusion module, a DRCA module, a batch normalization layer, and global average pooling. The following will provide a detailed explanation of each part.

[0043] The conditional encoding module performs layer normalization on multidimensional auxiliary features and projects them onto a spatial dimension consistent with the DDM image through a linear layer to obtain a spatial attention map. This spatial attention map is then concatenated with the original DDM image along the channel dimension, and feature fusion is achieved through 3×3 convolution to obtain a two-dimensional fused image. Specifically, in GNSS-R soil salinity inversion, the DDM image can provide spatial distribution characteristics of target scattering in the time-delay-Doppler domain, but its sensitivity to environmental conditions is limited. One-dimensional auxiliary features such as surface temperature, surface reflectivity, and soil bulk density can supplement factors not directly represented by the DDM from the perspectives of physical properties and environmental background. Joint modeling of these two features not only achieves information complementarity but also provides stable prior constraints under noise interference and complex observation conditions, thereby improving inversion accuracy and model robustness. Based on this, this invention designs a conditional encoding module to achieve efficient fusion of DDM image features and auxiliary features. This module takes auxiliary features and the DDM image as input and achieves effective integration of multi-source information through feature projection and spatial fusion. The processing flow includes two key steps: First, auxiliary features are layer-normalized and projected onto a spatial dimension consistent with the DDM image through a linear layer, while the GELU activation function is introduced to enhance non-linear expressive power. Then, the projected conditional features are concatenated with the original DDM image along the channel dimension, and feature fusion is achieved through 3×3 convolution, resulting in the final two-dimensional fused image. The structure of this module is as follows: Figure 5 As shown. The innovation of this design is reflected in: (1) converting auxiliary parameters into spatial attention maps through linear projection, preserving the correlation between parameters; (2) using 3×3 convolution kernels to achieve local feature fusion, effectively capturing feature interactions within the spatial neighborhood; (3) achieving deep integration of multi-source data with a simple and efficient structure.

[0044] A multi-scale residual enhancement module (MREM) is used to model the 2D fused image in parallel under different receptive fields using a multi-branch structure, and to concatenate the features extracted from each branch along the channel dimension to obtain multi-scale spatial information. Specifically, to effectively enhance the feature representation capability of the 2D fused image, this invention designs a multi-scale residual enhancement module (MREM), the structure of which is as follows: Figure 6As shown, this module employs a multi-branch architecture to simulate different receptive fields and capture rich spatial information. The overall structure comprises three main branches and a linear fusion channel, as follows: Branch 1 contains a 1×1 convolution and a 3×3 convolution for basic local feature extraction, preserving the original spatial information; Branch 2 consists of four convolutional layers: 1×1, 1×3, 3×1, and 3×3 dilated convolutions. Asymmetric convolutional kernels enhance long-range dependencies in the horizontal direction, while dilated convolutions expand the receptive field; Branch 3 consists of four convolutional layers: 1×1, 3×1, 1×3, and 3×3 dilated convolutions, symmetrical to Branch 2, enhancing vertical features and forming spatial complementarity. The features extracted by each branch are concatenated along the channel dimension and then compressed and fused using a 1×1 convolution. Simultaneously, to prevent information loss and gradient vanishing, residual connections are designed in this module.

[0045] In summary, MREM has the following advantages: (1) By using a multi-branch structure to model in parallel under different receptive fields, it effectively improves the ability to capture multi-scale spatial information; (2) By using asymmetric convolution and symmetric design to enhance feature extraction in the horizontal and vertical directions respectively, it achieves complementarity in spatial dimensions; (3) By introducing dilated convolution to expand the receptive field, it obtains richer contextual information without significantly increasing the amount of computation; (4) By combining residual connections, it ensures the integrity of feature transmission, alleviates the gradient vanishing problem, and thus enhances the stability and robustness of network training.

