Carnallite pool deposition thickness classification method and system based on salt pan remote sensing image

Abstract

The invention relates to a carnallite pool deposition thickness classification method and system based on a salt pan remote sensing image. The method comprises the following steps: acquiring an optical remote sensing image containing a carnallite pool area, performing correction and cutting treatment, extracting optical features, and performing logarithmic transformation to obtain a processed image; marking a collected area and an uncollected area of the carnallite pool, and generating a label data file; using a self-attention mechanism to identify a line target of a salt mining ship trajectory, generating an attention weight value of a center pixel of the line target in combination with a line target weight method of physical experience, then carrying out splicing, fusion and convolution processing, and carrying out remodeling to generate a multi-dimensional feature map; and performing classification processing by using the target pixel-level classification model to obtain a deposition thickness classification image of the carnallite pool. By collecting an optical remote sensing image containing a carnallite pool area, combining with area labeling, and utilizing a salt mining ship to carry out line target identification, the carnallite pool area and the deposition state thereof can be accurately identified, and fine classification is realized.

Description

Technical Field

[0001] This invention relates to the field of optical remote sensing image environmental monitoring technology, and in particular to a method and system for classifying the deposition thickness of carnallite pools based on remote sensing images of salt fields. Background Technology

[0002] Carnallite ponds in salt pans are a crucial part of the salt production process, primarily used to extract abundant potassium, magnesium, and other minerals. Typically, workers regularly monitor crystal formation in the ponds, observing the color of the brine deposits and measuring the thickness of the carnallite deposits to determine the optimal harvesting time before using salt harvesting boats. However, this method is not only inefficient but also susceptible to measurement errors, leading to misjudgments of resources and reduced production efficiency. Optical remote sensing image classification is a key research direction in remote sensing technology, widely applied in land use / cover classification, agricultural resource monitoring, and disaster assessment. It offers significant advantages such as wide coverage, short monitoring cycles, and sensitivity to color changes in carnallite ponds, enabling dynamic monitoring of carnallite pond depositional characteristics and resource distribution, providing crucial support for refined management and assessment of harvesting efficiency. However, in the application of carnallite pools, classification research is still in its early stages. Currently, it mainly focuses on coarse classification, such as dividing the Qarhan Salt Lake area into major categories like salt lakes, sodium salt pools, carnallite pools, and other surface areas. There is a lack of refined classification research on uncollected and collected areas within the carnallite pools. This current deficiency in classification limits the potential of remote sensing technology in assessing the efficiency of carnallite resource collection and optimizing production.

[0003] Traditional monitoring of carnallite pond sediment thickness distribution primarily involves on-site measurements by workers from a ship in uncollected areas to determine if the sediment thickness meets the collection standard (20cm). On-site measurements require the salt harvesting vessel to advance gradually, measuring different areas one by one. This method is time-consuming and difficult to quickly complete large-scale carnallite pond sediment monitoring, especially in large salt fields where efficiency is particularly low. Due to complex environments and improper operation of measuring tools or personnel, measurement errors are prone to occur, leading to misjudgments. These errors can affect carnallite pond collection decisions, thus impacting production efficiency. Manual measurements can only cover partial areas and cannot obtain a real-time, comprehensive picture of the entire carnallite pond's sediment thickness distribution, potentially leading to omissions or inaccurate judgments. Furthermore, the complex environment of salt fields prevents workers from obtaining the sediment thickness distribution of carnallite pond areas at all times, potentially affecting the timeliness of collection and judgment.

[0004] Therefore, traditional monitoring of carnallite pool deposition thickness distribution is prone to errors due to the influence of human operation and environmental interference, which may lead to collection decision-making errors and reduced production efficiency. As a result, the accuracy of distribution classification is often low. Summary of the Invention

[0005] To address the aforementioned technical challenges, a method and system for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields are provided. This method can quickly and comprehensively acquire depositional thickness information of carnallite pools, accurately identify carnallite pool areas and their depositional states, and achieve fine-grained classification of carnallite depositional thickness.

[0006] A method for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields, the method comprising:

[0007] An optical remote sensing image containing the carnallite pool area is acquired, the optical remote sensing image is corrected and cropped, and the optical features are extracted and then logarithmically transformed to obtain the processed image.

[0008] Based on optical remote sensing images containing carnallite pool areas, the acquired and unacquired areas of the carnallite pools are marked, a label data file is generated, and the label data file is parsed to extract feature image block data.

[0009] Using a self-attention mechanism, the collection curve of the salt harvesting boat in the salt field is identified as a line target. Based on the feature image block data, a structural attention weight value for the center pixel of the line target is generated. Using a physical empirical model, an empirical attention weight value for the center pixel of the line target based on the empirical model is calculated. The structural attention weight value and the empirical attention weight value are concatenated, fused, and convolved to obtain the attention weight value for the center pixel of the line target and reconstruct a multi-dimensional feature map.

[0010] Select an image classification model, use the multidimensional feature map as a classification sample, use the image classification model to classify the classification sample and the processed image, and adjust the parameters of the image classification model according to the classification result to obtain a target pixel-level classification model;

[0011] The processed image is classified using the target pixel-level classification model to obtain a classification image of the deposition thickness of the carnallite pool.

[0012] In one embodiment, the optical remote sensing image is corrected and cropped, and after extracting optical features, a logarithmic transformation is performed to obtain a processed image, including:

[0013] Radiometric calibration and atmospheric correction are performed on the optical remote sensing effects to obtain the corrected image;

[0014] The carnallite pool region in the corrected image is determined, and the carnallite pool region is cropped to obtain the cropped image;

[0015] Obtain the feature dimension corresponding to the cropped image, extract optical features based on the feature dimension, obtain the image after feature extraction, and perform logarithmic transformation to obtain the processed image.

[0016] In one embodiment, optical features are extracted based on the feature dimension, including:

[0017] The feature dimensions include red light, green light, blue light, and infrared band;

[0018] Each land feature index is calculated based on the red, green, blue, and infrared bands.

[0019] The optical features are extracted by using each of the aforementioned land feature indices as the features of the label pixels.

[0020] In one embodiment, based on optical remote sensing imagery containing the carnallite pool area, the acquired and unacquired areas of the carnallite pool are labeled, generating a tag data file, including:

[0021] The location of the salt harvesting vessel is determined based on the optical remote sensing image containing the carnallite pool area, and the trajectory of the salt harvesting vessel is marked as the area of ​​the carnallite pool that has been collected based on the location of the salt harvesting vessel.

[0022] Based on the location of the salt harvesting vessel, the area not traversed by the salt harvesting vessel is marked as the area of ​​the carnallite pool that was not collected.

[0023] Use a new layer to fill the collected and uncollected areas of the carnallite pool with different colors, and save the layer;

[0024] Based on the saved layers, use ENVI to generate and save the label data file.

[0025] In one embodiment, parsing the label data file to extract feature image patch data includes:

[0026] Parse the label data file to read the defined polygon coordinate information;

[0027] Pixel regions are determined based on the polygon coordinate information, and feature image block data is extracted within the pixel regions.

