A cultivated land soil salinization inversion method and system based on remote sensing and deep learning

Abstract

This invention relates to the field of remote sensing monitoring technology for soil salinization, and discloses a method and system for inverting farmland soil salinization based on remote sensing and deep learning. The method includes: acquiring multimodal farmland data of a target area; constructing spatially consistent soil salinity content monitoring samples; constructing remote sensing pixel samples; screening sensitive features based on a SHAP-based feature importance assessment mechanism; using a deep learning regression modeling method to invert soil salinity content; and outputting soil salinization distribution results. Compared to existing methods that rely on empirical sampling, especially in regional-scale applications with strong soil spatial heterogeneity, this invention addresses the technical problem of accurately determining soil salinity content. By introducing multimodal consistency constraints and a SHAP-based sensitive feature screening mechanism, this invention effectively constrains the soil salinity inversion process, improving the accuracy of farmland soil salinization inversion.

Description

Technical Field

[0001] This invention relates to the field of remote sensing monitoring technology for soil salinization, and in particular to a method and system for inverting soil salinization in arable land based on remote sensing and deep learning. Background Technology

[0002] Currently, soil salinization is one of the most significant soil degradation problems affecting arable land quality and sustainable agricultural development. It exhibits significant spatial heterogeneity and scale differences under varying climatic conditions, soil types, and farming management practices. Accurate identification and spatial distribution monitoring of arable land soil salinization levels are crucial for implementing salinization prevention and control measures, assessing arable land quality, and making agricultural production decisions.

[0003] Existing technologies for monitoring and assessing soil salinization in arable land mainly include ground sampling and analysis methods and inversion methods based on remote sensing images. Ground sampling and analysis methods, which involve collecting soil samples in the field and measuring their physicochemical properties, can obtain relatively accurate data on soil salinity. However, this method is costly, time-consuming, and difficult to meet the needs of large-scale, continuous arable land monitoring. Inversion methods based on remote sensing images utilize multispectral or hyperspectral image information to establish a mapping relationship between remote sensing features and soil salinity, enabling regional-scale salinization monitoring. These methods offer advantages such as wide coverage and high acquisition efficiency.

[0004] However, constructing a remote sensing-based soil salinity inversion model is a key challenge in remote sensing monitoring of soil salinization. Traditional linear regression methods rely on a limited number of parameters, and due to their linear assumptions and inability to account for the nonlinear relationship between soil salinity and influencing factors, they fail to capture the complexity of soil salinization. This limitation reduces prediction accuracy and may overlook the interactions of key features. In contrast, machine learning techniques can effectively address the nonlinearity between soil properties and environmental factors, thereby improving feature selection efficiency and prediction accuracy.

[0005] Therefore, there is an urgent need to propose a method for retrieving farmland soil salinization based on remote sensing and deep learning, so as to effectively improve the efficiency and accuracy of remote sensing monitoring of soil salinization at the regional scale. Summary of the Invention

[0006] To address the aforementioned technical shortcomings, the purpose of this invention is to propose a method for inverting farmland soil salinization based on remote sensing and deep learning. This method aims to solve the technical problem that existing methods, which rely on empirical sampling, are unable to accurately determine soil salinity, especially in regional application scenarios with strong spatial heterogeneity of soil.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: The present invention provides a method for inverting farmland soil salinization based on remote sensing and deep learning.

[0008] The method for retrieving farmland soil salinization based on remote sensing and deep learning includes:

[0009] Step S10: Obtain multimodal farmland data of the target area, and perform sampling spatial planning task based on the sampling point layout mechanism based on multimodal consistency constraints using the multimodal farmland data, and output the set of soil sampling point locations P;

[0010] Step S20: Based on the set of soil sampling point locations P, a sample construction mechanism based on spatial constraint consistency is used to perform the task of generating salinity monitoring samples, and output the salinity sample set D;

[0011] Step S30: Based on the salt content sample set D, a pixel-level remote sensing feature alignment mechanism is used to perform a consistency alignment task, and a remote sensing pixel sample set is output.

[0012] Step S40: Based on the remote sensing pixel sample set, a statistical calculation method based on remote sensing band reflectance and spectral index is used to perform the feature construction process, and an initial remote sensing pixel-level feature vector is output; and based on the output initial feature vector, a sensitivity constraint screening is performed using a feature importance evaluation mechanism based on SHAP, and a set of sensitive feature vectors F is output.

[0013] Step S50: Based on the sensitive feature vector set F, a deep learning-based regression modeling method is used to perform the soil salinity inversion task and output the soil salinization distribution results.

[0014] Preferably, step S10, which involves acquiring multimodal farmland data of the target area, performing sampling spatial planning based on a sampling point layout mechanism based on multimodal consistency constraints using the multimodal farmland data, and outputting the soil sampling point location set P, specifically includes:

[0015] Step S101: Obtain multimodal farmland data for the target area. The multimodal farmland data includes farmland spatial distribution data, soil type data, and farmland management information. Based on the farmland spatial distribution data, divide the farmland into multiple basic spatial units.

