Soil saline-alkali soil monitoring method and system based on image data processing

Abstract

The invention relates to the technical field of saline-alkali soil monitoring, and discloses a soil saline-alkali soil monitoring method and system based on image data processing, and the method comprises the steps: generating image data and ground sample labels adaptive to a network codec after radiometric calibration, denoising and standardization processing; inputting the standardized image data into an encoder capable of dynamically feeding back and decoupling; inputting the double-feature graph into a self-adaptive physical constraint generator; calculating the deviation between the initial salinity distribution map and a ground sample label, the deviation between the double-feature map and an ideal feature, and the deviation between a generation result and a physical rule; and outputting the salt distribution diagram and the salt mist distribution diagram after multiple rounds of calibration through the generator. Through the collaborative design of dynamic feedback decoupling coding and self-adaptive physical constraint generation, the salt mist scattering characteristics and the salt absorption characteristics are accurately separated, in the scene of salt mist concentration interference, the salt inversion error is effectively reduced, and the problem of difficult decoupling caused by salt mist interference is thoroughly solved.

Description

Technical Field

[0001] This invention relates to the field of saline-alkali land monitoring technology, and more specifically, to a method and system for monitoring saline-alkali soil based on image data processing. Background Technology

[0002] As an important land resource, the salt content and distribution of saline-alkali land directly affect vegetation growth, soil quality and ecological environment stability. Accurate monitoring of relevant parameters of saline-alkali land is of great significance for land improvement, agricultural planning and ecological protection.

[0003] Currently, monitoring of saline-alkali land mainly relies on hyperspectral remote sensing technology. This technology achieves parameter inversion by capturing the differences in the spectral response of different substances and is widely used in large-scale saline-alkali land surveys.

[0004] However, in practical applications, existing hyperspectral saline-alkali land analysis techniques face several intractable challenges: First, salt fog interference makes decoupling difficult. In coastal and foggy saline-alkali land scenarios, the spectral characteristics of salt fog and salinity are severely coupled, and existing technologies cannot effectively separate their information, leading to a significant increase in salinity inversion errors. Second, poor adaptability to small sample scenarios. In some remote saline-alkali areas, due to inconvenient transportation and harsh environments, it is difficult to collect a large number of ground-based measurement samples. Existing technologies rely on massive sample training models, and their generalization performance drops sharply under small sample conditions, failing to meet actual monitoring needs. Third, insufficient adaptability to physical constraints. First, the physical constraints introduced in existing technologies are mostly fixed thresholds or static formulas, which cannot be dynamically adjusted according to the salt spray concentration, land cover type, etc. in different scenarios, and are prone to inversion results that are "visually consistent but mechanistically incorrect". Second, the problem of error accumulation is prominent. Existing technologies mostly adopt a "decoupling-generation" serial architecture, and the errors in the preceding links will continue to be transmitted to the subsequent links, resulting in a decrease in the accuracy of the final monitoring results. Third, the function is limited. Existing technologies can only output salt distribution data and cannot simultaneously obtain salt spray concentration distribution information. Salt spray is an important influencing factor in the evolution of saline-alkali land, and its distribution data has important reference value for the dynamic monitoring of saline-alkali land.

[0005] There are currently no effective solutions to the problems in the relevant technologies. Summary of the Invention

[0006] In view of the problems in related technologies, this invention proposes a method and system for monitoring soil salinity and alkali land based on image data processing, so as to overcome the above-mentioned technical problems existing in the existing related technologies.

[0007] Therefore, the specific technical solution adopted by the present invention is as follows:

[0008] Firstly, a method for monitoring soil salinity and alkali land based on image data processing is provided, including:

[0009] S1. Collect core band hyperspectral images of characteristic responses of saline-alkali land, and generate image data and ground sample labels adapted to network codecs after radiometric calibration, denoising and standardization.

[0010] S2. Input standardized image data into a dynamically decoupled encoder, receive decoupling bias signals and dynamically adjust the encoder's convolution kernel weights and adversarial loss weights, and output optimized dual feature maps; the dual feature maps include salt absorption feature maps and salt spray scattering feature maps.

[0011] S3. Input the dual feature map into the generator of adaptive physical constraints, call the remote sensing physical rule base, and after matching the spectral and spatial features of the dual feature map with the rule base, dynamically adjust the generator parameters and output the initial salinity distribution map.

[0012] S4. Calculate the deviation between the initial salinity distribution map and the ground sample label, the deviation between the dual feature map and the ideal feature, and the deviation between the generated result and the physical rules. Fuse them to generate a standardized decoupled deviation signal and transmit it to the encoder to complete multiple rounds of calibration.

[0013] S5. Output the salt distribution map and salt spray distribution map after multiple rounds of calibration through the generator.

[0014] Furthermore, standardized image data is input into a dynamically decoupled encoder that receives decoupling bias signals and dynamically adjusts the encoder's convolutional kernel weights and adversarial loss weights, outputting optimized dual feature maps including:

[0015] S21. Based on the U-shaped network, build a decoupled coding architecture for the encoder, and configure the input layer, downsampling layer, upsampling layer, output layer, and spectral attention module embedded between the input layer and the downsampling layer;

[0016] S22. Pre-set the decoupling bias signal composed of feature bias and mechanism bias, convert it into adjustment coefficient, and set the encoder parameter dynamic adjustment logic to build a two-stage dynamic decoupling process that connects the initial stage and the feedback stage, so as to realize the dynamic adjustment of convolution kernel weight and adversarial loss weight.

[0017] S23. Using multiple sets of cross-scene saline-alkali land samples as pre-training data, the encoder is pre-trained by model-independent element learning. Small sample feature constraint terms are embedded in the decoupling bias signal receiving channel and weights are set to complete the small sample adaptation optimization.

[0018] Furthermore, a decoupling deviation signal composed of characteristic deviation and mechanistic deviation is pre-set, converted into adjustment coefficients, and encoder parameter dynamic adjustment logic is set to build a two-stage dynamic decoupling process that connects the initial stage and the feedback stage, including:

[0019] S221. Set the Euclidean distance between the salt absorption characteristic map, the salt spray scattering characteristic map and the ideal characteristic map as the characteristic deviation, and set the deviation between the spectral curve of the dual characteristic map and the salt standard spectrum and the salt spray standard spectrum as the mechanism deviation, and combine them to form a decoupling deviation signal.

[0020] S222. Convert the decoupling bias signal into an adjustment coefficient within a preset numerical range. Dynamically adjust the decoupling resistance loss weights according to the product of the base weights and the adjustment coefficients. And dynamically correct the weights of each layer of convolution kernels through a 1×1 convolution kernel.

[0021] S223. Set the adjustment coefficient of the initial stage to the initial value, and the encoder outputs the initial dual feature map according to the default parameters; after receiving the decoupling deviation signal after calibration transmission in the feedback stage, calculate the adjustment coefficient in real time and adjust the encoder parameters, re-execute the decoupling operation, and output the optimized dual feature map.

[0022] Furthermore, the dual feature maps are input into an adaptive physical constraint generator, which calls a remote sensing physical rule base. After matching the spectral and spatial features of the dual feature maps with the rule base, the generator parameters are dynamically adjusted and an initial salinity distribution map is output, including:

[0023] S31. Build the generator's generation architecture, set the input and output data specifications, and insert a physical prior interaction link between the spatial adaptive normalization module and the convolutional layer to realize the real-time injection of physical rules.

[0024] S32. Build a dynamic remote sensing physical rule base, set the calculation logic of physical consistency score, and set dynamic adjustment rules for generator parameters based on physical consistency score. Combine feature map edge information to optimize the generation of initial salt distribution map and realize the real-time constraint of physical rules on the generation process.