[0028] In one embodiment, a self-attention mechanism is used to identify the collection curve of the salt harvesting boat in the salt field as a line target. A structural attention weight value for the center pixel of the line target is generated based on the feature image block data. An empirical attention weight value for the center pixel of the line target is calculated using a physical empirical model, including:

[0029] Based on the feature image block data, a query vector q, a key vector k, and a value vector v are generated for each pixel using a self-attention mechanism.

[0030] The score matrix of the center pixel and neighboring pixels in the feature image block data is calculated based on the query vector q, key vector k, and value vector v. The highest-scoring vector is determined based on the score matrix, and then a self-attention mechanism is used for iteration to complete the recognition of a line target of length N and generate a token set. The iteration formula for the token is: in Q, K, and V are the query matrix, key matrix, and value matrix of a 3x3 image patch within the feature image patch, respectively; the structural attention weights for the center pixel of the target line based on the structural model are...

[0031] Based on the aforementioned feature image patch data, a similarity metric is used to determine line targets of horizontal or vertical length N for the salt harvesting vessel. Arithmetic sequence weighting, Euclidean distance weighting, and probability distribution weighting are used to extract the straight-line features collected by the salt harvesting vessel. The empirical attention weight value of the center pixel of the line target is calculated; the calculation formula is: Where T is the set of target pixels for a horizontal or vertical line of length N; w i Weights for arithmetic sequences: N[·] is the normalization function.

[0032] In one embodiment, the structural attention weight values ​​and empirical attention weight values ​​are concatenated, fused, and convolved to obtain the attention weight value of the center pixel of the line target, and then reconstructed to generate a multi-dimensional feature map, including:

[0033] The attention weight value of the center pixel of the line target is obtained by fully splicing and fusing the structural attention weight value and the empirical attention weight value using a multilayer perceptron.

[0034] The attention weight values ​​are processed using one-dimensional convolution, and a multi-dimensional feature map is generated based on the center pixel data to achieve pixel-level classification.

[0035] The multidimensional feature map is processed by performing redundancy elimination through two-dimensional convolution to obtain the processed multidimensional feature map.

[0036] In one embodiment, an image classification model is selected, the multidimensional feature map is used as a classification sample, the image classification model is used to classify the classification sample and the processed image, and the parameters of the image classification model are adjusted according to the classification result to obtain a target pixel-level classification model, including:

[0037] The ViT model, which uses the ResNet and Transformer architectures in convolutional neural networks, is used as the image classification model. The classification samples are divided into training set, validation set and test set.

[0038] Set hyperparameters during model training, train the image classification model using the training set based on the hyperparameters, and update weights and adjust parameters through forward and backward propagation during training.

[0039] The performance of the image classification model was monitored using a validation set;

[0040] After the model training is completed, the trained image classification model is evaluated using a test set to obtain the final target pixel-level classification model;

[0041] The multidimensional feature map is input into the target pixel-level classification model to obtain a classification image of the deposition thickness of the carnallite pool.

[0042] In one embodiment, the target pixel-level classification model is a pixel-level remote sensing image classification model designed based on deep learning for scenarios of changes in the deposition thickness of carnallite pools. The target pixel-level classification model is trained on remote sensing sample data of the carnallite area and has the ability to distinguish different deposition thickness categories of carnallite pools based on the actual salt mining ship trajectory.

[0043] The principle of attention mechanism is used to automatically identify the trajectory features of salt harvesting boats in salt field production, and an attention weight distribution that fits the actual running direction of the salt harvesting boats is generated based on the trajectory features to enhance the ability of the target pixel-level classification model to distinguish the operating area of ​​the salt harvesting boats.

[0044] Based on the actual data collection method of the salt harvester, three line target weight allocation strategies are used to fuse line target information and enhance the target pixel-level classification model’s ability to identify the trajectory area of ​​the salt harvester.

[0045] The target pixel-level classification model uses deep learning to reconstruct a multi-dimensional feature map based on the features of the central pixel to achieve pixel-level classification.

[0046] A system for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields, the system comprising:

[0047] The image processing module is used to acquire optical remote sensing images containing the carnallite pool area, correct and crop the optical remote sensing images, extract optical features and perform logarithmic transformation to obtain the processed image.

[0048] The annotation module is used to annotate the acquired and unacquired areas of the carnallite pool based on optical remote sensing images containing the carnallite pool area, generate a label data file, and parse the label data file to extract feature image block data.

[0049] The line target recognition and attention weight value generation module is used to identify the collection curve of the salt harvesting boat in the salt field as a line target using a self-attention mechanism, generate the structural attention weight value of the center pixel of the line target based on the feature image block data, calculate the empirical attention weight value of the center pixel of the line target based on the empirical model using a physical empirical model, and perform splicing, fusion and convolution processing on the structural attention weight value and the empirical attention weight value to obtain the attention weight value of the center pixel of the line target and reshape it to generate a multi-dimensional feature map.

[0050] The model training module is used to select an image classification model, use the multidimensional feature map as a classification sample, use the image classification model to classify the classification sample and the processed image, and adjust the parameters of the image classification model according to the classification result to obtain a target pixel-level classification model.

[0051] The classification module is used to classify the processed image using the target pixel-level classification model to obtain a classification image of the deposition thickness of the carnallite pool.

[0052] The aforementioned method and system for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields can quickly and comprehensively acquire depositional thickness information of carnallite pools by acquiring and processing optical remote sensing images containing carnallite pool areas and combining this with regional annotation. Linear target identification and acquisition are performed using salt harvesting vessels in the carnallite pools of salt fields, and multi-dimensional feature maps are generated after convolutional reshaping. Then, a target pixel-level classification model is trained, which can accurately identify carnallite pool areas and their depositional states, achieving fine-grained classification and extraction of carnallite depositional thickness. Attached Figure Description

[0053] Figure 1 This is an application environment diagram of a method for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields in one embodiment.

[0054] Figure 2 This is a flowchart illustrating a method for classifying the deposition thickness of carnallite pools based on remote sensing images of salt fields in one embodiment.

[0055] Figure 3 This is a schematic diagram of the linear target attention module and classification in one embodiment;

[0056] Figure 4 This is a schematic diagram of the attention mechanism for iterative line target recognition in one embodiment;

[0057] Figure 5 This is a schematic diagram of Sentinel-2 image (10m) and tag in a carnallite pool area in one embodiment;

[0058] Figure 6This is a schematic diagram of the fine classification results of ResNet and ViT deposition thickness in one embodiment;

[0059] Figure 7 A comparison image of the deposition thickness classification of the carnallite pool obtained in one embodiment with the results of field survey;

[0060] Figure 8 This is a structural block diagram of a photonaphalite pool deposition thickness classification system based on remote sensing images of salt fields in one embodiment;

[0061] Figure 9 This is an internal structural diagram of a computer device in one embodiment. Detailed Implementation

[0062] To make the objectives, technical solutions, and advantages of this application clearer, the following detailed description is provided in conjunction with the accompanying drawings and embodiments. It should be understood that the specific embodiments described herein are merely illustrative and not intended to limit the scope of this application.