[0016] Step S102: For each basic spatial unit of cultivated land, based on soil type data and cultivated land management information, combined with a preset attribute and management joint judgment template, a consistency judgment method based on threshold driving is used to perform consistency judgment, and the template consistency score and basic spatial unit set are output.

[0017] Step S103: Under the premise of satisfying the preset minimum sampling interval constraint, the sampling density self-adjustment point distribution method based on template consistency score is used to perform the adaptive generation task of sampling point spatial location for the basic spatial unit set, and output the soil sampling point location set P.

[0018] Preferably, step S20, which involves using a sample construction mechanism based on spatial constraint consistency to generate salinity monitoring samples based on the set of soil sampling point locations P, and outputting a salinity sample set D, specifically includes:

[0019] Step S201: Collect target soil samples within a 0-20cm thickness range of the topsoil at each soil sampling point location in the soil sampling point location set P;

[0020] Step S202: Measure the soil electrical conductivity EC of the target soil sample, and calculate the corresponding soil salinity SSC based on the soil electrical conductivity EC. ;

[0021] Step S203: Bind the spatial coordinates of each soil sampling point to its corresponding soil salinity content SSC to form a salinity sample set D with spatial coordinates.

[0022] Preferably, step S30, which involves performing a consistency alignment task based on a pixel-level remote sensing feature alignment mechanism using the salt content sample set D, and outputting a remote sensing pixel sample set, specifically includes:

[0023] Step S301: Obtain ground sampling time information corresponding to the salt content sample set D, and obtain multi-temporal remote sensing image data; extract time range information from the multi-temporal remote sensing image data, calculate the time offset difference based on the time range information and sampling time information, filter remote sensing images from the multi-temporal remote sensing image data whose time offset difference is less than or equal to the preset time consistency threshold, and output the target remote sensing image data.

[0024] Step S302: Perform radiometric calibration, atmospheric correction, image fusion, and image cropping sequentially on the target remote sensing image data to output a standardized digital orthophoto image;

[0025] Step S303: Obtain the spatial coordinates of the salt content sample set D. Based on the standardized digital orthophoto and spatial coordinates, perform pixel-level indexing and positioning using a point sampling method based on nearest neighbor pixel positioning, and output the remote sensing pixel sample set.

[0026] Preferably, step S40, which involves performing a feature construction process based on a statistical calculation method using remote sensing band reflectance and spectral index, using a set of remote sensing pixel samples, and outputting an initial remote sensing pixel-level feature vector; and then performing sensitivity constraint screening based on a SHAP-based feature importance evaluation mechanism according to the output initial feature vector, and outputting a set of sensitive feature vectors F, specifically includes:

[0027] Step S401: Obtain the blue band reflectance (B), green band reflectance (G), red band reflectance (R), and near-infrared band reflectance (NIR) corresponding to each remote sensing pixel from the remote sensing pixel sample set; calculate the corresponding Normalized Difference Vegetation Index (NDVI), Soil Salinity Index (SI), and Canopy Response Salinity Index (CRSI) based on the blue band reflectance (B), green band reflectance (G), red band reflectance (R), and near-infrared band reflectance (NIR) according to the preset spectral index calculation model;

[0028] Step S402: Based on the blue band reflectance B, green band reflectance G, red band reflectance R, near-infrared band reflectance NIR, normalized difference vegetation index NDVI, soil salinity index SI, and canopy response salinity index CRSI, an initial remote sensing pixel-level feature vector is constructed using a feature stitching method.

[0029] Step S403: Preset a machine learning model for predicting salinity content, obtain historical initial remote sensing pixel-level feature vectors and historical soil salinity content (SSC), use the historical initial remote sensing pixel-level feature vectors as input to the machine learning model for predicting salinity content, use the historical soil salinity content (SSC) as output to the machine learning model for predicting salinity content, and use the SHAP feature importance evaluation method to calculate the feature importance value of each feature in the process of predicting soil salinity content (SSC) in the machine learning model for predicting salinity content.

[0030] Step S404: Select features whose importance value is greater than the preset importance threshold to form a set of sensitive feature vectors F.

[0031] Preferably, in step S40, the formula for the Normalized Difference Vegetation Index (NDVI) is expressed as follows: The formula for the soil salinity index (SI) is: The formula for the canopy response salinity index (CRSI) is as follows: .

[0032] Preferably, step S50, which involves using a deep learning-based regression modeling method to perform soil salinity inversion based on the sensitive feature vector set F and outputting the soil salinization distribution results, specifically includes:

[0033] Step S501: Design a soil salinity inversion model based on a deep neural network regression model in advance; use the sensitive feature vector set F as the input of the soil salinity inversion model, and use the soil salinity SSC as the supervision label of the soil salinity inversion model. Then, use grid search and ten-fold cross-validation to determine the optimal combination of model parameters for the soil salinity inversion model.