[0025] S33. Construct a dual-core loss function architecture and use a training strategy of pre-training and fine-tuning for new scenarios to train the generator to learn the mapping relationship between remote sensing physical rules and generation parameters.

[0026] Furthermore, the generator's generation architecture is constructed, input and output data specifications are defined, and a physical prior interaction link is inserted between the spatial adaptive normalization module and the convolutional layer, including:

[0027] S311. Construct a generation architecture that includes multiple downsampling layers and upsampling layers. Set the input to be the spliced ​​data of salt absorption feature map and salt fog scattering feature map, and the output to be a salt pseudo-color map.

[0028] S312. Physical prior interaction links are inserted between the convolutional kernels and spatial adaptive normalization modules in the downsampling and upsampling layers, respectively, to provide a physical channel for the adaptive injection of physical constraints; the physical prior interaction links include feature extraction, rule invocation, parameter conversion and transmission interface.

[0029] Furthermore, a dynamic remote sensing physical rule base is built, the calculation logic for the physical consistency score is set, and the generator parameter dynamic adjustment rules based on the physical consistency score are defined. The initial salinity distribution map is then optimized by combining feature map edge information, including:

[0030] S321. Build a dynamic remote sensing physical rule base containing multiple scalable rules, including salt absorption rules, salt fog scattering rules, and ground feature association rules.

[0031] S322. Extract the spectral and spatial features of the intermediate feature map of the generator through the physical prior interaction link, and compare the features with the dynamic remote sensing rule base to calculate the physical consistency score.

[0032] If the physical consistency score is less than the preset score, the normalized mean and variance of the spatial adaptive normalization module are dynamically adjusted, and the corresponding constraint rules are strengthened; if the physical consistency score is greater than or equal to the preset threshold, the generator parameters are kept unchanged.

[0033] S323. Using the edge features of the salt absorption feature map and the geographic association rules in the dynamic remote sensing physical rule base, perform physical rationalization optimization on the edge areas of the salt pseudo-color map and output the initial salt distribution map.

[0034] Furthermore, a dual-core loss function architecture is constructed, and a training strategy of pre-training and fine-tuning for new scenarios is used to train the generator to learn the mapping relationship between remote sensing physical rules and generation parameters, including:

[0035] S331. Construct a dual-core loss function architecture that combines physical consistency loss and adversarial loss, and combine the dual-core loss function to generate a total loss function to form an effective constraint on physical rules.

[0036] S332. In the pre-training stage, the generator is iteratively trained using different types of cross-scenario saline-alkali land data as training samples; and in the new scenario fine-tuning stage, the channel parameters of the physical prior interaction link are automatically adjusted according to the newly added monitoring scenario, and the pre-training and fine-tuning are determined to be completed based on the changing trend of the physical consistency score and the salinity inversion index.

[0037] Furthermore, the deviations between the initial salinity distribution map and the ground sample labels, the deviations between the dual-feature map and the ideal features, and the deviations between the generated results and the physical rules are calculated. These are then fused to generate a standardized decoupled deviation signal, which is transmitted to the encoder to complete multiple rounds of calibration, including:

[0038] S41. Establish a deviation calculation and feedback link for multi-unit serial and parallel interfaces. Based on the received output results of the encoder and generator, calculate the sample deviation, feature deviation and physical deviation respectively, and perform weighted summation according to the preset weight ratio. After standardization, generate a decoupled deviation signal.

[0039] S42. Set up an iterative calibration process that includes signal transmission, parameter adjustment, re-decoupling generation, and deviation calculation. Set the iteration trigger conditions, termination criteria, and calibration operations to achieve gradual convergence of errors and output the salt distribution map and salt spray distribution map after iterative calibration.

[0040] S43. A bilateral filter is used to perform a light smoothing process on the salt distribution map and salt spray distribution map to remove abnormal pixels. The abnormal areas are then corrected using a neighborhood physical weighted interpolation method to obtain the final salt distribution map and salt spray distribution map.

[0041] Furthermore, an iterative calibration process is defined, including signal transmission, parameter adjustment, re-decoupling generation, and deviation calculation. The iteration trigger conditions, termination criteria, and calibration operations are also defined.

[0042] S421. When the initial salinity distribution map is generated for the first time, the first round of calibration is automatically triggered. Subsequent rounds of calibration are triggered by the deviation calculation results of the previous round. When the decoupling deviation signal is greater than the preset signal threshold, the next round of calibration is automatically entered.

[0043] S422. The decoupling bias signal is transmitted to the encoder through a bidirectional transmission interface. The encoder adjusts the convolution kernel weights and adversarial loss weights according to the decoupling bias signal and synchronizes the adjustment instructions to the generator to adjust the rule matching threshold of the physical prior interaction link. The encoder and generator regenerate new output results and recalculate and generate the decoupling bias signal.

[0044] S423. When the decoupling deviation signal is less than the preset signal threshold or the number of iterations reaches the maximum number of iterations, the iteration is terminated.

[0045] Secondly, a soil salinity monitoring system based on image data processing is also provided, the system comprising:

[0046] The image acquisition and preprocessing module is used to acquire core band hyperspectral images of the characteristic response of saline-alkali land. After radiometric calibration, denoising and standardization, image data and ground sample labels adapted to the network codec are generated.

[0047] The dynamic feedback decoupling encoding module is used to input standardized image data into a dynamically decoupled encoder, receive decoupling bias signals and dynamically adjust the convolution kernel weights and adversarial loss weights of the encoder, and output optimized dual feature maps; the dual feature maps include salt absorption feature maps and salt spray scattering feature maps;

[0048] The adaptive physical constraint generation module is used to input the dual feature map into the adaptive physical constraint generator, call the remote sensing physical rule base, and after matching the spectral and spatial features of the dual feature map with the rule base, dynamically adjust the generator parameters and output the initial salinity distribution map.

[0049] The closed-loop deviation calibration module is used to calculate the deviation between the initial salinity distribution map and the ground sample label, the deviation between the dual feature map and the ideal feature, and the deviation between the generated result and the physical rules. It fuses and generates a standardized decoupled deviation signal and transmits it to the encoder to complete multiple rounds of calibration.

[0050] The result output verification module is used to generate and output the salt distribution map and salt spray distribution map after multiple rounds of calibration.

[0051] The beneficial effects of this invention are as follows:

[0052] 1. Through a collaborative design of dynamic feedback decoupling encoding and adaptive physical constraint generation, the salt spray scattering features and salt absorption features are accurately separated. In scenarios with salt spray concentration interference, the salt inversion error is effectively reduced, and the decoupling difficulty caused by salt spray interference is completely solved. At the same time, by combining meta-learning pre-training and small sample feature constraint design, high-precision monitoring can be achieved with only a small number of ground samples. It is suitable for monitoring saline-alkali land in remote and harsh environments, and significantly reduces the cost and difficulty of sample collection.

[0053] 2. Through a dynamic remote sensing physical rule base and real-time consistency verification design, the physical consistency score of the generated results meets the standard requirements, avoiding mechanistic errors. The constraint strategy can be dynamically adjusted according to the salt spray concentration and land cover type of different scenarios, adapting to various saline-alkali land scenarios such as coastal areas and arid and semi-arid areas. At the same time, the closed-loop deviation calibration architecture realizes the iterative optimization of decoupling-generation-calibration, corrects the errors of each link in real time, effectively reduces the accumulation of errors in the serial architecture, and ensures accuracy attenuation under extreme salt spray scenarios. Attached Figure Description

[0054] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0055] Figure 1 This is a flowchart of a method for monitoring soil salinity based on image data processing according to an embodiment of the present invention;

[0056] Figure 2 This is an architecture diagram of the encoder in a soil saline-alkali land monitoring method based on image data processing according to an embodiment of the present invention;

[0057] Figure 3 This is an architecture diagram of the generator in a soil salinity monitoring method based on image data processing according to an embodiment of the present invention.

