A method for image transformation of large remote sensing models of long-distance pipelines based on quicksand dynamics

Abstract

The present invention provides a method for transforming images of a large-scale remote sensing model of a long-distance pipeline based on quicksand dynamics, comprising the following steps: preprocessing source domain images and preprocessing target domain images; establishing a quicksand dynamics model; inputting the preprocessed target domain image, its geospatial coordinates, and the acquisition time of the target domain image into the quicksand dynamics model to perform quicksand dynamics transformation; establishing a large-scale remote sensing model of the long-distance pipeline, training the large-scale remote sensing model using the preprocessed source domain image, its geospatial coordinates, and the acquisition time of the source domain image; and inputting the target domain image that has undergone quicksand dynamics transformation into the large-scale remote sensing model of the long-distance pipeline to obtain remote sensing data of the long-distance pipeline. A technical effect of the present invention is that it has high computational efficiency and can effectively retain key information, thereby accurately interpreting images.

Description

Technical Field

[0001] The invention belongs to the technical field of remote sensing image processing, and in particular relates to a method for transforming a large remote sensing model image of a long-distance pipeline based on quicksand dynamics. Background Art

[0002] With the rapid development of remote sensing technology, more and more application scenarios require efficient and accurate image interpretation. However, existing remote sensing image processing technologies still face many challenges when dealing with image differences in complex terrain and different climatic conditions.

[0003] Currently, remote sensing image processing primarily relies on traditional image processing algorithms, such as feature extraction-based and pixel-based classification methods. These methods can achieve good results under specific conditions, but they often fail to maintain stable performance across diverse geographic environments and temporal variations. In particular, the domain gap problem arises. In practical applications, differences in feature distribution between the source domain (training data) and the target domain (data to be inferred) lead to degraded model performance. This is particularly true in long-distance pipeline monitoring, where existing models often fail to accurately interpret imagery due to topographical and seasonal variations.

[0004] Furthermore, traditional methods often lose or distort image features when processing different feature regions, failing to effectively retain key information. Furthermore, computational efficiency is a significant issue. Existing technologies often consume significant computing resources during feature extraction and classification, making it difficult to meet the needs of real-time monitoring. Summary of the Invention

[0005] The present invention aims to solve at least one of the technical problems existing in the prior art and provide a new technical solution for a method for transforming the image of a large remote sensing model of a long-distance pipeline based on quicksand dynamics.

[0006] According to a first aspect of the present invention, a method for converting a large remote sensing model image of a long-distance pipeline based on quicksand dynamics is provided, comprising the following steps:

[0007] Step S100, pre-processing the source domain image and pre-processing the target domain image;

[0008] Step S200: Establishing a quicksand dynamics model; inputting a preprocessed target domain image, the geospatial coordinates of the target domain image, and the acquisition time of the target domain image into the quicksand dynamics model to perform quicksand dynamics transformation; wherein, the transformation result of the target domain image is controlled by controlling the flipping frequency, angle, and number of bubbles of the quicksand in the quicksand dynamics model; Step S300: Establishing a large remote sensing model of a long-distance pipeline; and training the large remote sensing model of the long-distance pipeline using the preprocessed source domain image, the geospatial coordinates of the source domain image, and the acquisition time of the source domain image.

[0009] Step S400: inputting the target domain image after the quicksand dynamics transformation into the long-distance pipeline remote sensing large model to obtain remote sensing data of the long-distance pipeline.

[0010] Optionally, the pre-processed target domain image, the geospatial coordinates of the target domain image, and the acquisition time of the target domain image are input into the quicksand dynamics model to perform quicksand dynamics transformation, including:

[0011] Obtain pre-processed target domain images and classify them according to the geospatial coordinates of the target domain images and the acquisition time of the target domain images;

[0012] inputting the classified target domain image into the quicksand dynamics model;

[0013] Using the principles of quicksand dynamics, control parameters of the quicksand dynamics model are designed; wherein the control parameters include the turnover frequency, angle and number of bubbles of the quicksand;

[0014] The quicksand dynamics model is adaptive according to the geographic space coordinates of the target domain image and the acquisition time of the target domain image, and transforms the classified target domain image.

[0015] Optionally, the method for transforming a large remote sensing model image of a long-distance pipeline based on quicksand dynamics further includes:

[0016] Utilize reinforcement learning methods to optimize the mapping process of quicksand dynamics model.

[0017] Optionally, a reinforcement learning method is used to optimize the mapping process of the quicksand dynamics model, including:

[0018] Setting a reinforcement learning framework; wherein the reinforcement learning framework includes a state space, an action space, and a reward mechanism, and calculating the reward mechanism based on the state space and the action space;

[0019] The KL divergence is used to quantify the difference in the probability distribution of pixel space between the source domain and the target domain, a threshold is set, and the KL divergence is compared with the threshold to determine the similarity of the features of the source domain and the target domain;

[0020] During the reinforcement learning process, the quicksand dynamics model continuously adjusts the control parameters to calculate a new KL divergence in each round of learning, and adjusts the control parameters according to the reward mechanism until the set threshold conditions are met;

[0021] Selective state space patterns are introduced to optimize the performance of the quicksand dynamics model by focusing on specific state features. The selection criteria of the selective state space patterns are based on the similarity of spatial and temporal features, and a similarity metric is used to evaluate and select states.

