A thin cloud continuous transition modeling and boundary segmentation method and system for remote sensing images
By using multi-source evidence modeling and energy function optimization, the problems of boundary breaks and misjudgments in thin cloud segmentation of remote sensing images were solved, and the continuity and uncertainty of thin cloud boundaries were characterized, thus improving the accuracy and interpretability of the segmentation results.
Patent Information
- Authority / Receiving Office
- CN · China
- Patent Type
- Applications(China)
- Current Assignee / Owner
- NORTHEAST FORESTRY UNIV
- Filing Date
- 2026-02-14
- Publication Date
- 2026-05-29
AI Technical Summary
Existing methods for thin cloud segmentation in remote sensing images suffer from boundary breaks, jagged edges, high false positive rates, and a lack of characterization of boundary continuity and uncertainty in thin cloud regions. Furthermore, they cannot effectively represent the gradual change in the degree of thin cloud coverage.
A multi-source evidence modeling method is adopted to construct a strip-shaped evidence representation of thin cloud boundaries by fusing brightness variation and texture attenuation features. An energy function constrained by data consistency, smoothness and geometric continuity is introduced to optimize the thin cloud coverage intensity field and output a continuous thin cloud coverage intensity map.
It effectively suppresses boundary breaks and misjudgments, improves the accuracy and interpretability of thin cloud segmentation, provides a clear indication of the gradual change process of thin cloud coverage and the range of boundary uncertainty, and is suitable for remote sensing image processing with complex surface backgrounds.
Smart Images
Figure FT_1 
Figure SMS_10 
Figure SMS_17
Abstract
Description
Technical Field
[0001] This invention relates to the field of remote sensing image processing technology, specifically to a method and system for modeling and segmenting thin cloud transitions and boundaries in remote sensing images. Background Technology
[0002] Remote sensing imagery serves as a core data source for fields such as ground feature information extraction, resource surveys, and environmental monitoring; its imaging quality directly determines the accuracy of subsequent interpretation and analysis. Visible light remote sensing images are highly susceptible to interference from thin atmospheric clouds during acquisition. These clouds, appearing as a semi-transparent layer covering the Earth's surface, alter its brightness and weaken the texture details of ground features, posing a significant challenge to subsequent remote sensing image processing. Therefore, accurate segmentation of thin cloud areas is a crucial step in remote sensing image preprocessing.
[0003] Existing methods for cloud segmentation in remote sensing images mostly model the problem as a pixel-level binary classification task, attempting to make an absolute "cloud / non-cloud" determination at each pixel location. While these methods achieve good results in thick cloud segmentation scenarios, they suffer from numerous intractable drawbacks in thin cloud segmentation. The core problem lies in the fact that existing methods ignore the essential characteristics of the boundary between thin clouds and the land surface—thin clouds and the land surface are not a clear dichotomy; their boundary is a continuous transition zone with a finite width, exhibiting gradual changes in brightness and texture, and the boundary region is wide and unstable. Specific drawbacks include:
[0004] Pixel-level binary classification models forcibly create "sharp" thin cloud boundaries, leading to severe boundary breaks and jagged edges in the segmentation results, which contradicts the actual imaging mechanism of thin clouds. Existing methods often use single visual features such as single-scale gradients for boundary detection, which easily leads to misjudgments of strong structures or high-reflectivity areas such as roads and buildings on the ground, resulting in a persistently high misjudgment rate in thin cloud segmentation. There is a lack of effective characterization of the continuity and uncertainty of thin cloud boundaries; they only attempt to detect precise boundary lines at the single-pixel level, which cannot characterize the gradual process of thin cloud coverage or indicate the uncertain range of the boundary. Relying solely on local brightness or texture features for boundary discrimination without introducing global geometric constraints leads to problems such as boundary fragmentation and false bifurcations in areas with complex surface textures or under noise interference, resulting in poor global connectivity and topological consistency. Summary of the Invention
[0005] The purpose of this invention is to provide a method and system for thin cloud continuous transition modeling and boundary segmentation for remote sensing images, so as to solve the problems mentioned in the background art.
[0006] To achieve the above objectives, the present invention provides the following technical solution: a method for thin cloud continuity transition modeling and boundary segmentation for remote sensing images, comprising the following steps:
[0007] S1. Preprocess the input visible light remote sensing image, convert the multi-channel image into a brightness channel, and perform radiometric normalization and mild smoothing and denoising on the brightness channel to obtain the preprocessed brightness field.
[0008] S2. The brightness field is scale-separated and decomposed into low-frequency background components and high-frequency residual components by Gaussian filtering. Multi-scale brightness change evidence is extracted and normalized based on the low-frequency background components. Relative texture intensity is calculated and texture attenuation evidence is constructed based on the high-frequency residual components. The brightness change evidence and texture attenuation evidence are weighted and fused and obtained by S-shaped mapping to obtain a boundary evidence map. The boundary evidence map is then subjected to Gaussian smoothing banding to generate a banded evidence representation of the thin cloud boundary.
[0009] S3. Based on the strip evidence expression, an initial thin cloud cover intensity field with a value range of [0,1] is constructed so that the intensity field shows a monotonic change trend in the high response area of the strip evidence and remains stable in the non-evidence area, where the intensity value 0 represents a clear surface and 1 represents complete thin cloud cover;
[0010] S4. Construct an energy function that includes a data consistency term, a smoothness constraint term, and a geometric continuity constraint term. Optimize the initial thin cloud cover intensity field with the goal of minimizing this energy function to obtain the final thin cloud cover intensity field. The data consistency term constrains the deviation between the intensity field and the initial estimate, the smoothness constraint term suppresses boundary jaggedness, and the geometric continuity constraint term ensures the connectivity and topological consistency of the boundary.
