Polarization prior constraint-based unmanned aerial vehicle rain-day remote sensing image enhancement method and system

Abstract

The application discloses a UAV rain-day remote sensing image enhancement method and system based on polarization prior constraint, belongs to the technical field of remote sensing image processing, and solves the problems of target structure weakening and contrast reduction in low-altitude UAV remote sensing images under rainfall conditions.The method comprises the following steps: organizing UAV polarization remote sensing products after completing attitude normalization and quality screening into input units; constructing rain-day degradation prior, polarization credibility prior and target structure preservation prior according to linear polarization degree, normalized Stokes component, polarization angle consistency and rain curtain scattering index; using the polarization prior to guide the enhancement network, performing contrast improvement, rain strip interference suppression, rain curtain graying alleviation and edge texture recovery on the UAV remote sensing image in the rain day; and evaluating whether the enhanced result preserves target information from the aspects of edge preservation, texture preservation, polarization structure consistency and target and background separability change.The application is suitable for enhancing the UAV remote sensing image in the rain day.

Description

Technical Field

[0001] This application relates to the field of remote sensing image processing technology, and in particular to low-altitude UAV remote sensing image processing, polarization imaging, rainy weather image enhancement, visual perception of complex weather, and remote sensing target information preservation. Background Technology

[0002] Rainfall significantly degrades the imaging quality of outdoor vision systems and UAV remote sensing systems. Traditional research on rainy image processing typically focuses on rain streak removal, raindrop removal, rain-fog mixed degradation recovery, and heavy rain image restoration. Existing research has recognized that real rainfall degradation is not simply the superposition of rain streaks, but encompasses multiple degradation forms, including rain streak occlusion, lens raindrops, rain curtain accumulation, background graying, decreased contrast, edge blurring, and reduced long-distance transmittance. Existing rain removal benchmark studies categorize rainy images into different degradation types, such as rain streaks, raindrops, and rain-fog mixtures, indicating that rainy image restoration requires distinguishing between different degradation mechanisms and cannot rely solely on a single rain streak model.

[0003] In recent years, deep learning-based image deraining and image enhancement methods have developed rapidly. Network structures have evolved from early convolutional networks, residual networks, and multi-scale networks to attention mechanisms, Transformers, diffusion models, and physically guided networks. These methods have achieved good results in deraining ordinary RGB images, but they mainly rely on image intensity, color, texture, and statistical priors, lacking additional physical observation constraints. In real-world UAV rainy remote sensing scenarios, training data is scarce, rain patterns are complex, and flight altitude and oblique viewpoints vary significantly. Ordinary RGB deraining networks are prone to problems such as over-smoothing, pseudo-textures, target edge drift, and decreased target separability after enhancement. Furthermore, there are domain differences between synthetic and real rain data. Existing benchmark studies on real rainy days have also pointed out that deraining models trained on synthetic data still have limitations in generalizing to real-world scenes.

[0004] Image processing of drones in rainy weather is a newly emerging subfield in recent years. Compared with ground cameras, drone platforms have characteristics such as varying flight altitude, changing viewpoint, rapid movement, platform attitude perturbations, small-scale ground objects, and a wide background range. Therefore, rain degradation in drone imagery presents a more complex spatial distribution. Existing benchmarks for drone raindrop removal specifically point out that raindrops adhering to drone lenses can obscure the background, reduce visibility, and affect subsequent image processing tasks. Research in this area has begun to focus on dedicated datasets and algorithm evaluations for drone aerial images under rainy conditions. Further research on the impact of rainfall on drone target detection shows that severe weather and image distortion significantly reduce the performance of vision-based drone detection models; heavy rainfall, noise, and motion blur all lead to decreased detection performance. These studies indicate that drone image enhancement in rainy weather is not only a visual quality issue but also a matter of preserving target information and ensuring the reliability of subsequent perception.