[0022] Optionally, the KL divergence is calculated as follows:

[0023] ;

[0024] In the above formula, is the KL divergence difference between the source domain image P and the target domain image Q, is the source domain image P at position The probability of The target domain image Q is located at probability.

[0025] Alternatively, in the quicksand dynamics model, the calculation formula for the overturning frequency is as follows:

[0026] ;

[0027] In the above formula, For the mapping state, is the flip frequency, t is the time, and d represents the differential.

[0028] Alternatively, in the quicksand dynamics model, the calculation formula for the flip angle is as follows:

[0029] ;

[0030] In the above formula, is the vector of initial image features; After flipping angle Adjusted image feature vector; is the rotation transformation function.

[0031] Optionally, in the quicksand dynamics model, the bubble effect The calculation formula is as follows:

[0032] ;

[0033] In the above formula, is the adjusted velocity field; is the velocity field before adjustment.

[0034] Optionally, the probability distribution of the target domain image is calculated using the following formula:

[0035] ;

[0036] In the above formula, For different feature areas, is the probability value of the corresponding area; is the probability distribution function in two-dimensional space;

[0037] The calculation formula for the quicksand dynamics model transformation is as follows:

[0038] ;

[0039] In the above formula, is the transformation function, is the probability distribution of the current target domain image; is the probability distribution of the transformed target domain image.

[0040] Optionally, preprocess the source domain image, including:

[0041] First, the source domain image is orthorectified;

[0042] Secondly, geometric correction is performed on the orthorectified source domain image;

[0043] Secondly, the cloud and shadow are removed from the geometrically corrected source domain image;

[0044] Preprocess the target domain image, including:

[0045] First, the target domain image is processed;

[0046] Secondly, the brightness of the geometrically aligned target domain image is adjusted.

[0047] A technical effect of the present invention is:

[0048] In an embodiment of the present application, the long-distance pipeline remote sensing large model image transformation method based on quicksand dynamics optimizes the feature mapping of remote sensing images under different environmental conditions by introducing a quicksand dynamics model to ensure that the target domain image is highly similar to the source domain image in pixel space probability distribution, thereby improving the interpretation accuracy of the model in practical applications, and achieving the technical effect of effectively reducing the feature differences between the source domain and the target domain.

[0049] Furthermore, this method for transforming large-scale remote sensing models of long-distance pipelines based on quicksand dynamics can enhance the ability to retain image features. Through an adaptive control mechanism, combined with geographic coordinates (and the image's geospatial coordinates) and time information (i.e., the image acquisition time), it ensures that key information can be effectively retained during the image transformation process, reducing the loss or deformation of features.

[0050] In addition, through the transformation process of the quicksand dynamics model, not only the computational efficiency is significantly improved, but also the computational complexity of traditional image processing methods in feature extraction and classification is reduced, thereby realizing fast and real-time remote sensing image processing to meet the real-time needs of long-distance pipeline monitoring.

[0051] Moreover, the quicksand dynamics model of the long-distance pipeline remote sensing large-scale model image transformation method based on quicksand dynamics provides a flexible mapping framework, making this method widely applicable to remote sensing image processing in different regions and climatic conditions, thereby meeting the needs of different application scenarios. BRIEF DESCRIPTION OF THE DRAWINGS

[0052] Figure 1 This is a schematic structural diagram of a method for converting a large remote sensing model image of a long-distance pipeline based on quicksand dynamics according to an embodiment of the present invention;

[0053] Figure 2 is a schematic diagram of a source domain image according to an embodiment of the present invention;

[0054] Figure 3 is a schematic diagram of a target domain image according to an embodiment of the present invention;

[0055] Figure 4 for Figure 3 Schematic diagram of the target domain image after transformation using the long-distance pipeline remote sensing large model image transformation method based on quicksand dynamics. DETAILED DESCRIPTION

[0056] Various exemplary embodiments of the present application will now be described in detail with reference to the accompanying drawings. It should be noted that unless otherwise specifically stated, the relative arrangements of components and steps, numerical expressions and numerical values ​​set forth in these embodiments do not limit the scope of the present application.

[0057] The embodiments of the present application will be described in detail below. Examples of the embodiments are shown in the accompanying drawings, wherein the same or similar reference numerals throughout represent the same or similar elements or elements having the same or similar functions. The embodiments described below with reference to the accompanying drawings are exemplary and are only used to explain the present application and are not to be construed as limiting the present application. Based on the embodiments in the present application, all other embodiments obtained by ordinary technicians in this field without making creative efforts are within the scope of protection of this application.

[0058] The terms "first" and "second" in the specification and claims of this application may explicitly or implicitly refer to one or more of the features. Throughout the description of this application, unless otherwise specified, "plurality" means two or more. Furthermore, "and / or" in the specification and claims refers to at least one of the connected entities, and the character " / " generally indicates an "or" relationship between the connected entities.