[0011] S5. Extract the isomorphic structure or gradient ridge from the final thin cloud coverage intensity field as the thin cloud boundary, determine the region with intensity value in the middle range as the boundary uncertainty region, and output a continuous thin cloud coverage intensity map to complete the thin cloud continuous transition modeling and boundary segmentation.
[0012] Furthermore, in S2, the multi-source evidence modeling composed of brightness change evidence and texture attenuation evidence is used in visible light remote sensing images. Thin clouds typically cover the surface as a semi-transparent layer, and their imaging results show slow changes in surface brightness and significant attenuation of high-frequency texture information. The boundary between thin clouds and the surface is not a typical strong contrast transition, but rather a continuous transition region with a finite width. The discrimination of thin cloud boundaries is extended from "strong gradient point detection" to "transition region modeling supported by multi-source evidence." By jointly characterizing brightness change features and texture attenuation features, a multi-source evidence representation of thin cloud boundaries is constructed. Let the input visible light remote sensing image be:
[0013]
[0014] Where x represents the pixel coordinates and Ω is the image domain; if I is a multi-channel image, it is first converted to the luminance channel in a linear or perceptually consistent manner. This is done to eliminate the interference of color differences on subsequent gradient and texture measurements; subsequently... Radiation normalization and mild smoothing are performed to suppress sensing noise and ensure the comparability of feature amplitudes across different scenes. Based on this, a scale separation approach is introduced to decompose the brightness field into low-frequency background components and high-frequency texture components. Through a scale of... Gaussian filtering is used to obtain low-frequency background:
[0015] ,
[0016] It primarily reflects the slow changing trend of reflection from thin cloud cover and the Earth's surface; it also defines high-frequency residuals:
[0017]
[0018] Used to depict the texture and detailed structure of the ground.
[0019] Furthermore, the brightness variation features are used to characterize the gradual spatial position of thin cloud boundaries. However, considering that thin cloud boundaries typically correspond to brightness variations at mid-to-low frequency scales rather than sharp transitions, this paper does not use single-scale gradients as the criterion, but rather uses multiple scales. upper low frequency background Calculate the gradient magnitude:
[0020]
[0021] Subsequently, evidence of brightness changes was obtained through scale aggregation:
[0022]
[0023] It is then linearly normalized to the interval [0,1]; this operation can adapt to the boundary scale corresponding to thin clouds of different thicknesses, so that G(x) forms a continuous response in the thin cloud-surface transition region, while avoiding over-enhancement in the region with strong surface structure but no thin cloud cover.
[0024] Furthermore, the texture attenuation feature is used to characterize the attenuation effect of thin clouds on high-frequency information of the ground surface; in the high-frequency residual Above, in a local window Calculate texture energy:
[0025]
[0026] in and These represent the local variance and mean operators, respectively; to avoid bias caused by inherent texture differences among different land surface types, a larger-scale reference window is introduced. Constructing relative texture intensity:
[0027]
[0028] Where ε is a minimal constant to prevent numerical instability; based on this relative metric, a texture attenuation evidence map is defined:
[0029]
[0030] And also normalized to [0,1]; in areas covered by thin clouds, the surface texture is uniformly weakened, Larger values will be obtained at bright but textured surface structures.
[0031] Furthermore, the evidence of brightness change Evidence of texture attenuation A multi-source evidence representation for constructing thin cloud boundaries is used in conjunction; to highlight regions that simultaneously satisfy brightness gradation and texture attenuation, while suppressing false boundaries that are significant only on a single feature, the following fusion method is adopted:
[0032]
[0033] Where a, b, and c are feature weights, and τ is the bias term. This represents the sigmoid mapping function; the fusion result Instead of directly representing the single-pixel-level boundary location, it characterizes the "strength of spatial evidence for the possible existence of thin cloud boundaries"; to extend the thin cloud boundary from a single response location to a transition region with a finite width, it... Further stripping treatment is required.
[0034] Through the scale The smoothing operator is obtained
[0035]
[0036] in Control the typical width of the transition region. The zonal evidence used as the boundary of thin clouds corresponds to a continuous transition zone between thin clouds and the ground surface in high-response regions, rather than a single strong gradient pixel. This zonal evidence not only provides a stable spatial support domain for subsequent transition zone modeling under geometric continuity constraints, but also naturally characterizes the uncertainty range of the thin cloud boundary, providing an interpretable input basis for subsequent analysis and experimental design.
[0037] Furthermore, the energy function constructed in step S4 is:
[0038]
[0039] in For data consistency items, For smoothness constraint terms, For geometric continuity constraints, , These are the weighting coefficients;
[0040] The data consistency item is:
[0041]
[0042] Weight Expressed by strip evidence The decision, in Low values are used for high-response regions, and high values are used for non-evidence regions.
[0043] The smoothness constraint term is:
[0044]
[0045] in Take a smaller value within the strip of evidence and a larger value in non-evidence areas;
[0046] The geometric continuity constraint term is achieved by introducing curvature smoothing penalties on the contour lines of the thin cloud cover intensity field, performing connectivity analysis on the transition region of the intensity field, and applying penalties or backfill constraints to isolated regions.
