Defogging method and system for multispectral remote sensing satellite image

Abstract

The invention relates to the technical field of image defogging, discloses a multi-spectral remote sensing satellite image defogging method and system, and particularly provides a multi-spectral remote sensing satellite image defogging scheme which combines an atmospheric scattering model and deep learning and is self-adaptively optimized for a specific application scene. The transmissivity and the atmospheric light value are estimated through an atmospheric scattering model, and physical constraints are provided for the deep learning network. In the scene recognition stage, the scene type of the image is accurately recognized through multi-scale feature extraction, adaptive feature aggregation and a self-supervised learning mechanism. In the defogging stage, the physical constraint weight and the multi-band fusion weight are dynamically adjusted according to the identified scene, and it is ensured that the defogging result conforms to the atmospheric physical law and scene features. Therefore, detail retention, color rendition and physical consistency of the defogged image are effectively improved, the adaptive capacity of the deep learning network to complex scenes and diversified remote sensing images is enhanced, scene recognition errors are effectively reduced, and the quality and consistency of defogging results are further improved.

Description

Technical Field

[0001] The present invention relates to the technical field of image dehazing, and particularly to a dehazing method and system for multi-spectral remote sensing satellite images. Background Art

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

[0003] In recent years, dehazing methods based on deep learning have gradually been applied to the task of remote sensing image dehazing. 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. Ultimately, the dehazing effect of the output satellite images is still not ideal enough. Summary of the Invention

[0004] In view of this, the purpose of the present invention is to provide a dehazing method and system for multi-spectral remote sensing satellite images to solve the problem that the current dehazing effect on multi-spectral remote sensing satellite images is not ideal enough.

[0005] The first aspect of the present invention discloses a dehazing method for multi-spectral remote sensing satellite images, and the method includes the following steps:

[0006] S1. Collect multi-spectral remote sensing satellite image data to be processed and perform data preprocessing operations to obtain 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 through an atmospheric scattering model;

[0008] S3. Input the first satellite image data and the transmittance and atmospheric light value into a deep learning network with scene-adaptive physical constraints, and perform an automatic scene type recognition operation 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 inside 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 ocean scenes.

[0010] Further, the operation of automatically identifying the scene type based on the deep learning network in step S3 and determining the scene type of the first satellite image data includes the following sub-steps:

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

[0012] S302. Fuse the extracted multi-spectral features with the transmittance and atmospheric light value estimated in step S2 to generate atmospheric physical constraint features;

[0013] S303. Perform preliminary encoding operations on the atmospheric physical constraint features through the encoder of the deep learning network to obtain the high-level semantic features of the first satellite image;

[0014] S304. Perform feature aggregation operations on the multi-spectral 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. Input the aggregated adaptive features into the scene classification unit of the deep learning network, and identify the scene type of the first satellite image data through the scene classification unit.

[0016] Further, the operation of automatically identifying the scene type based on the deep learning network in step S3 and determining the scene type of the first satellite image data further includes the following sub-steps:

[0017] S306. Perform a confidence evaluation operation on the scene classification result output by the scene classification unit, judge the reliability of the scene classification result according to the confidence evaluation result, and correct the scene classification result through a self-supervised learning mechanism when the confidence is lower than the preset confidence threshold; the scene classification result includes a scene category label and a corresponding confidence score.

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

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

[0020] S3042. Adaptively adjust the initial feature weights according to the target scene and the requirements of the defogging task to generate a final feature weight allocation scheme;

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

[0022] S3044. Perform feature enhancement on the aggregated adaptive features; the feature enhancement includes operations of strengthening key features and suppressing redundant information.

[0023] Further, the confidence evaluation operation on 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 a feature consistency calculation result, and generate a correction factor based on the feature consistency calculation result.

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

[0028] Further, when the confidence is lower than the preset confidence threshold in step S306, the correction of the scene classification result through the self-supervised learning mechanism includes the following sub-steps:

[0029] S3065. Based on the feature consistency calculation result, select the scene category with the highest consistency with the features from the preset scene category set, and automatically generate a pseudo-label, which is used to replace the original classification label with low confidence.