[0059] According to the first aspect of the present invention, see Figure 1 , provides a long-distance pipeline remote sensing large model image transformation method based on quicksand dynamics, aiming to solve the challenges faced in existing remote sensing image processing technology through innovative quicksand dynamics mapping methods, and improve the processing effect and application value of remote sensing images.

[0060] Specifically, the method for transforming the image of a large remote sensing model of a long-distance pipeline based on quicksand dynamics includes the following steps:

[0061] Step S100, pre-processing the source domain image and pre-processing the target domain image;

[0062] Step S200: Establishing a quicksand dynamics model; inputting a preprocessed target domain image, the geospatial coordinates of the target domain image, and the acquisition time of the target domain image into the quicksand dynamics model to perform quicksand dynamics transformation; wherein, the transformation result of the target domain image is controlled by controlling the flipping frequency, angle, and number of bubbles of the quicksand in the quicksand dynamics model; Step S300: Establishing a large remote sensing model of a long-distance pipeline; and training the large remote sensing model of the long-distance pipeline using the preprocessed source domain image, the geospatial coordinates of the source domain image, and the acquisition time of the source domain image.

[0063] Step S400: inputting the target domain image after the quicksand dynamics transformation into the long-distance pipeline remote sensing large model to obtain remote sensing data of the long-distance pipeline.

[0064] In an embodiment of the present application, the long-distance pipeline remote sensing large model image transformation method based on quicksand dynamics optimizes the feature mapping of remote sensing images under different environmental conditions by introducing a quicksand dynamics model to ensure that the target domain image is highly similar to the source domain image in pixel space probability distribution, thereby improving the interpretation accuracy of the model in practical applications, and achieving the technical effect of effectively reducing the feature differences between the source domain and the target domain.

[0065] Furthermore, this method for transforming large-scale remote sensing models of long-distance pipelines based on quicksand dynamics can enhance the ability to retain image features. Through an adaptive control mechanism, combined with geographic coordinates (and the image's geospatial coordinates) and time information (i.e., the image acquisition time), it ensures that key information can be effectively retained during the image transformation process, reducing the loss or deformation of features.

[0066] In addition, through the transformation process of the quicksand dynamics model, not only the computational efficiency is significantly improved, but also the computational complexity of traditional image processing methods in feature extraction and classification is reduced, thereby realizing fast and real-time remote sensing image processing to meet the real-time needs of long-distance pipeline monitoring.

[0067] Moreover, the quicksand dynamics model of the long-distance pipeline remote sensing large-scale model image transformation method based on quicksand dynamics provides a flexible mapping framework, making this method widely applicable to remote sensing image processing in different regions and climatic conditions, thereby meeting the needs of different application scenarios.

[0068] For example, the large-scale image transformation method of remote sensing of long-distance pipelines based on quicksand dynamics can effectively solve the problem of gaps between fields and improve the accuracy and efficiency of image processing.

[0069] It's important to note that in remote sensing image processing, the source domain refers to the image data used for model training, while the target domain refers to the imagery not used for training but to be inferred. Due to different regional environmental and topographical characteristics, the texture representation of the same feature in remote sensing imagery can vary significantly, creating a spatial domain gap. To address this issue, the quicksand dynamics model proposes a mapping method that controls the flipping frequency, angle, and number of bubbles in the quicksand, ensuring that the mapped target domain data closely resembles the imagery of the source domain data.

[0070] During the mapping process (also known as the transformation process), geospatial coordinates are incorporated as crucial information to ensure the full utilization of the spatial features of the image. The mapping function adjusts the image's probability distribution to more closely align with the source data. This process not only effectively handles unlearned source image texture features but also provides a new data source and research perspective for remote sensing interpretation models.

[0071] Furthermore, within the same geographic area, the closer the imagery is acquired, the more similar the texture features of the same features tend to be. At the same time, seasonal variations also lead to significant differences in the imagery of the same features. Therefore, by incorporating temporal information, the model can more effectively exploit seasonal similarities during the inference phase, thereby addressing temporal domain gaps. This combination of spatial and temporal information gives the model greater adaptability and reasoning capabilities.

[0072] In the embodiments of the present application, quicksand dynamics mapping is used as a preprocessing step to transform the target domain image so that its pixel-space probability distribution is highly similar to that of the source domain image. The model structure of the present invention consists of two major components: the aforementioned model training and inference components, which serve as the main model (i.e., the large-scale long-distance pipeline remote sensing model); and a module based on quicksand dynamics transformation mapping (i.e., the quicksand dynamics model) introduced before the main model. The quicksand dynamics model aims to achieve maximum similarity between the pixel-space probability distribution of the target domain image and that of the source domain image for the same feature by controlling the data transformation. The target domain transformation is adaptively controlled by two primary factors: space (geospatial coordinates) and time (image acquisition time).