[0047] Furthermore, in step S4, the optimization of the initial thin cloud cover intensity field with the goal of minimizing the energy function is achieved through numerical iteration. The iteration stops when the energy function converges or when a preset number of iterations is reached, thus obtaining the final thin cloud cover intensity field. The final thin cloud cover intensity field value is still [0,1]. Its isomorphic structure or gradient ridge is the core position of the thin cloud boundary. The region where the intensity value is in the middle range is the gradual transition region between the thin cloud and the ground surface, that is, the boundary uncertainty region.
[0048] Furthermore, it also includes a step of weakly supervised training of the thin cloud continuous transition modeling and boundary segmentation method:
[0049] Visible light remote sensing images were used as training samples, and regional binary thin cloud masks were used as weakly supervised annotations. The annotations only indicated thin cloud and non-thin cloud areas and had no precise boundary information.
[0050] Perform steps S1-S4 on the training samples to obtain the continuous thin cloud cover intensity field of the training samples, and construct the regional consistency loss function:
[0051]
[0052] The constraint intensity field converges to 1 in the thin cloud region and to 0 in the non-thin cloud region; the region consistency loss and the geometric regularization term are weighted to form the total loss, and the parameters can be learned by updating the model through backpropagation or numerical optimization; no annotation information is required in the testing phase, and steps S1-S5 are directly executed to complete the thin cloud segmentation.
[0053] A thin cloud continuous transition modeling and boundary segmentation system for remote sensing images includes an image preprocessing module, a multi-source evidence construction module, an initial intensity field construction module, an intensity field optimization module, and a result output module.
[0054] The image preprocessing module is used to preprocess the input visible light remote sensing image, complete the conversion from multi-channel to brightness channel, radiometric normalization and mild smoothing and noise reduction, and output the preprocessed brightness field.
[0055] The multi-source evidence construction module is used to perform scale separation on the brightness field to obtain low-frequency background components and high-frequency residual components, extract and normalize brightness change evidence, construct texture attenuation evidence, and output a banded evidence expression of thin cloud boundary after weighted fusion and banding processing.
[0056] The initial intensity field construction module is used to construct and output an initial thin cloud cover intensity field with a value range of [0,1] based on the strip evidence expression.
[0057] The intensity field optimization module is used to construct an energy function that includes a data consistency term, a smoothness constraint term, and a geometric continuity constraint term, and optimizes the initial thin cloud cover intensity field with the goal of minimizing the energy function, and outputs the final thin cloud cover intensity field.
[0058] The result output module is used to extract thin cloud boundaries from the final thin cloud coverage intensity field, determine the uncertain regions of the boundaries, and output a continuous thin cloud coverage intensity map to complete the thin cloud continuous transition modeling and boundary segmentation.
[0059] Furthermore, the multi-source evidence construction module includes a brightness feature extraction unit, a texture feature extraction unit, an evidence fusion unit, and a banding processing unit;
[0060] The brightness feature extraction unit is used to perform scale separation on the brightness field to obtain low-frequency background components, extract multi-scale brightness change evidence and normalize it.
[0061] The texture feature extraction unit is used to perform scale separation on the brightness field to obtain high-frequency residual components, calculate the relative texture intensity, and construct evidence of texture attenuation.
[0062] The evidence fusion unit is used to fuse brightness change evidence and texture attenuation evidence according to weights and obtain a boundary evidence map through S-shaped mapping.
[0063] The banding processing unit is used to perform Gaussian smoothing banding processing on the boundary evidence map to generate a banded evidence representation of the thin cloud boundary.
[0064] Compared with the prior art, the beneficial effects of the present invention are:
[0065] Abandoning the traditional pixel-level binary classification approach, this paper reformulates thin cloud segmentation as a modeling problem of continuous transition regions. The thin cloud-surface transition region is explicitly represented as a continuous cover intensity field with values [0,1]. The numerical values directly reflect the degree of thin cloud coverage, rather than discrete category labels, allowing the thin cloud boundary to naturally correspond to the gradual change region of the intensity field. This completely solves the boundary breakage and jaggedness problems caused by binary classification. By jointly extracting visual cues with complementary physical meanings—brightness change and texture attenuation—a multi-source evidence representation of thin cloud boundaries is constructed. Only the true thin cloud boundary regions that simultaneously satisfy "brightness gradual change + texture attenuation" are highlighted, effectively suppressing false boundaries of strong surface structures with only a single significant feature. This significantly reduces the misjudgment rate of thin cloud segmentation and improves the accuracy of boundary detection. Furthermore, by constructing an energy function with constraints on smoothness and geometric continuity, the thin cloud cover intensity field is optimized. This achieves continuous and gradually changing natural forms for thin cloud boundaries through curvature smoothing penalties, while suppressing isolated fragmented boundaries through connectivity analysis. Simultaneously, reasonable morphological changes in local regions are allowed to adapt to the irregular structure of real thin clouds, thus realizing… This method achieves a unified approach to segmentation results, ensuring global connectivity, smoothness, and local realism. The final output is not a single binary segmentation mask, but rather contains three core types of information: a continuous thin cloud coverage intensity map, thin cloud boundary locations, and uncertain boundary regions. This clearly indicates the gradual change in thin cloud coverage and the uncertain range of the boundaries, avoiding the overconfident hard judgments of traditional methods and significantly improving the interpretability of the thin cloud segmentation results. The generated continuous coverage intensity field, zonal boundary evidence, and uncertain region representation provide richer and more reliable intermediate information and prior knowledge for subsequent downstream tasks such as thin cloud morphology statistics, thin cloud influence intensity modeling, thin cloud removal from remote sensing images, and simulation studies. This effectively expands the application scenarios of this method in the field of high-precision remote sensing image processing. Through preprocessing and feature extraction designs such as radiometric normalization, multi-scale feature aggregation, and weighted smoothing, interference from sensing noise, exposure differences, and complex surface textures is effectively suppressed. This allows the method to maintain high stability in thin cloud segmentation of remote sensing images with weak contrast and complex surface backgrounds, making it suitable for various practical visible light remote sensing image processing scenarios. Attached Figure Description
[0066] Figure 1 This is a flowchart of the thin cloud continuous transition modeling and boundary segmentation method for remote sensing images according to the present invention. Detailed Implementation
[0067] Please see Figure 1 This invention provides a technical solution: a method for modeling and segmenting the continuous transition of thin clouds in remote sensing images, comprising the following steps:
[0068] S1. Preprocess the input visible light remote sensing image, convert the multi-channel image into a brightness channel, and perform radiometric normalization and mild smoothing and denoising on the brightness channel to obtain the preprocessed brightness field.