[0034] Step S502: Obtain historical soil sampling point sample data and divide the historical soil sampling point sample data into training sample dataset and validation sample dataset; based on the training sample dataset, apply the optimized model parameter combination to perform a pre-training process on the soil salinity inversion model;

[0035] After training, the coefficient of determination is used based on the validation sample dataset. Root mean square error and mean absolute error The model prediction accuracy is quantitatively evaluated, and the model performance evaluation score is output. The model performance evaluation score is used to determine the effectiveness of the soil salinity inversion model after training.

[0036] Step S503: For model parameter combinations whose performance evaluation scores are greater than the preset model performance evaluation threshold, select the model parameter combination with the highest score as the optimal model parameter combination; apply the optimal model parameter combination to obtain the optimized soil salinity inversion model; input the sensitive feature vector set F into the optimized soil salinity inversion model, and the optimized soil salinity inversion model outputs the soil salinity inversion content; classify the soil salinity level of cultivated land according to the soil salinity inversion content and the preset salinity classification standard template, and finally output the soil salinity distribution results.

[0037] This invention also provides a system for retrieving farmland soil salinization based on remote sensing and deep learning, comprising:

[0038] The sampling space planning module is used to acquire multimodal farmland data of the target area. Based on the multimodal farmland data, it adopts a sampling point layout mechanism based on multimodal consistency constraints to perform sampling space planning tasks and outputs the set of soil sampling point locations P.

[0039] The salinity content monitoring sample construction module is used to perform the task of generating salinity content monitoring samples based on the spatial constraint consistency sample construction mechanism based on the set of soil sampling point locations P, and outputs the salinity content sample set D;

[0040] The pixel-level remote sensing feature alignment module is used to perform a consistency alignment task based on the salt content sample set D using a pixel-level remote sensing feature alignment mechanism, and outputs a remote sensing pixel sample set.

[0041] The remote sensing feature construction and sensitivity screening module is used to perform the feature construction process based on the remote sensing pixel sample set and the statistical calculation method based on the remote sensing band reflectance and spectral index, and output the initial remote sensing pixel-level feature vector; and to perform sensitivity constraint screening based on the output initial feature vector using the SHAP-based feature importance evaluation mechanism, and output the sensitive feature vector set F.

[0042] The soil salinity inversion module is used to perform soil salinity inversion tasks based on a set of sensitive feature vectors F and a deep learning-based regression modeling method, and outputs the soil salinization distribution results.

[0043] The present invention also provides a farmland soil salinization inversion device based on remote sensing and deep learning, comprising: a memory, a processor, and a farmland soil salinization inversion program based on remote sensing and deep learning stored in the memory and executable on the processor. When the farmland soil salinization inversion program based on remote sensing and deep learning is executed by the processor, a farmland soil salinization inversion method based on remote sensing and deep learning is realized.

[0044] The present invention also provides a computer program product, including a farmland soil salinization inversion program based on remote sensing and deep learning, wherein the farmland soil salinization inversion program based on remote sensing and deep learning implements the farmland soil salinization inversion method based on remote sensing and deep learning when executed by a processor.

[0045] The beneficial effects of this invention are as follows: By employing the conventional red, green, blue, and near-infrared reflectance bands, and the vegetation index and soil salinity index calculated from them, this invention improves the universality and generalization ability of the soil salinity remote sensing model. Simultaneously, by using an ensemble learning model to mine the nonlinear relationship between multi-dimensional features and soil salinity content, rapid and accurate inversion of soil salinity content can be achieved.

[0046] This invention introduces a multimodal consistency constraint sampling point layout mechanism and a pixel-level remote sensing feature alignment mechanism during the inversion process of cultivated land soil salinization. This effectively constrains the correspondence between ground sampling data and remote sensing images in the spatial and temporal dimensions, thereby avoiding inversion bias caused by insufficient sampling representativeness or sample-pixel mismatch in traditional methods. This improves the stability and reliability of the regional scale cultivated land soil salinity inversion results.

[0047] This invention constructs an initial feature vector based on remote sensing band reflectance and spectral index, and quantifies and filters the feature contribution through a SHAP-based feature importance evaluation mechanism. Under the constraint of sensitive features, a deep learning regression inversion model is constructed, which can effectively reduce the interference of redundant features on model training, improve the model's ability to represent the nonlinear relationship between soil salinity and multi-source environmental factors, and thus improve the accuracy of farmland soil salinization inversion and the reliability of its spatial distribution results. Attached Figure Description

[0048] To more clearly illustrate the technical solutions in the embodiments of the present invention or the prior art, the drawings used in the description of the embodiments or the prior art will be briefly introduced below. Obviously, the drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0049] Figure 1 This is a flowchart illustrating the first embodiment of a method for retrieving farmland soil salinization based on remote sensing and deep learning according to the present invention.