[0073] Optionally, the pre-processed target domain image, the geospatial coordinates of the target domain image, and the acquisition time of the target domain image are input into the quicksand dynamics model to perform quicksand dynamics transformation, including:

[0074] Obtain pre-processed target domain images and classify them according to the geospatial coordinates of the target domain images and the acquisition time of the target domain images;

[0075] inputting the classified target domain image into the quicksand dynamics model;

[0076] Using the principles of quicksand dynamics, control parameters of the quicksand dynamics model are designed; wherein the control parameters include the turnover frequency, angle and number of bubbles of the quicksand;

[0077] The quicksand dynamics model is adaptive according to the geographic space coordinates of the target domain image and the acquisition time of the target domain image, and transforms the classified target domain image.

[0078] In the above implementation, the target domain image is treated as quicksand. By adjusting the movement of the quicksand, the image features are altered, resulting in a high degree of similarity to the source domain image. Furthermore, the model enables adaptive control, identifying spatial similarities between images and prioritizing images with similar seasons, thereby improving mapping accuracy.

[0079] Optionally, the method for transforming a large remote sensing model image of a long-distance pipeline based on quicksand dynamics further includes:

[0080] Utilize reinforcement learning methods to optimize the mapping process of quicksand dynamics model.

[0081] In the above embodiment, the reinforcement learning method is used to optimize the mapping process of the quicksand dynamics model, aiming to reduce the difference in spatial probability distribution of pixel texture features of the same ground feature image in the source domain and the target domain.

[0082] Optionally, a reinforcement learning method is used to optimize the mapping process of the quicksand dynamics model, including:

[0083] Setting a reinforcement learning framework; wherein the reinforcement learning framework includes a state space, an action space, and a reward mechanism, and calculating the reward mechanism based on the state space and the action space;

[0084] The KL divergence is used to quantify the difference in the probability distribution of pixel space between the source domain and the target domain. A threshold is set and the KL divergence is compared with the threshold to determine the similarity of the features of the source domain and the target domain. When the KL divergence is less than the threshold, the features of the source domain and the target domain are considered to be sufficiently similar.

[0085] During the reinforcement learning process, the quicksand dynamics model continuously adjusts the control parameters to calculate a new KL divergence in each round of learning, and adjusts the control parameters according to the reward mechanism until the set threshold conditions are met;

[0086] Selective state space patterns are introduced to optimize the performance of the quicksand dynamics model by focusing on specific state features. The selection criteria of the selective state space patterns are based on the similarity of spatial and temporal features, and a similarity metric is used to evaluate and select states.

[0087] Define the state space is the probability distribution feature of the source domain and target domain image features, then:

[0088] ;

[0089] Select the state pair with the most similar features according to the similarity ,but:

[0090] .

[0091] In a specific embodiment, the convergence condition of the quicksand dynamics model is:

[0092] .

[0093] That is, the optimization process of the quicksand dynamics model continues until the convergence condition is met, indicating that the feature differences between the source and target domains have been effectively reduced.

[0094] Optionally, the KL divergence is calculated as follows:

[0095] ;

[0096] In the above formula, is the KL divergence difference between the source domain image P and the target domain image Q, is the source domain image P at position The probability of The target domain image Q is located at probability.

[0097] In the above embodiment, by minimizing , thus reducing the differences between image fields.

[0098] In a specific implementation, the calculation formula of the state space is:

[0099] .in, The current selection state.

[0100] The calculation formula of action space is:

[0101] .

[0102] The calculation formula of the reward mechanism is:

[0103] .

[0104] For example, assuming that the movement of sand follows the influence of gravity and liquid viscosity, the following formula is used to simulate gravity and flow:

[0105] ;

[0106] In the above formula, It represents the rate of change of the image pixel space probability distribution function P(x,y) (or higher-dimensional P(x,y,z,…)) over time t. Represents the divergence operator, which is used to describe the "divergence" of the vector field, that is, the local diffusion or aggregation of the probability distribution caused by sand movement (or image feature changes). Indicates the movement speed of quicksand, which is affected by gravity and liquid viscosity and can be further decomposed into the base speed Vo and the bubble effect correction term .

[0107] Alternatively, in the quicksand dynamics model, the calculation formula for the overturning frequency is as follows:

[0108] ;

[0109] In the above formula, For the mapping state, is the flip frequency, t is the time, and d represents the differential.

[0110] In the above embodiment, the flip frequency Influencing the temporal characteristics of the mapping process: Increasing the frequency can speed up image state updates and enhance mapping flexibility. In the transformation function, the flip frequency can be incorporated into the mapping speed adjustment. The speed at which the mapping state changes over time is affected, which in turn affects image updates.

[0111] It should be noted that the expression in the figure In the equation, d is the differential symbol, which means "describing the infinitesimal change of a variable". Specifically:

[0112] Overall representative function The derivative with respect to time t (i.e., the rate of change of S over time). dS is the small change in the mapping state S, and dt is the small change in time t. The ratio of the two is Describes the instantaneous speed of S changing with time.

[0113] Alternatively, in the quicksand dynamics model, the calculation formula for the flip angle is as follows:

[0114] ;

[0115] In the above formula, is the vector of initial image features; After flipping angle Adjusted image feature vector; is the rotation transformation function.

[0116] In the above embodiment, the flip angle Affects the flow direction. The change of flip angle changes the flow characteristics, thereby affecting the relative position between pixels. This is in the transformation function The preservation and adjustment of image features are ensured during the implementation process.