[0058] Figure 4 This is a system principle block diagram of a soil saline-alkali land monitoring system based on image data processing according to an embodiment of the present invention.

[0059] In the picture:

[0060] 1. Image acquisition and preprocessing module; 2. Dynamic feedback decoupling coding module; 3. Adaptive physical constraint generation module; 4. Closed-loop deviation calibration module; 5. Result output verification module. Detailed Implementation

[0061] According to an embodiment of the present invention, a method for monitoring soil salinity based on image data processing is provided.

[0062] The present invention will now be further described in conjunction with the accompanying drawings and specific embodiments, such as... Figure 1 As shown, the soil salinity monitoring method based on image data processing according to an embodiment of the present invention includes:

[0063] S1. Collect core band hyperspectral images of characteristic responses of saline-alkali land, and generate image data and ground sample labels adapted to network codecs after radiometric calibration, denoising and standardization.

[0064] Specifically, UAVs equipped with ResononPikaL hyperspectral cameras or Hyperion satellites were used to acquire images in five core bands: 450nm, 850nm, 1450nm, 1900nm, and 2200nm, with consistent bandwidth across each band. The UAV acquisition resolution was set to 0.5m, and the satellite acquisition resolution was set to 30m. Atmospheric visibility and solar altitude angle were recorded simultaneously during the acquisition process. Only 30-50 sets of ground samples were needed for a single scene, including measured values ​​of salt content and salt spray concentration, with measured errors conforming to relevant industry standards. Multi-source data registration employed a simplified SIFT algorithm, reducing the number of Gaussian pyramid layers to 4 and setting the feature point threshold to 0.05 to ensure that at least 80 matching feature points were used, with a registration RMSE ≤ 0.3 pixels, thus improving processing efficiency while ensuring registration accuracy.

[0065] Radiometric calibration was performed using a laboratory calibration board to convert reflectance, ensuring a conversion error of ≤3% without the need for complex atmospheric correction. Subsequently, atmospheric effects were compensated for during the generation process by calling the remote sensing physical rule library. Lightweight denoising employed a 3×3 guided filter with the 1900nm band as the guide image, resulting in a signal-to-noise ratio (SNR) ≥42dB for the denoised image. Reflectance was normalized to the [0,1] interval and processed into blocks of 256×256 size, with mirror filling used for edge regions. Finally, network input data and a small number of ground sample labels were generated for subsequent initial bias calculations.

[0066] S2. Input standardized image data into a dynamically decoupled encoder, receive decoupling bias signals and dynamically adjust the encoder's convolution kernel weights and adversarial loss weights, and output optimized dual feature maps; the dual feature maps include salt absorption feature maps and salt fog scattering feature maps.

[0067] In the description of this invention, standardized image data is input to a dynamically decoupled encoder, decoupling bias signals are received, and the kernel weights and adversarial loss weights of the encoder are dynamically adjusted. The output optimized dual feature map includes:

[0068] S21. Based on the U-shaped network, build a decoupled coding architecture for the encoder and configure the input layer, downsampling layer, upsampling layer, output layer, and spectral attention module embedded between the input layer and the downsampling layer.

[0069] Specifically, such as Figure 2 As shown, this invention uses U-Net as its basic framework, setting up 4 downsampling layers and 4 upsampling layers. The downsampling layers adopt a combination structure of "3×3 convolutional kernel + batch normalization + LeakyReLU". After each downsampling layer, the feature map size is halved and the number of channels is doubled, realizing feature extraction and dimensionality compression. The upsampling layers adopt a structure of "4×4 transposed convolutional kernel + batch normalization". After each upsampling layer, the feature map size is doubled and the number of channels is halved, realizing feature map dimensionality restoration and detail fusion. The upsampling and downsampling units are connected by skip connections to retain low-level spatial detail features.

[0070] Meanwhile, this invention introduces a spectral attention module between the encoder input layer and the first downsampling layer. This module consists of a channel attention submodule and a spatial attention submodule connected in series: the channel attention submodule converts the input feature map into a channel feature vector through global average pooling, and learns the core band weights through a fully connected layer; the spatial attention submodule extracts spatial features through a 7×7 convolution kernel to generate a spatial weight map; finally, adaptive weight allocation is achieved through the element-wise product of "channel weight × spatial weight", which focuses on strengthening the feature response of core bands such as 450nm, 1900nm, and 2200nm, suppressing noise interference in non-core bands, and improving the distinguishability between salt spray and salt features to over 0.85.

[0071] S22. A decoupling bias signal consisting of feature bias and mechanism bias is pre-set, converted into adjustment coefficients, and encoder parameter dynamic adjustment logic is set to build a two-stage dynamic decoupling process that connects the initial stage and the feedback stage, so as to realize the dynamic adjustment of convolution kernel weights and adversarial loss weights.

[0072] Specifically, a decoupling bias signal consisting of feature bias and mechanism bias is defined and converted into an adjustment coefficient α in the range of 0.1-0.9 using a min-max normalization algorithm. A dynamic adjustment logic for encoder convolution kernel weights and loss weights based on the adjustment coefficient is designed, and a two-stage dynamic decoupling process of initial decoupling, bias feedback and secondary decoupling is built to ensure that the decoupling accuracy gradually improves with feedback iteration.

[0073] In the description of this invention, a decoupling deviation signal composed of characteristic deviation and mechanism deviation is pre-set, converted into an adjustment coefficient, and encoder parameter dynamic adjustment logic is set to build a two-stage dynamic decoupling process that connects the initial stage and the feedback stage, including:

[0074] S221. Set the Euclidean distance between the salt absorption characteristic map, the salt spray scattering characteristic map and the ideal characteristic map as the characteristic deviation, and set the deviation between the spectral curve of the dual characteristic map and the salt standard spectrum and the salt spray standard spectrum as the mechanism deviation, and combine them to form a decoupling deviation signal.

[0075] Specifically, the decoupling bias signal is composed of a 1:1 weighted average of characteristic bias and mechanistic bias. The characteristic bias is defined as the L2 distance (Euclidean distance) between the salt absorption characteristic map and salt fog scattering characteristic map output by the encoder and the ideal characteristic map. The ideal characteristic map is calibrated by bilinear interpolation using 30-50 sets of ground sample tags, with a calculation accuracy error ≤0.001. The mechanistic bias is defined as the Euclidean distance between the spectral curves corresponding to the above dual characteristic maps and the standard salt spectrum and standard salt fog spectrum. The standard spectra are obtained by fitting multiple sets of samples in the laboratory, with a wavelength matching error ≤2nm.

[0076] S222. Convert the decoupling bias signal into an adjustment coefficient within a preset numerical range. Adjust the decoupling resistance loss weights dynamically by multiplying the base weights by the adjustment coefficients. Dynamically correct the weights of each convolution kernel layer using a 1×1 convolution kernel.

[0077] Specifically, the decoupling deviation signal is converted into an adjustment coefficient α in the range of 0.1-0.9 using the min-max normalization algorithm. The normalization formula is α=(decoupling deviation signal-minimum deviation value) / (maximum deviation value-minimum deviation value), where the minimum deviation value is preset to 0 and the maximum deviation value is preset to 0.5.

[0078] The base weight of the decoupling adversarial loss is set to 0.6, and it is dynamically adjusted according to "base weight × adjustment coefficient α", where α increases as the bias increases. The dynamic adjustment of the weights of each convolutional kernel is achieved through a 1×1 convolutional kernel, and the adjustment formula is as follows: ;in, To adjust the weights of the convolutional kernel, These are the original convolutional kernel weights. This is the deviation value. This represents the maximum deviation value.