[0050] Figure 2 This is a schematic diagram of the remote sensing inversion results of soil salinity in arable land according to the first embodiment of the method for inverting soil salinity based on remote sensing and deep learning of the present invention.

[0051] Figure 3 This is a schematic diagram of soil salinization distribution results from the first embodiment of a method for retrieving farmland soil salinization based on remote sensing and deep learning according to the present invention.

[0052] Figure 4 This is a schematic diagram of the equipment used in the present invention to propose a method for retrieving farmland soil salinization based on remote sensing and deep learning. Detailed Implementation

[0053] 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.

[0054] Example 1: As Figure 1 The diagram shown is a flowchart of the first embodiment of the method for inverting farmland soil salinization based on remote sensing and deep learning according to the present invention. The first embodiment of the method for inverting farmland soil salinization based on remote sensing and deep learning according to the present invention is presented.

[0055] In the first embodiment, the method for retrieving farmland soil salinization based on remote sensing and deep learning includes:

[0056] Step S10: Obtain multimodal farmland data of the target area, and perform sampling spatial planning task based on the sampling point layout mechanism based on multimodal consistency constraints using the multimodal farmland data, and output the set of soil sampling point locations P;

[0057] It should be noted that the "multimodal cultivated land data" in this step includes at least cultivated land spatial distribution data, soil type data, and cultivated land management information. Specifically, the cultivated land spatial distribution data is used to characterize the spatial boundaries and morphological features of cultivated land within the target area; the soil type data is used to represent the differences in soil physicochemical properties between different cultivated land units; and the cultivated land management information reflects management factors that significantly affect soil salinity distribution, such as farming methods, irrigation conditions, and fertilization management. By jointly acquiring the above multimodal cultivated land data, a multidimensional constraint basis is provided for the subsequent layout of sampling points.

[0058] It is understandable that the sampling point layout mechanism based on multimodal consistency constraints does not simply rely on geometric spatial location for uniform or random sampling. Instead, it comprehensively considers the spatial distribution of cultivated land, the consistency of soil types, and the similarity of cultivated land management conditions. Under the premise of satisfying the preset spatial uniformity constraint and minimum sampling interval constraint, the target area is divided into several basic spatial units with relatively consistent attribute characteristics, and representative soil sampling points are laid out in each basic spatial unit, thereby improving the ability of sampling points to characterize the spatial characteristics of soil salinity in the target area.

[0059] Step S20: Based on the set of soil sampling point locations P, a sample construction mechanism based on spatial constraint consistency is used to perform the task of generating salinity monitoring samples, and output the salinity sample set D;

[0060] Understandably, the sample construction mechanism based on spatial constraint consistency uses the set of soil sampling point locations P output in step S10 as a foundation. While maintaining the spatial distribution structure of the sampling points, it maps and pairs the spatial coordinates of each sampling point with its corresponding soil salinity content, avoiding spatial mismatches or label confusion during sample construction. This ensures the consistency and traceability of the supervised samples in the spatial dimension. The salinity sample set D obtained through this mechanism retains the spatial distribution characteristics of the sampling points within the target area and ensures that each sample has a clear and reliable salinity content label. This provides a standardized supervised data foundation for subsequent remote sensing pixel-level feature alignment and the training of the soil salinity inversion model, improving the stability of the model training process and the reliability of the inversion results.

[0061] Step S30: Based on the salinity sample set D, a pixel-level remote sensing feature alignment mechanism is used to perform a consistency alignment task, outputting a remote sensing pixel sample set; Step S30, the step of performing a consistency alignment task based on the salinity sample set D using a pixel-level remote sensing feature alignment mechanism to output a remote sensing pixel sample set, specifically includes: Step S301: Obtaining ground sampling time information corresponding to the salinity sample set D, and obtaining multi-temporal remote sensing image data; extracting time range information from the multi-temporal remote sensing image data, and calculating based on the time range information and sampling time information. The time offset difference is used to select remote sensing images from multi-temporal remote sensing image data whose time offset difference is less than or equal to a preset time consistency threshold, and output the target remote sensing image data; Step S302: Radiometric calibration, atmospheric correction, image fusion and image cropping are performed on the target remote sensing image data in sequence to output a standardized digital orthophoto; Step S303: The spatial location coordinates in the salt content sample set D are obtained, and pixel-level indexing and positioning are performed using a point sampling method based on nearest neighbor pixel positioning based on the standardized digital orthophoto and spatial location coordinates, and the remote sensing pixel sample set is output.