[0005] Existing UAV rain enhancement methods primarily follow the approach of combining rain removal, dehazing, or target detection enhancement with ordinary RGB images. These methods typically learn the mapping from rainy to clear images through a network constrained by image intensity, color, texture, edge detection, and task loss. While these methods can improve visible light image quality to some extent, they do not fully utilize the physical information provided by polarization imaging. Raindrop scattering, rain curtain graying, wet surface reflection, and specular reflection from target surfaces alter the polarization state of light in rainy environments; the degree of polarization, polarization angle, and Stokes components provide information on scattering and reflection differences beyond the light intensity image. Therefore, simple RGB rain removal methods struggle to determine which areas' structures originate from the target itself and which from rain streaks, rain curtains, or scattered background, and also find it difficult to distinguish between reliable and false structures during the enhancement process.

[0006] Polarization imaging has a well-established research foundation in imaging scattering media, dehazing, underwater imaging, material differentiation, and weak target enhancement. By analyzing information such as Stokes parameters, linear polarization degree, polarization angle, and Mueller matrix, polarization imaging provides an additional physical dimension for distinguishing target light from scattered light. Related reviews indicate that polarization imaging can improve image degradation in scattering and absorbing media and has been combined with deep learning, computational imaging, and physical models to form a new trend of polarization-assisted image restoration. In dehazing, underwater, and turbulent media, existing studies have utilized polarization information to estimate scattering background, enhance transmitted targets, or recover Stokes-related information, demonstrating that polarization priors effectively support imaging in complex media.

[0007] However, existing polarization image restoration research mainly focuses on environments with fog, underwater, smoke, or general scattering media, with insufficient attention paid to low-altitude UAV remote sensing images under rainfall conditions. Rain and fog have different media morphologies and image degradation mechanisms: fog typically manifests as relatively uniform small particle scattering and reduced visibility, while rainfall simultaneously includes larger raindrops, directional rain streaks, non-uniform rain curtains, moving occlusion, lens raindrops, localized high-brightness reflections, and changes in reflections from wet surfaces. Therefore, polarization image restoration methods for foggy or underwater conditions cannot be directly transferred to UAV remote sensing image enhancement for rainy weather. Especially in low-altitude grassland remote sensing scenarios, rain degradation is superimposed on changes in surface reflections from wet grasslands, wet bare soil, water accumulation, degraded patches, and sandy patches. Ordinary polarization dehazing models struggle to simultaneously handle the combined effects of rain streaks, rain curtains, low-altitude observation geometry, and wet surface reflections.

[0008] A significant problem with existing deep learning enhancement methods is the lack of "credible prior adjustment" in the enhancement process. Polarization information is not reliable in all regions. Polarization parameters in regions with low linear polarization, overexposed areas, low light intensity areas, areas heavily obscured by rain streaks, and areas with unstable polarization angles are easily affected by noise and imaging errors. If polarization degree, polarization angle, or Stokes components are directly used as network input, the enhancement model may introduce false edges, false textures, or incorrect contrast in low-confidence areas. Therefore, polarization remote sensing image enhancement for UAVs in rainy weather needs to address two issues simultaneously: first, how to utilize polarization information to enhance target structures; and second, how to determine in which regions polarization information is reliable and in which regions it should be weakened or masked.

[0009] Furthermore, existing image enhancement evaluations often focus on visual quality, peak signal-to-noise ratio, structural similarity, or subjective sharpness, while UAV remote sensing pays more attention to whether target information is preserved. For low-altitude remote sensing of grasslands, enhancement results need to serve subsequent analysis of grassland-bare soil boundaries, waterlogged edges, moist bare soil, degraded patches, desertified patches, and small-scale ground features. Therefore, rain-induced enhancement results cannot be evaluated solely on whether they are "brighter and clearer," but also on edge preservation, texture preservation, changes in target-background separability, and the consistency between the enhanced structure and the reliable polarization structure. Existing common de-raining and de-fogging methods rarely include target-background separability as an evaluation criterion for enhancement effects and lack a target information preservation evaluation mechanism that combines prior polarization reliability.