[0079] S223. Set the adjustment coefficient of the initial stage to the initial value, and the encoder outputs the initial dual feature map according to the default parameters; after receiving the decoupling deviation signal after calibration transmission in the feedback stage, calculate the adjustment coefficient in real time and adjust the encoder parameters, re-execute the decoupling operation, and output the optimized dual feature map.

[0080] Specifically, in the initial stage, the adjustment coefficient α=0.5 is set, and the encoder is set to default parameters, that is, the initial weights of the convolution kernel follow a normal distribution N(0,0.02) and the decoupling adversarial loss weight=0.3, and the initial salt absorption feature map and salt fog scattering feature map are output.

[0081] After receiving the decoupling deviation signal transmitted from the closed-loop calibration link during the feedback phase, the adjustment coefficient α is calculated and the encoder parameters are adjusted within 10ms. After adjustment, the decoupling operation is re-executed, and the optimized salt absorption characteristic map and salt spray scattering characteristic map are output to ensure that the decoupling error is reduced by more than 30%.

[0082] S23. Using multiple sets of cross-scene saline-alkali land samples as pre-training data, the encoder is pre-trained by model-independent element learning. Small sample feature constraint terms are embedded in the decoupling bias signal receiving channel and weights are set to complete the small sample adaptation optimization.

[0083] Specifically, this invention uses 100 sets of cross-scene saline-alkali land samples as pre-training data, including light, moderate, and heavy saline-alkali land and different salt fog concentration scenarios, to perform model-independent element learning pre-training on the dynamic feedback decoupling encoder; the pre-training uses the Adam optimizer, with a learning rate of 2e-4, β1=0.5, a batch size of 8, and 10,000 iterations, and the stopping condition is that the decoupling error of the validation set fluctuates ≤0.001 for 100 consecutive iterations; through pre-training, the encoder learns the general decoupling rules across scenarios and has the ability to quickly adapt to new scenarios with a small amount of data.

[0084] Furthermore, a small-sample feature constraint term is embedded in the encoder's deviation signal receiving channel. The formula for the constraint term is as follows: Where F is the current feature, The prototype features are defined as N, the number of samples, and the weight of the small sample feature constraint term is set to 0.2. Based on the overfitting risk assessment in the small sample scenario, the constraint term weight is set to 0.2, which is co-optimized with the decoupled adversarial loss. Overfitting is suppressed through gradient backpropagation, effectively improving the test set accuracy in the small sample scenario.

[0085] S3. Input the dual feature map into the generator of adaptive physical constraints, call the remote sensing physical rule base, and after matching the spectral and spatial features of the dual feature map with the rule base, dynamically adjust the generator parameters and output the initial salinity distribution map.

[0086] In the description of this invention, the process of inputting a dual-feature map into an adaptive physical constraint generator, calling a remote sensing physical rule base, matching the spectral and spatial features of the dual-feature map with the rule base, dynamically adjusting the generator parameters, and outputting an initial salinity distribution map includes:

[0087] S31. Build the generator's generation architecture, set the input and output data specifications, and insert a physical prior interaction link between the spatial adaptive normalization module and the convolutional layer to realize the real-time injection of physical rules.

[0088] In the description of this invention, the generation architecture of the generator, the input and output data specifications, and the insertion of a physical prior interaction link between the spatial adaptive normalization module and the convolutional layer include:

[0089] S311. Construct a generation architecture that includes multiple downsampling layers and upsampling layers. Set the input as the stitched data of salt absorption feature map and salt fog scattering feature map, and the output as a salt pseudo-color image.

[0090] Specifically, such as Figure 3 As shown, this invention uses SPADE-U-Net as its basic framework, specifically configured as follows: it includes 4 downsampling layers and 4 upsampling layers. The downsampling layers adopt a structure of "3×3 convolution kernel + batch normalization (BN) + LeakyReLU activation function" with a stride of 2 to achieve feature dimensionality reduction; the upsampling layers adopt a structure of "transposed convolution + BN + ReLU activation function" to achieve feature dimensionality increase; the input is set to the concatenated data of salt absorption feature map and salt fog scattering feature map, and the size after the number of channels is superimposed is 256×256×2; the output is a 256×256×3 salt pseudo-color image, with 3 channels corresponding to the RGB color space, respectively mapping low, medium and high salt levels, to meet the technical requirements of subsequent result processing and verification.

[0091] SPADE-U-Net represents a high-level U-shaped network based on spatial adaptive normalization. U-Net is a classic deep learning model widely used for tasks such as image segmentation. It consists of an encoder and a decoder, and uses skip connections to pass detailed information from shallow layers to deeper layers to generate refined results. Traditional U-Nets may produce blurry or inaccurate images if the semantic map is directly input, because conventional normalization operations wash away important information from the semantic map. The core improvement of SPADE-U-Net is that it replaces the traditional normalization layer with a SPADE module (spatial adaptive normalization module), thereby better utilizing and preserving the input semantic information.

[0092] S312. Physical prior interaction links are inserted between the convolutional kernels and spatial adaptive normalization modules in the downsampling and upsampling layers, respectively, to provide a physical channel for the adaptive injection of physical constraints. The physical prior interaction links include feature extraction, rule invocation, parameter transformation, and transmission interfaces.

[0093] Specifically, a physical prior interaction link is designed between the SPADE module of the generator and the convolutional layer. This link is a serial multi-unit architecture, consisting of four parts: a feature extraction unit, a rule calling unit, a parameter transformation unit, and a transmission interface. Each unit is connected in series via a high-speed data bus to achieve real-time data interaction: 1) Feature Extraction Unit: A 1×1 convolutional kernel is used to reduce the dimensionality and aggregate features of the intermediate feature map output by the SPADE module, extracting spectral feature vectors (dimension = number of core bands) and spatial feature vectors. The spectral feature vector dimension = number of core bands, and the spatial feature vector dimension = 4, corresponding to the upper, lower, left, and right neighborhood information; 2) Rule Calling Unit: A built-in API interface establishes communication with the dynamic remote sensing physical rule base, calling corresponding rules according to the logic of "feature type matching rule type," such as matching spectral features with salt absorption / salt fog scattering rules, and so on. 3) Parameter conversion unit: The fully connected layer converts the rule matching results into parameter adjustment signals that can be directly applied to the SPADE module. The input dimension of the fully connected layer = feature vector dimension + rule parameter dimension, and the output dimension = SPADE module normalized parameter dimension; 4) Transmission interface: The parallel data transmission protocol is used to transmit the adjustment signal to the normalized parameter control terminal of the SPADE module in real time, with a transmission rate ≥100Mbps. This link provides a reliable channel support for the adaptive injection of physical constraints, and each unit supports modular replacement to adapt to different scenarios.

[0094] S32. Build a dynamic remote sensing physical rule base, set the calculation logic of physical consistency score, and set dynamic adjustment rules for generator parameters based on physical consistency score. Combine feature map edge information to optimize the generation of initial salt distribution map and realize the real-time constraint of physical rules on the generation process.

[0095] In the description of this invention, the construction of a dynamic remote sensing physical rule base, the setting of the calculation logic for physical consistency scores, and the dynamic adjustment rules for generator parameters based on physical consistency scores, combined with feature map edge information, optimizes the generation of an initial salinity distribution map, including:

[0096] S321. Build a dynamic remote sensing physical rule base containing multiple scalable rules, including salt absorption rules, salt fog scattering rules, and ground feature association rules.