[0063] The photonaphalite pool deposition thickness classification method based on remote sensing images of salt fields provided in this application can be applied to, for example... Figure 1 The application environment shown. For example... Figure 1 As shown, the application environment includes computer equipment 110. Computer equipment 110 can acquire optical remote sensing images containing areas of carnallite pools, perform correction and cropping processing on the optical remote sensing images, extract optical features, and then perform logarithmic transformation to obtain the processed image; computer equipment 110 can label the acquired and unacquired areas of carnallite pools based on the optical remote sensing images containing carnallite pool areas, generate a label data file, and parse the label data file to extract feature image block data; computer equipment 110 can use a self-attention mechanism to identify the acquisition curve of the salt harvesting boat in the salt field as a line target, and generate a structural attention weight value for the center pixel of the line target based on the feature image block data; using a physical empirical model, it calculates... The empirical attention weight value of the center pixel of the line target is calculated based on the empirical model; the structural attention weight value and the empirical attention weight value are concatenated, fused, and convolved to obtain the attention weight value of the center pixel of the line target and reconstruct a multi-dimensional feature map; the computer device 110 can select an image classification model, use the multi-dimensional feature map as a classification sample, use the image classification model to classify the classification sample and the processed image, and adjust the parameters of the image classification model according to the classification result to obtain a target pixel-level classification model; the computer device 110 can use the target pixel-level classification model to classify the processed image to obtain a classification image of the deposition thickness of the carnallite pool. The computer device 110 can be, but is not limited to, various personal computers, laptops, smartphones, robots, tablets, etc.

[0064] In one embodiment, such as Figure 2 As shown, a method for classifying the deposition thickness of carnallite pools based on remote sensing images of salt fields is provided, including the following steps:

[0065] Step 202: Obtain an optical remote sensing image containing the carnallite pool area, perform correction and cropping on the optical remote sensing image, extract optical features and perform logarithmic transformation to obtain the processed image.

[0066] Computer equipment acquires optical remote sensing images of the carnallite pool area, requiring the download of optical remote sensing data from a website. In this embodiment, as shown... Figure 3 As shown in the figure, the classification results of carnallite pool sediment thickness extraction are illustrated using Sentinel-2 data as an example.

[0067] Computer equipment can perform preprocessing on acquired optical remote sensing images, including radiometric and geometric corrections. This primarily involves correcting the optical remote sensing images to eliminate external environmental interference and cropping the preprocessed carnallite pool area to reduce computational load during classification. Next, the computer equipment can perform feature extraction and logarithmic transformation on the carnallite pool images to enhance details in dark areas. Feature extraction increases data dimensionality, and logarithmic transformation improves overall image contrast and detail, resulting in the processed image. For Sentinel-2 L2A data, the correction step can be skipped. For Sentinel-2 L1C data, sen2cor software is used for radiometric calibration and atmospheric correction. Other optical remote sensing data can be preprocessed using corresponding optical remote sensing software such as SNAP and ENVI.

[0068] In one embodiment, a method for classifying the carnallite pool deposition thickness based on remote sensing images of salt fields may further include a process of processing the optical remote sensing image. The specific process includes: performing radiometric calibration and atmospheric correction on the optical remote sensing image to obtain a corrected image; determining the carnallite pool region in the corrected image and cropping the carnallite pool region to obtain a cropped image; obtaining the feature dimension corresponding to the cropped image, extracting optical features based on the feature dimension, obtaining the image after feature extraction, and performing a logarithmic transformation to obtain the processed image.

[0069] After acquiring optical remote sensing images covering the carnallite pool area, computer equipment can use specialized processing software (such as Sen2Cor, SNAP, etc.) to perform precise corrections such as radiometric calibration and atmospheric correction on the images, eliminating the influence of sensors and the atmosphere.

[0070] Next, the computer equipment can crop the carnallite pool region in the radiometrically and atmospherically corrected image. Specifically, the cropping process defines the spatial extent of the carnallite pool and uses image processing algorithms or software to precisely crop the target region, ensuring that subsequent analysis is performed only within the relevant area. After cropping the carnallite pool region, the vertex coordinates of the carnallite pool rectangle or irregular polygon can be recorded, and indexed cropping can be performed using code and saved, or the region of interest cropping and saving function can be used with ENVI.

[0071] For the cropped image, optical features are extracted by obtaining feature dimensions, and then a logarithmic transformation is performed to obtain the processed image. Feature extraction can also be performed using specialized remote sensing image processing software. The purpose of performing a logarithmic transformation on the feature-extracted image is to enhance details in low-brightness areas while compressing contrast in high-brightness areas, thus optimizing the image's brightness distribution. Therefore, the logarithmic transformation can effectively improve the visual effect of the image, especially in low-brightness areas, enhancing details and making weak features in the image more prominent, facilitating subsequent processing. In this embodiment, the formula for the logarithmic transformation can be expressed as: s = c·log(1+r); Where s is the transformed pixel value (the brightness value of the output image), r is the pixel value of the input image, c is a constant, and r max This is the maximum pixel value in the image, ensuring that the transformed image brightness value does not exceed the display range. After applying the logarithmic transformation, the brightness distribution of the image will be more uniform, thereby improving image visibility, especially in low-light and low-contrast areas, making subsequent classification and feature extraction processes more accurate.

[0072] After performing a logarithmic transformation, two-dimensional convolution can be further performed to extract deeper features. Taking Sentinel-2 data as an example, 256 features can be extracted through two-dimensional convolution.

[0073] In one embodiment, a method for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields may further include a process of extracting optical features. The specific process includes: feature dimensions including red light, green light, blue light, and infrared bands; obtaining soil brightness factors and calculating various land cover indices based on red light, green light, blue light, and infrared bands; and using each land cover index as a label pixel feature to extract optical features.

[0074] In this embodiment, the cropped image typically includes red, green, blue, and near-infrared bands, therefore its feature dimension is 4. To further enrich the features of the tag pixels, some commonly used optical features can be extracted, such as Normalized Difference Vegetation Index (NDVI), Water Index (NDWI), and Soil Adjusted Vegetation Index (SAVI), as tag pixel features. The formulas for each can be expressed as follows: Here, NIR represents the near-infrared band, Red represents the red light band, Green represents the green light band, and L is the soil brightness factor, typically set to 0.5. These indices effectively provide rich information about vegetation, water bodies, and soil, further improving the accuracy of classification and detection.

[0075] Step 204: Based on the optical remote sensing image containing the carnallite pool area, mark the acquired and unacquired areas of the carnallite pool, generate a label data file, and parse the label data file to extract feature image block data.