[0010] In summary, existing technologies have the following main shortcomings: First, ordinary RGB rain image enhancement methods lack prior knowledge of polarization physics, making it difficult to distinguish target structures, rain streak interference, and rain curtain graying; Second, existing UAV rain image processing methods mainly focus on raindrop removal, rain removal, or detection performance, lacking enhancement constraints based on Stokes parameters, linear polarization degree, polarization angle consistency, and polarization channel differences; Third, existing polarization restoration methods are mostly geared towards fog, underwater, or general scattering media, and have not established specific methods for preserving rain streaks, rain curtains, attitude changes, and wet grassland surface target information in low-altitude UAV rain remote sensing; Fourth, existing enhancement networks typically lack polarization confidence adjustment mechanisms, posing a risk of false enhancement in low-confidence polarization regions; Fifth, existing evaluation methods emphasize visual quality and lack verification processes for preserving the target structure and texture and the target-background separability changes in UAV remote sensing. Summary of the Invention

[0011] The purpose of this invention is to solve the technical problems of target structure weakening, contrast reduction, rain streak occlusion, rain curtain graying, and ineffective utilization of polarization information in low-altitude UAV remote sensing images under rainfall conditions. It provides a method and system for enhancing UAV remote sensing images in rainy weather with polarization prior constraints.

[0012] The technical solution adopted by this invention to solve the above problems is: a polarization prior constraint-based UAV remote sensing image enhancement method for rainy weather, the method comprising:

[0013] Step 1: Construct rain-enhanced input unit; the original polarization image first undergoes polarization response correction, attitude normalization and quality screening to obtain the total light intensity image, normalized Stokes component, linear polarization degree, polarization angle, polarization quality mask, rain degradation index and observation geometry information under a unified reference coordinate;

[0014] Robust normalization is performed on the total light intensity image, bi-angle sine and cosine coding is used for the polarization angle, and mask constraint is applied to the low-confidence polarization region to form a standardized enhanced input unit.

[0015] Step 2: Construct a polarization prior map; generate rain degradation priors based on rain stripe occlusion index, rain curtain scattering index, and structure sharpness index; generate polarization confidence priors based on linear polarization degree, polarization angle consistency, normalized Stokes stability, and polarization channel differences; generate target structure preservation priors based on light intensity gradient, polarization structure gradient, and polarization confidence; these three types of priors are used to control enhancement intensity, polarization participation weight, and target edge texture protection, respectively.

[0016] Step 3: Perform image enhancement with polarization prior constraints; input the total light intensity image of rainy day, rain degradation prior, polarization confidence prior, and target structure preservation prior into a lightweight enhancement network; the network includes a degradation guidance branch, a polarization confidence guidance branch, a structure preservation branch, and a fusion reconstruction module; the degradation guidance branch extracts degradation features related to rain streaks and rain curtains; the polarization confidence guidance branch extracts reliable polarization auxiliary features; the structure preservation branch extracts target edge and texture features; the fusion reconstruction module outputs the enhancement residual, which is then superimposed with the input total light intensity image to generate the enhancement result;

[0017] Step 4: Perform target information preservation evaluation; calculate edge intensity gain, edge direction consistency, texture preservation index, and polarization structure consistency for the images before and after enhancement; further utilize typical grassland target and background sample pairs to calculate the changes in target and background separability before and after enhancement; the evaluation results include enhanced image, edge preservation results, texture preservation results, polarization structure consistency results, and target and background separability improvement results; this evaluation is used to determine whether the enhancement results preserve and improve target information.

[0018] Further, step 1 includes:

[0019] Constructing a standardized polarization remote sensing data source before enhancement: For the k-th frame of the UAV polarization remote sensing image in rainy weather, the basic input includes the total light intensity image S under a unified reference coordinate system. 0,k Linear polarization component S 1,k and S 2,kThe linear polarization degree DoLPk and polarization angle AoPk calculated from it are also input; simultaneously, the structure sharpness index Q obtained from the preceding rainy day quality assessment method is input. S,k Rain strip shading index Q L,k Rain curtain scattering index Q M,k Polarization reliability index Q P,k and the overall availability index Q U,k The observed geometric information includes the UAV's flight altitude h. k Pixel observation angle ϑ k And an effective observation mask after attitude normalization;

[0020] Constructing the normalized linear polarization component q k and u k :

[0021]

[0022] Where (x, y) represents the image pixel position, and ε is a very small positive number to avoid the denominator from approaching zero.