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

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

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

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

[0034] S401. Dynamically adjust the physical constraint weights inside the deep learning network according to the target scene type to match the optical characteristics and atmospheric conditions of the target scene; the physical constraint weights include the transmittance constraint weight and the atmospheric light value constraint weight.

[0035] S402. Select the most relevant band combination from the bands of the first satellite image according to the target scene type and dynamically adjust the fusion weights of different bands.

[0036] S403. Combine the multi - spectral features with the adjusted physical constraint weights and band fusion weights for feature fusion, and input the fused features into the deep learning network to output the de - fogged image through the deep learning network.

[0037] S404. Evaluate the quality of the generated de - fogged image to determine whether it meets the visual and physical requirements of the target scene; if not meeting the requirements, re - adjust the physical constraint weights and band fusion weights and repeat the de - fogging process.

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

[0039] S4031. Determine the most relevant band combination in the first satellite image according to the target scene type, and perform band - weighted fusion on the multi - spectral features in combination with the adjusted band fusion weights to generate band - fused features.

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

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

[0042] S4034. Output the de - fogged multi - spectral remote sensing satellite image through the output layer of the deep learning network.

[0043] Furthermore, the physical constraint weights further include a radiative transfer equation constraint weight and a scene prior knowledge constraint weight.

[0044] The second aspect of the present invention discloses a haze removal system for multi-spectral remote sensing satellite images. This system is implemented based on the method disclosed in the first aspect. The system includes a data processing module, an estimation module, and a deep learning network construction module; wherein,

[0045] The data processing module is used to collect multi-spectral remote sensing satellite image data to be processed and perform data preprocessing operations to obtain 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 through an atmospheric scattering model;

[0047] The deep learning network construction module is used to construct a deep learning network with scene-adaptive physical constraints;

[0048] Input the first satellite image data and the transmittance and atmospheric light value into the deep learning network;

[0049] The deep learning network further includes a scene recognition unit and is used to perform an automatic scene type recognition operation 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 haze-removed image; wherein, the scene types include agricultural, urban, forest, wetland, mountain, desert, and ocean scenes.

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

[0052] The present invention specifically proposes a haze removal solution for multi-spectral remote sensing satellite images that combines an atmospheric scattering model and deep learning and is adaptively optimized for specific application scenarios. The atmospheric scattering model is used to estimate the transmittance and atmospheric light value to provide physical constraints for the deep learning network. In the scene recognition stage, through multi-scale feature extraction, adaptive feature aggregation, and self-supervised learning mechanisms, the scene type of the image is accurately recognized. In the haze removal stage, the physical constraint weights and multi-band fusion weights are dynamically adjusted according to the recognized scene to ensure that the haze removal result conforms to atmospheric physical laws and scene characteristics. Thereby effectively improving the detail retention, color restoration, and physical consistency of the haze-removed image, enhancing the adaptability of the deep learning network to complex scenes and diverse remote sensing images, effectively reducing scene recognition errors, and further improving the quality and consistency of the haze removal result. Description of the Drawings

[0053] The accompanying drawings described herein are used to provide a further understanding of the embodiments of the present invention, and constitute a part of this economic application, and do not limit the embodiments of the present invention. In the drawings:

[0054] Figure 1 It is a schematic flowchart of a method for dehazing multi-spectral remote sensing satellite images disclosed in an embodiment of the present invention;

[0055] Figure 2 It is a schematic structural diagram of a dehazing system for multi-spectral remote sensing satellite images disclosed in another embodiment of the present invention. Detailed implementation manners

[0056] In order to enable those skilled in the art to better understand the solution of the present invention, the technical solutions in the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings in the embodiments of the present invention. Obviously, the described embodiments are only a part of the embodiments of the present invention, rather than all the embodiments.

