A method and system for removing fog from multispectral remote sensing satellite imagery

Abstract

This invention relates to the field of image dehazing technology, and discloses a method and system for dehazing multispectral remote sensing satellite imagery. Specifically, it proposes a dehazing scheme for multispectral remote sensing satellite imagery that combines atmospheric scattering models with deep learning and is adaptively optimized for specific application scenarios. Transmittance and atmospheric light values ​​are estimated using an atmospheric scattering model to provide physical constraints for the deep learning network. In the scene recognition stage, the scene type of the image is accurately identified through multi-scale feature extraction, adaptive feature aggregation, and a self-supervised learning mechanism. In the dehazing stage, the physical constraint weights and multi-band fusion weights are dynamically adjusted according to the identified scene to ensure that the dehazing result conforms to atmospheric physical laws and scene characteristics. This effectively improves the detail preservation, color restoration, and physical consistency of the dehazed imagery, enhances the adaptability of the deep learning network to complex scenes and diverse remote sensing images, effectively reduces scene recognition errors, and further improves the quality and consistency of the dehazing result.

Description

Technical Field

[0001] This invention relates to the field of image dehazing technology, and in particular to a method and system for dehazing multispectral remote sensing satellite images. Background Technology

[0002] With the rapid development of remote sensing technology, multispectral remote sensing satellite imagery has been widely used in agricultural monitoring, urban planning, environmental monitoring, and marine observation. However, due to interference factors such as haze, clouds, and aerosols in the atmosphere, remote sensing images are often severely degraded during acquisition, resulting in reduced image contrast, blurred details, and color distortion. This atmospheric interference not only reduces the quality of remote sensing data but also affects the subsequent image analysis and application effects.

[0003] In recent years, deep learning-based dehazing methods have been increasingly applied to remote sensing image dehazing tasks. However, existing deep learning dehazing methods often ignore the unique physical characteristics of remote sensing images, resulting in deficiencies in physical consistency and scene adaptability of the dehazing results, and the final dehazing effect of the output satellite images is still not ideal. Summary of the Invention

[0004] In view of this, the purpose of this invention is to provide a method and system for dehazing multispectral remote sensing satellite images, so as to solve the problem that the current dehazing effect of multispectral remote sensing satellite images is not ideal.

[0005] The first aspect of this invention discloses a method for dehazing multispectral remote sensing satellite imagery, the method comprising the following steps:

[0006] S1. Acquire multispectral remote sensing satellite image data to be processed and perform data preprocessing operations to obtain the first satellite image data; the data preprocessing operations include radiometric correction, geometric correction, multi-band registration, and atmospheric correction;

[0007] S2. Estimate the transmittance and atmospheric light value of the first satellite image data using an atmospheric scattering model;

[0008] S3. Input the first satellite image data, as well as the transmittance and atmospheric light value, into a deep learning network with scene adaptive physical constraints, and perform automatic scene type recognition based on the deep learning network to determine the scene type of the first satellite image data as the target scene.

[0009] S4. Adaptively adjust the physical constraints and multi-band weight allocation within the deep learning network according to the target scene to generate a dehazed image; wherein, the scene types include agricultural, urban, forest, wetland, mountain, desert and marine scenes.

[0010] Furthermore, the automatic scene type identification operation performed based on the deep learning network in step S3 to determine the scene type of the first satellite image data includes the following sub-steps:

[0011] S301. Perform multi-scale feature extraction on the first satellite image data through multi-layer convolutional units of a deep learning network, and extract feature information of visible light band and infrared band respectively as multispectral features;

[0012] S302. The extracted multispectral features are fused with the transmittance and atmospheric light value estimated in step S2 to generate atmospheric physical constraint features.

[0013] S303. The encoder of the deep learning network performs a preliminary encoding operation on the atmospheric physical constraint features to obtain the high-level semantic features of the first satellite image.

[0014] S304. Perform feature aggregation operations on the multispectral features, atmospheric physical constraint features, and high-level semantic features through an adaptive feature aggregation mechanism to generate aggregated adaptive features; wherein, the adaptive feature aggregation mechanism includes an attention mechanism;

[0015] S305. The aggregated adaptive features are input into the scene classification unit of the deep learning network, and the scene classification unit identifies the scene type of the first satellite image data.

[0016] Furthermore, step S3, which involves automatically identifying the scene type based on the deep learning network to determine the scene type of the first satellite image data, also includes the following sub-steps:

[0017] S306. Perform a confidence assessment on the scene classification results output by the scene classification unit, determine the reliability of the scene classification results based on the confidence assessment results, and correct the scene classification results through a self-supervised learning mechanism when the confidence is lower than a preset confidence threshold; the scene classification results include scene category labels and corresponding confidence scores.