[0062] Understandably, calculating the time offset difference between the ground sampling time and the remote sensing image acquisition time, and filtering target remote sensing image data based on a preset time consistency threshold, is to avoid inconsistencies between the ground sampling data and the remote sensing image representation caused by changes in soil salinity over short timescales due to factors such as rainfall, irrigation, or evaporation. This time consistency filtering process ensures that the remote sensing image used for subsequent alignment accurately reflects the soil salinity state near the ground sampling time. Using a point sampling method based on nearest neighbor pixel localization for pixel-level indexing, while ensuring consistency between the spatial reference of the remote sensing image and the coordinates of the ground sampling points, maps each salinity sample to a specific pixel location in the remote sensing image, thus establishing a one-to-one correspondence between salinity samples and remote sensing pixels. The remote sensing pixel sample set obtained in this way maintains spatial consistency with the salinity sample set, providing reliable input data for subsequent remote sensing feature construction and soil salinity inversion model training.

[0063] Step S40: Based on the remote sensing pixel sample set, a statistical calculation method based on remote sensing band reflectance and spectral index is used to perform the feature construction process, and an initial remote sensing pixel-level feature vector is output; and based on the output initial feature vector, a sensitivity constraint screening is performed using a feature importance evaluation mechanism based on SHAP, and a set of sensitive feature vectors F is output.

[0064] It should be noted that the "statistical calculation method for remote sensing band reflectance and spectral indices" in this step refers to calculating the corresponding vegetation index and soil salinity correlation index for each remote sensing pixel sample based on the reflectance data of each band in the remote sensing image, according to the preset spectral index calculation formula. The band reflectance values ​​are then combined with the calculated spectral indices to form an initial feature vector at the remote sensing pixel level. This initial feature vector is used to comprehensively characterize the spectral response characteristics of remote sensing pixels. The corresponding vegetation index and soil salinity correlation index can be specifically shown in the table below:

[0065]

[0066]

[0067] For example, taking Gaofen-1, Gaofen-2, Sentinel-2, and Landsat-9 remote sensing images covering the target area as an example, preprocessing is performed on the multi-source remote sensing image data, including radiometric calibration, atmospheric correction, image fusion, image mosaicking, and image cropping, to produce a standard digital orthophoto. Based on the produced digital orthophoto, the reflectance of each soil sampling point in the red, green, blue, and near-infrared bands is extracted using the latitude and longitude coordinates recorded by a handheld GPS, and the corresponding vegetation index and soil salinity index are calculated to construct the characteristic variable level for soil salinity inversion. In this embodiment, the calculation results of Sentinel-2 remote sensing image are used as an example for illustration, as shown in the table below:

[0068]

[0069]

[0070] It is understandable that the initial remote sensing pixel-level feature vector obtained through the above feature construction process typically contains multiple band reflectance features and their derived spectral index features. Different features exhibit varying degrees of response to soil salinity. If all features are used directly for model training without differentiation, redundant or low-relevance features are easily introduced, thus affecting the stability of the subsequent soil salinity inversion model. The SHAP-based feature importance assessment mechanism, while maintaining the initial feature vector structure, quantifies the contribution of each feature to the model's prediction of soil salinity, ranks the features by importance, and selects sensitive features with high contributions to soil salinity inversion, forming a sensitive feature vector set F. This sensitivity-constrained selection process effectively reduces the interference of redundant features on the model training process and improves the model's efficiency in utilizing key features. For example, using the calculation results of Sentinel-2 remote sensing imagery as an example, based on the calculated SHAP values, three band reflectance, five vegetation indices, and seven soil salinity indices were selected, as shown in the table below:

[0071]

[0072] Step S50: Based on the sensitive feature vector set F, a deep learning-based regression modeling method is used to perform the soil salinity inversion task and output the soil salinization distribution results.

[0073] It should be noted that the soil salinization distribution results are determined based on the soil salinity inversion task combined with the preset soil salinization classification standards, which are shown in the table below:

[0074]

[0075] Understandably, taking region A as an example, after completing the deep learning-based regression prediction, the predicted soil salinity content of each remote sensing pixel within the target area can be divided into different salinity ranges according to a preset soil salinization classification standard, thus forming a spatial distribution result reflecting the degree of soil salinization in the target area. This spatial distribution result can intuitively characterize the distribution characteristics of soil salinization at the regional scale. For example, as... Figure 2 As shown, the inversion results indicate that the soil salinity in area A ranges from 0.24 g / kg to 10.20 g / kg, with an average of 1.74 g / kg. However, as... Figure 3 As shown, based on the inverted soil salinity content, the degree of soil salinization in area A is divided into 5 levels. Specifically, the statistical results of soil salinization levels can be shown in the table below:

[0076]

[0077] It should be understood that, since the present invention effectively constrains the model input features through sampling consistency constraints, pixel-level alignment, and SHAP-based sensitive feature screening in the preceding steps, the soil salinity inversion performed in this step can more stably characterize the nonlinear relationship between soil salinity and remote sensing features while reducing redundant feature interference, thereby improving the reliability of the inversion results and the consistency of spatial distribution.