[0023] Further, in step 1, the method for forming the standardized enhanced input unit includes:

[0024] Total light intensity image S 0,k Robust normalization is performed before entering the enhancement model; the intensity quantile within the effective observation area is used as the normalization benchmark to obtain the enhanced input intensity map I. k :

[0025]

[0026] Among them, P1(S 0,k ) and P 99 (S 0,k ) represent S within the effective observation area respectively 0,k The 1st percentile and the 99th percentile;

[0027] Polarization angle AoP k Represented as a double angle in sine and cosine form:

[0028]

[0029] Among them, A c,k and A s,k A continuous encoding format representing the polarization direction;

[0030] Two types of masks are constructed; the first type is the enhanced effective mask M. E,k The first type is used to mark pixels with basic imaging quality; the second type is a polarization-reliable mask M. P,kUsed to mark regions where polarization information participates in enhanced constraints:

[0031]

[0032]

[0033] Among them, M 0,k τ represents the basic effective mask for the output of pre-processing polarization calibration and attitude normalization. U τ represents the overall availability threshold. P τ represents the polarization confidence threshold. D Indicates the effective threshold for linear polarization;

[0034] Subsequently, polarization intensity, polarization direction, and quality assessment information were organized into standardized input channels for the enhancement model; the polarization-related channels in M P,k Entering the model under constraints:

[0035]

[0036] The polarization direction channel also applies a reliable mask constraint:

[0037] .

[0038] Further, step 2 includes:

[0039] Constructing a priori graph P of degradation in rainy weather D,k :

[0040]

[0041] Where, ω L ω M and ω S The weights are non-negative, and the sum of the three is 1; P D,k The larger (x, y) is, the stronger the impact of rainfall degradation on that pixel or local area;

[0042] Constructing a priori graph of polarization confidence P C,k ;

[0043] Polarization angle consistency is achieved using bi-angular circle statistics within a local window. Internal calculation:

[0044]

[0045] Among them, R A,k The closer (x, y) is to 1, the more stable the local polarization direction; R A,k The lower the (x, y) value, the more significantly the polarization angle is affected by low polarization degree, noise, rain scattering, or local mixed pixels.

[0046] Normalized Stokes stability is determined by The degree of local discretization is given as follows:

[0047]

[0048] Among them, R S,k The larger (x, y) is, the more stable the normalized linear polarization component is in the local region; σ S For stability scale parameters;

[0049] The polarization channel difference diagram is denoted as D. C,k It is calculated from the local structural response differences of the four-directional corrected polarization images; D C,k The larger the value, the stronger the difference in structural response between different polarization channels in this region;

[0050]

[0051] Where, α Q α D α A α S and α C The weights are non-negative; P C,k The larger the value of P, the more suitable the polarization information in that region is as an enhancement prior; C,k The lower the value, the more likely the polarization information in that region is used as a weak reference or does not participate in the constraint.

[0052] Construct the target structure while preserving the prior graph P T,k ;

[0053] The intensity-structured response is given by the normalized total intensity gradient, and the polarization-structured response is given by the normalized Stokes component and the linear polarization gradient.

[0054]

[0055] Among them, G I,k G represents the light intensity structural response. P,k The polarization structure response is represented by ∇, which represents the spatial gradient operator in the image plane, consisting of first-order differences in the horizontal and vertical directions, used to calculate the intensity of changes in the image or polarization parameters in the local neighborhood; the target structure preservation prior is defined as:

[0056]

[0057] in, β represents the normalization operator based on the effective region. I and β P The weights are non-negative, and their sum is 1; the multiplier is 1 − Q.L,k (x, y) is used to suppress the misidentification of rain streak directional structures as target edges; P C,k (x, y) is used to ensure that the polarization structure only participates in target structure preservation within the polarization confidence region; P T,k The larger the value, the more structural details should be preserved in that area during the enhancement process.