[0018] Furthermore, step S304 specifically includes the following sub-steps:

[0019] S3041. Calculate the correlation between the multispectral features, atmospheric physical constraint features, and high-level semantic features based on the attention mechanism, and generate initial feature weights;

[0020] S3042. Based on the target scene and dehazing task requirements, the initial feature weights are adaptively adjusted to generate the final feature weight allocation scheme.

[0021] S3043. Based on the final feature weight allocation scheme, multi-scale features, atmospheric physical constraint features and high-level semantic features are weighted and fused to generate aggregated adaptive features.

[0022] S3044. Perform feature enhancement on the aggregated adaptive features; the feature enhancement includes operations to strengthen key features and suppress redundant information.

[0023] Furthermore, the confidence evaluation operation of the scene classification result output by the scene classification unit in step S306 includes the following sub-steps:

[0024] S3061. Obtain the output classification probability distribution from the scene classification unit; the classification probability distribution represents the prediction probability of the scene classification unit for each scene category;

[0025] S3062. Based on the classification probability distribution, obtain the category label with the highest probability and its corresponding probability value, obtain the highest probability value, and use the highest probability value as the initial confidence score;

[0026] S3063. Calculate the consistency between the scene classification result output by the scene classification unit and the multi-scale features, atmospheric physical constraint features, and high-level semantic features to obtain the feature consistency calculation result, and generate a correction factor based on the feature consistency calculation result.

[0027] S3064. Based on the correction factor, the initial confidence score is corrected, and the final confidence score is output as the confidence evaluation result of the scene classification result.

[0028] Furthermore, step S306, when the confidence level is lower than a preset confidence threshold, corrects the scene classification result through a self-supervised learning mechanism, including the following sub-steps:

[0029] S3065. Based on the feature consistency calculation results, select the scene category with the highest feature consistency from the preset scene category set, and automatically generate pseudo labels. The pseudo labels are used to replace the original classification labels with low confidence.

[0030] S3066. Based on the pseudo-label, extract multi-scale features, atmospheric physical constraint features, and high-level semantic features from the first satellite image again, and match these features with the sample features in the pseudo-label feature library, calculate the matching similarity, and verify the rationality of the pseudo-label.

[0031] S3067. The loss function based on self-supervised learning evaluates the difference between the scene classification result of the current scene classification unit and the pseudo label, and adaptively updates the parameters of the scene classification unit by minimizing the loss function; wherein, the loss function of self-supervised learning includes feature consistency loss and confidence constraint loss.

[0032] S3068. Reclassify the first satellite image using the updated scene classification unit, output the corrected scene category label and the corresponding confidence score, and determine whether the corrected confidence score is higher than the preset confidence threshold. If the threshold is not reached, re-execute the feature consistency check and pseudo-label generation, and perform iterative correction until the confidence score reaches the threshold or the preset maximum number of iterations is reached.

[0033] Further, step S4 includes the following sub-steps:

[0034] S401. Based on the target scene type, dynamically adjust the physical constraint weights within the deep learning network to match the optical characteristics and atmospheric conditions of the target scene; the physical constraint weights include transmittance constraint weights and atmospheric light value constraint weights.

[0035] S402. Based on the target scene type, select the band combination with the highest relevance from the bands of the first satellite image, and dynamically adjust the fusion weights of different bands.

[0036] S403. Multispectral features are combined with adjusted physical constraint weights and band fusion weights to perform feature fusion, and the fused features are input into a deep learning network to output a dehazed image.

[0037] S404. Evaluate the quality of the generated dehazed image to determine whether it meets the visual and physical requirements of the target scene. If it does not meet the requirements, readjust the physical constraint weights and band fusion weights, and repeat the dehazing process.

[0038] Further, step S403 includes the following sub-steps:

[0039] S4031. Based on the target scene type, determine the band combination with the highest relevance in the first satellite image, and combine it with the adjusted band fusion weights to perform band-weighted fusion of multispectral features to generate band fusion features.

[0040] S4032. Input the band fusion features into a deep learning network, extract features at different scales through a multi-layer convolutional module, and aggregate the features at different scales based on an attention mechanism to generate multi-scale fusion features.

[0041] S4033. Introduce the adjusted physical constraint weights into the mapping process of multi-scale fusion features to generate physical constraint fusion features;

[0042] S4034. Output dehazed multispectral remote sensing satellite images through the output layer of the deep learning network.

[0043] Furthermore, the physical constraint weights also include radiative transfer equation constraint weights and scenario prior knowledge constraint weights.

[0044] The second aspect of this invention discloses a dehazing system for multispectral remote sensing satellite imagery. This system is based on the method disclosed in the first aspect and includes a data processing module, an estimation module, and a deep learning network construction module.