[0069] S2. The brightness field is scale-separated and decomposed into low-frequency background components and high-frequency residual components by Gaussian filtering. Multi-scale brightness change evidence is extracted and normalized based on the low-frequency background components. Relative texture intensity is calculated based on the high-frequency residual components and texture attenuation evidence is constructed. The brightness change evidence and texture attenuation evidence are weighted and fused and obtained by S-shaped mapping to obtain the boundary evidence map. The boundary evidence map is then subjected to Gaussian smoothing and banding to generate a banded evidence representation of the thin cloud boundary.
[0070] S3. Based on the expression of strip evidence, an initial thin cloud cover intensity field with a value range of [0,1] is constructed so that the intensity field shows a monotonic change trend in the high response area of strip evidence and remains stable in the non-evidence area, where the intensity value 0 represents a clear surface and 1 represents complete thin cloud cover;
[0071] S4. Construct an energy function containing a data consistency term, a smoothness constraint term, and a geometric continuity constraint term. Optimize the initial thin cloud cover intensity field with the goal of minimizing this energy function to obtain the final thin cloud cover intensity field. The data consistency term constrains the deviation between the intensity field and the initial estimate, the smoothness constraint term suppresses boundary jaggedness, and the geometric continuity constraint term ensures the connectivity and topological consistency of the boundary.
[0072] S5. Extract the isomorphic structure or gradient ridge from the final thin cloud coverage intensity field as the thin cloud boundary, determine the region with intensity value in the middle range as the boundary uncertainty region, and output a continuous thin cloud coverage intensity map to complete the thin cloud continuous transition modeling and boundary segmentation.
[0073] The boundary of thin clouds is not the location of a single pixel, but rather a spatial region of finite width within which the thin cloud coverage gradually transitions from the ground surface to complete cloud cover. Based on this understanding, the method first extracts two types of visual cues from visible light imagery that are closely related to the thin cloud boundary but have complementary physical meanings: one is brightness variation features, used to describe the overall brightness gradient caused by the thin cloud cover; the other is texture attenuation features, used to characterize the attenuation effect of thin clouds on surface details and high-frequency textures. By jointly analyzing these two types of cues, the method does not directly output a boundary line in the sense of "strongest gradient," but instead constructs a banded boundary evidence region to indicate the possible spatial extent of the thin cloud boundary. After obtaining this banded evidence, the method further introduces geometric continuity modeling, explicitly representing the transition region between thin clouds and the ground as a continuous intensity field, the value of which reflects the degree of thin cloud coverage rather than discrete category labels. Under this representation, the thin cloud boundary naturally corresponds to a transition region in a continuous intensity field, rather than a binary abrupt change location. Subsequently, by introducing geometric constraints such as smoothness and connectivity, this continuous representation is optimized, ensuring that the thin cloud boundary remains coherent and smooth at the overall scale, while allowing reasonable morphological variations in local regions to adapt to the irregular structure of real thin clouds. The fundamental difference between this method and traditional cloud segmentation methods lies in its explicit modeling of the continuity and uncertainty of thin cloud boundaries, rather than forcibly pursuing pixel-level hard boundary determination. This modeling approach avoids artificially creating "overly sharp" segmentation results in thin cloud scenes, making the boundaries more stable and interpretable under complex surface backgrounds and weak contrast conditions, and better conforming to the actual imaging mechanism of thin clouds. Ultimately, the method not only outputs the segmentation results of thin cloud regions but also provides a continuous transition zone representation, offering richer and more reliable intermediate information for subsequent thin cloud morphology analysis, influence intensity modeling, and simulation studies.
[0074] In S2, multi-source evidence modeling, consisting of brightness variation evidence and texture attenuation evidence, addresses the issue that thin clouds in visible light remote sensing images typically appear as a semi-transparent layer covering the Earth's surface. The imaging results show slow changes in surface brightness and significant attenuation of high-frequency texture information. The boundary between thin clouds and the surface is not a typical strong contrast transition, but rather resembles a continuous transition region with a finite width. The determination of thin cloud boundaries is expanded from "strong gradient point detection" to "transition region modeling supported by multi-source evidence." By jointly characterizing brightness variation features and texture attenuation features, a multi-source evidence representation of thin cloud boundaries is constructed. The input visible light remote sensing image is:
[0075]
[0076] Where x represents the pixel coordinates and Ω is the image domain; if I is a multi-channel image, it is first converted to the luminance channel in a linear or perceptually consistent manner. This is done to eliminate the interference of color differences on subsequent gradient and texture measurements; subsequently... Radiation normalization and mild smoothing are performed to suppress sensing noise and ensure the comparability of feature amplitudes across different scenes. Based on this, a scale separation approach is introduced to decompose the brightness field into low-frequency background components and high-frequency texture components. Through a scale of... Gaussian filtering is used to obtain low-frequency background:
[0077] ,
[0078] It primarily reflects the slow changing trend of reflection from thin cloud cover and the Earth's surface; it also defines high-frequency residuals:
[0079]
[0080] Used to depict the texture and detailed structure of the ground.