[0078] Example 2: Furthermore, the present invention provides a farmland soil salinization inversion system based on remote sensing and deep learning, employing a farmland soil salinization inversion method based on remote sensing and deep learning from the above embodiments, which can solve the technical problem of farmland soil salinization inversion based on remote sensing and deep learning. The beneficial effects of the farmland soil salinization inversion system based on remote sensing and deep learning provided by the present invention are the same as those of the farmland soil salinization inversion method based on remote sensing and deep learning provided in the above embodiments, and other technical features of the farmland soil salinization inversion system based on remote sensing and deep learning are the same as those disclosed in the methods of the above embodiments, and will not be repeated here.

[0079] Example 3: This invention provides a farmland soil salinization inversion device based on remote sensing and deep learning. Please refer to... Figure 4A remote sensing and deep learning-based farmland soil salinization inversion device includes: at least one processor; and a memory communicatively connected to the at least one processor; wherein the memory stores instructions executable by the at least one processor, which are executed by the at least one processor to enable the at least one processor to execute the remote sensing and deep learning-based farmland soil salinization inversion method described in Embodiment 1 above. The remote sensing and deep learning-based farmland soil salinization inversion device in this embodiment of the invention may include, but is not limited to, mobile terminals such as mobile phones, laptops, digital radio receivers, PDAs (Personal Digital Assistants), PADs (Portable Application Descriptions), PMPs (Portable Media Players), vehicle terminals (e.g., vehicle navigation terminals), and fixed terminals such as digital TVs and desktop computers. This remote sensing and deep learning-based farmland soil salinization inversion device is merely an example and should not impose any limitations on the functionality and scope of use of the embodiments of the invention. A farmland soil salinization inversion device based on remote sensing and deep learning may include a processing unit 1001 (e.g., a central processing unit, a graphics processing unit, etc.), which can perform various appropriate actions and processes according to a program stored in a read-only memory 1002 or a program loaded from a storage device 1003 into a random access memory 1004. The random access memory 1004 also stores various programs and data required for the operation of the farmland soil salinization inversion device based on remote sensing and deep learning. The processing unit 1001, the read-only memory 1002, and the random access memory 1004 are interconnected via a bus 1005. An I / O interface 1006 is also connected to the bus. Typically, the following systems can be connected to I / O interface 1006: input devices 1007 including, for example, touchscreens, touchpads, keyboards, mice, image sensors, microphones, accelerometers, gyroscopes, etc.; output devices 1008 including, for example, liquid crystal displays (LCDs), speakers, vibrators, etc.; storage devices 1003 including, for example, magnetic tapes, hard disks, etc.; and communication devices 1009. Communication device 1009 allows a remote sensing and deep learning-based farmland soil salinization inversion device to wirelessly or wiredly communicate with other devices to exchange data. Although the figure shows a remote sensing and deep learning-based farmland soil salinization inversion device with various systems, it should be understood that it is not required to implement or possess all the systems shown. More or fewer systems can be implemented alternatively.

[0080] Example 4: This invention also provides a computer program product, including a computer program that, when executed by a processor, implements the steps of the above-described method for retrieving farmland soil salinization based on remote sensing and deep learning. The computer program product provided by this invention can solve the technical problem of retrieving farmland soil salinization based on remote sensing and deep learning. Compared with the prior art, the beneficial effects of the computer program product provided by this invention are the same as those of the above-described method for retrieving farmland soil salinization based on remote sensing and deep learning, and will not be repeated here.

[0081] In particular, according to the embodiments disclosed in this invention, the processes described above with reference to the flowcharts can be implemented as computer software programs. For example, embodiments of this invention include a computer program product comprising a computer program carried on a computer-readable medium, the computer program containing program code for performing the methods shown in the flowcharts. In such embodiments, the computer program can be downloaded and installed from a network via a communication device, or installed from storage device 1003, or installed from read-only memory 1002. When the computer program is executed by processing device 1001, it performs the functions defined in the methods of the embodiments disclosed in this invention.

[0082] It should be understood that the various parts disclosed in this invention can be implemented using hardware, software, firmware, or a combination thereof. In the description of the above embodiments, specific features, structures, materials, or characteristics may be combined in any suitable manner in one or more embodiments or examples.

[0083] Obviously, those skilled in the art can make various modifications and variations to this invention without departing from its spirit and scope. Therefore, if these modifications and variations fall within the scope of the claims of this invention and their equivalents, this invention also intends to include these modifications and variations.