[0045] The data processing module is used to acquire multispectral remote sensing satellite image data to be processed and perform data preprocessing operations to obtain the first satellite image data; the data preprocessing operations include radiometric correction, geometric correction, multi-band registration and atmospheric correction;

[0046] The estimation module is used to estimate the transmittance and atmospheric light value of the first satellite image data using an atmospheric scattering model.

[0047] The deep learning network building module is used to build a deep learning network that adapts to scene-adaptive physical constraints;

[0048] The first satellite image data, along with the transmittance and atmospheric light values, are input into the deep learning network;

[0049] The deep learning network further includes a scene recognition unit, which is used to perform automatic scene type recognition operations to determine the scene type of the first satellite image data as the target scene;

[0050] The deep learning network adaptively adjusts the internal physical constraints and multi-band weight allocation according to the target scene to generate a dehazed image; wherein the scene types include agricultural, urban, forest, wetland, mountain, desert and marine scenes.

[0051] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0052] This invention proposes a dehazing scheme for multispectral remote sensing satellite imagery that combines atmospheric scattering models with deep learning and is adaptively optimized for specific application scenarios. The scheme utilizes atmospheric scattering models to estimate transmittance and atmospheric light values, providing physical constraints for the deep learning network. In the scene recognition stage, multi-scale feature extraction, adaptive feature aggregation, and self-supervised learning mechanisms accurately identify the scene type of the image. In the dehazing stage, the physical constraint weights and multi-band fusion weights are dynamically adjusted based on the identified scene to ensure that the dehazing results conform to atmospheric physical laws and scene characteristics. This effectively improves the detail preservation, color restoration, and physical consistency of the dehazed imagery, enhances the adaptability of the deep learning network to complex scenes and diverse remote sensing imagery, effectively reduces scene recognition errors, and further improves the quality and consistency of the dehazing results. Attached Figure Description

[0053] The accompanying drawings, which are included to provide a further understanding of embodiments of the invention and constitute a part of this application, do not limit the scope of the invention. In the drawings:

[0054] Figure 1 This is a flowchart illustrating a method for dehazing multispectral remote sensing satellite imagery disclosed in an embodiment of the present invention.

[0055] Figure 2 This is a schematic diagram of the structure of a defogging system for multispectral remote sensing satellite imagery, as disclosed in another embodiment of the present invention. Detailed Implementation

[0056] To enable those skilled in the art to better understand the present invention, the technical solutions of the present invention will be clearly and completely described below with reference to the accompanying drawings of the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments.

[0057] Example 1

[0058] The first aspect of this invention discloses a method for dehazing multispectral remote sensing satellite imagery. Please refer to [link to relevant documentation]. Figure 1 , Figure 1 This is a flowchart illustrating a method for dehazing multispectral remote sensing satellite imagery disclosed in an embodiment of the present invention. The method includes the following steps:

[0059] S1. Acquire multispectral remote sensing satellite image data to be processed and perform data preprocessing operations to obtain the first satellite image data; the data preprocessing operations include radiometric correction, geometric correction, multi-band registration and atmospheric correction.

[0060] It is understood that, in this embodiment of the invention, the purpose of preprocessing multispectral remote sensing satellite image data based on operations such as radiometric correction, geometric correction, multiband registration, and atmospheric correction is to provide high-quality and accurate input data for subsequent dehazing processing. Radiometric correction eliminates sensor radiometric response errors, ensuring that the image's radiometric values ​​truly reflect the characteristics of ground features; geometric correction corrects geometric distortions caused by sensor imaging angles and satellite attitude changes, ensuring the accuracy of the image's spatial location; multiband registration ensures precise spatial alignment of data across bands, guaranteeing the consistency of multispectral information; and atmospheric correction removes the effects of atmospheric scattering and absorption, restoring the true reflectivity of the Earth's surface. This embodiment of the invention does not limit the specific implementation of the above preprocessing operations.

[0061] S2. Estimate the transmittance and atmospheric light value of the first satellite image data using an atmospheric scattering model.

[0062] Transmittance refers to the proportion of energy retained by light as it passes through the atmosphere, while atmospheric light intensity represents the light intensity caused by atmospheric scattering. These two parameters are key physical quantities in atmospheric scattering models and can be estimated using the dark channel prior method, radiative transfer model, or other physical models. In this embodiment of the invention, the estimation method is not limited.

[0063] Because remote sensing images are affected by atmospheric scattering and absorption during the imaging process, relying solely on deep learning methods can easily overlook atmospheric physical characteristics, leading to artifacts or non-physical discrepancies in the dehazing results. Therefore, step S2 establishes a physical constraint basis by introducing transmittance and atmospheric light values ​​into the dehazing network to provide accurate atmospheric scattering constraints for the deep learning network, ensuring that the dehazing process conforms to actual physical phenomena.

[0064] S3. Input the first satellite image data, as well as the transmittance and atmospheric light value, into a deep learning network with scene adaptive physical constraints, and perform automatic scene type recognition based on the deep learning network to determine the scene type of the first satellite image data as the target scene.