[0058] Furthermore, step 3 includes:

[0059] The augmentation network input for the k-th frame image is defined as:

[0060]

[0061] Among them, I k Provide information on total solar intensity during rainy weather that needs improvement, P D,k P indicates the spatial distribution of rain streaks, rain curtains, and structural degradation. C,k Indicating the reliable region of polarization information, P T,k Indicates the target edges and texture structure that should be preserved, M E,k Indicates the effective area for enhancement; the network output is the enhanced total light intensity image. ;

[0062] The enhancement network consists of a degradation-guided branch, a polarization confidence-guided branch, a structure-preserving branch, and a fusion reconstruction module; the degradation-guided branch uses I... k and P D,k As input, extract features such as rain streak occlusion, rain curtain graying, and local low-contrast degradation; polarization confidence guides the branch with I k and P C,k As input, enhance auxiliary features are extracted from reliable polarization prior regions; structure-preserving branches are used with I k and P T,k As input, it extracts target boundary, texture, and local structural features; all three branches employ lightweight convolutional residual structures.

[0063] The characteristics of the three branches are represented as follows:

[0064]

[0065] in, This indicates the degradation-guided feature extraction module. This indicates a polarization confidence-guided feature extraction module. This indicates the structure-preserving feature extraction module; F D, k Primarily used to guide enhancement of areas where rain stripes and rain curtains degrade, F C, k The weights F used to control the participation of polarization priors in enhancement. T, k Used to constrain and preserve target structural information;

[0066] The fusion reconstruction module fuses the three types of branch features according to prior weights; the degenerate region is determined by P. D,k Adjusting the enhancement intensity, the reliable polarization region is determined by P C,k Adjusting the polarization prior contribution, the target structure region is determined by P T,k Adjusting edge and texture weights; the fusion feature is defined as:

[0067]

[0068] Where ⊙ represents pixel-by-pixel weighting, This indicates a feature fusion module;

[0069] The network outputs an enhanced image using residual enhancement; the fused features are used by the reconstruction module to generate an intensity residual ∆I. k The final enhancement result is defined as:

[0070]

[0071] Among them, clip(·) is used to limit the enhancement result to the effective intensity range; M E,k The enhancement is ensured to be applied only to the effective imaging area; the invalid area retains the original input and continues to be marked in the output, thus avoiding the network from incorrectly enhancing low-quality areas.

[0072] Furthermore, the training objectives of the augmentation network in step 3 include:

[0073] During the training phase, data from artificial rainfall platforms and simulated rainfall samples are used to create a clear reference image. The overall loss function is defined as:

[0074]

[0075] in, For strength recovery loss, To maintain loss for the target structure, To preserve fidelity in low-degradation areas, For the consistency loss of the polarization confidence region; λ T , λ I and λ C Non-negative weights;

[0076] The strength recovery loss constraint enhancement result is close to the reference clear image:

[0077]

[0078] Where ||1 represents the L1 norm, which is the sum of the absolute errors of all valid pixels within the parentheses;

[0079] Structure preservation loss in the target structure preservation prior map P T,k The calculation is performed under constraints to protect the target edges and textures.

[0080]

[0081] Low-degradation region fidelity loss is used to suppress unnecessary and spurious enhancements; the low-degradation region consists of 1 − P D,k express:

[0082] .

[0083] Furthermore, step 4 includes:

[0084] Input includes an enhanced image of total light intensity before rain (I). k Enhanced image Enhance the effective mask M E, k Rainy day degradation priori diagram P D,k Polarization confidence prior diagram P C,k Target structure preserves prior graph P T,k and typical targets and background samples of grassland Evaluation is only in M E,k Perform within the marked valid area;

[0085] The edge intensity gain within the target structural region is defined as:

[0086]

[0087] Among them, G E,k This indicates the degree of improvement in the edge strength of the target structural region after enhancement compared to before enhancement;

[0088] Further calculation of gradient direction consistency:

[0089]

[0090] C E,k The higher the value, the closer the enhanced edge position and orientation are to the unenhanced observable structure;

[0091] set up If the fixed bandpass filter operator is used, then the texture preservation index is defined as:

[0092]

[0093] G T,k This metric is used to describe the degree of preservation and enhancement of texture response within a target structural region; it is only used in P... T,kThe reliable region of the constrained target structure is calculated to avoid mistaking rain streaks or noise for real textures.