[0057] Embodiment 1

[0058] The first aspect of the present invention discloses a method for dehazing multi-spectral remote sensing satellite images. Please refer to Figure 1 , Figure 1 is a schematic flowchart of a method for dehazing multi-spectral remote sensing satellite images disclosed in an embodiment of the present invention. The method includes the following steps:

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

[0060] It can be understood that in the embodiments of the present invention, the purpose of preprocessing multi-spectral remote sensing satellite image data based on operations such as radiometric calibration, geometric calibration, multi-band registration, and atmospheric correction is to provide high-quality and accurate input data for subsequent dehazing processing. Through radiometric calibration, the sensor radiation response error is eliminated to ensure that the radiation value of the image truly reflects the characteristics of the ground object; based on geometric calibration, geometric distortions caused by the imaging angle of the sensor, satellite attitude changes, etc. are corrected to ensure the accuracy of the spatial position of the image; based on multi-band registration, the data of each band are accurately aligned in space to ensure the consistency of multi-spectral information; based on atmospheric correction, the influence of atmospheric scattering and absorption is removed to restore the true reflectance of the ground surface. The specific implementation manners of the above preprocessing operations in the embodiments of the present invention are not limited.

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

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

[0063] Due to the influence of atmospheric scattering and absorption during the imaging process of remote sensing images, simply relying on deep learning methods is likely to ignore the atmospheric physical characteristics, resulting in artifacts or non-compliance with physical laws in the defogging results. Therefore, by constructing a physical constraint basis in step S2 and introducing transmittance and atmospheric light value into the defogging network, accurate atmospheric scattering constraints can be provided for the deep learning network to ensure that the defogging 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 an automatic scene type recognition operation 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 inside the deep learning network according to the target scene to generate a defogged image; wherein, the scene types include agricultural, urban, forest, wetland, mountain, desert, and ocean scenes.

[0066] Further, the automatic scene type recognition operation 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:

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

[0068] S302. Fuse the extracted multi-spectral features with the transmittance and atmospheric light value estimated in step S2 to generate atmospheric physical constraint features.

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

[0070] S304. Perform a feature aggregation operation on the multi-spectral 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. Input the aggregated adaptive features into the scene classification unit of the deep learning network, and identify the scene type of the first satellite image data through the scene classification unit.

[0072] In the embodiments of the present invention, the high-level semantic features of satellite images refer to the features that reflect the overall structure and semantic information of the scenes in satellite images, and are used to abstractly express the texture, shape, and spatial relationships in the images, providing important semantic information support for subsequent scene recognition and adaptive haze removal, and helping the network to more accurately understand the image content and perform corresponding haze removal operations in different scenes. For example, in an agricultural scene, the high-level semantic features may include farmland distribution patterns, vegetation coverage types, and cultivation area boundaries; in an urban scene, they may include building outlines, road networks, and artificial facility layouts; in a forest scene, they may include canopy density, forest types, and forest area edges.

[0073] Further, the step S3 of performing the automatic scene type recognition operation based on the deep learning network to determine the scene type of the first satellite image data further includes the following sub-steps:

[0074] S306. Perform a confidence evaluation operation on the scene classification result output by the scene classification unit, judge the reliability of the scene classification result according to the confidence evaluation result, and correct the scene classification result through a self-supervised learning mechanism when the confidence is lower than the preset confidence threshold; the scene classification result includes a scene category label and a corresponding confidence score.

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

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

[0077] S3042. Adaptively adjust the initial feature weights according to the target scene and the requirements of the haze removal task to generate a final feature weight allocation scheme.

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

[0079] S3044. Perform feature enhancement on the aggregated adaptive features; the feature enhancement includes operations of strengthening key features and suppressing redundant information.

[0080] Further, the confidence evaluation operation on the scene classification result 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, 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, correct the initial confidence score, and output the final confidence score as the confidence evaluation result of the scene classification result.