[0065] S4. Adaptively adjust the physical constraints and multi-band weight allocation within the deep learning network according to the target scene to generate a dehazed image; wherein, the scene types include agricultural, urban, forest, wetland, mountain, desert and marine scenes.

[0066] Furthermore, step S3, which involves performing automatic scene type identification based on a deep learning network to determine the scene type of the first satellite image data, includes the following sub-steps:

[0067] S301. Perform multi-scale feature extraction on the first satellite image data through multi-layer convolutional units of a deep learning network, and extract feature information of visible light band and infrared band respectively as multispectral features;

[0068] S302. The extracted multispectral features are fused with the transmittance and atmospheric light value estimated in step S2 to generate atmospheric physical constraint features.

[0069] S303. The encoder of the deep learning network performs a preliminary encoding operation on the atmospheric physical constraint features to obtain the high-level semantic features of the first satellite image.

[0070] S304. Perform feature aggregation operations on the multispectral features, atmospheric physical constraint features, and high-level semantic features through an adaptive feature aggregation mechanism to generate aggregated adaptive features; wherein, the adaptive feature aggregation mechanism includes an attention mechanism;

[0071] S305. The aggregated adaptive features are input into the scene classification unit of the deep learning network, and the scene classification unit identifies the scene type of the first satellite image data.

[0072] In this embodiment of the invention, high-level semantic features of satellite imagery refer to features that reflect the overall structure and semantic information of a scene in the satellite imagery. These features are used to abstractly express the texture, shape, and spatial relationships in the imagery, providing crucial semantic information support for subsequent scene recognition and adaptive dehazing. This helps the network more accurately understand image content and perform corresponding dehazing operations in different scenarios. For example, in agricultural scenarios, high-level semantic features may include farmland distribution patterns, vegetation cover types, and cultivated area boundaries; in urban scenarios, they may include building outlines, road networks, and the layout of artificial facilities; and in forest scenarios, they may include canopy density, forest type, and forest edges.

[0073] Furthermore, step S3, which involves automatically identifying the scene type based on a deep learning network to determine the scene type of the first satellite image data, also includes the following sub-steps:

[0074] S306. Perform a confidence assessment on the scene classification results output by the scene classification unit, determine the reliability of the scene classification results based on the confidence assessment results, and correct the scene classification results through a self-supervised learning mechanism when the confidence is lower than a preset confidence threshold; the scene classification results include scene category labels and corresponding confidence scores.

[0075] Furthermore, step S304 specifically includes the following sub-steps:

[0076] S3041. Calculate the correlation between multispectral features, atmospheric physical constraint features, and high-level semantic features based on the attention mechanism, and generate initial feature weights;

[0077] S3042. Based on the target scene and dehazing task requirements, the initial feature weights are adaptively adjusted to generate the final feature weight allocation scheme.

[0078] S3043. Based on the final feature weight allocation scheme, multi-scale features, atmospheric physical constraint features and high-level semantic features are weighted and fused to generate aggregated adaptive features.

[0079] S3044. Perform feature enhancement on the aggregated adaptive features; the feature enhancement includes operations to strengthen key features and suppress redundant information.

[0080] Furthermore, the confidence evaluation operation of the scene classification results output by the scene classification unit in step S306 includes the following sub-steps:

[0081] S3061. Obtain the output classification probability distribution from the scene classification unit; the classification probability distribution represents the prediction probability of the scene classification unit for each scene category;

[0082] S3062. Based on the classification probability distribution, obtain the category label with the highest probability and its corresponding probability value, obtain the highest probability value, and use the highest probability value as the initial confidence score;

[0083] S3063. Calculate the consistency between the scene classification result output by the scene classification unit and the multi-scale features, atmospheric physical constraint features, and high-level semantic features to obtain the feature consistency calculation result, and generate a correction factor based on the feature consistency calculation result.

[0084] S3064. Based on the correction factor, the initial confidence score is corrected, and the final confidence score is output as the confidence evaluation result of the scene classification result.

[0085] Preferably, the classification probability distribution P = {p1, p2, ..., p3} obtained from the scene classification unit is set as follows: i ,…,p n}, where p i This represents the probability of belonging to scene category i, where n is the total number of preset scene categories:

[0086]

[0087] Where Z = {z1, z2} j ,…,z n} represents the unnormalized classification score of the output.

[0088] Select the highest probability value P from the probability distribution P. max The category label C with the highest probability is obtained. pred and P max As the initial confidence score.

[0089] The process of calculating feature consistency and generating correction factors is as follows:

[0090]

[0091] λ corr =exp(-α(1-S))

[0092] Where S represents feature consistency, F multi For multi-scale features, F phys Due to atmospheric physical constraints, F sem For high-level semantic features, F cls λ is the feature vector extracted from the scene classification result output by the scene classification unit. corr α is the correction factor, and α is the adjustment parameter.