[0094] Let the polarization structure response be G P,k It consists of normalized Stokes components and the gradient of DoLP; the enhanced image structure response is Polarization structure consistency is defined as:

[0095]

[0096] C P,k The higher the value, the more consistent the enhanced strength structure is with the reliable polarization structure; this index is used to suppress the generation of spurious edges with inconsistent orientations in the polarization reliable structure region of the enhancement network;

[0097] For the m-th target and background sample pair, the enhancement is based on... Extract light intensity / texture features and calculate the separability index. Enhanced based on Extract similar features and calculate the separability index The enhancement of target-context separability is defined as follows:

[0098]

[0099] Where, ∆C m,k > 0 indicates that the separability of the target and background sample pairs in the intensity / texture feature space is improved after enhancement; ∆C m,k = 0 indicates that the enhancement does not change the separability; ∆C m,k < 0 indicates that the enhancement reduces the separability of the sample pairs; this index is used to determine whether the enhancement is truly beneficial to the preservation of typical grassland target information.

[0100] The overall separability improvement of the k-th frame image is defined as:

[0101]

[0102] Among them, w m,k The evaluation weight of the m-th sample pair is determined by the proportion of valid pixels, the quality confidence level, and the sample size; ∆C k Used to measure the improvement in the overall separability of the target and background of an enhanced image;

[0103] Finally, construct the target information to maintain a comprehensive score:

[0104]

[0105] in, Represents the normalization function, γ E γC γ T γ P and γ S The weights are non-negative, and the sum of the weights is 1; Q O,k The higher the value, the better the enhancement results are at maintaining and improving the consistency of target structure, texture, polarization, and the separability between target and background.

[0106] Secondly, the present invention provides a polarization prior constraint UAV rainy weather remote sensing image enhancement system, which has a program module corresponding to the steps of the polarization prior constraint UAV rainy weather remote sensing image enhancement method described above, and executes the steps in the polarization prior constraint UAV rainy weather remote sensing image enhancement method when running.

[0107] Thirdly, the present invention provides a computer device, including a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, it executes the steps of the polarization prior constraint UAV rainy weather remote sensing image enhancement method described above.

[0108] Fourthly, the present invention provides a computer-readable storage medium for storing a computer program that executes a polarization prior constraint-based UAV rainy weather remote sensing image enhancement method as described above.

[0109] The beneficial effects of this invention are:

[0110] This invention addresses polarization-based remote sensing images acquired by low-altitude UAVs under rainfall conditions. It utilizes attitude-normalized Stokes parameters, linear polarization degree, polarization angle consistency, polarization channel differences, rain degradation quality index, and observational geometry to construct rain degradation priors, polarization confidence priors, and target structure preservation priors. Based on these priors, it enhances UAV remote sensing images acquired in rainy conditions. Its direct application aims to improve the contrast, edge sharpness, texture visibility, and target-background separability of low-altitude remote sensing images acquired in rainy conditions, while suppressing rain streak occlusion, rain curtain graying, and pseudo-enhancement in low-confidence polarization regions.

[0111] First, improve the structural visibility of UAV remote sensing images in rainy weather. Enhancement improves the visibility of target edges, local textures, and low-contrast target areas, making grassland-bare soil boundaries, waterlogged edges, degraded patch boundaries, and sandy patch textures easier for subsequent algorithms to utilize.

[0112] Second, reduce the interference of rain streaks and rain curtains on the target structure. By adjusting the enhancement process through rain streak occlusion priors and rain curtain scattering priors, the interference of directional rain streaks on the edge structure is reduced, and the low contrast and texture attenuation caused by rain curtain graying are alleviated.

[0113] Third, reduce the risk of spurious enhancement in low-confidence polarization regions. By using polarization confidence priors to limit the regions where polarization information participates in enhancement, erroneous structures can be avoided in low polarization degrees, unstable polarization angles, and abnormal channel regions.

[0114] Fourth, improve the ability to retain target information. The enhanced effect is not only reflected in subjective clarity, but also in edge direction consistency, texture preservation, polarization structure consistency, and improved target-background separability.

[0115] Fifth, the enhancement results are traceable. While outputting the enhanced image, the priors of rain degradation, polarization confidence, target structure preservation, and effective masking are retained, facilitating subsequent analysis of the source and reliability of the enhancement results.