[0117] Optionally, in the quicksand dynamics model, the bubble effect The calculation formula is as follows:

[0118] ;

[0119] In the above formula, is the adjusted velocity field; is the velocity field before adjustment.

[0120] In the above embodiment, the buffering effect of bubbles in the liquid can be modeled as an interference term , the velocity field is corrected so that the velocity field presents more realistic flow characteristics. The effect of bubbles causes the velocity field to change, affecting the dynamic stability of the quicksand and thus affecting the accuracy of the mapping.

[0121] Optionally, the probability distribution of the target domain image is calculated using the following formula:

[0122] ;

[0123] In the above formula, For different feature areas, is the probability value of the corresponding area; is the probability distribution function in two-dimensional space;

[0124] The calculation formula for the quicksand dynamics model transformation is as follows:

[0125] ;

[0126] In the above formula, is the transformation function, is the probability distribution of the current target domain image; is the probability distribution of the transformed target domain image.

[0127] It should be noted that the explicit expression of the transformation function T is as follows:

[0128] Combined with the dynamic parameters of quicksand (flip frequency f, angle , bubble effect ) and the probability distribution evolution process, define:

[0129] ;

[0130] In the above formula: is an exponential time evolution term controlled by the flip frequency f, which represents the cumulative change of the state over time (refer to the formula in the image Solution ).

[0131] is the rotation matrix, corresponding to the flip angle θ, which adjusts the movement direction of the quicksand (image feature):

[0132] ;

[0133] Base velocity field Superimposed bubble disturbance , corrected flow velocity.

[0134] is the probability distribution of the source domain image at position (x, y).

[0135] Therefore, The time integration result of is embedded into T to realize the dynamic evolution of probability distribution.

[0136] In the above embodiment, transforming the target domain image through the quicksand dynamics model helps to make it highly similar to the source domain image in terms of pixel space probability distribution.

[0137] For example, the two-dimensional distribution is extended to three dimensions and higher dimensions, and a new coordinate system is defined. , the corresponding extended probability distribution function , then the corresponding extended probability distribution function as follows:

[0138] ;

[0139] The final probability distribution is as follows:

[0140] ;

[0141] In the above formula, represents a specific variant of the quicksand dynamics mapping function, which is different from the previous transformation function There are some differences. Specifically:

[0142] Multiple Different values ​​of may correspond to different control parameter settings or to the quicksand dynamics mapping function applied in different contexts. This allows the model to adapt the mapping strategy based on specific image features, environmental factors, or task requirements.

[0143] Functional Diversity: Each application may employ different parameter combinations (e.g., flipping frequency, angle, etc.), allowing it to process different types of image data or address specific domain gaps. This enhances the flexibility and adaptability of the mapping, enabling the model to perform better when faced with diverse data.

[0144] Optionally, preprocess the source domain image, including:

[0145] First, the source domain image is orthorectified;

[0146] Secondly, geometric correction is performed on the orthorectified source domain image.

[0147] Secondly, the cloud and shadow are removed from the geometrically corrected source domain image;

[0148] In the above implementation, preprocessing of source domain images helps to accurately train a large remote sensing model of a long-distance pipeline.

[0149] Preprocess the target domain image, including:

[0150] First, the target domain image is processed;

[0151] Secondly, the brightness of the geometrically aligned target domain image is adjusted.

[0152] In the above implementation, preprocessing the target source domain image helps to accurately obtain the transformed image.

[0153] In the embodiments of the present application, the remote sensing image transformation method based on quicksand dynamics of the present invention exhibits significant advantages and positive effects in many aspects compared with the prior art.

[0154] First, by effectively narrowing the domain gap between the source and target domains, the present invention achieves a high degree of similarity between the pixel spatial probability distribution of target domain images and source domain images, thereby improving the accuracy of image interpretation. Experiments have shown that using this method in long-distance pipeline monitoring can increase image interpretation accuracy to over 90%, an improvement of approximately 15% compared to traditional methods.

[0155] Secondly, the enhanced ability to retain image features significantly improves the stability of image processing under different climate and terrain conditions. Specific data shows that by introducing an adaptive control mechanism for geographic coordinates and time information, the retention rate of image features has increased by 20%, significantly reducing the risk of feature loss or deformation.

[0156] Furthermore, in terms of computational efficiency, the quicksand dynamics mapping process of this invention reduces the computational time for feature extraction and classification by approximately 30%. This means that in practical applications, large amounts of remote sensing data can be processed more quickly, meeting the needs of real-time monitoring and providing technical support for the safe operation of long-distance pipelines.

[0157] Finally, the flexible mapping framework provided by this invention is widely applicable to remote sensing image processing in different regions and climate conditions, showing strong adaptability. Case studies have shown that the mapping method of this invention demonstrates consistent efficiency in applications across different regions, further validating its value in diverse application scenarios.

[0158] In summary, the present invention not only significantly improves the accuracy of image interpretation and feature retention capabilities, but also demonstrates strong advantages in computational efficiency and adaptability, promoting the development of remote sensing image processing technology.