[0093] Correct the initial confidence score:

[0094] P final =P max ×λ corr

[0095] Among them, P final This is the final confidence score.

[0096] In this embodiment of the invention, the core of feature consistency calculation lies in evaluating the degree of matching between the scene classification result output by the scene classification unit and the image features. By comparing the classification result with multi-scale features, atmospheric physical constraint features, and high-level semantic features extracted from the image, it is ensured that the classification result matches the actual image features. For example, assuming the scene classification unit outputs the scene category label "forest," the feature vector corresponding to this category label is compared with multi-scale features (such as texture, edge details, etc.), atmospheric physical constraint features (such as transmittance and atmospheric light value), and high-level semantic features (such as forest canopy density and vegetation cover pattern), and the similarity between these features is calculated. When the consistency is low, a correction factor is generated through a deep learning network to reduce the initial confidence score.

[0097] Furthermore, step S306, when the confidence level is lower than a preset confidence threshold, corrects the scene classification result through a self-supervised learning mechanism, including the following sub-steps:

[0098] S3065. Based on the feature consistency calculation results, select the scene category with the highest feature consistency from the preset scene category set, and automatically generate pseudo labels. The pseudo labels are used to replace the original classification labels with low confidence.

[0099] Specifically, in this step, the previously calculated feature consistency scores are used to rank all possible scene categories (such as agriculture, city, forest, wetland, etc.), and the scene category with the highest matching degree with the current image features is selected as the pseudo-label. For example, if the original classification result is "wetland," but the matching degree with wetland features is low, while the feature consistency with the "forest" scene is higher, then "forest" will be selected as the pseudo-label. The mechanism of scene category correction through consistency calculation can effectively correct the misjudgment of scene classification units under low confidence conditions and enhance the classification accuracy of the network in complex or mixed scenes.

[0100] S3066. Based on the pseudo-label, extract multi-scale features, atmospheric physical constraint features, and high-level semantic features from the first satellite image again, and match these features with the sample features in the pseudo-label feature library, calculate the matching similarity, and verify the rationality of the pseudo-label.

[0101] S3067. The loss function based on self-supervised learning evaluates the difference between the scene classification result of the current scene classification unit and the pseudo label, and adaptively updates the parameters of the scene classification unit by minimizing the loss function; wherein, the loss function of self-supervised learning includes feature consistency loss and confidence constraint loss.

[0102] S3068. Reclassify the first satellite image using the updated scene classification unit, output the corrected scene category label and the corresponding confidence score, and determine whether the corrected confidence score is higher than the preset confidence threshold. If the threshold is not reached, re-execute the feature consistency check and pseudo-label generation, and perform iterative correction until the confidence score reaches the threshold or the preset maximum number of iterations is reached.

[0103] In this embodiment of the invention, scene recognition is a crucial step in achieving adaptive dehazing. By automatically identifying the scene type of the first satellite image data using a deep learning network, the scene category to which the image belongs can be accurately determined, such as agriculture, urban areas, forests, wetlands, mountains, deserts, and oceans. For different scene categories, the atmospheric scattering and optical properties in the images differ significantly. Scene classification can provide targeted optimization strategies for the subsequent dehazing process. Specifically, the scene classification result determines the dynamic adjustment of physical constraint weights (such as transmittance and atmospheric light values) and the adaptive allocation of multi-band fusion weights, ensuring that the dehazing network can generate the optimal dehazing effect based on the characteristics of a specific scene.

[0104] Furthermore, confidence assessment and self-supervised learning correction mechanisms are employed to improve the reliability and accuracy of classification results. By aggregating multi-scale features, atmospheric physical constraint features, and high-level semantic features, the deep learning network can generate rich adaptive features and automatically generate pseudo-labels for self-supervised learning correction when confidence is low. This enhances the deep learning network's adaptability to complex scenes and diverse remote sensing images, effectively reduces scene recognition errors, and further improves the quality and consistency of dehazing results.

[0105] Further, step S4 includes the following sub-steps:

[0106] S401. Based on the target scene type, dynamically adjust the physical constraint weights within the deep learning network to match the optical characteristics and atmospheric conditions of the target scene; the physical constraint weights include transmittance constraint weights and atmospheric light value constraint weights.

[0107] S402. Based on the target scene type, select the band combination with the highest relevance from the bands of the first satellite image, and dynamically adjust the fusion weights of different bands.

[0108] S403. Multispectral features are combined with adjusted physical constraint weights and band fusion weights to perform feature fusion, and the fused features are input into a deep learning network to output a dehazed image.

[0109] S404. Evaluate the quality of the generated dehazed image to determine whether it meets the visual and physical requirements of the target scene. If it does not meet the requirements, readjust the physical constraint weights and band fusion weights, and repeat the dehazing process.