[0081] Brightness variation features are used to characterize the gradual spatial position of thin cloud boundaries. However, considering that thin cloud boundaries typically correspond to brightness variations at mid-to-low frequency scales rather than sharp transitions, this paper does not use single-scale gradients as a criterion, but rather uses multiple scales. upper low frequency background Calculate the gradient magnitude:
[0082]
[0083] Subsequently, evidence of brightness changes was obtained through scale aggregation:
[0084]
[0085] It is then linearly normalized to the interval [0,1]; this operation can adapt to the boundary scale corresponding to thin clouds of different thicknesses, so that G(x) forms a continuous response in the thin cloud-surface transition region, while avoiding over-enhancement in the region with strong surface structure but no thin cloud cover.
[0086] Texture attenuation features are used to characterize the attenuation effect of thin clouds on high-frequency information of the Earth's surface; in high-frequency residuals Above, in a local window Calculate texture energy:
[0087]
[0088] in and These represent the local variance and mean operators, respectively; to avoid bias caused by inherent texture differences among different land surface types, a larger-scale reference window is introduced. Constructing relative texture intensity:
[0089]
[0090] Where ε is a minimal constant to prevent numerical instability; based on this relative metric, a texture attenuation evidence map is defined:
[0091]
[0092] And also normalized to [0,1]; in areas covered by thin clouds, the surface texture is uniformly weakened, Larger values will be obtained at bright but textured surface structures.
[0093] Evidence of brightness changes Evidence of texture attenuation A multi-source evidence representation for constructing thin cloud boundaries is used in conjunction; to highlight regions that simultaneously satisfy brightness gradation and texture attenuation, while suppressing false boundaries that are significant only on a single feature, the following fusion method is adopted:
[0094]
[0095] Where a, b, and c are feature weights, and τ is the bias term. This represents the sigmoid mapping function; the fusion result Instead of directly representing the single-pixel-level boundary location, it characterizes the "strength of spatial evidence for the possible existence of thin cloud boundaries"; to extend the thin cloud boundary from a single response location to a transition region with a finite width, it... Further stripping treatment is required.
[0096] Through the scale The smoothing operator is obtained
[0097]
[0098] in Control the typical width of the transition region. The zonal evidence used as the boundary of thin clouds corresponds to a continuous transition zone between thin clouds and the ground surface in high-response regions, rather than a single strong gradient pixel. This zonal evidence not only provides a stable spatial support domain for subsequent transition zone modeling under geometric continuity constraints, but also naturally characterizes the uncertainty range of the thin cloud boundary, providing an interpretable input basis for subsequent analysis and experimental design.
[0099] The expression of multi-source evidence for thin cloud boundaries constructed in the previous section While it's possible to spatially characterize the potentially continuous transition region between thin clouds and the surface boundary, this evidence primarily reflects the statistical response of local brightness and texture features, failing to explicitly guarantee the connectivity, smoothness, and topological consistency of the boundary at a global scale. Since thin clouds typically exhibit a generally continuous, locally gradational cape structure in their spatial distribution, relying solely on local evidence for boundary determination can easily lead to boundary fragmentation, false bifurcation, or unreasonable local abrupt changes in areas with complex surface textures or under noise interference. Therefore, it's necessary to introduce geometric continuity constraints during the transition zone modeling stage to structurally represent and optimize the thin cloud boundary. To this end, the thin cloud-surface transition region is explicitly modeled as a continuous scalar field. Its physical meaning is thin cloud cover intensity or thin cloud-surface mixing ratio, among which Indicates a clearly defined area of the earth's surface. This indicates a completely thin cloud-covered area, while the thin cloud boundary corresponds to... A continuous transitional region in space. Based on the strip-shaped evidence obtained in the first part. , can be constructed initial estimate For example, by... Linear or nonlinear mappings are applied to make it exhibit a monotonic variation trend within the boundary evidence zone, while maintaining relative stability in the non-evidence region. Based on this, the modeling and optimization of the thin cloud transition zone is formulated as an energy minimization problem with geometric constraints, the objective of which is to minimize energy while maintaining consistency with the initial evidence. Spatially satisfying the required geometric continuity of thin cloud boundaries
[0100] The energy function constructed in step S4 is:
[0101]
[0102] in For data consistency items, For smoothness constraint terms, For geometric continuity constraints, , These are the weighting coefficients;
[0103] The data consistency term is used to constrain the optimized α(x) from deviating from the initial transition band structure obtained from multi-source evidence. The data consistency term is:
[0104]
[0105] Weight Expressed by strip evidence The decision, in Low values are used for high-response regions, and high values are used for non-evidence regions.