[0076] Computer equipment can label the collected and uncollected areas around the salt-collecting vessel in the carnallite pool area, outputting label data in XML or TIF format, i.e., label data files, which are then parsed to extract pixel data.

[0077] In one embodiment, a method for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields may further include a process of generating a label data file. The specific process includes: determining the location of salt harvesters based on optical remote sensing images containing carnallite pool areas, and marking the salt harvester's trajectory as the collected area of ​​the carnallite pool based on the salt harvester's location; marking the area where the salt harvester did not travel as the uncollected area of ​​the carnallite pool based on the salt harvester's location; filling the collected and uncollected areas of the carnallite pool with different colors using a new layer, and saving the layer; and generating a label data file using ENVI based on the saved layer and saving it.

[0078] Computer equipment can help mark the salt harvester's trajectory as the collected area based on the salt harvester's location, and mark the non-traveled area as the uncollected area of ​​the carnallite pool.

[0079] In this embodiment, the selection of label regions follows these basic criteria: 1. Labels of different categories should have significant visual color differences to enhance feature discrimination and improve model recognition performance; 2. Category subdivision should be rationally planned according to actual needs, satisfying classification accuracy requirements while avoiding sample imbalance or feature overlap due to too many categories; 3. The labeled region should cover typical feature areas and their boundary transition areas (around the salt harvester) to ensure the model can comprehensively learn category characteristics; 4. The spatial distribution of labels should be as uniform as possible to avoid concentration in local areas affecting the model's generalization ability; 5. The scale of the labeled region should be consistent with the image resolution and the spatial scale of the classification task to ensure the completeness and effectiveness of the labeled information; 6. Label labeling should be based on field survey results to ensure a high degree of consistency between the label location and the deposition thickness type. These basic criteria will directly affect the performance and convergence efficiency of the classification model.

[0080] After labeling is complete, the generated label data can be output as XML format (each pixel represents category information) or TIF format, which facilitates the training and evaluation of subsequent classification models.

[0081] Specifically, computer devices can use Photoshop to create a new layer to select the label area based on the basic standards for label annotation, and assign different fill colors. The layer can then be exported and saved. Alternatively, ENVI can be used to annotate the label using the division of the region of interest, generate an XML file, and save it.

[0082] In one embodiment, a method for classifying the deposition thickness of carnallite pools based on remote sensing images of salt fields may further include a process of extracting feature image block data. The specific process includes: parsing the label data file and reading the defined polygon coordinate information; determining the pixel region based on the polygon coordinate information; and extracting feature image block data within the pixel region.

[0083] Computer equipment can parse XML or TIF files, read the polygon coordinate information defined within them, and then extract feature image block data from the corresponding pixel regions in the remote sensing image.

[0084] Step 206: Using a self-attention mechanism, the collection curve of the salt harvesting boat in the salt field is identified as a line target. Based on the feature image block data, the structural attention weight value of the center pixel of the line target is generated. Using a physical empirical model, the empirical attention weight value of the center pixel of the line target based on the empirical model is calculated. The structural attention weight value and the empirical attention weight value are spliced, fused and convolved to obtain the attention weight value of the center pixel of the line target and reconstructed to generate a multi-dimensional feature map.

[0085] The computer device can extract line targets based on a semi-empirical, semi-model approach within the labels, and then perform one-dimensional convolution on the pixels using line target attention to generate feature images, which serve as samples for subsequent image classification. The computer device labels the logarithmically transformed images, applies line target attention, and finally uses one-dimensional convolution to generate feature images, outputting label data and feature image classification samples in XML or TIF format.

[0086] In one embodiment, a method for classifying the deposition thickness of carnallite pools based on remote sensing images of salt fields may further include a process of implementing a self-attention mechanism. The specific process includes: generating a query vector q, a key vector k, and a value vector v for each pixel using a self-attention mechanism based on feature image block data; calculating the score matrix of the center pixel and neighboring pixels in the feature image block data based on the query vector q, key vector k, and value vector v, and determining the highest-scoring vector based on the score matrix; iterating using the self-attention mechanism to complete the identification of line targets of length N and generate a token set; outputting the structural attention weight value of the center pixel of the line target based on the token set and a structural model; and using a similarity metric to determine line targets of length N for salt harvesting vessels, extracting the straight-line features collected by the salt harvesting vessels using an arithmetic sequence weighting method, an Euclidean distance weighting method, and a probability distribution weighting method, and calculating the empirical attention weight value of the center pixel of the line target.

[0087] like Figure 4 As shown, the computer device can flatten the label pixel with its surrounding pixels and generate Q, K, and V vectors for each pixel in the attention mechanism. It then calculates the score matrix of the center pixel and its 3x3 neighboring elements using a self-attention mechanism, outputting the two highest-scoring attention vectors and the token vector of the center element. Next, the computer device can input the highest-scoring attention vector into the next self-attention mechanism, continuing to output a highest-scoring attention vector and its own token vector. Finally, through the above iterations, it achieves line target recognition of the center pixel based on an empirical model, such as... Figure 5 As shown, this method can effectively extract the curve features collected by salt harvesting vessels. The line target attention output for the central element is the sum of all relevant token values, calculated as follows:

[0088] The formula for the Token set can be expressed as:

[0089]

[0090] Where N is the length of the desired line target. q i The query vector centered at element k j v j Let the key vector and value vector of the neighboring pixel be respectively, then the structural attention weight value of the center pixel of the line target output by the structural model is...

[0091] In this embodiment, line targets are extracted based on a model. According to practical research, salt harvesting vessels mostly use horizontal or vertical sampling depending on the location of the salt fields. Therefore, this embodiment proposes a physical model method for horizontal and vertical line targets that conforms to the actual situation. First, based on the formula: Similarity metrics determine whether the salt harvester's trajectory is horizontal or vertical, and based on this, a physical model for the horizontal or vertical line target is selected. Attention weights are then allocated based on practical experience. Here, A is the central element vector, B is the element vectors of the horizontal and vertical line targets, and ||*|| represents the L2 norm of the * vector, with dimension d, where d is the feature dimension after the FE module. The similarity metric method is not unique; any method that can determine which type of horizontal or vertical element is similar to the central element is acceptable, such as Euclidean distance or Pearson correlation coefficient.

[0092] Computer equipment can determine the horizontal or vertical line target of the salt harvesting ship with a length of N based on the feature image patch data and use similarity measurement. It can also extract the straight line features collected by the salt harvesting ship using the arithmetic sequence weight method, the Euclidean distance weight method, and the probability distribution weight method, and calculate the attention weight value of the center pixel of the line target.

[0093] In this embodiment, three linear target attention methods are proposed: arithmetic sequence weighting, Euclidean distance weighting, and probability distribution weighting. Among them:

[0094] In the arithmetic sequence weighting method, it is assumed that during the linear data collection process by the salt harvester, the middle pixel value on the straight line has the largest weight, while the contribution of the two side pixel values ​​to the middle pixel value will decrease sequentially.