[0110] Further, step S403 includes the following sub-steps:

[0111] S4031. Based on the target scene type, determine the band combination with the highest relevance in the first satellite image, and combine it with the adjusted band fusion weights to perform band-weighted fusion of multispectral features to generate band fusion features.

[0112] S4032. Input the band fusion features into a deep learning network, extract features at different scales through a multi-layer convolutional module, and aggregate the features at different scales based on an attention mechanism to generate multi-scale fusion features.

[0113] S4033. Introduce the adjusted physical constraint weights into the mapping process of multi-scale fusion features to generate physical constraint fusion features;

[0114] S4034. Output dehazed multispectral remote sensing satellite images through the output layer of the deep learning network.

[0115] Dehazing is the process of restoring multispectral remote sensing satellite imagery from a degraded state to a clear state. In this embodiment of the invention, by dynamically adjusting the physical constraint weights and multi-band fusion weights within the deep learning network, the dehazing operation can adapt to the specific characteristics of different scenarios, such as agriculture, cities, forests, wetlands, mountains, deserts, and oceans.

[0116] By setting physical constraint weights, the dehazing results are ensured to conform to the actual atmospheric scattering patterns and scene characteristics, reducing artifacts and unnatural color deviations. Meanwhile, multi-band fusion weights dynamically allocate weights to different spectral bands according to scene requirements, maximizing the complementarity of information from different bands and improving image detail preservation and color reproduction.

[0117] In addition, by introducing a quality assessment mechanism, the visual and physical consistency of the generated dehazed images can be evaluated. If the expected results are not achieved, the physical constraint weights and band fusion weights are automatically adjusted to iteratively optimize the dehazing process, thereby improving the clarity, detail retention and physical consistency of the dehazing results and ensuring that high-quality dehazed images can still be generated in complex and ever-changing remote sensing scenarios.

[0118] Furthermore, the physical constraint weights also include the radiative transfer equation constraint weights and the scenario prior knowledge constraint weights.

[0119] Specifically, to further optimize the defogging process of the present invention, constraint weights of the radiative transfer equation and scene prior knowledge are introduced to enrich the physical constraints in the defogging process.

[0120] The radiative transfer equation constraint weights are used to describe the scattering, absorption, and transmission characteristics of light as it propagates through the atmosphere. By dynamically introducing these weights, the feature mapping process can be adjusted in the deep learning network to ensure that the dehazing results conform to the physical laws of radiative transfer. This further reduces optical errors caused by atmospheric scattering and improves the brightness and color reproduction of images. Especially under complex atmospheric conditions, such as haze, clouds, and high humidity environments, it can significantly improve the quality and reliability of dehazed images.

[0121] Furthermore, the scene prior knowledge constraint weights incorporate the optical and structural features of specific scenes (such as agriculture, cities, forests, and wetlands), with this prior knowledge derived from extensive remote sensing imagery experience data and domain knowledge. By introducing scene prior knowledge into the deep learning network, the feature learning focus within the network can be dynamically adjusted. For example, in agricultural scenes, constraints on vegetation reflectivity are enhanced; in urban scenes, constraints on building outlines and road structures are emphasized. This scene prior knowledge constraint can guide the network to generate dehazed images that better reflect the actual scene, effectively preserving key details and reducing artifacts and information loss.

[0122] Example 2

[0123] The second aspect of this invention discloses a dehazing system for multispectral remote sensing satellite imagery. Please refer to [link to relevant documentation]. Figure 2 , Figure 2 This is a schematic diagram of the structure of a dehazing system for multispectral remote sensing satellite imagery, disclosed in another embodiment of the present invention. The system includes a data processing module, an estimation module, and a deep learning network construction module; wherein,

[0124] The data processing module is used to acquire multispectral remote sensing satellite image data to be processed and perform data preprocessing operations to obtain the first satellite image data; the data preprocessing operations include radiometric correction, geometric correction, multi-band registration and atmospheric correction;

[0125] The estimation module is used to estimate the transmittance and atmospheric light value of the first satellite image data using an atmospheric scattering model;

[0126] The deep learning network building module is used to build deep learning networks that adapt to scene-adaptive physical constraints;

[0127] The first satellite image data, along with the transmittance and atmospheric light values, are input into the deep learning network;

[0128] The deep learning network also includes a scene recognition unit, which performs automatic scene type recognition to determine the scene type of the first satellite image data as the target scene;

[0129] The deep learning network adaptively adjusts the internal physical constraints and multi-band weight allocation according to the target scene to generate a dehazed image; wherein the scene types include agricultural, urban, forest, wetland, mountain, desert and marine scenes.

[0130] It should be noted that the specific implementation process of Example 2 is similar to that of Example 1, and will not be repeated in Example 2.