[0046] The downsampling layer employs a differentiated downsampling strategy to downsample multi-scale spatial information. This strategy includes both conservative and aggressive downsampling approaches to more effectively adapt to the modeling needs of receptive fields at different scales. The local attention module uses a conservative downsampling strategy to downsample multi-scale spatial information, preserving more original spatial information to facilitate local detail modeling and obtain local texture features. The conservative downsampling layer consists of two consecutive convolutional modules: first, a 3×3 convolutional kernel (stride=1, padding=1) is used to maintain the input spatial resolution for initial feature extraction; then, another 3×3 convolution (stride=2, padding=1) is used to achieve a 2x downsampling. Each convolutional operation is followed by a batch normalization layer and a GELU activation function to improve training stability and non-linear expressive power. Figure 7As shown, the global attention module employs an aggressive downsampling strategy to rapidly aggregate global contextual information across multiple spatial scales through larger-scale spatial compression, thereby obtaining global structural distribution features. The aggressive downsampling layer consists of two 5×5 convolutional layers stacked sequentially: the first convolutional layer compresses the input channels to half their original size (stride=2, padding=2), and the second layer further restores the channel dimensions to the target dimensions and performs another 2x downsampling, achieving a total spatial resolution compression of 4x. Similarly, each convolutional layer is followed by a batch normalization layer and a GELU activation function. Figure 8 As shown.

[0047] The core idea behind this design is that for local attention, higher resolution helps capture fine edge and texture information; while for global attention, lower spatial resolution can effectively expand the receptive field and reduce computational costs. By customizing specific downsampling strategies for the two types of attention mechanisms, the model can achieve more efficient utilization of computational resources while maintaining accuracy, thereby enhancing the overall multi-scale fusion capability and performance.

[0048] A local-global dual-branch attention mechanism is employed to achieve multi-scale modeling and fusion of features in 2D fused images. This structure consists of two parallel attention modules, focusing on local details and global semantics respectively, thereby enhancing the model's joint perception of spatial details and long-range dependencies. Both the local and global attention modules use a convolutional layer as a positional embedding layer to inject spatial information into the subsequent attention mechanism. Then, layer normalization is used for standardization to improve training stability. The core component is based on Local AwarenessAttention (LAA) and Multi-headed Self-attention (MHSA) mechanisms, followed by a Multilayer Perceptron (MLP) module; stable training is achieved through residual connections. Figure 9 , Figure 10 As shown.

[0049] Local Awareness Attention (LAA) is designed to efficiently capture local details while enabling cross-window information interaction and aggregation. Its core idea is to introduce a hierarchical computation mechanism between window-level coarse-grained routing and local multi-head self-attention, thereby effectively enhancing the detail representation capability of features while reducing overall computational complexity. Specifically, let the input features... First, divide it into Each of the following non-overlapping local windows contains Each spatial location (token). In the first stage (window-level routing), a window-level representation is extracted from the pixel-level query (Q) and key (K) vectors: (5) in, A window-level representation of the query vector. A window-level representation of the key vector. Display window The Middle A query vector for each token. Representative window The Middle The key vectors of each token. A similarity matrix is ​​calculated based on window-level Q / K. (6) The dimension of the key vector is represented by T, which represents the transpose.

[0050] And for each window Select the one with the highest similarity A target window is formed, creating a sparse route set. .

[0051] In the second stage (local multi-head self-attention), only within the routing window... Internal multi-head scaling dot product self-attention calculation: (7) in For attention head index, , , These are the query, key, and value vectors for this head, respectively. Unlike standard self-attention, the computational scope of this stage is limited to within a highly relevant window, thus effectively focusing on local high-frequency details (such as texture and edges) while maintaining multi-view feature modeling capabilities, avoiding the dilution effect of global attention.

[0052] This two-stage structure of "global routing and local multi-head precision calculation" enables the model to efficiently filter key contextual information globally, while simultaneously achieving detailed feature characterization locally. While significantly reducing computational costs, this mechanism effectively enhances the model's ability to capture local details.

[0053] The Multilayer Perceptron (MLP) module consists of two linear transformation layers, one depthwise separable convolutional layer, and a non-linear activation function. This structure aims to perform a linear mapping along the channel dimension for each input feature, while introducing local spatial awareness and enhancing the model's expressive power through non-linear activation, thereby strengthening its ability to model complex feature patterns. Based on this, the outputs of the local attention branch and the global attention branch are input to the first Dynamic Fusion Module (DFM) for adaptive fusion. DFM dynamically selects and aggregates the most discriminative features, optimizing the fusion effect of cross-branch features and reducing redundant computation through feature selection mechanisms, thus improving overall training efficiency and model performance. Its structure is as follows: Figure 11 As shown.