[0085] Preferably, set the classification probability distribution P obtained from the scene classification unit as P = {p 1 , p i , …, p n}, where p i represents the probability of belonging to the scene category i, and n is the total number of preset scene categories:

[0086]

[0087] where, Z = {z 1 , z j , …, z n} is the unnormalized classification score output.

[0088] Select the highest probability value P max from the probability distribution P, obtain the category label C pred corresponding to the highest probability, and use P max as the initial confidence score.

[0089] Then the process of calculating feature consistency and generating a correction factor is:

[0090]

[0091] λ corr = exp(-α(1 - S))

[0092] where, S is the feature consistency, F multi is the multi-scale feature, F phys is the atmospheric physical constraint feature, F sem is the high-level semantic feature, F cls is the feature vector extracted based on the scene classification result output by the scene classification unit, λcorr is the correction factor, and α is the adjustment parameter.

[0093] Correct the initial confidence score:

[0094] P final = P max × λ corr

[0095] where P final is the final confidence score.

[0096] In the embodiments of the present invention, the core of feature consistency calculation lies in evaluating the matching degree between the scene classification result output by the scene classification unit and the image features. By comparing the classification result with the multi-scale features, atmospheric physical constraint features, and high-level semantic features extracted from the image, it is ensured that the classification result conforms to the actual image features. For example, assuming that the scene classification unit outputs the scene category label of "forest", the feature vector corresponding to this category label is compared with the multi-scale features (such as texture, edge and other detailed features), atmospheric physical constraint features (such as transmittance and atmospheric light value), and high-level semantic features (such as forest canopy density, vegetation coverage 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 be used for reducing the initial confidence score.

[0097] Further, when the confidence is lower than the preset confidence threshold in step S306, correcting the scene classification result through a self-supervised learning mechanism includes the following sub-steps:

[0098] S3065. Based on the feature consistency calculation result, select the scene category with the highest feature consistency from the preset scene category set, and automatically generate a pseudo-label, where the pseudo-label is used to replace the original classification label with low confidence.

[0099] Specifically, in this step, using the previously calculated feature consistency score, all possible scene categories (such as agriculture, urban, forest, wetland, etc.) preset are sorted, 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 the wetland features is low, and the feature consistency with the "forest" scene is higher, then "forest" will be selected as the pseudo-label. The mechanism of correcting the scene category through consistency calculation can effectively correct the misjudgment of the scene classification unit in the case of low confidence, and enhance the classification accuracy of the network in complex or mixed scenes.

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

[0101] S3067. Evaluate the difference between the scene classification result of the current scene classification unit and the pseudo-labels based on the loss function of self-supervised learning, and adaptively update 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. Use the updated scene classification unit to re-classify the first satellite image, 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 not reaching the threshold, re-perform the feature consistency check and pseudo-label generation for iterative correction until the confidence score reaches the threshold or reaches the preset maximum number of iterations.

[0103] In the embodiments of the present invention, scene recognition is a key link to achieve adaptive dehazing. By automatically identifying the scene type of the first satellite image data through a deep learning network, the scene category to which the image belongs can be accurately determined, such as agriculture, city, forest, wetland, mountain, desert, and ocean, etc. For different scene categories, there are significant differences in the atmospheric scattering characteristics and optical characteristics in the image. Through scene classification, targeted optimization strategies can be provided for the subsequent dehazing process. Specifically, the scene classification result determines the dynamic adjustment of physical constraint weights (such as transmittance, atmospheric light value) and the adaptive allocation of multi-band fusion weights to ensure that the dehazing network can generate the optimal dehazing effect according to the characteristics of a specific scene.

[0104] In addition, the reliability and accuracy of the classification result are improved through confidence evaluation and self-supervised learning correction mechanisms. Through the aggregation of 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 the confidence is low, enhancing the adaptability of the deep learning network to complex scenes and diverse remote sensing images, effectively reducing scene recognition errors, and further improving the quality and consistency of the dehazing result.