[0106] To suppress boundary jaggedness and local oscillations caused by noise or high-frequency surface textures, a spatial smoothing constraint is introduced into α(x) during the optimization process. Considering that the thin cloud boundary itself should maintain a certain width and gradient variation, this smoothing term does not employ overly strong second-order smoothing, but rather a regularized form weighted by the gradient magnitude. The smoothness constraint term is:
[0107]
[0108] in Take a smaller value within the strip of evidence and a larger value in non-evidence areas;
[0109] The geometric continuity constraint is achieved by introducing curvature smoothing penalties onto the contour lines of the thin cloud cover intensity field, performing connectivity analysis on the transition regions of the intensity field, and applying penalties or backfill constraints to isolated regions. Specifically, this can be achieved by... The contour geometry introduces a curvature smoothing penalty, causing the boundaries to tend to form continuous, gradually varying geometric shapes, thereby suppressing boundary polygons and spurious abrupt changes caused by evidence of local anomalies. Simultaneously, by adjusting the contour geometry during the optimization iteration process... Located in the transition zone Connectivity analysis is performed on the region, and penalties or backfill constraints are imposed on transitional regions that are too small or isolated, thereby avoiding boundary fragmentation.
[0110] In step S4, minimizing the energy function is the objective of optimizing the initial thin cloud cover intensity field using a numerical iterative approach. The iteration stops when the energy function converges or when a preset number of iterations is reached, yielding the final thin cloud cover intensity field. The final value of the thin cloud cover intensity field is still [0,1]. Its isostructure or gradient ridge is the core location of the thin cloud boundary. The region where the intensity value is in the middle range is the gradual transition region between the thin cloud and the ground surface, that is, the boundary uncertainty region.
[0111] It also includes the step of weakly supervised training of the thin cloud continuous transition modeling and boundary segmentation method:
[0112] Visible light remote sensing images were used as training samples, and regional binary thin cloud masks were used as weakly supervised annotations. The annotations only indicated the thin cloud and non-thin cloud regions, without precise boundary information.
[0113] During the training phase, this method uses visible light remote sensing imagery as input and employs regional-level thin cloud annotation for weakly supervised learning, without relying on precise annotations of thin cloud boundaries or transition zones. Let the nth sample in the training set be a visible light remote sensing image. Where H and W represent the spatial dimensions of the image, and C represents the number of channels. The corresponding weak supervision annotation is a binary thin cloud mask. This mask only indicates thin-cloud regions and non-thin-cloud regions, and does not contain any geometric information about the location of thin-cloud boundaries or the width of transitions. The training objective is not to directly predict this binary mask, but to learn from the input image. The transition zone to continuous thin clouds indicates The mapping, where x represents the pixel coordinate, This indicates the intensity of thin cloud cover or the ratio of thin cloud to surface mixture. For any input image... Its forward processing flow remains completely consistent between the training and testing phases. First, following the aforementioned multi-source evidence modeling method, from... Evidence of brightness variation and texture attenuation was constructed and fused to obtain a banded evidence representation of thin cloud boundaries. This evidence is only used to characterize the spatial region where the thin cloud-surface boundary may exist, without directly determining the boundary. Based on this, according to... Initial estimation of the intensity field in the continuous transition zone This causes the values to tend to be high within the thin cloud region and low within the non-thin cloud region, exhibiting a smooth variation trend within the boundary evidence zone. Subsequently, geometric continuity constraints are introduced to... Optimization was performed to obtain the final representation of the continuous thin cloud transition zone. This optimization process is accomplished by minimizing an energy function that includes data consistency and geometric regularization terms, thereby ensuring... Maintain connectivity, smoothness, and topological consistency at the overall scale, while allowing reasonable morphological variations consistent with the evidence in local regions. Supervision information during the training phase is applied to the continuous representation only through region consistency constraints. Specifically, when the thin cloud mask satisfies Within the area, constraints Converging toward 1; in Within the area, constraints It converges to 0 without imposing hard constraints on the transition width or precise location at the thin cloud boundary.
[0114] Perform steps S1-S4 on the training samples to obtain the continuous thin cloud cover intensity field of the training samples, and construct the regional consistency loss function:
[0115]
[0116] The constraint intensity field converges to 1 in the thin cloud region and to 0 in the non-thin cloud region; the region consistency loss and the geometric regularization term are weighted to form the total loss, and the parameters can be learned by updating the model through backpropagation or numerical optimization; no annotation information is required in the testing phase, and steps S1-S5 are directly executed to complete the thin cloud segmentation.
[0117] A thin cloud continuous transition modeling and boundary segmentation system for remote sensing images includes an image preprocessing module, a multi-source evidence construction module, an initial intensity field construction module, an intensity field optimization module, and a result output module.
[0118] The image preprocessing module is used to preprocess the input visible light remote sensing images, complete the conversion from multi-channel to brightness channel, radiometric normalization and mild smoothing and noise reduction, and output the preprocessed brightness field.
[0119] The multi-source evidence construction module is used to perform scale separation of the brightness field to obtain low-frequency background components and high-frequency residual components, extract and normalize brightness change evidence, construct texture attenuation evidence, and output the banded evidence expression of thin cloud boundary after weighted fusion and banding processing.
[0120] The initial intensity field construction module is used to construct and output an initial thin cloud cover intensity field with values ranging from [0,1] based on the strip evidence expression.
[0121] The intensity field optimization module is used to construct an energy function that includes a data consistency term, a smoothness constraint term, and a geometric continuity constraint term. The initial thin cloud cover intensity field is optimized with the goal of minimizing this energy function, and the final thin cloud cover intensity field is output.
[0122] The results output module is used to extract thin cloud boundaries from the final thin cloud coverage intensity field, determine the uncertain regions of the boundaries, and output a continuous thin cloud coverage intensity map to complete the continuous transition modeling and boundary segmentation of thin clouds.