[0054] The first dynamic fusion module adaptively fuses local texture features with global structural distribution features to obtain spatial structural features in the two-dimensional fused image. The second dynamic fusion module uses a dynamic attention fusion mechanism to adaptively fuse one-dimensional fused features with spatial structural features across modalities to obtain fused features. In a further implementation, the second dynamic fusion module comprises three parts: differential feature extraction, dynamic convolution optimization, and adaptive attention fusion. In this embodiment, the purpose of dynamic fusion is to efficiently integrate feature information from different sources; its overall structure is as follows: Figure 11 As shown, the first dynamic fusion module and the second dynamic fusion module in this embodiment have the same structure. Compared with the traditional static fusion method, static weighting or simple concatenation often fails to fully characterize the unbalanced relationship between cross-modal features, easily leading to information redundancy or feature submersion. Therefore, the dynamic fusion module proposed in this invention focuses more on adaptive interaction and multi-scale modeling between features in its structural design. The core architecture of this module consists of three parts: differentiated feature extraction, dynamic convolution module, and adaptive attention fusion. In the feature fusion stage, this invention innovatively introduces a dynamic attention fusion mechanism: firstly, adaptive weight calibration of feature channels is achieved through channel attention, highlighting the contribution of key modalities; then, local features are extracted in parallel through multi-scale dynamic convolution, where dynamic convolution utilizes learnable attention weights to achieve adaptive fusion among the outputs of multiple convolution kernels, thereby dynamically adjusting the feature extraction method according to the differences in input features, significantly improving the flexibility and adaptability of feature representation. The specific structure of the dynamic convolution is as follows... Figure 12As shown, this module extracts global and salient response information by performing global average pooling (GAP) and global max pooling (GMP) on the input features. Subsequently, a set of normalized weight coefficients is generated through convolutional layers and Softmax, which are used to weight and combine multiple standard convolutional kernels, thereby adaptively generating input-related dynamic convolutional kernels to achieve adaptive feature enhancement and expression. Finally, the fused features are weighted and integrated through a fully connected layer to complete regression prediction, and supervised and optimized using the Loss2 loss function.

[0055] The advantages of this design are as follows: the dynamic convolution module enhances local feature modeling capabilities while maintaining model lightweightness through a parameter-sharing mechanism, effectively avoiding the problem of limited expressive power of fixed convolution kernels; the multi-scale attention fusion strategy breaks through the limitations of traditional single-scale modeling, achieving a dynamic balance between local details and global semantics. Experiments show that this architecture achieves efficient fusion of cross-modal features through multi-scale feature modeling and dynamic weight allocation, significantly improving prediction accuracy while maintaining computational efficiency.

[0056] In this invention, both the reconstruction loss Loss1 and the regression loss Loss2 during training use Mean Square Error (MSE) as the loss function. MSE is a commonly used numerical standard, and its calculation formula is as follows: (8) In the formula, n Represents the total number of sample points. Represents the actual observed value. This represents the model's predicted value.

[0057] The feature vectors F1 and F2 output from the two branches are fused and then input into the final fully connected layer for soil salinity prediction. To improve the model's generalization ability and structural constraints, this invention uses a joint loss function to optimize the model: the main loss term (Loss2) is the mean squared error (MSE) between the predicted and reference values; the auxiliary loss term (Loss1) is the reconstruction error from the Tucker decomposition, used to regularize the low-rank structure of the tensor and prevent overfitting during feature extraction. The final total loss function is as follows: (9) Among them, the weighting coefficient By controlling the contribution of auxiliary loss, the model can be effectively guided to maintain the interpretability and compactness of the input feature structure while ensuring prediction accuracy.