[0159] In a specific embodiment, see Figures 2 to 4 ,First, data preparation and preprocessing are as follows:

[0160] See also Figure 2 For the source image data (data collected by sensor A in Northwest China), the sensor type is sensor A (assuming it is a medium-resolution multispectral remote sensing device). The resolution is 10 meters per pixel. The geographical scope is the coverage area. , geographic coordinates are accurate to three decimal places to ensure geographic space accuracy. The acquisition time is the spring of 2023 (March to May). The spring images provide obvious vegetation features, which are helpful in capturing seasonal texture features. The spectral characteristics of the image data include red, green, blue and near-infrared bands (RGB+NIR). These bands are used to enhance the reflectance characteristics of different ground objects (such as vegetation, soil, water bodies, etc.) to support the model's recognition of ground object textures. Preprocessing steps: Orthorectification to orthorectify the image to correct tilt and geometric distortion; geometric correction to ensure spatial accuracy and reduce image deviation. Cloud and shadow removal to filter out pixels affected by clouds or shadows to ensure that the ground objects are clearly visible.

[0161] See also Figure 3 For the target domain image data (data of central China collected by sensor B), the sensor type is sensor B (assuming it is a high-resolution multispectral remote sensing device). The resolution is 1 meter / pixel. The geographical scope is the coverage area. , precisely aligned with the geographic coordinates of the source domain. The acquisition time is the autumn of 2023 (September to November). The difference between the autumn data characteristics and the spring data is used to verify the model's adaptability to the gap between time domains. The spectral characteristics also include red, green, blue and near-infrared bands (RGB+NIR), which are consistent with the source domain. The seasonal feature is that the B sensor image is collected in autumn, resulting in a decrease in vegetation cover, an increase in soil exposure, and more obvious water body features. This feature difference is used for model adaptive control. Preprocessing steps: Geometric alignment is used for precise alignment according to geographic coordinates to be consistent with the spatial distribution of the source domain image. Brightness adjustment is used to remove uneven lighting and local shadows to enhance the texture consistency of the image.

[0162] Secondly, the control parameters of the quicksand dynamics model are set.

[0163] For the flip frequency :

[0164] Range: 1 to 10 times / second, 5 times / second was optimized to balance the speed of image feature preservation and change.

[0165] Details: A frequency of 5 times / second works best in the test, and the converted B sensor image quickly adapts to the A sensor image in terms of texture.

[0166] For flip angle :

[0167] Range: 0° to 90°, and 45° was finally determined to be the optimal angle.

[0168] Details: Experiments have shown that a 45° flip angle can effectively match the differences in terrain features between northwest and central China, achieving consistency in the spatial distribution of source and target domain images.

[0169] For the bubble effect :

[0170] Range: 0 to 1, step size 0.1, 0.3 is selected as the bubble intensity.

[0171] Details: A bubble effect strength of 0.3 provides moderate random perturbations, simulating the natural flow of quicksand. This helps to smooth the transition of B-sensor image details and adapt to changes in different resolution features.

[0172] Again, the transformation process (i.e., mapping process) of the quicksand dynamics model is as follows:

[0173] The first step is to divide the B sensor image into blocks according to geographic coordinates, ensuring that each block is aligned with the corresponding part of the A sensor image. For the geographic coordinate range: the UTM coordinate system is used with a resolution of 1 meter to ensure alignment accuracy.

[0174] The second step, spatial adaptation, introduces geographic coordinate data from northwestern and central China and achieves image matching by identifying the spatial distribution characteristics of features. Spatial feature extraction extracts feature texture, edge information, and neighboring pixel associations from the target image to support feature transfer from the source image. Temporal adaptation builds a seasonal similarity model through seasonal feature classification to account for seasonal differences between sensor A and sensor B images. Texture recognition utilizes changes in autumn imagery to implement temporal compensation within the model, minimizing the difference between images.

[0175] Next, we strengthen the learning and training of the quicksand dynamics model. For the state space, we define the state space ,in and are the probability distributions of the images from sensor A and sensor B, respectively. For the sampling strategy, a 256×256 pixel block is used as the sampling unit to obtain representative distribution samples from the image data of sensors A and B. For the reward mechanism, the KL divergence metric is used to calculate the difference between the source domain (sensor A image) and the target domain (sensor B image). The reward value is defined as:

[0176] ;

[0177] Reward calculation: If , a high reward value is given, indicating that the conversion is successful; otherwise, the control parameters are adjusted and optimized continuously.

[0178] Furthermore, for action space and parameter adjustment, the action space is The parameter optimization strategy is to adopt a step-by-step optimization strategy, adjusting the parameter change amplitude by 0.1 each time to achieve the best convergence of the parameters.

[0179] Next, the quicksand dynamics model is refined:

[0180] Refinement of the flip operation formula:

[0181] Initial distribution: A sensor image Divided into different areas , each region is assigned a feature probability .

[0182] Post-mapping distribution: Controlling parameters using quicksand dynamics ,Will Convert to :

[0183] .

[0184] For the velocity field correction formula:

[0185] Introducing interference =0.3, correct the velocity field and enhance the stability of the B sensor image:

[0186] .