[0095] In the Euclidean distance weighting method, the center pixel is also considered to play a dominant role. Therefore, the Euclidean distance between edge pixels and the center pixel is calculated, and the weight is determined based on the distance. The closer the distance, the greater the weight; conversely, the farther the distance, the smaller the weight. Therefore, through e -x The function processes the data, and then normalizes the weights of the output elements, D. i,j This represents the Euclidean distance between pixel i and pixel j. The formula is as follows:

[0096] In the probability distribution weighting method, the depth acquisition area of ​​the carnallite pool is considered to play a dominant role. Therefore, a probability density function is used for weight allocation. The lower the probability density, the greater the weight; conversely, the higher the probability density, the smaller the weight. Therefore, it needs to be processed using the -ln(x) function, expressed by the formula: Where N is the length of the target line.

[0097] In this embodiment, based on the above three line target attention methods, the line target can be transformed into a weighted center pixel with upper and lower elements by combination. The combination of the three methods can effectively extract the straight line features collected by the salt harvesting ship. Finally, the output of the model-based line target attention method is the average of the three attention methods, expressed by the formula: Where T is the set of target pixels for a line with a horizontal or vertical length of N; w i w represents the weight of the arithmetic sequence. i The calculation formula can be expressed as:

[0098]

[0099] Where N[·] is the normalization function.

[0100] Taking Sentinel-2 data as an example, based on the label data file, pixel values ​​and their surrounding element values ​​within the label region are extracted. The elements in the image are flattened and input into the attention mechanism to identify line targets and output their token values ​​(256 dimensions). Simultaneously, based on the label data file, pixel values ​​and their surrounding element values ​​within the label region are extracted. Their linear characteristics are determined through similarity analysis. The average of three line attention methods is then used to form a new center pixel value with information about its upper and lower elements (256 dimensions). The pixels with line target information extracted by the two methods are concatenated to form a 512-dimensional vector. This vector is then fused with the line targets obtained from the two methods using a multilayer perceptron. Next, the 512-dimensional vector undergoes a one-dimensional convolution to reconstruct a feature image, followed by a two-dimensional convolution to reduce the redundancy generated by the one-dimensional convolution.

[0101] In this embodiment, the computer device can connect the two attention weight values ​​finally obtained from the semi-empirical and semi-model-based line target attention, and then pass them through a multilayer perceptron to better integrate the line target weight values ​​based on the semi-empirical and semi-model-based approach. This allows the model to combine physical prior knowledge and learn more complex distribution relationships through data-driven methods when calculating line target attention, thereby effectively improving the model's classification or detection capabilities.

[0102] In this embodiment, by providing a line target attention method, the actual travel path of the salt harvesting vessel is identified through the line target attention mechanism, thereby enhancing the refined classification of the carnallite pool area.

[0103] In one embodiment, a method for classifying the deposition thickness of carnallite pools based on remote sensing images of salt fields may further include a convolutional processing step. Specifically, this step includes: using a multilayer perceptron to fully stitch together and fuse structural attention weight values ​​and empirical attention weight values ​​to obtain the attention weight value of the center pixel of the line target; processing the attention weight values ​​using one-dimensional convolution and reshaping a multi-dimensional feature map based on the center pixel data to achieve pixel-level classification; and performing redundancy elimination processing on the multi-dimensional feature map using two-dimensional convolution to obtain the processed multi-dimensional feature map.

[0104] Computer equipment can perform convolutional processing on the attention weight values ​​P0', P0” of the center pixel of the line target, and reconstruct a multi-dimensional feature map based on the center pixel data.

[0105] The attention weights of the fused line after passing through a multilayer perceptron still represent channel-based vectorized features, which may lose some local information if directly used in subsequent tasks. Therefore, in this embodiment, one-dimensional convolution is used to more effectively model the relationship between adjacent attention weights, and then reconstruct a multi-dimensional feature map, thereby transforming it into a more spatially structured multi-dimensional feature map.

[0106] To further optimize the data structure and reduce the information redundancy that one-dimensional convolution may bring, computer devices can eliminate the redundancy of multiple one-dimensional convolutions through two-dimensional convolutions, enhance the local context awareness, enable the model to understand the distribution of attention weights at different scales, thereby better adapting to complex linear target structures, and can match existing mainstream classification network models, providing a new pixel-level classification model.

[0107] Step 208: Select an image classification model, use the multidimensional feature map as a classification sample, use the image classification model to classify the classification sample and the processed image, and adjust the parameters of the image classification model according to the classification results to obtain the target pixel-level classification model.

[0108] Computer equipment can select a good image classification model, then train the model, and then validate the model.

[0109] In one embodiment, a method for classifying the deposition thickness of carnallite pools based on remote sensing images of salt fields may further include a model training process. This process includes: using a ViT model with a ResNet and Transformer architecture from a convolutional neural network as the image classification model; dividing the classification samples into a training set, a validation set, and a test set; setting hyperparameters during model training; training the image classification model using the training set based on the hyperparameters; updating weights and adjusting parameters during training through forward and backward propagation; monitoring the performance of the image classification model using the validation set; and evaluating the trained image classification model using the test set after model training to obtain the final target pixel-level classification model.

[0110] The computer equipment can select a suitable image classification model. When the resolution of optical remote sensing images is high, a large number of data samples will exist in small areas, in which case the ViT model can be selected for pre-training. Therefore, in this embodiment, the ResNet and Transformer architecture ViT models from convolutional neural networks can be used as image classification models.

[0111] Before training the classification model, you need to use the transforms.Resize function in PyTorch to scale the image to a suitable size to match the model's input, train it on the training set, test the network's classification performance on the validation set, and end the training when a certain accuracy is achieved or when the training rounds are finished.

[0112] In this embodiment, all samples can be divided into training, validation, and test sets in a 6:2:2 ratio for training. Then, hyperparameters are set during training, including learning rate, batch size, optimizer (e.g., Adam), and loss function (e.g., cross-entropy). During training, network weights are updated through forward and backpropagation, and the optimizer adjusts its parameters based on the loss function value. A validation set is used to monitor model performance during training to prevent overfitting.

[0113] For example, using a ResNet18 network with a learning rate of 3e-4, an Adam optimizer, a batch size of 128, 100 training epochs, and a cross-entropy loss function, and fine-tuning experiments with ViT_b_16 with a learning rate of 3e-4, an Adam optimizer, a batch size of 64, 10 training epochs, and a cross-entropy loss function, the learning rate scheduling and momentum factor parameters in the Adam optimizer can be set according to requirements to complete model training.

[0114] After model training is complete, a test set can be used to perform a final performance evaluation. If the accuracy meets expectations, training ends and the model parameters are saved. Specifically, the model's effectiveness can be verified by calculating the accuracy using the following formula: Where T is the total number of correctly predicted samples and F is the total number of incorrectly predicted samples.

[0115] Step 210: Use the target pixel-level classification model to classify the processed image to obtain a classification image of the deposition thickness of the carnallite pool.