[0058] This embodiment also provides a specific model training and testing process. Model training: In the microwave band, changes in soil moisture and salinity affect soil conductivity, and consequently, its dielectric constant. Therefore, a soil dielectric model is fundamental for microwave remote sensing inversion of soil moisture and salinity. The real part of the dielectric constant of saline soil is primarily controlled by soil moisture, while the imaginary part is determined by both moisture and salinity. Therefore, this invention uses soil salinity obtained from soil moisture results provided by SMAP through the Wang empirical model and an improved Dobson-S model as the model reference value, and inverts the soil salinity through a cross-modal dynamic fusion network model. The Wang empirical model and the improved Dobson-S model are as follows: (10) (11) In the formula, The complex permittivity of soil, Represents soil moisture. j Represents the imaginary unit ( ), used to distinguish between the real and imaginary parts of the complex permittivity, Represents soil salinity. The imaginary part of the dielectric constant, parameter Depending on the types of salt ions contained in the soil, the value used in this invention is 0.78; and These represent soil bulk density and soil particle density, respectively. For frequency; Free space conductivity ( ); The shape factor is set to 0.65. The first-order fitting coefficient for soil solution conductivity and salinity is 0.14. The correction coefficients for the sand content and clay content in the soil are expressed by formula (12); The temperature compensation coefficient is expressed by formula (13): (12) (13) In the formula, , Indicates surface temperature; Indicates the proportion of sand in the soil; This indicates the proportion of clay in the soil.

[0059] The model input consisted of DDM images and auxiliary parameters such as corrected reflectance. The dataset based on SMAP data contained 15,000 sample points, and 80% of this dataset was used as the training set. During training, the model hyperparameters were set as follows: learning rate: 0.001; epoch: 150; batch size: 256; optimizer: AdamW; loss function: MSE; dropout parameter: 0.1. The experimental development environment used the PyTorch library.

[0060] The CDF-Net model first transmits input data to the output layer through layer-by-layer forward propagation to generate prediction results. Then, based on a gradient descent strategy, it adjusts network parameters to gradually reduce the error between the predicted output and the reference value. This study uses the AdamW optimization algorithm during training, an improved version of the Adam optimizer. This algorithm makes the optimization more robust by separating weight decay. The AdamW optimizer can adaptively adjust the learning rate and momentum values ​​according to the current training state. Simultaneously, by independently introducing a weight decay mechanism, it more accurately controls the scale of model parameters, thus significantly improving the model's training performance. Compared with other optimization algorithms, the AdamW optimizer not only obtains momentum by calculating the gradient mean but also separates the regularization process from the gradient update process, avoiding the adverse effects of regularization on momentum estimation. By introducing momentum, the AdamW optimizer can integrate historical gradient information during parameter updates, making the training process smoother and more stable. The adaptive learning rate automatically adjusts the learning rate and update step size of each parameter by dynamically estimating the gradient variance to adapt to the importance and variation of different parameters. Furthermore, AdamW performs weight decay before gradient correction during parameter updates, effectively improving training convergence speed and model generalization ability. The parameter update principle of the AdamW optimizer is as follows: (14) in, Represents the loss function; Indicates the current parameter The gradient below; and These represent the first and second moments of the gradient, respectively. and These are the decay factors of the first and second moments of the gradient, respectively; and This indicates the revised result of the estimate; This represents a numerically stable term, used to prevent the denominator from being zero; This represents the weight decay coefficient; This represents the number of iterations. This represents the learning rate.

[0061] Model Testing and Soil Salinity Inversion Results: 20% of the dataset was selected as the test set and input into the trained model. The inversion accuracy of the cross-modal dynamic fusion network model was evaluated using two performance metrics: correlation coefficient (R) and root mean square error (RMSE). The correlation coefficient (R) measures the linear relationship between two variables, ranging from -1 to 1; a value closer to 1 indicates better model performance. The root mean square error (RMSE), a commonly used metric, measures the error between predicted and actual values; a value closer to 0 indicates higher model accuracy. The calculation methods for RMSE and correlation coefficient (R) are as follows: (15) (16) in, These are actual observed values. These are the model's predicted values. Represents the total number of sample points. and These are the average of the predicted values ​​and the average of the actual observed values, respectively.

[0062] S3: Soil salinity inversion is performed on a preset area based on a cross-modal dynamic fusion model to obtain soil salinity inversion results. To demonstrate the advantages of this method, an economic and ecological zone (116°55′-120°19′E, 36°25′-38°14′N) is used as an example. This method is compared with the following three methods: the Transformer-based soil salinity inversion method, the XGBoost-based soil salinity inversion method, and the Artificial Neural Network-based soil salinity inversion method. The main focus is on the quantitative comparison and analysis of the inversion results. Figure 13 , Figure 14 As shown.