Claims

1. A method for inverting farmland soil salinization based on remote sensing and deep learning, characterized in that, The methods include: Step S10: Obtain multimodal farmland data of the target area, and perform sampling spatial planning task based on the sampling point layout mechanism based on multimodal consistency constraints using the multimodal farmland data, and output the set of soil sampling point locations P; Step S20: Based on the set of soil sampling point locations P, a sample construction mechanism based on spatial constraint consistency is used to perform the task of generating salinity monitoring samples, and output the salinity sample set D; Step S30: Based on the salt content sample set D, a pixel-level remote sensing feature alignment mechanism is used to perform a consistency alignment task, and a remote sensing pixel sample set is output. Step S40: Based on the remote sensing pixel sample set, a statistical calculation method based on remote sensing band reflectance and spectral index is used to perform the feature construction process, and an initial remote sensing pixel-level feature vector is output; and based on the output initial feature vector, a sensitivity constraint screening is performed using a feature importance evaluation mechanism based on SHAP, and a set of sensitive feature vectors F is output. Step S50: Based on the sensitive feature vector set F, a deep learning-based regression modeling method is used to perform the soil salinity inversion task and output the soil salinization distribution results.

2. The method for retrieving farmland soil salinization based on remote sensing and deep learning as described in claim 1, characterized in that, Step S10 involves acquiring multimodal farmland data for the target area, performing sampling spatial planning based on a sampling point layout mechanism using multimodal consistency constraints, and outputting the soil sampling point location set P. Specifically, this includes: Step S101: Obtain multimodal farmland data for the target area. The multimodal farmland data includes farmland spatial distribution data, soil type data, and farmland management information. Based on the farmland spatial distribution data, divide the farmland into multiple basic spatial units. Step S102: For each basic spatial unit of cultivated land, based on soil type data and cultivated land management information, combined with a preset attribute and management joint judgment template, a consistency judgment method based on threshold driving is used to perform consistency judgment, and the template consistency score and basic spatial unit set are output. Step S103: Under the premise of satisfying the preset minimum sampling interval constraint, the sampling density self-adjustment point distribution method based on template consistency score is used to perform the adaptive generation task of sampling point spatial location for the basic spatial unit set, and output the soil sampling point location set P.

3. The method for retrieving farmland soil salinization based on remote sensing and deep learning as described in claim 1, characterized in that, Step S20, which involves generating salinity monitoring samples based on the set of soil sampling point locations P using a spatially constrained consistency-based sample construction mechanism, and outputting a salinity sample set D, specifically includes: Step S201: Collect target soil samples within a 0-20cm thickness range of the topsoil at each soil sampling point location in the soil sampling point location set P; Step S202: Measure the soil electrical conductivity EC of the target soil sample, and calculate the corresponding soil salinity SSC based on the soil electrical conductivity EC. ; Step S203: Bind the spatial coordinates of each soil sampling point to its corresponding soil salinity content SSC to form a salinity sample set D with spatial coordinates.

4. The method for retrieving farmland soil salinization based on remote sensing and deep learning as described in claim 1, characterized in that, Step S30, which involves performing a consistency alignment task based on the salinity content sample set D using a pixel-level remote sensing feature alignment mechanism to output the remote sensing pixel sample set, specifically includes: Step S301: Obtain ground sampling time information corresponding to the salt content sample set D, and obtain multi-temporal remote sensing image data; extract time range information from the multi-temporal remote sensing image data, calculate the time offset difference based on the time range information and sampling time information, filter remote sensing images from the multi-temporal remote sensing image data whose time offset difference is less than or equal to the preset time consistency threshold, and output the target remote sensing image data. Step S302: Perform radiometric calibration, atmospheric correction, image fusion, and image cropping sequentially on the target remote sensing image data to output a standardized digital orthophoto image; Step S303: Obtain the spatial coordinates of the salt content sample set D. Based on the standardized digital orthophoto and spatial coordinates, perform pixel-level indexing and positioning using a point sampling method based on nearest neighbor pixel positioning, and output the remote sensing pixel sample set.

5. The method for retrieving farmland soil salinization based on remote sensing and deep learning as described in claim 3, characterized in that, In step S40, the feature construction process is performed based on the remote sensing pixel sample set using a statistical calculation method based on remote sensing band reflectance and spectral index, outputting an initial remote sensing pixel-level feature vector; and based on the output initial feature vector, a sensitivity constraint screening is performed using a feature importance evaluation mechanism based on SHAP, outputting a sensitive feature vector set F. Specifically, this includes: Step S401: Obtain the blue band reflectance (B), green band reflectance (G), red band reflectance (R), and near-infrared band reflectance (NIR) corresponding to each remote sensing pixel from the remote sensing pixel sample set; calculate the corresponding Normalized Difference Vegetation Index (NDVI), Soil Salinity Index (SI), and Canopy Response Salinity Index (CRSI) based on the blue band reflectance (B), green band reflectance (G), red band reflectance (R), and near-infrared band reflectance (NIR) according to the preset spectral index calculation model; Step S402: Based on the blue band reflectance B, green band reflectance G, red band reflectance R, near-infrared band reflectance NIR, normalized difference vegetation index NDVI, soil salinity index SI, and canopy response salinity index CRSI, an initial remote sensing pixel-level feature vector is constructed using a feature stitching method. Step S403: Preset a machine learning model for predicting salinity content, obtain historical initial remote sensing pixel-level feature vectors and historical soil salinity content (SSC), use the historical initial remote sensing pixel-level feature vectors as input to the machine learning model for predicting salinity content, use the historical soil salinity content (SSC) as output to the machine learning model for predicting salinity content, and use the SHAP feature importance evaluation method to calculate the feature importance value of each feature in the process of predicting soil salinity content (SSC) in the machine learning model for predicting salinity content. Step S404: Select features whose importance value is greater than the preset importance threshold to form a set of sensitive feature vectors F.