[0187] This interference term improves the conversion stability of the image, making it less likely for the conversion process to produce obvious texture noise.

[0188] Adjustments for high-dimensional expansion:

[0189] Adding a height dimension to the model , establish three-dimensional coordinates To adapt to different terrain heights.

[0190] The expanded formula is as follows:

[0191] .

[0192] After considering the height information, 3D images can be further generated, which helps to improve the fineness of the conversion.

[0193] Then, reinforcement learning is used with selective state-space pattern optimization.

[0194] For selective state-space refinement:

[0195] Use feature similarity measurement to select the feature area with the highest similarity from the A and B sensor images for optimization. The specific formula is as follows:

[0196] .

[0197] The optimal state pairs are selected by maximizing similarity to further narrow the gap between fields.

[0198] Then, set the convergence conditions and stopping criteria:

[0199] Setting the KL divergence threshold = 0.3 as the convergence condition to ensure that the image features of sensors A and B are similar.

[0200] The iteration is stopped when the KL divergence is lower than 0.05 and the parameter change is less than 1e-4.

[0201] Finally, the results and verification are refined:

[0202] After the model training is completed, the new A sensor image (source domain) will be converted into the B sensor (target domain) image based on the quicksand dynamics mapping model for inference and compared with the original B sensor image to verify the effectiveness of the model in reducing the gap between domains.

[0203] In addition, verify the conversion effect:

[0204] About test data:

[0205] The source domain image data is the A sensor image (northwest China) collected in April 2023, with an image resolution of 10 meters per pixel. The target domain image data is the B sensor image (central China) collected in October 2023, with a resolution of 1 meter per pixel.

[0206] About the actual data characteristics:

[0207] For vegetation cover:

[0208] The A sensor image shows that the vegetation coverage rate in spring is about 60%, and the reflection intensity in the red and near-infrared bands is obvious, reflecting the characteristics of the high growth period of vegetation.

[0209] The B sensor image shows autumn vegetation coverage of about 40%, with weakened reflective features, mainly showing bare soil and low vegetation after harvest.

[0210] For water reflections:

[0211] The image of sensor A shows a relatively small water area, but the near-infrared band reflectance is low, with a typical value below 0.1.

[0212] The image of sensor B shows an increase in the water area, but the near-infrared reflection is stronger, ranging from about 0.2 to 0.3, which is affected by seasonal and water level changes.

[0213] Regarding the mapping results: By converting the B sensor image through the quicksand dynamics model, the following actual data results were obtained:

[0214] Texture similarity: The pixel value distribution in the near-infrared band of the converted B sensor image is similar to that of the corresponding area of ​​the A sensor image. The difference in near-infrared reflectance values ​​is reduced from 0.1 in the original image to 0.03, indicating that the consistency of vegetation reflectance characteristics has improved.

[0215] KL divergence calculation: The KL divergence of the pixel distribution of the converted B image and A image was calculated, and the result was 0.045, which is lower than the set threshold of 0.05, indicating that the difference between the fields has been significantly reduced.

[0216] Vegetation coverage simulation: The difference in vegetation coverage between the converted B image and the A image was reduced from 20% before conversion to 5%, indicating that the model successfully simulated the vegetation coverage in the A sensor image through quicksand dynamics mapping.

[0217] Consistency of water characteristics: In the water area of ​​the converted B image, the near-infrared reflectance dropped from 0.2 before conversion to 0.12, close to the 0.1 water reflectance value of the A sensor image, reflecting the significant adjustment effect of the conversion model on the water reflectance characteristics.

[0218] For verification results, see Figure 4 Visual comparison is a visual analysis of the converted B-sensor image. It was found that the texture and tone of ground objects (such as vegetation and water bodies) are highly consistent with those of the A-sensor image. In particular, the bare soil features in autumn are effectively converted into the texture of spring vegetation cover, and the visual difference is significantly reduced. Quantitative evaluation: The mean absolute error (MAE) between the source domain A image and the converted B image is reduced to 0.02 (0.09 before conversion), indicating that the two are highly similar in pixel values ​​after conversion. Mean square error (MSE): The MSE is reduced from 0.08 before conversion to 0.03, proving that the conversion model accurately simulates the characteristics of the target domain image. Reasoning consistency: When using the converted B-sensor image as training data for the target object detection task, the detection accuracy of the model is improved from 78% before conversion to 92%, verifying that the converted target domain image can successfully narrow the gap between domains and meet the needs of practical applications.

[0219] In summary, through the image transformation method based on quicksand dynamics, the target image of sensor B shows a high degree of consistency with the source image of sensor A after transformation, effectively narrowing the gap between the two fields and providing higher generalization capability and adaptability for remote sensing interpretation of large remote sensing models of long-distance pipelines.

[0220] It will be understood that the above embodiments are merely exemplary embodiments for illustrating the principles of the present invention, and the present invention is not limited thereto. Those skilled in the art will appreciate that various modifications and improvements can be made without departing from the spirit and substance of the present invention, and such modifications and improvements are also considered to be within the scope of protection of the present invention.