[0063] Following the "inversion-comparison" method, soil salinity data were selected and inverted using the CDF-Net, Transformer, XGBoost, and ANN methods, respectively. The results were then statistically analyzed, and the final results are shown in Table 1. The table shows that the CDF-Net method yielded the best inversion results, with a correlation coefficient of 0.9538 between the inversion results and the soil salinity reference values, representing improvements of 6.6%, 8.9%, and 12.4% compared to the Transformer, XGBoost, and ANN methods, respectively. Its RMSE was the lowest at 0.6047, decreasing by 32.0%, 37.7%, and 53.0% compared to the other three methods, respectively. This indicates that the proposed method significantly outperforms the comparative methods in both inversion accuracy and stability.

[0064] Table 1

[0065] By applying this technical solution to specific examples, the present invention can draw the following two conclusions: (1) The inversion accuracy of the CDF-Net method was verified using SMAP satellite data. The average RMSE was 0.6047, and the average correlation coefficient between the predicted data and the actual data was 0.9538, which proved the effectiveness of the method. (2) Compared with the Transformer method, XGBoost method and ANN method, the CDF-Net method has higher inversion accuracy, indicating that the method can effectively improve the inversion accuracy.

[0066] Example 2: This invention relates to a GNSS-R soil salinity inversion system based on cross-modal fusion, and provides a method for implementing this method, including: The data processing module is used to perform data fusion and quality control on the Tianmu-1 satellite observation dataset and the SMAP satellite auxiliary dataset to obtain DDM images and multi-dimensional auxiliary features; the model building module is used to construct a cross-modal dynamic fusion model based on the DDM images and multi-dimensional auxiliary features, combined with the T-GateFusionNet module, the HybridFormer module, and the second dynamic fusion module; the salinity inversion module is used to invert the soil salinity of a preset area based on the cross-modal dynamic fusion model to obtain the soil salinity inversion results.

[0067] A further implementation method is that, in the model building module, the cross-modal dynamic fusion model includes: The T-GateFusionNet module extracts physical prior features from multidimensional auxiliary features and uses Tucker tensor decomposition to analyze the core tensor statistics and singular value distribution of the factor matrix of the DDM image. It then combines a gated attention mechanism to fuse the physical prior features, core tensor statistics, and singular value distribution of the factor matrix to obtain a one-dimensional fused feature. The HybridFormer module performs feature projection and spatial fusion on the DDM image and multidimensional auxiliary features to obtain a two-dimensional fused image. It also employs a parallel dual-path Transformer structure to jointly model the local texture features and global structural distribution features of the two-dimensional fused image, obtaining the spatial structural features. The second dynamic fusion module uses a dynamic attention fusion mechanism to adaptively fuse the one-dimensional fused feature and the spatial structural feature across modalities to obtain the fused feature.

[0068] The embodiments described above are merely preferred embodiments of the present invention and are not intended to limit the scope of the present invention. Various modifications and improvements made to the technical solutions of the present invention by those skilled in the art without departing from the spirit of the present invention should fall within the protection scope defined by the claims of the present invention.

Claims

1. A GNSS-R soil salinity inversion method based on cross-modal fusion, characterized in that, include: Data fusion and quality control were performed on the Tianmu-1 satellite observation dataset and the SMAP satellite auxiliary dataset to obtain DDM images and multidimensional auxiliary features. Based on DDM images and multidimensional auxiliary features, a cross-modal dynamic fusion model is constructed by combining the T-GateFusionNet module, the HybridFormer module, and the second dynamic fusion module. Based on the cross-modal dynamic fusion model, the soil salinity of the preset area is inverted to obtain the soil salinity inversion results; The T-GateFusionNet module includes: The residual module is used to effectively characterize and enhance the multidimensional auxiliary features to obtain physical prior features; The Tucker low-rank decomposition module is used to perform Tucker decomposition on the DDM image to obtain the core tensor statistics and the singular value distribution of the factor matrix, and to concatenate the core tensor statistics and the singular value distribution of the factor matrix to obtain a one-dimensional feature vector. The gated fusion module is used to adaptively fuse the physical prior features and the one-dimensional feature vector using a dual-path collaborative mechanism to obtain a one-dimensional fused feature. The HybridFormer module includes: The conditional encoding module is used to perform layer normalization on the multidimensional auxiliary features, and project them onto the same spatial dimension as the DDM image through a linear layer to obtain a spatial attention map. The spatial attention map is then concatenated with the original DDM image in the channel dimension, and feature fusion is completed through 3×3 convolution to obtain a two-dimensional fused image. The multi-scale residual enhancement module is used to model the two-dimensional fused image in parallel under different receptive fields using a multi-branch structure, and to stitch the features extracted by each branch in the channel dimension to obtain multi-scale spatial information. A downsampling layer is used to downsample the multi-scale spatial information using a differentiated downsampling strategy; the differentiated downsampling strategy includes a conservative downsampling strategy and an aggressive downsampling strategy; The local attention module is used to downsample the multi-scale spatial information using the conservative downsampling strategy to obtain local texture features; The global attention module is used to aggregate the global context information of the multi-scale spatial information using the aggressive downsampling strategy to obtain global structural distribution features. The first dynamic fusion module is used to adaptively fuse the local texture features with the global structural distribution features to obtain the spatial structural features in the two-dimensional fused image.