[0116] Computer equipment can use a target pixel-level classification model to classify multidimensional feature maps, obtaining a classification map of carnallite deposition thickness in the study area. Using existing deep learning image classification models, a refined classification of the carnallite pool region is achieved, and the classification model parameters are saved to obtain the target classification model. This model is then used to classify the processed image, ultimately outputting a classification result map reflecting the carnallite pool deposition thickness. Specifically, the ViT model using the ResNet and Transformer architectures from convolutional neural networks is used as the image classification model to achieve a refined classification result of deposition thickness, as shown in the image. Figure 6 As shown.

[0117] In one embodiment, a method for classifying the deposition thickness of carnallite pools based on remote sensing images of salt fields may further include a classification process, specifically including: inputting a multidimensional feature map into a target pixel-level classification model, obtaining the category of each pixel through the target pixel-level classification model; and rearranging the multidimensional feature map according to its spatial structure based on the category of each pixel to obtain a classification image of the deposition thickness of the carnallite pool.

[0118] Computer equipment can use a target classification model to predict the multidimensional feature map, output the category of each pixel, and rearrange them according to the spatial structure of the image to output the final classification image of the deposition thickness of the carnallite pool.

[0119] In this embodiment, as Figure 7 As shown, the obtained classification images of carnallite pool sediment thickness can be compared with the results of field surveys to preliminarily verify the classification results. The field surveys can be conducted using drones.

[0120] In one embodiment, the target pixel-level classification model is a pixel-level remote sensing image classification model based on deep learning, designed to address the characteristics of varying deposition thickness in carnallite pools. The model is trained on remote sensing sample data of the carnallite region, enabling it to distinguish different deposition thickness categories in carnallite pools based on actual salt-harvesting vessel trajectories. Specifically, this includes:

[0121] The system automatically identifies the trajectory features of salt harvesting vessels in actual salt field production using the principle of attention mechanism, and generates an attention weight distribution that fits the actual operating direction of the salt harvesting vessels based on the trajectory, thereby enhancing the classification model's ability to distinguish the operating area of ​​the salt harvesting vessels. According to the actual working (horizontal or vertical) data collection method of the salt harvesting vessels, three line target weight allocation strategies are used to fuse line target information and enhance the model's ability to identify the trajectory area of ​​the salt harvesting vessels. Deep learning is used to reconstruct the multidimensional feature map of the center pixel features to achieve pixel-level classification.

[0122] This application presents a method for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields. To predict the depositional thickness classification of carnallite pool areas, the optical remote sensing images of the carnallite pool areas must undergo the same preprocessing steps as the training set. This includes feature extraction and pixel reshaping. For example, after logarithmic transformation, the input image extracts the same optical features (such as NDVI, NDWI, etc.) as the training set, and preprocesses each pixel to match the input distribution during model training. These steps ensure the consistency between the input data and the training data, thereby effectively enabling model prediction and outputting a detailed classification result map of the depositional thickness in the carnallite area.

[0123] In this application, considering the unique characteristics of linear data collection by salt harvesters in carnallite ponds, a semi-model, semi-empirical linear target attention method is proposed. This method primarily addresses the discontinuity in sediment thickness classification caused by the discontinuous trajectory of the salt harvesters, enhancing the distinguishability of different sediment thicknesses. To address the challenges of narrow salt harvester trajectories, low remote sensing image resolution, and high pixel similarity between collected and uncollected areas, a method is proposed to reconstruct individual pixels using one-dimensional convolution. This method preserves the original pixel value features while utilizing mainstream classification networks to classify pixels with high pixel similarity, achieving pixel-level classification and thus achieving refined classification. Furthermore, addressing the current lack of refined classification for the crucial carnallite pond region in salt lake chemical production, this application is the first to apply deep learning classification methods to the refined classification of carnallite pond areas, and its effectiveness has been verified through field research.

[0124] It should be understood that although the steps in the flowcharts above are shown sequentially as indicated by the arrows, these steps are not necessarily executed in the order indicated by the arrows. Unless explicitly stated herein, there is no strict order restriction on the execution of these steps, and they can be executed in other orders. Moreover, at least some steps in the flowcharts above may include multiple sub-steps or multiple stages. These sub-steps or stages are not necessarily completed at the same time, but can be executed at different times. The execution order of these sub-steps or stages is not necessarily sequential, but can be performed alternately or in turn with other steps or at least some of the sub-steps or stages of other steps.

[0125] In one embodiment, such as Figure 8 As shown, a system for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields is provided, including: an image processing module 810, an annotation module 820, a line target recognition module 830, a model training module 840, and a classification module 850, wherein:

[0126] Image processing module 810 is used to acquire optical remote sensing images containing the carnallite pool area, correct and crop the optical remote sensing images, extract optical features and perform logarithmic transformation to obtain the processed image.

[0127] The annotation module 820 is used to annotate the acquired and unacquired areas of the carnallite pool based on optical remote sensing images containing the carnallite pool area, generate a label data file, and parse the label data file to extract feature image block data.

[0128] The line target recognition and attention weight value generation module 830 is used to identify the collection curve of the salt harvesting boat in the salt field as a line target using a self-attention mechanism, generate the structural attention weight value of the center pixel of the line target based on the feature image block data, calculate the empirical attention weight value of the center pixel of the line target based on the empirical model using a physical empirical model, and perform concatenation, fusion and convolution processing on the structural attention weight value and the empirical attention weight value to obtain the attention weight value of the center pixel of the line target and reshape it to generate a multi-dimensional feature map.

[0129] The model training module 840 is used to select an image classification model, use multi-dimensional feature maps as classification samples, use the image classification model to classify the classification samples and the processed images, and adjust the parameters of the image classification model according to the classification results to obtain the target pixel-level classification model.

[0130] The classification module 850 is used to classify the processed image using a target pixel-level classification model to obtain a classification image of the deposition thickness of the carnallite pool.

[0131] In one embodiment, the image processing module 810 is further configured to perform radiometric calibration and atmospheric correction on the optical remote sensing effects to obtain a corrected image; determine the carnallite pool region in the corrected image and crop the carnallite pool region to obtain a cropped image; obtain the feature dimension corresponding to the cropped image, extract optical features based on the feature dimension, obtain the image after feature extraction, and perform logarithmic transformation to obtain the processed image.

[0132] In one embodiment, the feature dimensions include red light, green light, blue light, and infrared bands; the image processing module 810 is also used to calculate each land feature index based on the red light, green light, blue light, and infrared bands; and to extract optical features by using each land feature index as a label pixel feature.

[0133] In one embodiment, the annotation module 820 is further configured to determine the location of the salt harvesting vessel based on the optical remote sensing image containing the carnallite pool area, and to annotate the salt harvesting vessel's trajectory as the area of ​​the carnallite pool that has been collected based on the location of the salt harvesting vessel; to annotate the area of ​​the carnallite pool that the salt harvesting vessel has not traversed as the area of ​​the carnallite pool that has not been collected based on the location of the salt harvesting vessel; to fill the collected area and the uncollected area of ​​the carnallite pool with different colors using a new layer, and to save the layer; and to generate a label data file using ENVI based on the saved layer and save it.