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

[0106] S401. Dynamically adjust the physical constraint weights inside the deep learning network according to the target scene type 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. Select the band combination with the highest relevance from the bands of the first satellite image according to the target scene type, and dynamically adjust the fusion weights of different bands;

[0108] S403. Perform feature fusion by combining the multi-spectral features with the adjusted physical constraint weights and band fusion weights, and input the fused features into the deep learning network to output the defogged image through the deep learning network;

[0109] S404. Evaluate the quality of the generated defogged image to determine whether it meets the visual and physical requirements of the target scene; if not, readjust the physical constraint weights and band fusion weights and repeat the defogging process.

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

[0111] S4031. Determine the band combination with the highest relevance in the first satellite image according to the target scene type, and perform band weighted fusion on the multi-spectral features in combination with the adjusted band fusion weights to generate band fusion features;

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

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

[0114] S4034. Output the defogged multi-spectral remote sensing satellite image through the output layer of the deep learning network.

[0115] Defogging is the process of restoring a multi-spectral remote sensing satellite image from a degraded state to a clear state. In the embodiments of the present invention, by dynamically adjusting the physical constraint weights and multi-band fusion weights inside the deep learning network, the defogging operation can adapt to the specific features of different scenes, such as agriculture, city, forest, wetland, mountain, desert, and ocean, etc.

[0116] By setting the physical constraint weights, it is ensured that the defogging result conforms to the actual atmospheric scattering law and scene characteristics, reducing artifacts and unnatural color deviations. The multi-band fusion weights dynamically allocate the weights of different spectral bands according to the scene requirements to maximize the complementarity of different band information and improve the detail retention and color restoration effects of the image.

[0117] In addition, by introducing a quality assessment mechanism, visual and physical consistency assessment can be carried out on the generated haze-removed images. If the expected effect is not achieved, the physical constraint weight and the band fusion weight are automatically adjusted, and the haze-removing process is iteratively optimized, improving the clarity, detail retention, and physical consistency of the haze-removing results, ensuring that high-quality haze-removed images can still be generated in complex and changing remote sensing scenarios.

[0118] Furthermore, the physical constraint weight also includes a radiative transfer equation constraint weight and a scene prior knowledge constraint weight.

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

[0120] The radiative transfer equation constraint weight is used to describe the scattering, absorption, and transmission characteristics of light when propagating in the atmosphere. By dynamically introducing this weight, the feature mapping process can be adjusted in the deep learning network to ensure that the haze-removing results conform to the physical laws of radiative transfer, further reducing the optical errors caused by atmospheric scattering, and improving the brightness and color restoration of the images. Especially in complex atmospheric conditions, such as haze, clouds, and high-humidity environments, it can significantly improve the quality and credibility of the haze-removed images.

[0121] In addition, the scene prior knowledge constraint weight combines the optical and structural characteristics of specific scenes (such as agriculture, urban, forest, wetland, etc.). These prior knowledge are derived from a large amount of remote sensing image experience data and domain knowledge. By introducing the scene prior knowledge into the deep learning network, the focus of feature learning inside the network can be dynamically adjusted. For example, in the agricultural scene, the constraint on the vegetation reflection characteristics is enhanced; in the urban scene, the constraints on the building outline and road structure are highlighted. This kind of scene prior knowledge constraint can guide the network to generate haze-removed images that are more in line with the actual scene, effectively retaining key details and reducing artifacts and information loss.

[0122] Embodiment 2

[0123] The second aspect of the present invention discloses a haze-removing system for multi-spectral remote sensing satellite images. Please refer to Figure 2 , Figure 2 which is a schematic structural diagram of a haze-removing system for multi-spectral remote sensing satellite images 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 collect multi-spectral remote sensing satellite image data to be processed and perform data preprocessing operations to obtain 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 through an atmospheric scattering model;

[0126] The deep learning network construction module is used to construct a deep learning network with scene-adaptive physical constraints;

[0127] Input the first satellite image data and the transmittance and atmospheric light value into the deep learning network;

[0128] The deep learning network also includes a scene recognition unit, which is used to perform an automatic scene type recognition operation 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 defogged image; wherein, the scene types include agricultural, urban, forest, wetland, mountain, desert and ocean scenes.