2. The method according to claim 1, characterized in that, The DDM images include raw count DDM images, analog power DDM images, bistatic radar cross section DDM images, and effective scattering area DDM images; The multidimensional auxiliary features include features related to DDM and features related to the land surface; The DDM-related features include: corrected surface reflectance, incident angle, DDM mean, DDM peak value, signal-to-noise ratio, and the row and column of the DDM peak value; The surface-related characteristics include: time, longitude of the specular reflection point, latitude of the specular reflection point, vegetation optical thickness, surface roughness, surface temperature, vegetation moisture content, soil bulk density, and the proportion of clay to sand in the soil.

3. The method according to claim 1, characterized in that, In the cross-modal dynamic fusion model The T-GateFusionNet module is used to extract physical prior features from the multidimensional auxiliary features, and to use Tucker tensor decomposition and parsing of the core tensor statistics and the singular value distribution of the factor matrix of the DDM image, combined with a gated attention mechanism to fuse the physical prior features, the core tensor statistics and the singular value distribution of the factor matrix to obtain a one-dimensional fused feature. The HybridFormer module is used to perform feature projection and spatial fusion on the DDM image and multi-dimensional auxiliary features to obtain a two-dimensional fused image. It also uses a parallel dual-path Transformer structure to jointly model the local texture features and global structural distribution features of the two-dimensional fused image to obtain the spatial structural features in the two-dimensional fused image. The second dynamic fusion module is used to adaptively fuse the one-dimensional fusion feature and the spatial structure feature across modal features using a dynamic attention fusion mechanism to obtain the fusion feature.

4. The method according to claim 3, characterized in that, The second dynamic fusion module comprises three parts: differential feature extraction, dynamic convolution optimization, and adaptive attention fusion.

5. A GNSS-R soil salinity inversion system based on cross-modal fusion, used to implement the method described in any one of claims 1-4, characterized in that, include: The data processing module is used to perform data fusion and quality control on the Tianmu-1 satellite observation dataset and the SMAP satellite auxiliary dataset to obtain DDM images and multidimensional auxiliary features. The model building module is used to construct a cross-modal dynamic fusion model based on DDM images and multi-dimensional auxiliary features, combined with the T-GateFusionNet module, HybridFormer module and second dynamic fusion module. The salinity inversion module is used to invert the soil salinity of a preset area based on the cross-modal dynamic fusion model to obtain the soil salinity inversion results.

6. The system according to claim 5, characterized in that, In the model building module, the cross-modal dynamic fusion model includes: The T-GateFusionNet module is used to extract physical prior features from the multidimensional auxiliary features, and to use Tucker tensor decomposition and parsing of the core tensor statistics and the singular value distribution of the factor matrix of the DDM image, combined with a gated attention mechanism to fuse the physical prior features, the core tensor statistics and the singular value distribution of the factor matrix to obtain a one-dimensional fused feature. The HybridFormer module is used to perform feature projection and spatial fusion on the DDM image and multi-dimensional auxiliary features to obtain a two-dimensional fused image. It also uses a parallel dual-path Transformer structure to jointly model the local texture features and global structural distribution features of the two-dimensional fused image to obtain the spatial structural features in the two-dimensional fused image. The second dynamic fusion module is used to adaptively fuse the one-dimensional fusion feature and the spatial structure feature across modal features using a dynamic attention fusion mechanism to obtain the fusion feature.

Images