[0131] Finally, it should be noted that the method and system for dehazing multispectral remote sensing satellite images disclosed in the embodiments of the present invention are merely preferred embodiments of the present invention and are only used to illustrate the technical solutions of the present invention, not to limit it. Although the present invention has been described in detail with reference to the foregoing embodiments, those skilled in the art should understand that modifications can still be made to the technical solutions described in the foregoing embodiments, or equivalent substitutions can be made to some of the technical features. Such modifications or substitutions do not cause the essence of the corresponding technical solutions to deviate from the spirit and scope of the technical solutions of the embodiments of the present invention.

Claims

1. A method for defogging multispectral remote sensing satellite imagery, characterized in that, The method includes the following steps: S1. Acquire multispectral remote sensing satellite image data to be processed, and perform data preprocessing operations to obtain the first satellite image data; the data preprocessing operations include radiometric correction, geometric correction, multi-band registration, and atmospheric correction; S2. Estimate the transmittance and atmospheric light value of the first satellite image data using an atmospheric scattering model; S3. Input the first satellite image data, along with the transmittance and atmospheric light values, into a deep learning network with scene adaptive physical constraints, and perform automatic scene type identification based on the deep learning network to determine the scene type of the first satellite image data as the target scene; wherein, step S3, which involves performing automatic scene type identification based on the deep learning network to determine the scene type of the first satellite image data, includes the following sub-steps: S301. Perform multi-scale feature extraction on the first satellite image data through multi-layer convolutional units of a deep learning network, and extract feature information of visible light band and infrared band respectively as multispectral features; S4. Adaptively adjust physical constraints and multi-band weight allocation according to the target scene to generate the dehazed image; wherein, step S4 includes the following sub-steps: S401. Based on the target scene type, dynamically adjust the physical constraint weights to match the optical characteristics and atmospheric conditions of the target scene; the physical constraint weights include transmittance constraint weights and atmospheric light value constraint weights. S402. Based on the target scene type, select the band combination with the highest relevance from the bands of the first satellite image and adjust the fusion weights of different bands. S403. Multispectral features are combined with adjusted physical constraint weights and band fusion weights to perform feature fusion, and the fused features are input into the second deep learning network to output the dehazed image. The scene types include agricultural, urban, forest, wetland, mountain, desert and marine scenes.

2. The method for dehazing multispectral remote sensing satellite imagery according to claim 1, characterized in that, The step S3, which involves automatically identifying the scene type based on the deep learning network to determine the scene type of the first satellite image data, also includes the following sub-steps: S302. The extracted multispectral features are fused with the transmittance and atmospheric light value estimated in step S2 to generate atmospheric physical constraint features. S303. The encoder of the deep learning network performs a preliminary encoding operation on the atmospheric physical constraint features to obtain the high-level semantic features of the first satellite image. S304. Perform feature aggregation operations on the multispectral features, atmospheric physical constraint features, and high-level semantic features through an adaptive feature aggregation mechanism to generate aggregated adaptive features; wherein, the adaptive feature aggregation mechanism includes an attention mechanism; S305. The aggregated adaptive features are input into the scene classification unit of the deep learning network, and the scene classification unit identifies the scene type of the first satellite image data.

3. The method for dehazing multispectral remote sensing satellite imagery according to claim 2, characterized in that, Step S3 further includes the following sub-steps: S306. Perform a confidence assessment on the scene classification results output by the scene classification unit, determine the reliability of the scene classification results based on the confidence assessment results, and correct the scene classification results through a self-supervised learning mechanism when the confidence is lower than a preset confidence threshold; the scene classification results include scene category labels and corresponding confidence scores.

4. The method for dehazing multispectral remote sensing satellite imagery according to any one of claims 2-3, characterized in that, Step S304 specifically includes the following sub-steps: S3041. Calculate the correlation between the multispectral features, atmospheric physical constraint features, and high-level semantic features based on the attention mechanism, and generate initial feature weights; S3042. Based on the target scene and dehazing task requirements, the initial feature weights are adaptively adjusted to generate the final feature weight allocation scheme. S3043. Based on the final feature weight allocation scheme, multi-scale features, atmospheric physical constraint features and high-level semantic features are weighted and fused to generate aggregated adaptive features. S3044. Perform feature enhancement on the aggregated adaptive features; The feature enhancements include strengthening key features and suppressing redundant information operations.

5. The method for dehazing multispectral remote sensing satellite imagery according to claim 3, characterized in that, The confidence evaluation operation of the scene classification result output by the scene classification unit in step S306 includes the following sub-steps: S3061. Obtain the output classification probability distribution from the scene classification unit; the classification probability distribution represents the prediction probability of the scene classification unit for each scene category; S3062. Based on the classification probability distribution, obtain the category label with the highest probability and its corresponding probability value, obtain the highest probability value, and use the highest probability value as the initial confidence score; S3063. Calculate the consistency between the feature vector extracted based on the scene classification result and the multi-scale features, atmospheric physical constraint features, and high-level semantic features to obtain the feature consistency calculation result, and generate a correction factor based on the feature consistency calculation result. S3064. Based on the correction factor, the initial confidence score is corrected, and the final confidence score is output as the confidence evaluation result of the scene classification result.