[0123] The multi-source evidence construction module includes a brightness feature extraction unit, a texture feature extraction unit, an evidence fusion unit, and a banding processing unit;
[0124] The brightness feature extraction unit is used to perform scale separation on the brightness field to obtain the low-frequency background component, extract multi-scale brightness change evidence and normalize it;
[0125] The texture feature extraction unit is used to perform scale separation on the brightness field to obtain high-frequency residual components, calculate the relative texture intensity, and construct evidence of texture attenuation.
[0126] The evidence fusion unit is used to fuse brightness change evidence and texture attenuation evidence according to weights and obtain the boundary evidence map through S-shaped mapping.
[0127] The banding processing unit is used to perform Gaussian smoothing banding on the boundary evidence map to generate a banded evidence representation of the thin cloud boundary.
Claims
1. A method for modeling and segmenting the continuous transition of thin clouds in remote sensing images, characterized in that, Includes the following steps: S1. Preprocess the input visible light remote sensing image, convert the multi-channel image into a brightness channel, and perform radiometric normalization and mild smoothing and denoising on the brightness channel to obtain the preprocessed brightness field. S2. The brightness field is scale-separated and decomposed into low-frequency background components and high-frequency residual components by Gaussian filtering. Multi-scale brightness change evidence is extracted and normalized based on the low-frequency background components. Relative texture intensity is calculated and texture attenuation evidence is constructed based on the high-frequency residual components. The brightness change evidence and texture attenuation evidence are weighted and fused and obtained by S-shaped mapping to obtain a boundary evidence map. The boundary evidence map is then subjected to Gaussian smoothing banding to generate a banded evidence representation of the thin cloud boundary. S3. Based on the strip evidence expression, an initial thin cloud cover intensity field with a value range of [0,1] is constructed so that the intensity field shows a monotonic change trend in the high response area of the strip evidence and remains stable in the non-evidence area, where the intensity value 0 represents a clear surface and 1 represents complete thin cloud cover; S4. Construct an energy function that includes a data consistency term, a smoothness constraint term, and a geometric continuity constraint term. Optimize the initial thin cloud cover intensity field with the goal of minimizing this energy function to obtain the final thin cloud cover intensity field. The data consistency term constrains the deviation between the intensity field and the initial estimate, the smoothness constraint term suppresses boundary jaggedness, and the geometric continuity constraint term ensures the connectivity and topological consistency of the boundary. S5. Extract the isomorphic structure or gradient ridge from the final thin cloud coverage intensity field as the thin cloud boundary, determine the region with intensity value in the middle range as the boundary uncertainty region, and output a continuous thin cloud coverage intensity map to complete the thin cloud continuous transition modeling and boundary segmentation.
2. A system for modeling and segmenting continuous transitions of thin clouds in remote sensing imagery, characterized in that, It includes an image preprocessing module, a multi-source evidence construction module, an initial intensity field construction module, an intensity field optimization module, and a result output module; The image preprocessing module is used to preprocess the input visible light remote sensing image, complete the conversion from multi-channel to brightness channel, radiometric normalization and mild smoothing and noise reduction, and output the preprocessed brightness field. The multi-source evidence construction module is used to perform scale separation on the brightness field to obtain low-frequency background components and high-frequency residual components, extract and normalize brightness change evidence, construct texture attenuation evidence, and output a banded evidence expression of thin cloud boundary after weighted fusion and banding processing. The initial intensity field construction module is used to construct and output an initial thin cloud cover intensity field with a value range of [0,1] based on the strip evidence expression. The intensity field optimization module is used to construct an energy function that includes a data consistency term, a smoothness constraint term, and a geometric continuity constraint term, and optimizes the initial thin cloud cover intensity field with the goal of minimizing the energy function, and outputs the final thin cloud cover intensity field. The result output module is used to extract thin cloud boundaries from the final thin cloud coverage intensity field, determine the uncertain regions of the boundaries, and output a continuous thin cloud coverage intensity map to complete the thin cloud continuous transition modeling and boundary segmentation.
3. The method for thin cloud continuity transition modeling and boundary segmentation based on remote sensing imagery according to claim 1, characterized in that: In S2, the multi-source evidence modeling, consisting of brightness change evidence and texture attenuation evidence, addresses the issue that thin clouds in visible light remote sensing images typically appear as a semi-transparent layer covering the surface. The imaging results show slow changes in surface brightness and significant attenuation of high-frequency texture information. The boundary between the thin cloud and the surface is not a typical strong contrast transition, but rather resembles a continuous transition region with a finite width. The determination of the thin cloud boundary is expanded from "strong gradient point detection" to "transition region modeling supported by multi-source evidence." By jointly characterizing brightness change features and texture attenuation features, a multi-source evidence representation of the thin cloud boundary is constructed. The input visible light remote sensing image is: Where x represents the pixel coordinates and Ω is the image domain; if I is a multi-channel image, it is first converted to the luminance channel in a linear or perceptually consistent manner. This is done to eliminate the interference of color differences on subsequent gradient and texture measurements; subsequently... Radiation normalization and mild smoothing are performed to suppress sensing noise and ensure the comparability of feature amplitudes across different scenes. Based on this, a scale separation approach is introduced to decompose the brightness field into low-frequency background components and high-frequency texture components. Through a scale of... Gaussian filtering is used to obtain low-frequency background: , It primarily reflects the slow changing trend of reflection from thin cloud cover and the Earth's surface; it also defines high-frequency residuals: Used to depict the texture and detailed structure of the ground.