[0097] Specifically, a dynamic remote sensing physical rule base containing three core scalable rules is constructed, employing a classification index + key-value pair storage structure: 1) Storage structure: The top layer establishes an index categorized by "salt absorption, salt fog scattering, and ground object association," while the bottom layer stores specific rules in key-value pair format of "rule ID - rule content - applicable conditions - weight coefficient," supporting quick retrieval by rule ID; 2) Core rule details: ① Salinity absorption rule: The intensity ratio of the spectral absorption peaks in the 1900nm and 2200nm bands (I1900 / I2200) and the salt content (C, unit g / kg) satisfy a linear relationship C=k×(I1900 / I2200)+b, where k=85.2 and b=-12.3. The applicable conditions are: salt content 0-50g / kg; ② Salt spray scattering rule: the reflectivity (R450) of the 450nm band and the salt spray concentration (Cfog, unit mg / m³) satisfy Cfog=120×R450, and it is clear that the scattering of this band does not affect the position of the 1900nm and 2200nm salt absorption peaks; ③ Land feature association rule: the salt content in the vegetation area is close to 0, and the salt distribution in the bare soil area satisfies spatial continuity; 3) Expansion mechanism: a reserved interface for adding rules is provided, which supports adding custom rules according to the principle of "adding new rule IDs in the same category". After adding, the classification index is automatically updated; 4) Calling priority: land feature association rule > salt absorption rule > salt spray scattering rule to avoid rule conflicts.

[0098] S322. Extract the spectral and spatial features of the intermediate feature map of the generator through the physical prior interaction link, and compare the features with the dynamic remote sensing rule base to calculate the physical consistency score. If the physical consistency score is less than the preset score, dynamically adjust the normalized mean and variance of the spatial adaptive normalization module and strengthen the corresponding constraint rules. If the physical consistency score is greater than or equal to the preset threshold, ensure that the generator parameters remain unchanged.

[0099] Specifically, after feature extraction and rule invocation are completed through the physical prior interaction link, physical consistency verification and parameter adjustment are implemented according to the following steps: 1) Feature comparison: Feature comparison is performed using the method of "vector cosine similarity + threshold screening". When the cosine similarity between the spectral feature vector and the standard spectral vector in the rule base is ≥0.85 and the deviation between the spatial feature vector and the rule constraint vector is ≤0.1, it is judged as a preliminary match; 2) Score calculation: The physical consistency score S is calculated using the weighted summation formula, S=ω1×S1+ω2×S2+ω3×S3, where ω1=0.4, representing the weight of the salt absorption rule; ω2=0.3, representing the weight of the salt fog scattering rule; ω3=0.3, representing the weight of the ground feature association rule. S1, S2, and S3 are the matching scores of each rule, with a value range of 0-1. The higher the similarity, the higher the score; 3) Parameter adjustment: If S<0.8, it is determined that the generated feature violates the physical rules. First, the parameter is adjusted using the formula. , The normalized mean and variance of the SPADE module are dynamically adjusted, and the weight coefficients of the corresponding rules are increased by 20% to strengthen the constraints; if S≥0.8, the generated features are determined to conform to the physical rules, and the generation parameters and rule weights remain unchanged; among them, the initial normalized mean... ,variance , which is the preset baseline value.

[0100] S323. Using the edge features of the salt absorption feature map and the geographic association rules in the dynamic remote sensing physical rule base, perform physical rationalization optimization on the edge areas of the salt pseudo-color map and output the initial salt distribution map.

[0101] Specifically, in the final output stage of the adaptive physical constraint generator, a combined algorithm of edge detection, physical rule verification, and weighted interpolation correction is used to optimize the physical rationality of the generated results: 1) Edge detection: The Canny operator is used to extract edges from the salt absorption feature map to obtain an edge mask map of salt distribution, thus identifying the edge areas that need to be optimized; 2) Rule verification: Pixels covered by the edge mask map are verified according to the land cover association rules. If the salt difference between adjacent edge pixels is >2g / kg or the salt content in the vegetation area is >0.5g / kg, it is judged as a "physically unreasonable edge"; 3) Weighted interpolation correction: For unreasonable edge pixels, a 3×3 neighborhood weighted interpolation algorithm is used for correction. The weights are set so that the closer the distance, the greater the weight, and the greater the weight of pixels that conform to the rules. Through the above optimization, visually realistic but mechanistically incorrect generated results are avoided, effectively maintaining the physical consistency compliance rate of edge areas.

[0102] S33. Construct a dual-core loss function architecture and use a training strategy of pre-training and fine-tuning for new scenarios to train the generator to learn the mapping relationship between remote sensing physical rules and generation parameters.

[0103] In the description of this invention, a dual-core loss function architecture is constructed, and a training strategy of pre-training and fine-tuning for new scenarios is used to train the generator to learn the mapping relationship between remote sensing physical rules and generation parameters, including:

[0104] S331. Construct a dual-core loss function architecture that combines physical consistency loss and adversarial loss, and combine the dual-core loss function to generate a total loss function to form an effective constraint on physical rules.

[0105] Specifically, the physical consistency loss is calculated based on the physical consistency score S, using the formula L. phy =1-S, when S<0.8, the weight ω is automatically adjusted. phy The weights are linearly increased from a base value of 0.3 to 0.8; when S ≥ 0.8, the weights remain at the base value of 0.3. The adversarial loss is calculated using the Wasserstein distance between the generator and the discriminator, with the formula L. adv =E[D(G(z))]-E[D(x)], where G(z) is the generator output, x is the real sample, D is the discriminator, and its weight ω adv It is fixed at 0.7.

[0106] The total loss function is L total =ω phy ×L phy +ω adv ×L adv This ensures the effective constraint of physical rules while enhancing the visual realism of the generated results.

[0107] S332. In the pre-training phase, the generator is iteratively trained using different types of cross-scenario saline-alkali land data as training samples. In the new scenario fine-tuning phase, the channel parameters of the physical prior interaction link are automatically adjusted according to the newly added monitoring scenario. The pre-training and fine-tuning are determined to be completed based on the changing trends of the physical consistency score and the salinity inversion index.

[0108] Specifically, in the pre-training phase, five types of cross-scene saline-alkali land data were used as training samples, including coastal, arid and semi-arid, slightly saline-alkali, moderately saline-alkali, and severely saline-alkali land, with a total of 1000 samples and a training batch size of 8. The Adam optimizer was used, with a learning rate of 2e-4, β1=0.5, β2=0.999, and 20000 training iterations. The stopping condition was that the total loss function value of the validation set fluctuated by ≤0.001 for 100 consecutive iterations. This phase focused on learning the mapping relationship between "remote sensing physical rules and generation parameters" and saved the pre-trained model weights as a benchmark. In the new scene fine-tuning phase, only three types of parameters of the physical prior interaction link were fine-tuned for new monitoring scenarios, including the 1×1 convolution kernel weights of the feature extraction unit, the priority coefficients of the rule calling unit, and the weights of the fully connected layer of the parameter transformation unit. The fine-tuning sample size only required 30-50 sets, the learning rate was reduced to 1e-5, and the number of iterations was 3000. When the physical consistency score S≥0.9 and the salt inversion RMSE≤0.4g / kg of the new scenario validation set are determined to be fine-tuned, this strategy does not require retraining the entire generator, which greatly reduces training costs and time.

[0109] S4. Calculate the deviation between the initial salinity distribution map and the ground sample label, the deviation between the dual feature map and the ideal feature, and the deviation between the generated result and the physical rules. Fuse them to generate a standardized decoupled deviation signal and transmit it to the encoder to complete multiple rounds of calibration.

[0110] In the description of this invention, the deviations between the initial salinity distribution map and the ground sample label, the deviations between the dual feature map and the ideal features, and the deviations between the generated result and the physical rules are calculated. These are then fused to generate a standardized decoupled deviation signal, which is transmitted to the encoder to complete multiple rounds of calibration, including:

[0111] S41. Establish a deviation calculation and feedback link for multi-unit serial and parallel interfaces. Based on the received output results of the encoder and generator, calculate the sample deviation, feature deviation and physical deviation respectively, and perform weighted summation according to the preset weight ratio. After standardization, generate a decoupled deviation signal.