6. The method for dehazing multispectral remote sensing satellite imagery according to claim 5, characterized in that, Step S306, which involves correcting the scene classification result through a self-supervised learning mechanism when the confidence level is lower than a preset confidence threshold, includes the following sub-steps: S3065. Based on the feature consistency calculation results, select the scene category with the highest feature consistency from the preset scene category set, and automatically generate pseudo labels. The pseudo labels are used to replace the original classification labels with low confidence. S3066. Based on the pseudo-label, extract multi-scale features, atmospheric physical constraint features, and high-level semantic features from the first satellite image again, and match these features with the sample features in the pseudo-label feature library, calculate the matching similarity, and verify the rationality of the pseudo-label. S3067. The loss function based on self-supervised learning evaluates the difference between the scene classification result of the current scene classification unit and the pseudo label, and adaptively updates the parameters of the scene classification unit by minimizing the loss function; wherein, the loss function of self-supervised learning includes feature consistency loss and confidence constraint loss. S3068. Reclassify the first satellite image using the updated scene classification unit, output the corrected scene category label and the corresponding confidence score, and determine whether the corrected confidence score is higher than the preset confidence threshold. If the threshold is not reached, re-execute the feature consistency check and pseudo-label generation, and perform iterative correction until the confidence score reaches the threshold or the preset maximum number of iterations is reached.

7. The method for dehazing multispectral remote sensing satellite imagery according to claim 2, characterized in that, Following step S403, the method further includes: S404. Evaluate the quality of the generated dehazed image to determine whether it meets the visual and physical requirements of the target scene. If it does not meet the requirements, readjust the physical constraint weights and band fusion weights, and repeat the dehazing process.

8. The method for dehazing multispectral remote sensing satellite imagery according to claim 7, characterized in that, Step S403 includes the following sub-steps: S4031. Based on the target scene type, determine the band combination with the highest relevance in the first satellite image, and combine it with the adjusted band fusion weights to perform band-weighted fusion of multispectral features to generate band fusion features. S4032. Input the band fusion features into the second deep learning network, extract features of different scales through multi-layer convolution modules, and aggregate the features of different scales based on the attention mechanism to generate multi-scale fusion features. S4033. Introduce the adjusted physical constraint weights into the mapping process of multi-scale fusion features to generate physical constraint fusion features; S4034. The dehazed multispectral remote sensing satellite imagery is output through the output layer of the second deep learning network.

9. The method for dehazing multispectral remote sensing satellite imagery according to any one of claims 7-8, characterized in that, The physical constraint weights also include the radiative transfer equation constraint weights and the scenario prior knowledge constraint weights.

10. A dehazing system for multispectral remote sensing satellite imagery, characterized in that, The system is implemented based on the method described in any one of claims 1-9, and the system includes a data processing module, an estimation module, and a deep learning network construction module; wherein... The data processing module is used to acquire multispectral remote sensing satellite image data to be processed and perform data preprocessing operations to obtain the first satellite image data; the data preprocessing operations include radiometric correction, geometric correction, multi-band registration and atmospheric correction; The estimation module is used to estimate the transmittance and atmospheric light value of the first satellite image data using an atmospheric scattering model. The deep learning network building module is used to build a deep learning network that adapts to scene-adaptive physical constraints; The first satellite image data, along with the transmittance and atmospheric light values, are input into the deep learning network; The deep learning network further includes a scene recognition unit, which is used to perform automatic scene type recognition operation to determine the scene type of the first satellite image data as the target scene; wherein, when performing the automatic scene type recognition operation, it includes: performing multi-scale feature extraction operation on the first satellite image data through the multi-layer convolutional units of the deep learning network to extract feature information of the visible light band and the infrared band respectively as multispectral features; The deep learning network building module is also used to build a second deep learning network; Multispectral features are combined with adjusted physical constraint weights and band fusion weights for feature fusion, and the fused features are then input into a second deep learning network, which outputs a dehazed image. The process of adjusting physical constraint weights includes: Based on the target scene type, the physical constraint weights are adjusted to match the optical characteristics and atmospheric conditions of the target scene; the physical constraint weights include transmittance constraint weights and atmospheric light value constraint weights. The adjustment process for band fusion weights includes: Based on the target scene type, select the band combination with the highest relevance from the bands of the first satellite image, and adjust the fusion weights of different bands; The scene types include agricultural, urban, forest, wetland, mountain, desert, and marine scenes.

Images