4. The method for thin cloud continuity transition modeling and boundary segmentation based on remote sensing imagery according to claim 3, characterized in that: The brightness variation features are used to characterize the gradual spatial position of thin cloud boundaries. However, considering that thin cloud boundaries usually correspond to brightness variations at mid-to-low frequency scales rather than sharp transitions, this paper does not use single-scale gradients as a criterion, but rather uses multiple scales. upper low frequency background Calculate the gradient magnitude: Subsequently, evidence of brightness changes was obtained through scale aggregation: It is then linearly normalized to the interval [0,1]; this operation can adapt to the boundary scale corresponding to thin clouds of different thicknesses, so that G(x) forms a continuous response in the thin cloud-surface transition region, while avoiding over-enhancement in the region with strong surface structure but no thin cloud cover.
5. The method for thin cloud continuity transition modeling and boundary segmentation based on remote sensing imagery according to claim 4, characterized in that: The texture attenuation feature is used to characterize the attenuation effect of thin clouds on high-frequency information of the ground surface; in the high-frequency residual Above, in a local window Calculate texture energy: in and These represent the local variance and mean operators, respectively; to avoid bias caused by inherent texture differences among different land surface types, a larger-scale reference window is introduced. Constructing relative texture intensity: Where ε is a minimal constant to prevent numerical instability; based on this relative metric, a texture attenuation evidence map is defined: And also normalized to [0,1]; in areas covered by thin clouds, the surface texture is uniformly weakened, Larger values will be obtained at bright but textured surface structures.
6. The method for thin cloud continuity transition modeling and boundary segmentation based on remote sensing imagery according to claim 5, characterized in that: Evidence of brightness change Evidence of texture attenuation A multi-source evidence representation for constructing thin cloud boundaries is used in conjunction; to highlight regions that "simultaneously satisfy brightness gradation and texture attenuation" while suppressing false boundaries that are significant only on a single feature, the following fusion method is adopted: Where a, b, and c are feature weights, and τ is the bias term. This represents the sigmoid mapping function; the fusion result Instead of directly representing the single-pixel-level boundary location, it characterizes the "spatial evidence strength of the possible existence of thin cloud boundaries"; to extend the thin cloud boundary from a single response location to a transition region with a finite width, it... Further stripping treatment is required. Through the scale The smoothing operator is obtained in Control the typical width of the transition region. The zonal evidence used as the boundary of thin clouds corresponds to a continuous transition zone between thin clouds and the ground surface in high-response regions, rather than a single strong gradient pixel. This zonal evidence not only provides a stable spatial support domain for subsequent transition zone modeling under geometric continuity constraints, but also naturally characterizes the uncertainty range of the thin cloud boundary, providing an interpretable input basis for subsequent analysis and experimental design.
7. The method for thin cloud continuity transition modeling and boundary segmentation based on remote sensing imagery according to claim 6, characterized in that, The energy function constructed in step S4 is: in For data consistency items, For smoothness constraint terms, For geometric continuity constraints, , These are the weighting coefficients; The data consistency item is: Weight Expressed by strip evidence The decision, in Low values are used for high-response regions, and high values are used for non-evidence regions. The smoothness constraint term is: in Take a smaller value within the strip of evidence and a larger value in non-evidence areas; The geometric continuity constraint term is achieved by introducing curvature smoothing penalties on the contour lines of the thin cloud cover intensity field, performing connectivity analysis on the transition region of the intensity field, and applying penalties or backfill constraints to isolated regions.
8. The method for thin cloud continuity transition modeling and boundary segmentation based on remote sensing imagery according to claim 7, characterized in that, In step S4, minimizing the energy function is the objective of optimizing the initial thin cloud cover intensity field using a numerical iterative method. The iteration stops when the energy function converges or when a preset number of iterations is reached, thus obtaining the final thin cloud cover intensity field. The final thin cloud cover intensity field value is still [0,1]. Its isomorphic structure or gradient ridge is the core position of the thin cloud boundary. The region where the intensity value is in the middle range is the gradual transition region between the thin cloud and the ground surface, that is, the boundary uncertainty region.
9. The method for thin cloud continuity transition modeling and boundary segmentation based on remote sensing imagery according to claim 1, characterized in that, It also includes a step of weakly supervised training of the thin cloud continuous transition modeling and boundary segmentation method: Visible light remote sensing images were used as training samples, and regional binary thin cloud masks were used as weakly supervised annotations. The annotations only indicated thin cloud and non-thin cloud areas and had no precise boundary information. Perform steps S1-S4 on the training samples to obtain the continuous thin cloud cover intensity field of the training samples, and construct the regional consistency loss function: The constraint intensity field converges to 1 in the thin cloud region and to 0 in the non-thin cloud region; the region consistency loss and the geometric regularization term are weighted to form the total loss, and the parameters can be learned by updating the model through backpropagation or numerical optimization; no annotation information is required in the testing phase, and steps S1-S5 are directly executed to complete the thin cloud segmentation.
10. The thin cloud continuous transition modeling and boundary segmentation system for remote sensing imagery according to claim 2, characterized in that, The multi-source evidence construction module includes a brightness feature extraction unit, a texture feature extraction unit, an evidence fusion unit, and a banding processing unit; The brightness feature extraction unit is used to perform scale separation on the brightness field to obtain low-frequency background components, extract multi-scale brightness change evidence and normalize it. The texture feature extraction unit is used to perform scale separation on the brightness field to obtain high-frequency residual components, calculate the relative texture intensity, and construct evidence of texture attenuation. The evidence fusion unit is used to fuse brightness change evidence and texture attenuation evidence according to weights and obtain a boundary evidence map through S-shaped mapping. The banding processing unit is used to perform Gaussian smoothing banding processing on the boundary evidence map to generate a banded evidence representation of the thin cloud boundary.