[0112] Specifically, the link adopts a modular serial architecture, consisting of a data input interface, a deviation calculation unit, a signal fusion unit, a standardization processing unit, and a bidirectional transmission interface from top to bottom. Each unit interacts with the other via a PCIe 4.0 bus. The data input interface reserves three types of input channels: the initial salt distribution map output by the adaptive physical constraint generator, the salt absorption feature map and salt fog scattering feature map output by the dynamic feedback decoupling encoder, and the ground sample label input channel. The bidirectional transmission interface establishes communication with the deviation signal receiving channel of the dynamic feedback decoupling encoder and the physical prior interaction link parameter adjustment terminal of the adaptive physical constraint generator, respectively, to achieve bidirectional signal interaction. The deviation calculation unit contains three independent calculation subunits, corresponding to sample deviation, feature deviation, and physical deviation, respectively.

[0113] The sample bias is calculated using the mean absolute error to determine the difference between the generated and measured values. The calculation object is the pixel value of the initial salinity distribution map output by the adaptive physical constraint generator and the corresponding measured pixel value of the ground sample label. The formula is: Sample bias = (1 / total number of pixels of ground sample label) × Σ|initial salinity distribution map pixel value - measured pixel value of ground sample label|; where the total number of pixels of ground sample label is the number of pixels corresponding to 30-50 groups of samples, and the allowable error is ≤0.05g / kg.

[0114] The feature deviation is calculated from the salt absorption feature map, salt fog scattering feature map, and ideal feature map output by the dynamic feedback decoupled encoder. These features are calibrated by ground sample tags and denoted as the L2 distance between the ideal salt absorption feature map and the ideal salt fog scattering feature map. The formula is: Feature Deviation = √[(1 / Total number of feature map pixels) × Σ(||Salt Absorption Feature Map - Ideal Salt Absorption Feature Map||² + ||Salt Fog Scattering Feature Map - Ideal Salt Fog Scattering Feature Map||²)]; where the total number of feature map pixels is 256 × 256 = 65536.

[0115] The physical deviation is directly related to the physical consistency score of the adaptive physical constraint generator. The formula is physical deviation = 1 - physical consistency score. The value range is the same as that of the physical consistency score (0-1). When the physical consistency score is ≥0.9, the physical deviation is ≤0.1, and it is judged that the physical deviation meets the standard.

[0116] The three types of biases are weighted and summed according to a preset weight ratio to obtain the fused signal. For example, the preset formula is: fused signal = 0.3 × sample bias + 0.2 × feature bias + 0.5 × physical bias.

[0117] S42. Set up an iterative calibration process that includes signal transmission, parameter adjustment, re-decoupling generation, and deviation calculation. Set the iteration trigger conditions, termination criteria, and calibration operations to achieve gradual convergence of errors and output the salt distribution map and salt spray distribution map after iterative calibration.

[0118] In the description of this invention, an iterative calibration process is defined, including signal transmission, parameter adjustment, re-decoupling generation, and deviation calculation. The iterative triggering conditions, termination criteria, and calibration operations are also defined.

[0119] S421. When the initial salinity distribution map is generated for the first time, the first round of calibration is automatically triggered. Subsequent rounds of calibration are triggered by the deviation calculation results of the previous round. When the decoupling deviation signal is greater than the preset signal threshold, the next round of calibration is automatically entered.

[0120] S422. The decoupling bias signal is transmitted to the encoder through the bidirectional transmission interface. The encoder adjusts the convolution kernel weights and adversarial loss weights according to the decoupling bias signal and synchronizes the adjustment instructions to the generator to adjust the rule matching threshold of the physical prior interaction link. The encoder and generator regenerate new output results and recalculate the decoupling bias signal.

[0121] Specifically, the decoupling bias signal is transmitted in real time to the bias signal receiving channel of the dynamic feedback decoupling encoder through a bidirectional transmission interface; the encoder adjusts the convolution kernel weights and decoupling adversarial loss weights according to the decoupling bias signal, and simultaneously synchronizes the adjustment instructions to the adaptive physical constraint generator to fine-tune the rule matching threshold of its physical prior interaction link; the encoder re-executes decoupling based on the adjusted parameters and outputs a new salt absorption feature map and a new salt fog scattering feature map; the generator regenerates a new initial salt distribution map based on the new dual feature map; the link recalculates the three types of biases on the newly generated initial salt distribution map and the new dual feature map to generate a new decoupling bias signal.

[0122] S423. When the decoupling deviation signal is less than the preset signal threshold or the number of iterations reaches the maximum number of iterations, the iteration is terminated.

[0123] S43. A bilateral filter is used to perform a light smoothing process on the salt distribution map and salt spray distribution map to remove abnormal pixels. The abnormal areas are then corrected using a neighborhood physical weighted interpolation method to obtain the final salt distribution map and salt spray distribution map.

[0124] Specifically, the final output dual-feature map and the final initial salinity distribution map are lightly optimized. For the dual-feature map optimization, a 5×5 bilateral filtering algorithm is used to smooth the final salinity absorption feature map and the final salt fog scattering feature map. The filtering parameters are set to spatial domain standard deviation = 2 and grayscale domain standard deviation = 0.1, preserving feature edge information while reducing noise. For the salinity distribution map optimization, land cover association rules based on a dynamic remote sensing physical rule base are used to remove abnormal pixels from the final initial salinity distribution map. The criteria are: salinity difference between adjacent pixels in bare soil areas > 2 g / kg and salinity in vegetated areas > 0.5 g / kg. Abnormal pixels are corrected using 3×3 neighborhood weighted interpolation to ensure that the proportion of abnormal areas is ≤ 0.3%. The optimized output includes the final salinity absorption feature map, the final salt fog scattering feature map, and the final salinity distribution map.

[0125] S5. Output the salt distribution map and salt spray distribution map after multiple rounds of calibration through the generator.

[0126] Please see Figure 4 Furthermore, a soil salinity monitoring system based on image data processing is also provided, the system comprising:

[0127] Image acquisition and preprocessing module 1 is used to acquire core band hyperspectral images of characteristic responses of saline-alkali land. After radiometric calibration, denoising and standardization, image data and ground sample labels adapted to network codec are generated.

[0128] The dynamic feedback decoupling encoding module 2 is used to input standardized image data into a dynamically feedback decoupling encoder, receive decoupling bias signals, dynamically adjust the encoder's convolution kernel weights and adversarial loss weights, and output optimized dual feature maps. The dual feature maps include a salt absorption feature map and a salt fog scattering feature map.

[0129] The adaptive physical constraint generation module 3 is used to input the dual feature map into the adaptive physical constraint generator, call the remote sensing physical rule base, and after matching the spectral and spatial features of the dual feature map with the rule base, dynamically adjust the generator parameters and output the initial salinity distribution map.

[0130] The closed-loop deviation calibration module 4 is used to calculate the deviation between the initial salinity distribution map and the ground sample label, the deviation between the dual feature map and the ideal feature, and the deviation between the generated result and the physical rules. It integrates and generates a standardized decoupled deviation signal and transmits it to the encoder to complete multiple rounds of calibration.

[0131] The result output verification module 5 is used to output the salt distribution map and salt spray distribution map after multiple rounds of calibration through the generator.

[0132] The following analysis, based on specific implementation examples and comparative experimental data, examines the performance verification of salt inversion under different scenarios. The comparison objects cover traditional machine learning, traditional deep learning, and variant schemes without core modules.