Claims

1. A method for transforming large-scale remote sensing models of long-distance pipelines based on quicksand dynamics, characterized by: The steps include: Step S100, pre-processing the source domain image and pre-processing the target domain image; Step S200: Establishing a quicksand dynamics model; inputting a preprocessed target domain image, the geospatial coordinates of the target domain image, and the acquisition time of the target domain image into the quicksand dynamics model to perform quicksand dynamics transformation; wherein, the transformation result of the target domain image is controlled by controlling the flipping frequency, angle, and number of bubbles of the quicksand in the quicksand dynamics model; Step S300: Establishing a large remote sensing model of a long-distance pipeline; and training the large remote sensing model of the long-distance pipeline using the preprocessed source domain image, the geospatial coordinates of the source domain image, and the acquisition time of the source domain image. Step S400: inputting the target domain image after the quicksand dynamics transformation into the long-distance pipeline remote sensing large model to obtain remote sensing data of the long-distance pipeline.

2. The method for converting a large remote sensing model of a long-distance pipeline based on quicksand dynamics according to claim 1 is characterized in that: Inputting the pre-processed target domain image, the geospatial coordinates of the target domain image, and the acquisition time of the target domain image into the quicksand dynamics model to perform quicksand dynamics transformation, including: Obtain pre-processed target domain images and classify them according to the geospatial coordinates of the target domain images and the acquisition time of the target domain images; inputting the classified target domain image into the quicksand dynamics model; Using the principles of quicksand dynamics, control parameters of the quicksand dynamics model are designed; wherein the control parameters include the turnover frequency, angle and number of bubbles of the quicksand; The quicksand dynamics model is adaptive according to the geographic space coordinates of the target domain image and the acquisition time of the target domain image, and transforms the classified target domain image.

3. The method for converting a large remote sensing model of a long-distance pipeline based on quicksand dynamics according to claim 2 is characterized in that: Also includes: Utilize reinforcement learning methods to optimize the mapping process of quicksand dynamics model.

4. The method for converting a large remote sensing model of a long-distance pipeline based on quicksand dynamics according to claim 3 is characterized in that: Utilize reinforcement learning methods to optimize the mapping process of the quicksand dynamics model, including: Setting a reinforcement learning framework; wherein the reinforcement learning framework includes a state space, an action space, and a reward mechanism, and calculating the reward mechanism based on the state space and the action space; The KL divergence is used to quantify the difference in the probability distribution of pixel space between the source domain and the target domain, a threshold is set, and the KL divergence is compared with the threshold to determine the similarity of the features of the source domain and the target domain; During the reinforcement learning process, the quicksand dynamics model continuously adjusts the control parameters to calculate a new KL divergence in each round of learning, and adjusts the control parameters according to the reward mechanism until the set threshold conditions are met; Selective state space patterns are introduced to optimize the performance of the quicksand dynamics model by focusing on specific state features. The selection criteria of the selective state space patterns are based on the similarity of spatial and temporal features, and a similarity metric is used to evaluate and select states.

5. The method for converting a large remote sensing model of a long-distance pipeline based on quicksand dynamics according to claim 4 is characterized in that: The calculation formula of KL divergence is as follows: ； In the above formula, is the KL divergence difference between the source domain image P and the target domain image Q, is the source domain image P at position The probability of The target domain image Q is located at probability.

6. The method for converting a large remote sensing model of a long-distance pipeline based on quicksand dynamics according to claim 5 is characterized in that: In the quicksand dynamics model, the calculation formula for the overturning frequency is as follows: ； In the above formula, For the mapping state, is the flip frequency, t is the time, and d represents the differential.

7. The method for converting a large remote sensing model of a long-distance pipeline based on quicksand dynamics according to claim 6 is characterized in that: In the quicksand dynamics model, the calculation formula for the flip angle is as follows: ； In the above formula, is the vector of initial image features; After flipping angle Adjusted image feature vector; is the rotation transformation function.

8. The method for converting a large remote sensing model of a long-distance pipeline based on quicksand dynamics according to claim 7 is characterized in that: In the quicksand dynamics model, the bubble effect The calculation formula is as follows: ； In the above formula, is the adjusted velocity field; is the velocity field before adjustment.

9. The method for converting a large remote sensing model of a long-distance pipeline based on quicksand dynamics according to claim 8 is characterized in that: The probability distribution of the target domain image is calculated using the following formula: ； In the above formula, For different feature areas, is the probability value of the corresponding area; is the probability distribution function in two-dimensional space; The calculation formula for the quicksand dynamics model transformation is as follows: ； In the above formula, is the transformation function, is the probability distribution of the current target domain image; is the probability distribution of the transformed target domain image.

10. The method for converting a large remote sensing model image of a long-distance pipeline based on quicksand dynamics according to claim 9 is characterized in that: Preprocess the source domain image, including: First, the source domain image is orthorectified; Secondly, geometric correction is performed on the orthorectified source domain image; Secondly, the cloud and shadow are removed from the geometrically corrected source domain image; Preprocess the target domain image, including: First, the target domain images are geometrically aligned; Secondly, the brightness of the geometrically aligned target domain image is adjusted.

Images