[0130] It should be noted that the specific implementation process of the second embodiment is similar to that of the first embodiment, and will not be elaborated in the second embodiment.

[0131] Finally, it should be noted that: a method and system for defogging multi-spectral remote sensing satellite images disclosed in the embodiments of the present invention only disclose the preferred embodiments of the present invention, and are only used to illustrate the technical solutions of the present invention, rather than limiting them; although the present invention has been described in detail with reference to the foregoing embodiments, those of ordinary skill in the art should understand that they can still modify the technical solutions recorded in the foregoing embodiments, or perform equivalent replacements on some of the technical features; and these modifications or replacements do not make the essence of the corresponding technical solutions 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 images, characterized in that: The method comprises the following steps: S1, collecting multispectral remote sensing satellite image data to be processed, and performing data preprocessing operations to obtain first satellite image data; the data preprocessing operations include radiation correction, geometric correction, multi-band registration and atmospheric correction; S2, estimating the transmittance and atmospheric light value of the first satellite image data by using an atmospheric scattering model; S3, inputting the first satellite image data and the transmittance and the atmospheric light value into a deep learning network of scene adaptive physical constraints, and performing an automatic scene type recognition operation based on the deep learning network to determine the scene type of the first satellite image data as the target scene; S4. Adaptively adjust the physical constraints and multi-band weight distribution within the deep learning network according to the target scene to generate a dehazed image; wherein the scene types include agriculture, city, forest, wetland, mountain, desert and ocean scenes.

2. The method for defogging multispectral remote sensing satellite images according to claim 1, characterized in that: In step S3, the automatic scene type recognition operation is performed based on the deep learning network to determine the scene type of the first satellite image data, which includes the following sub-steps: S301, performing a multi-scale feature extraction operation on the first satellite image data through a multi-layer convolution unit of a deep learning network, and extracting feature information of a visible light band and an infrared band respectively as multispectral features; S302, fusing the extracted multispectral features with the transmittance and atmospheric light value estimated in step S2 to generate atmospheric physical constraint features; S303, performing a preliminary encoding operation on the atmospheric physical constraint features through an encoder of the deep learning network to obtain high-level semantic features of the first satellite image; S304, performing a feature aggregation operation on the multispectral features, the atmospheric physical constraint features, and the 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: Input the aggregated adaptive features into a scene classification unit of a deep learning network, and identify the scene type of the first satellite image data through the scene classification unit.

3. The method for defogging multispectral remote sensing satellite images according to claim 2, characterized in that: In step S3, the automatic scene type recognition operation is performed based on the deep learning network to determine the scene type of the first satellite image data, and the following sub-steps are further included: S306. Perform a confidence assessment operation on the scene classification result output by the scene classification unit, determine the reliability of the scene classification result based on the confidence assessment result, and when the confidence is lower than a preset confidence threshold, correct the scene classification result through a self-supervised learning mechanism; the scene classification result includes a scene category label and a corresponding confidence score.

4. The method for defogging a multispectral remote sensing satellite image according to any one of claims 2 to 3, characterized in that: The step S304 specifically includes the following sub-steps: S3041, calculating the correlation between the multispectral features, the atmospheric physical constraint features, and the high-level semantic features based on the attention mechanism, and generating initial feature weights; S3042, adaptively adjusting the initial feature weights according to the target scene and the defogging task requirements, and generating a final feature weight allocation scheme; S3043, based on the final feature weight allocation scheme, weighted fusion of multi-scale features, atmospheric physical constraint features and high-level semantic features to generate aggregated adaptive features; S3044, performing feature enhancement on the aggregated adaptive features; The feature enhancement includes strengthening key features and suppressing redundant information operations.