[0133] Example: Coastal high salt spray scenario (salt spray concentration 20-30 mg / m³)

[0134] This invention selects a coastal saline-alkali land area in a delta of a certain province as the study area, covering an area of ​​approximately 50 km². This region has a temperate monsoon climate with an average annual precipitation of 550 mm and an annual evaporation of 1900 mm. The soil type is mainly coastal saline soil, with a surface soil (0-20 cm) salt content of 0.1%-2.0%, encompassing light (0.1%-0.3%), moderate (0.3%-0.6%), and severe (>0.6%) saline-alkali land types. Furthermore, influenced by the maritime climate, it experiences medium to high salt spray interference of 20-30 mg / m³ year-round, meeting the requirements of the experimental scenario.

[0135] In accordance with the NY / T 1121.1-2006 standard "Soil Testing Part 1: Soil Sample Collection, Processing and Storage", sampling points were set up using a grid method with a grid spacing of 500m × 500m, totaling 200 sampling points. At each sampling point, a five-point sampling method was used to collect soil samples from the 0-20cm surface layer. Each sample weighed ≥1kg. The latitude, longitude, altitude, and land cover type of each sampling point were recorded simultaneously. Immediately after collection, impurities such as stones and plant debris were removed, and the samples were air-dried indoors. After grinding, the samples were sieved through a 2mm sieve. The soil salinity was determined using a gravimetric method and used as the sample label value, with a measurement accuracy error ≤0.01g / kg.

[0136] The data source was Sentinel-2 multispectral imagery with a spatial resolution of 10m. Ten bands were selected for the experiment, including 490nm, 560nm, 665nm, 705nm, 740nm, 783nm, 842nm, 865nm, 1610nm, and 2190nm. Preprocessing generated 256×256 pixel image patches for model input. All comparison methods were implemented in the same hardware and software environment. The core parameters are as follows: 1. Random Forest: 100 decision trees, unlimited maximum tree depth, using Euclidean distance as the distance function; 2. Support Vector Machine: linear kernel function, complexity parameter C=1.0; 3. U-Net: consistent with the encoder architecture of this invention, without a spectral attention module and dynamic feedback mechanism, the optimizer is Adam (learning rate 2e-4); 4. This invention (variant without physical constraints): the physical rule base in the adaptive physical constraint generator is removed, and the remaining parameters are consistent with this method. The core performance comparison experimental data are shown in Table 1.

[0137] Table 1. Experimental data comparing core performance

[0138] Comparison Methods Salt content inversion RMSE (g / kg) Coefficient of determination R² Physical consistency score Precision attenuation rate under salt spray interference 30 small sample R² Processing time per scene (min) This invention 0.35 0.96 0.92 0.8% 0.94 18 Random Forest 1.13 0.65 0.59 18.3% 0.62 32 Support Vector Machine 1.05 0.68 0.62 16.7% 0.65 30 U-Net 0.82 0.78 0.68 12.5% 0.73 45 This invention (variant without physical constraints) 0.61 0.85 0.71 8.2% 0.81 16

[0139] As can be seen, the salt inversion RMSE of this invention is reduced by 69.0% compared to random forest, 66.7% compared to support vector machine, and 57.3% compared to U-Net, with the core accuracy indicators significantly better than existing technologies. When facing medium to high salt spray interference, the accuracy decay rate of this invention is only 1 / 10 to 1 / 20 of that of traditional methods, demonstrating outstanding anti-interference ability. The performance advantage is obvious in small sample scenarios, with R² improved by 51.6% compared to traditional machine learning for 30 sets of samples.

[0140] In summary, by utilizing the technical solutions described above, the collaborative design of dynamic feedback decoupling encoding and adaptive physical constraint generation accurately separates salt fog scattering features and salt absorption features. In scenarios with salt fog concentration interference, it effectively reduces salt inversion errors and completely solves the decoupling difficulties caused by salt fog interference. Simultaneously, by combining meta-learning pre-training and small-sample feature constraint design, high-precision monitoring can be achieved with only a small number of ground samples, making it suitable for monitoring saline-alkali land in remote and harsh environments, significantly reducing sample collection costs and difficulties. Through a dynamic remote sensing physical rule base and real-time consistency verification design, the generated results achieve physical consistency scores that meet standard requirements, avoiding mechanistic errors. The constraint strategy can be dynamically adjusted according to the salt fog concentration and land cover type of different scenarios, adapting to various saline-alkali land scenarios such as coastal, arid, and semi-arid regions. Furthermore, the closed-loop deviation calibration architecture achieves iterative optimization of decoupling-generation-calibration, correcting errors in each stage in real time, effectively reducing the accumulation of errors in the cascaded architecture, and ensuring accuracy degradation under extreme salt fog scenarios.

[0141] The above description is only a preferred embodiment of the present invention and is not intended to limit the present invention. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for monitoring soil salinity and alkali land based on image data processing, characterized in that, include: S1. Collect core band hyperspectral images of characteristic responses of saline-alkali land, and generate image data and ground sample labels adapted to network codecs after radiometric calibration, denoising and standardization. S2. Input standardized image data into a dynamically decoupled encoder, receive decoupling bias signals and dynamically adjust the encoder's convolution kernel weights and adversarial loss weights, and output an optimized dual feature map; the dual feature map includes a salt absorption feature map and a salt spray scattering feature map. S3. Input the dual feature map into the generator of adaptive physical constraints, call the remote sensing physical rule base, and after matching the spectral and spatial features of the dual feature map with the rule base, dynamically adjust the generator parameters and output the initial salinity distribution map. S4. Calculate the deviation between the initial salinity distribution map and the ground sample label, the deviation between the dual feature map and the ideal feature, and the deviation between the generated result and the physical rules. Fuse them to generate a standardized decoupled deviation signal and transmit it to the encoder to complete multiple rounds of calibration. S5. Output the salt distribution map and salt spray distribution map after multiple rounds of calibration through the generator.

2. The method for monitoring soil salinity and alkali land based on image data processing according to claim 1, characterized in that, The process of inputting standardized image data into a dynamically decoupled encoder, receiving decoupling bias signals and dynamically adjusting the encoder's convolutional kernel weights and adversarial loss weights, and outputting an optimized dual feature map includes: S21. Based on the U-shaped network, build a decoupled coding architecture for the encoder, and configure the input layer, downsampling layer, upsampling layer, output layer, and spectral attention module embedded between the input layer and the downsampling layer; S22. Pre-set the decoupling bias signal composed of feature bias and mechanism bias, convert it into adjustment coefficient, and set the encoder parameter dynamic adjustment logic to build a two-stage dynamic decoupling process that connects the initial stage and the feedback stage, so as to realize the dynamic adjustment of convolution kernel weight and adversarial loss weight. S23. Using multiple sets of cross-scene saline-alkali land samples as pre-training data, the encoder is pre-trained by model-independent element learning. Small sample feature constraint terms are embedded in the decoupling bias signal receiving channel and weights are set to complete the small sample adaptation optimization.

3. The method for monitoring soil salinity and alkali land based on image data processing according to claim 2, characterized in that, The pre-set decoupling deviation signal, composed of characteristic deviation and mechanism deviation, is converted into an adjustment coefficient, and the encoder parameter dynamic adjustment logic is set to establish a two-stage dynamic decoupling process that connects the initial stage and the feedback stage, including: S221. Set the Euclidean distance between the salt absorption characteristic map, the salt spray scattering characteristic map and the ideal characteristic map as the characteristic deviation, and set the deviation between the spectral curve of the dual characteristic map and the salt standard spectrum and the salt spray standard spectrum as the mechanism deviation, and combine them to form a decoupling deviation signal. S222. Convert the decoupling bias signal into an adjustment coefficient within a preset numerical range. Dynamically adjust the decoupling resistance loss weights according to the product of the base weights and the adjustment coefficients. And dynamically correct the weights of each layer of convolution kernels through a 1×1 convolution kernel. S223. Set the adjustment coefficient of the initial stage to the initial value, and the encoder outputs the initial dual feature map according to the default parameters; after receiving the decoupling deviation signal after calibration transmission in the feedback stage, calculate the adjustment coefficient in real time and adjust the encoder parameters, re-execute the decoupling operation, and output the optimized dual feature map.