[0134] In one embodiment, the annotation module 820 is further configured to parse the label data file, read the defined polygon coordinate information, determine the pixel region based on the polygon coordinate information, and extract feature image block data within the pixel region.

[0135] In one embodiment, the line target recognition and attention weight value generation module 830 is further configured to generate a query vector q, a key vector k, and a value vector v for each pixel based on feature image block data using a self-attention mechanism; calculate the score matrix of the center pixel and neighboring pixels in the feature image block data based on the query vector q, key vector k, and value vector v, and determine the highest score vector based on the score matrix, then iterate using the self-attention mechanism to complete the recognition of a line target of length N and generate a token set; output the structural attention weight value of the center pixel of the line target based on the structural model according to the token set; and determine the line targets of length N of the salt harvesting ship using a similarity metric based on the feature image block data, extract the straight line features collected by the salt harvesting ship using the arithmetic sequence weight method, the Euclidean distance weight method, and the probability distribution weight method, and calculate the empirical attention weight value of the center pixel of the line target.

[0136] In one embodiment, the line target recognition and attention weight value generation module 830 is further configured to use a multilayer perceptron to fully splice and fuse structural attention weight values ​​and empirical attention weight values ​​to obtain the attention weight value of the center pixel of the line target; use one-dimensional convolution to process the attention weight value, and reshape and generate a multi-dimensional feature map based on the center pixel data to achieve pixel-level classification; and perform redundancy elimination processing on the multi-dimensional feature map through two-dimensional convolution to obtain the processed multi-dimensional feature map.

[0137] In one embodiment, the model training module 840 is further configured to use the ViT model with the ResNet and Transformer architecture in a convolutional neural network as an image classification model, divide the classification samples into a training set, a validation set, and a test set; set hyperparameters during model training, train the image classification model using the training set based on the hyperparameters, and update weights and adjust parameters through forward and backward propagation during training; monitor the performance of the image classification model using the validation set; and evaluate the trained image classification model using the test set after model training is completed to obtain the final target pixel-level classification model.

[0138] In one embodiment, the classification module 850 is further configured to input the multidimensional feature map into the target pixel-level classification model, obtain the category of each pixel through the target pixel-level classification model, and rearrange the multidimensional feature map according to the spatial structure of the multidimensional feature map based on the category of each pixel to obtain the deposition thickness classification image of the carnallite pool.

[0139] In one embodiment, a computer device is provided, which may be a terminal, and its internal structure diagram may be as follows: Figure 9 As shown, the computer device includes a processor, memory, network interface, display screen, and input devices connected via a system bus. The processor provides computing and control capabilities. The memory includes non-volatile storage media and internal memory. The non-volatile storage media stores the operating system and computer programs. The internal memory provides an environment for the operation of the operating system and computer programs stored in the non-volatile storage media. The network interface is used for communication with external terminals via a network connection. When executed by the processor, the computer program implements a method for classifying the deposition thickness of photonaphthalene ponds based on remote sensing images of salt fields. The display screen can be a liquid crystal display (LCD) or an e-ink display. The input devices can be a touch layer covering the display screen, buttons, a trackball, or a touchpad mounted on the computer device casing, or an external keyboard, touchpad, or mouse.

[0140] Those skilled in the art will understand that Figure 9The structure shown is merely a block diagram of a portion of the structure related to the present application and does not constitute a limitation on the computer device to which the present application is applied. Specific computer devices may include more or fewer components than those shown in the figure, or combine certain components, or have different component arrangements.

[0141] In one embodiment, a computer device is provided, including a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of a method for classifying the deposition thickness of carnallite pools based on remote sensing images of salt fields.

[0142] In one embodiment, a computer-readable storage medium is provided having a computer program stored thereon, which, when executed by a processor, implements the steps of a method for classifying the deposition thickness of carnallite pools based on remote sensing images of salt fields.

[0143] Those skilled in the art will understand that all or part of the processes in the methods of the above embodiments can be implemented by a computer program instructing related hardware. The computer program can be stored in a non-volatile computer-readable storage medium, and when executed, it can include the processes of the embodiments of the above methods. Any references to memory, storage, databases, or other media used in the embodiments provided in this application can include non-volatile and / or volatile memory. Non-volatile memory can include read-only memory (ROM), programmable ROM (PROM), electrically programmable ROM (EPROM), electrically erasable programmable ROM (EEPROM), or flash memory. Volatile memory can include random access memory (RAM) or external cache memory. By way of illustration and not limitation, RAM is available in various forms, such as static RAM (SRAM), dynamic RAM (DRAM), synchronous DRAM (SDRAM), dual data rate SDRAM (DDRSDRAM), enhanced SDRAM (ESDRAM), synchronous link DRAM (SLDRAM), Rambus direct RAM (RDRAM), direct memory bus dynamic RAM (DRDRAM), and memory bus dynamic RAM (RDRAM), etc.

[0144] The technical features of the above embodiments can be combined in any way. For the sake of brevity, not all possible combinations of the technical features in the above embodiments are described. However, as long as there is no contradiction in the combination of these technical features, they should be considered to be within the scope of this specification.

[0145] The embodiments described above are merely illustrative of several implementation methods of this application, and while the descriptions are relatively specific and detailed, they should not be construed as limiting the scope of the invention patent. It should be noted that those skilled in the art can make various modifications and improvements without departing from the concept of this application, and these all fall within the protection scope of this application. Therefore, the protection scope of this patent application should be determined by the appended claims.

Claims

1. A method for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields, characterized in that, The method includes: An optical remote sensing image containing the carnallite pool area is acquired, the optical remote sensing image is corrected and cropped, and the optical features are extracted and then logarithmically transformed to obtain the processed image. Based on optical remote sensing images containing carnallite pool areas, the acquired and unacquired areas of the carnallite pools are marked, a label data file is generated, and the label data file is parsed to extract feature image block data. Using a self-attention mechanism, the collection curve of the salt harvesting boat in the salt field is identified as a line target. Based on the feature image block data, a structural attention weight value for the center pixel of the line target is generated. Using a physical empirical model, an empirical attention weight value for the center pixel of the line target based on the empirical model is calculated. The structural attention weight value and the empirical attention weight value are concatenated, fused, and convolved to obtain the attention weight value for the center pixel of the line target and reconstruct a multi-dimensional feature map. Select an image classification model, use the multidimensional feature map as a classification sample, use the image classification model to classify the classification sample and the processed image, and adjust the parameters of the image classification model according to the classification result to obtain a target pixel-level classification model; The processed image is classified using the target pixel-level classification model to obtain a classification image of the deposition thickness of the carnallite pool.