5. The method for defogging a multispectral remote sensing satellite image according to any one of claims 2 to 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, obtaining an 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. Obtain the category label with the highest probability and its corresponding probability value based on the classification probability distribution, obtain the highest probability value, and use the highest probability value as the initial confidence score; S3063, calculating the consistency between the scene classification result output by the scene classification unit and the multi-scale features, the atmospheric physical constraint features, and the high-level semantic features, obtaining a feature consistency calculation result, and generating a correction factor based on the feature consistency calculation result; S3064: Correct the initial confidence score based on the correction factor, and output a final confidence score as a confidence evaluation result of the scene classification result.

6. The method for defogging multispectral remote sensing satellite images according to claim 5, characterized in that: In step S306, when the confidence level is lower than a preset confidence threshold, the scene classification result is corrected by the self-supervised learning mechanism, including the following sub-steps: S3065: Based on the feature consistency calculation result, a scene category with the highest feature consistency is selected from a preset scene category set, and a pseudo label is automatically generated, where the pseudo label is used to replace the original classification label with low confidence; S3066, re-extracting multi-scale features, atmospheric physical constraint features, and high-level semantic features from the first satellite image based on the pseudo-label, and matching these features with sample features in the pseudo-label feature library, calculating matching similarity, and verifying the rationality of the pseudo-label; S3067, evaluating the difference between the scene classification result of the current scene classification unit and the pseudo label based on the loss function of self-supervised learning, and adaptively updating 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. Use the updated scene classification unit to reclassify the first satellite image, output the corrected scene category label and the corresponding confidence score, and determine whether the corrected confidence score is higher than a 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 defogging multispectral remote sensing satellite images according to claim 2, characterized in that: The step S4 comprises the following sub-steps: S401. Dynamically adjust the physical constraint weights within the deep learning network according to the target scene type 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, selecting a band combination with the highest correlation from the bands of the first satellite image according to the target scene type, and dynamically adjusting the fusion weights of different bands; S403, combining the multispectral features with the adjusted physical constraint weights and band fusion weights to perform feature fusion, and inputting the fused features into a deep learning network, and outputting a dehazed image through the deep learning network; S404: Perform a quality assessment on the generated dehazed image to determine whether it meets the visual and physical requirements of the target scene; if not, readjust the physical constraint weights and band fusion weights, and repeat the dehazing process.

8. The method for defogging multispectral remote sensing satellite images according to claim 7, characterized in that: The step S403 includes the following sub-steps: S4031. Determine the band combination with the highest correlation in the first satellite image according to the target scene type, and perform band weighted fusion on the multispectral features in combination with the adjusted band fusion weights to generate band fusion features; S4032, inputting the band fusion features into a deep learning network, extracting features of different scales through a multi-layer convolution module, and aggregating the features of different scales based on an attention mechanism to generate multi-scale fusion features; S4033, introducing the adjusted physical constraint weight into the mapping process of the multi-scale fusion feature to generate the physical constraint fusion feature; S4034, output the dehazed multispectral remote sensing satellite image through the output layer of the deep learning network.

9. The method for defogging a multispectral remote sensing satellite image according to any one of claims 7 to 8, characterized in that: The physical constraint weights also include radiation transfer equation constraint weights and scene prior knowledge constraint weights.

10. A multi-spectral remote sensing satellite image defogging system, characterized in that: The system is implemented based on the method described in any one of claims 1 to 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 collect multispectral remote sensing satellite image data to be processed and perform data preprocessing operations to obtain first satellite image data; the data preprocessing operations include radiation 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 through an atmospheric scattering model; The deep learning network construction module is used to construct a deep learning network of scene adaptive physical constraints; Inputting the first satellite image data and the transmittance and atmospheric light value into the deep learning network; The deep learning network also includes a scene recognition unit and is used to perform an automatic scene type recognition operation to determine the scene type of the first satellite image data as a target scene; The deep learning network is used to adaptively adjust the internal physical constraints and multi-band weight distribution according to the target scene to generate a dehazed image; wherein the scene types include agricultural, urban, forest, wetland, mountain, desert and ocean scenes.

Images