4. The method for monitoring soil salinity and alkali land based on image data processing according to claim 1, characterized in that, The process of inputting the dual-feature map into the adaptive physical constraint generator, calling the remote sensing physical rule base, matching the spectral and spatial features of the dual-feature map with the rule base, dynamically adjusting the generator parameters, and outputting the initial salinity distribution map includes: S31. Build the generator's generation architecture, set the input and output data specifications, and insert a physical prior interaction link between the spatial adaptive normalization module and the convolutional layer to realize the real-time injection of physical rules. S32. Build a dynamic remote sensing physical rule base, set the calculation logic of physical consistency score, and set dynamic adjustment rules for generator parameters based on physical consistency score. Combine feature map edge information to optimize the generation of initial salt distribution map and realize the real-time constraint of physical rules on the generation process. S33. Construct a dual-core loss function architecture and use a training strategy of pre-training and fine-tuning for new scenarios to train the generator to learn the mapping relationship between remote sensing physical rules and generation parameters.

5. The method for monitoring soil salinity and alkali land based on image data processing according to claim 4, characterized in that, The generation architecture for building the generator, which sets the input and output data specifications and inserts a physical prior interaction link between the spatial adaptive normalization module and the convolutional layer, includes: S311. Construct a generation architecture that includes multiple downsampling layers and upsampling layers. Set the input to be the spliced ​​data of salt absorption feature map and salt fog scattering feature map, and the output to be a salt pseudo-color map. S312. Insert physical prior interaction links between the convolutional kernels and spatial adaptive normalization modules in the downsampling layer and the upsampling layer respectively, to provide a physical channel for the adaptive injection of physical constraints; the physical prior interaction links include feature extraction, rule invocation, parameter conversion and transmission interface.

6. The method for monitoring soil salinity and alkali land based on image data processing according to claim 4, characterized in that, The process of building a dynamic remote sensing physical rule base, setting the calculation logic for physical consistency scores, and dynamically adjusting generator parameters based on physical consistency scores, combined with feature map edge information to optimize the generation of an initial salinity distribution map, includes: S321. Construct a dynamic remote sensing physical rule base containing multiple scalable rules, including salt absorption rules, salt fog scattering rules, and ground feature association rules. S322. Extract the spectral and spatial features of the intermediate feature map of the generator through the physical prior interaction link, and compare the features with the dynamic remote sensing rule base to calculate the physical consistency score. If the physical consistency score is less than the preset score, the normalized mean and variance of the spatial adaptive normalization module are dynamically adjusted, and the corresponding constraint rules are strengthened; if the physical consistency score is greater than or equal to the preset threshold, the generator parameters are kept unchanged. S323. Using the edge features of the salt absorption feature map and the geographic association rules in the dynamic remote sensing physical rule base, perform physical rationalization optimization on the edge areas of the salt pseudo-color map and output the initial salt distribution map.

7. The method for monitoring soil salinity and alkali land based on image data processing according to claim 4, characterized in that, The construction of a dual-core loss function architecture, utilizing a training strategy of pre-training and fine-tuning for new scenarios, trains the generator to learn the mapping relationship between remote sensing physical rules and generation parameters, including: S331. Construct a dual-core loss function architecture that combines physical consistency loss and adversarial loss, and combine the dual-core loss function to generate a total loss function to form an effective constraint on physical rules. S332. In the pre-training stage, the generator is iteratively trained using different types of cross-scenario saline-alkali land data as training samples; and in the new scenario fine-tuning stage, the channel parameters of the physical prior interaction link are automatically adjusted according to the newly added monitoring scenario, and the pre-training and fine-tuning are determined to be completed based on the changing trend of the physical consistency score and the salinity inversion index.

8. The method for monitoring soil salinity and alkali land based on image data processing according to claim 1, characterized in that, The process involves calculating the deviations between the initial salinity distribution map and the ground sample label, the deviations between the dual-feature map and the ideal features, and the deviations between the generated results and the physical rules. These are then fused to generate a standardized decoupled deviation signal, which is transmitted to the encoder to complete multiple rounds of calibration. S41. Establish a deviation calculation and feedback link for multi-unit serial and parallel interfaces. Based on the received output results of the encoder and generator, calculate the sample deviation, feature deviation and physical deviation respectively, and perform weighted summation according to the preset weight ratio. After standardization, generate a decoupled deviation signal. S42. Set up an iterative calibration process that includes signal transmission, parameter adjustment, re-decoupling generation, and deviation calculation. Set the iteration trigger conditions, termination criteria, and calibration operations to achieve gradual convergence of errors and output the salt distribution map and salt spray distribution map after iterative calibration. S43. A bilateral filter is used to perform a light smoothing process on the salt distribution map and salt spray distribution map to remove abnormal pixels. The abnormal areas are then corrected using a neighborhood physical weighted interpolation method to obtain the final salt distribution map and salt spray distribution map.

9. A method for monitoring soil salinity and alkali land based on image data processing according to claim 8, characterized in that, The setup includes an iterative calibration process encompassing signal transmission, parameter adjustment, re-decoupling generation, and deviation calculation, and sets the iteration trigger conditions, termination criteria, and calibration operations, including: S421. When the initial salinity distribution map is generated for the first time, the first round of calibration is automatically triggered. Subsequent rounds of calibration are triggered by the deviation calculation results of the previous round. When the decoupling deviation signal is greater than the preset signal threshold, the next round of calibration is automatically entered. S422. The decoupling bias signal is transmitted to the encoder through a bidirectional transmission interface. The encoder adjusts the convolution kernel weights and adversarial loss weights according to the decoupling bias signal and synchronizes the adjustment instructions to the generator to adjust the rule matching threshold of the physical prior interaction link. The encoder and generator regenerate new output results and recalculate and generate the decoupling bias signal. S423. When the decoupling deviation signal is less than the preset signal threshold or the number of iterations reaches the maximum number of iterations, the iteration is terminated.

10. A soil salinity monitoring system based on image data processing, used to implement the soil salinity monitoring method based on image data processing according to any one of claims 1-9, characterized in that, The system includes: The image acquisition and preprocessing module is used to acquire core band hyperspectral images of the characteristic response of saline-alkali land. After radiometric calibration, denoising and standardization, image data and ground sample labels adapted to the network codec are generated. The dynamic feedback decoupling encoding module is used to input standardized image data into a dynamically feedback decoupling encoder, receive decoupling deviation signals and dynamically adjust the convolution kernel weights and adversarial loss weights of the encoder, and output an optimized dual feature map; the dual feature map includes a salt absorption feature map and a salt spray scattering feature map. The adaptive physical constraint generation module is used to input the dual feature map into the adaptive physical constraint generator, call the remote sensing physical rule base, and after matching the spectral and spatial features of the dual feature map with the rule base, dynamically adjust the generator parameters and output the initial salinity distribution map. The closed-loop deviation calibration module is used to calculate the deviation between the initial salinity distribution map and the ground sample label, the deviation between the dual feature map and the ideal feature, and the deviation between the generated result and the physical rules. It fuses and generates a standardized decoupled deviation signal and transmits it to the encoder to complete multiple rounds of calibration. The result output verification module is used to generate and output the salt distribution map and salt spray distribution map after multiple rounds of calibration.

Images