[0116] This invention does not belong to a simple ordinary image deraining method, nor does it belong to a complete rain field physical inversion method; its technical positioning is to use polarization information as a reliable prior to constrain the process of UAV rainy weather remote sensing image enhancement, so that the enhancement results are more conducive to subsequent grassland moist surface target characterization, ecological anomaly target analysis and complex weather low-altitude remote sensing perception tasks. Attached Figure Description

[0117] To more clearly illustrate the technical solution of this application, the drawings used in the embodiments will be briefly introduced below. Obviously, for those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0118] Figure 1 This is a flowchart illustrating a polarization prior constraint-based UAV remote sensing image enhancement method for rainy weather according to the present invention. Detailed Implementation

[0119] Existing image enhancement methods for common rainy days mainly rely on visible light intensity, color, texture, and deep learning image priors to remove rain or enhance images. While these methods can improve visual clarity to some extent, they struggle to distinguish between real target structures and rain streaks, rain curtains, scattered backgrounds, and noisy pseudo-structures. For low-altitude remote sensing images from UAVs, the target scale is small, the background range is wide, and there are significant variations in flight altitude and observation angle. Conventional enhancement methods are prone to edge drift, overly smoothed textures, local pseudo-enhancements, and decreased target-background separability.

[0120] The algorithm and computer processing flow involved in this invention can accomplish the following tasks: First, the UAV polarization remote sensing products after attitude normalization and quality screening are organized into standardized enhancement input units; second, rain degradation priors, polarization confidence priors, and target structure preservation priors are constructed based on linear polarization degree, normalized Stokes component, polarization angle consistency, polarization channel difference, rain streak occlusion index, and rain curtain scattering index; third, the polarization priors are used to guide the enhancement network to improve contrast, suppress rain streak interference, alleviate rain curtain graying, and restore edge textures in rainy UAV remote sensing images; finally, the enhancement results are evaluated from aspects such as edge preservation, texture preservation, polarization structure consistency, and target-background separability changes to determine whether the enhancement results preserve target information.

[0121] Therefore, the technical problem solved by this invention is not the physical inversion of the complete rain field, nor is it target detection or classification. Instead, it is to establish a UAV rainy weather remote sensing image enhancement process under polarization prior constraints. This process enables constrained image enhancement based on the type of rainfall degradation, the reliability of polarization information, and the location of the target structure, avoiding common problems in ordinary rainy weather enhancement methods such as pseudo-enhancement, over-smoothing, and target structure damage.

[0122] In one specific implementation, a multi-rotor UAV equipped with a focal plane array-split linear polarization camera is used to acquire low-altitude polarization remote sensing images of rainy weather in artificially simulated rainfall fields or controlled outdoor rainfall environments. Experimental scenarios include grasslands, bare soil, moist bare soil, waterlogged areas, degraded patches, sandy patches, and artificially calibrated targets. During UAV flight, flight altitude, attitude, gimbal angle, exposure parameters, observation angle, and rainfall intensity information are recorded simultaneously.

[0123] like Figure 1 As shown in this embodiment, a method for enhancing UAV remote sensing images in rainy weather with polarization prior constraints specifically includes:

[0124] The first step is to construct the rain-enhanced input unit. The original polarization image first undergoes polarization response correction, attitude normalization, and quality screening to obtain the total light intensity image, normalized Stokes component, linear polarization degree, polarization angle, polarization quality mask, rain degradation index, and observation geometry information under a unified reference coordinate system. Robust normalization is performed on the total light intensity image, bi-angle sine and cosine coding is used for the polarization angle, and mask constraints are applied to low-confidence polarization regions to form a standardized enhanced input unit.

[0125] The second step involves constructing a polarization prior map. Rain degradation priors are generated based on the rain streak occlusion index, rain curtain scattering index, and structure sharpness index; polarization confidence priors are generated based on linear polarization degree, polarization angle consistency, normalized Stokes stability, and polarization channel differences; and target structure preservation priors are generated based on light intensity gradient, polarization structure gradient, and polarization confidence. These three types of priors are used to control enhancement intensity, polarization participation weights, and target edge texture protection, respectively.