6. The method for retrieving farmland soil salinization based on remote sensing and deep learning as described in claim 5, characterized in that, In step S40, the formula for the Normalized Difference Vegetation Index (NDVI) is expressed as follows: ; The formula for the soil salinity index (SI) is as follows: The formula for the canopy response salinity index (CRSI) is as follows: .

7. The method for inverting farmland soil salinization based on remote sensing and deep learning as described in claim 3, characterized in that, Step S50, which involves using a deep learning-based regression modeling method to perform soil salinity inversion based on the sensitive feature vector set F and outputting the soil salinization distribution results, specifically includes: Step S501: Design a soil salinity inversion model based on a deep neural network regression model in advance; use the sensitive feature vector set F as the input of the soil salinity inversion model, and use the soil salinity SSC as the supervision label of the soil salinity inversion model. Then, use grid search and ten-fold cross-validation to determine the optimal combination of model parameters for the soil salinity inversion model. Step S502: Obtain historical soil sampling point sample data and divide the historical soil sampling point sample data into training sample dataset and validation sample dataset; based on the training sample dataset, apply the optimized model parameter combination to perform a pre-training process on the soil salinity inversion model; After training, the coefficient of determination is used based on the validation sample dataset. Root mean square error and mean absolute error The model prediction accuracy is quantitatively evaluated, and the model performance evaluation score is output. The model performance evaluation score is used to determine the effectiveness of the soil salinity inversion model after training. Step S503: For model parameter combinations whose performance evaluation scores are greater than the preset model performance evaluation threshold, select the model parameter combination with the highest score as the optimal model parameter combination; apply the optimal model parameter combination to obtain the optimized soil salinity inversion model; input the sensitive feature vector set F into the optimized soil salinity inversion model, and the optimized soil salinity inversion model outputs the soil salinity inversion content; classify the soil salinity level of cultivated land according to the soil salinity inversion content and the preset salinity classification standard template, and finally output the soil salinity distribution results.

8. A farmland soil salinization inversion system based on remote sensing and deep learning, applied to the farmland soil salinization inversion method based on remote sensing and deep learning as described in any one of claims 1 to 7, characterized in that, The farmland soil salinization inversion system based on remote sensing and deep learning includes: The sampling space planning module is used to acquire multimodal farmland data of the target area. Based on the multimodal farmland data, it adopts a sampling point layout mechanism based on multimodal consistency constraints to perform sampling space planning tasks and outputs the set of soil sampling point locations P. The salinity content monitoring sample construction module is used to perform the task of generating salinity content monitoring samples based on the spatial constraint consistency sample construction mechanism based on the set of soil sampling point locations P, and outputs the salinity content sample set D; The pixel-level remote sensing feature alignment module is used to perform a consistency alignment task based on the salt content sample set D using a pixel-level remote sensing feature alignment mechanism, and outputs a remote sensing pixel sample set. The remote sensing feature construction and sensitivity screening module is used to perform the feature construction process based on the remote sensing pixel sample set and the statistical calculation method based on the remote sensing band reflectance and spectral index, and output the initial remote sensing pixel-level feature vector; and to perform sensitivity constraint screening based on the output initial feature vector using the SHAP-based feature importance evaluation mechanism, and output the sensitive feature vector set F. The soil salinity inversion module is used to perform soil salinity inversion tasks based on a set of sensitive feature vectors F and a deep learning-based regression modeling method, and outputs the soil salinization distribution results.

9. A farmland soil salinization inversion device based on remote sensing and deep learning, characterized in that, The farmland soil salinization inversion device based on remote sensing and deep learning includes: a memory, a processor, and a farmland soil salinization inversion program based on remote sensing and deep learning stored in the memory and executable on the processor. When the farmland soil salinization inversion program based on remote sensing and deep learning is executed by the processor, it implements a farmland soil salinization inversion method based on remote sensing and deep learning as described in any one of claims 1 to 7.

10. A computer program product, characterized in that, The computer program product includes a farmland soil salinization inversion program based on remote sensing and deep learning. When the farmland soil salinization inversion program based on remote sensing and deep learning is executed by the processor, it implements a farmland soil salinization inversion method based on remote sensing and deep learning as described in any one of claims 1 to 7.

Images