2. The method for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields according to claim 1, characterized in that, The optical remote sensing image is corrected and cropped, and its optical features are extracted and then logarithmically transformed to obtain the processed image, including: Radiometric calibration and atmospheric correction are performed on the optical remote sensing effects to obtain the corrected image; The carnallite pool region in the corrected image is determined, and the carnallite pool region is cropped to obtain the cropped image; Obtain the feature dimension corresponding to the cropped image, extract optical features based on the feature dimension, obtain the image after feature extraction, and perform logarithmic transformation to obtain the processed image.

3. The method for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields according to claim 2, characterized in that, Optical features are extracted based on the aforementioned feature dimensions, including: The feature dimensions include red light, green light, blue light, and infrared band; Each land feature index is calculated based on the red, green, blue, and infrared bands. The optical features are extracted by using each of the aforementioned land feature indices as the features of the label pixels.

4. The method for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields according to claim 1, characterized in that, Based on optical remote sensing images containing the carnallite pool area, the acquired and unacquired areas of the carnallite pool are labeled, generating a labeled data file, including: The location of the salt harvesting vessel is determined based on the optical remote sensing image containing the carnallite pool area, and the trajectory of the salt harvesting vessel is marked as the area of ​​the carnallite pool that has been collected based on the location of the salt harvesting vessel. Based on the location of the salt harvesting vessel, the area not traversed by the salt harvesting vessel is marked as the area of ​​the carnallite pool that was not collected. Use a new layer to fill the collected and uncollected areas of the carnallite pool with different colors, and save the layer; Based on the saved layers, use ENVI to generate and save the label data file.

5. The method for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields according to claim 1, characterized in that, The process of parsing and extracting feature image block data from the label data file includes: Parse the label data file to read the defined polygon coordinate information; Pixel regions are determined based on the polygon coordinate information, and feature image block data is extracted within the pixel regions.

6. The method for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields according to claim 1, characterized in that, Using a self-attention mechanism, the collection curve of the salt harvesting boat in the salt field is identified as a line target, and a structural attention weight value of the center pixel of the line target is generated based on the feature image block data. Using a physical empirical model, the empirical attention weight values ​​for the center pixel of the line target are calculated, including: Based on the feature image block data, a query vector q, a key vector k, and a value vector v are generated for each pixel using a self-attention mechanism. The score matrix of the center pixel and neighboring pixels in the feature image block data is calculated based on the query vector q, key vector k, and value vector v. The highest-scoring vector is determined based on the score matrix, and then a self-attention mechanism is used for iteration to complete the recognition of a line target of length N and generate a token set. The iteration formula for the token is: in Q, K, and V are the query matrix, key matrix, and value matrix of a 3x3 image patch within the feature image block, respectively; the structural attention weights for the center pixel of the target line based on the structural model are... Based on the aforementioned feature image patch data, a similarity metric is used to determine line targets of horizontal or vertical length N for the salt harvesting vessel. Arithmetic sequence weighting, Euclidean distance weighting, and probability distribution weighting are used to extract the straight-line features collected by the salt harvesting vessel. The empirical attention weight value of the center pixel of the line target is calculated; the calculation formula is: Where T is the set of target pixels for a horizontal or vertical line of length N; w i Weights for arithmetic sequences: N[·] is the normalization function.

7. The method for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields according to claim 1, characterized in that, The structural attention weights and empirical attention weights are concatenated, fused, and convolved to obtain the attention weights of the center pixels of the line target, and then reconstructed to generate a multi-dimensional feature map, including: The attention weight value of the center pixel of the line target is obtained by fully splicing and fusing the structural attention weight value and the empirical attention weight value using a multilayer perceptron. The attention weight values ​​are processed using one-dimensional convolution, and a multi-dimensional feature map is generated based on the center pixel data to achieve pixel-level classification. The multidimensional feature map is processed by performing redundancy elimination through two-dimensional convolution to obtain the processed multidimensional feature map.

8. The method for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields according to claim 1, characterized in that, Selecting an image classification model, using the multidimensional feature map as classification samples, classifying the classification samples and the processed image using the image classification model, and adjusting the parameters of the image classification model based on the classification results to obtain a target pixel-level classification model, including: The ViT model, which uses the ResNet and Transformer architectures in convolutional neural networks, is used as the image classification model. The classification samples are divided into training set, validation set and test set. Set hyperparameters during model training, train the image classification model using the training set based on the hyperparameters, and update weights and adjust parameters through forward and backward propagation during training. The performance of the image classification model was monitored using a validation set; After the model training is completed, the trained image classification model is evaluated using a test set to obtain the final target pixel-level classification model; The multidimensional feature map is input into the target pixel-level classification model to obtain a classification image of the deposition thickness of the carnallite pool.

9. The method for classifying the depositional thickness of carnallite pools based on remote sensing images of salt fields according to claim 1, characterized in that, The target pixel-level classification model is a pixel-level remote sensing image classification model designed based on deep learning for scenarios of changes in the deposition thickness of carnallite pools. The target pixel-level classification model is trained on remote sensing sample data of the carnallite area and has the ability to distinguish different deposition thickness categories of carnallite pools based on the actual salt mining ship trajectory. The principle of attention mechanism is used to automatically identify the trajectory features of salt harvesting boats in salt field production, and an attention weight distribution that fits the actual running direction of the salt harvesting boats is generated based on the trajectory features to enhance the ability of the target pixel-level classification model to distinguish the operating area of ​​the salt harvesting boats. Based on the actual data collection method of the salt harvesting vessel, three line target weight allocation strategies are used to fuse line target information and enhance the target pixel-level classification model’s ability to identify the trajectory area of ​​the salt harvesting vessel. The target pixel-level classification model uses deep learning to reconstruct a multi-dimensional feature map based on the features of the central pixel to achieve pixel-level classification.

10. A system for classifying the thickness of carnallite pool sediments based on remote sensing images of salt fields, characterized in that, The system includes: The image processing module is used to acquire optical remote sensing images containing the carnallite pool area, correct and crop the optical remote sensing images, extract optical features and perform logarithmic transformation to obtain the processed image. The annotation module is used to annotate the acquired and unacquired areas of the carnallite pool based on optical remote sensing images containing the carnallite pool area, generate a label data file, and parse the label data file to extract feature image block data. The line target recognition and attention weight value generation module is used to identify the collection curve of the salt harvesting boat in the salt field as a line target using a self-attention mechanism, generate the structural attention weight value of the center pixel of the line target based on the feature image block data, calculate the empirical attention weight value of the center pixel of the line target based on the empirical model using a physical empirical model, and perform splicing, fusion and convolution processing on the structural attention weight value and the empirical attention weight value to obtain the attention weight value of the center pixel of the line target and reshape it to generate a multi-dimensional feature map. The model training module is used to select an image classification model, use the multidimensional feature map as a classification sample, use the image classification model to classify the classification sample and the processed image, and adjust the parameters of the image classification model according to the classification result to obtain a target pixel-level classification model. The classification module is used to classify the processed image using the target pixel-level classification model to obtain a classification image of the deposition thickness of the carnallite pool.

Images