[0126] The third step involves image enhancement constrained by polarization priors. The total light intensity image under rainy conditions, the rain degradation prior, the polarization confidence prior, and the target structure preservation prior are input into a lightweight enhancement network. The network includes a degradation-guided branch, a polarization confidence-guided branch, a structure preservation branch, and a fusion reconstruction module. The degradation-guided branch extracts degradation features related to rain streaks and rain curtains; the polarization confidence-guided branch extracts reliable polarization auxiliary features; and the structure preservation branch extracts target edge and texture features. The fusion reconstruction module outputs the enhancement residual, which is then superimposed on the input total light intensity image to generate the enhancement result. Invalid regions retain their original input and are still marked, but are not included in the enhancement evaluation.

[0127] The fourth step is to evaluate the preservation of target information. Edge intensity gain, edge orientation consistency, texture preservation index, and polarization structure consistency are calculated for the images before and after enhancement. Furthermore, using typical grassland target-background sample pairs, the change in target-background separability before and after enhancement is calculated. The evaluation results include the enhanced image, edge preservation results, texture preservation results, polarization structure consistency results, and target-background separability improvement results. This evaluation does not output target categories or generate classification maps; it is only used to determine whether the enhancement results have preserved and improved target information.

[0128] In this implementation, the system output includes enhanced UAV remote sensing images in rainy weather, rain-degraded prior maps, polarization confidence prior maps, target structure preservation prior maps, effective region masks, and target information preservation evaluation results. This output is used for subsequent target characterization on wet grassland surfaces, image quality analysis under rainfall conditions, and validation of robust sensing methods.

[0129] In a limited simulation implementation, a rainless UAV polarized remote sensing image is used as the reference image. Training samples are constructed by synthesizing rain streaks, rain curtain graying, and low-contrast degradation. Simultaneously, real rainy day samples are collected using artificial rain fields for validation. During the training phase, the reference clear image is used to constrain strength recovery and structure preservation; during the testing phase, only the rainy day polarized remote sensing image and prior image are input, without using the reference image. The experimental comparison objects include a standard RGB deraining network, an enhancement network without polarization prior, and the method of this invention. Evaluation metrics include structural similarity, edge preservation index, texture preservation index, polarization structure consistency, and target-background separability improvement. This simulation implementation is used to verify the contribution of polarization prior to rainy day UAV remote sensing image enhancement and target information preservation.

[0130] Example:

[0131] This embodiment addresses the problems of rain streak occlusion, rain curtain graying, decreased contrast, edge weakening, and reduced target-background separability in low-altitude UAV polarization remote sensing images under rainfall conditions. It proposes a polarization prior constraint-based UAV rainy-day remote sensing image enhancement method. This method takes the UAV polarization remote sensing product after attitude normalization and quality screening as input, and utilizes the total intensity image, normalized Stokes component, linear polarization degree, polarization angle consistency, polarization channel difference, and rain degradation quality index to construct a polarization prior map and degradation cue map for the enhancement task.

[0132] First, the method extracts rain enhancement input units from standardized polarization remote sensing products, including total light intensity images, polarization intensity information, polarization direction stability information, polarization confidence masks, and quality assessment results such as rain streak occlusion, rain curtain scattering, and structure degradation. Then, based on this information, three types of priors are constructed: first, a rainfall degradation prior, used to indicate the spatial distribution of rain streaks, rain curtains, and local graying; second, a polarization confidence prior, used to limit the effective area for polarization information to participate in enhancement; and third, a target structure preservation prior, used to constrain the edges, textures, and local contrast of grassland features from being over-smoothed or falsely enhanced during the enhancement process.

[0133] In the enhancement phase, the method inputs the total light intensity image of the rainy day and the polarization prior image into the enhancement model. Through degradation type guidance, polarization confidence adjustment, and target structure preservation constraints, it enhances the contrast, suppresses rain streak interference, mitigates rain curtain graying, and restores edge textures in the UAV remote sensing image of the rainy day. This enhancement process does not perform a complete rain field physical inversion. Instead, it uses polarization priors to enhance the target structure in regions with reliable polarization information and reduces the weight of polarization priors in regions with unstable polarization information, thus avoiding misleading image enhancement by low-confidence polarization parameters.

[0134] Finally, the method evaluates the target information preservation of the enhancement results, comparing the structural sharpness, edge preservation, texture retention, and typical target-background separability changes before and after enhancement. It outputs the enhanced image, the structural preservation evaluation results, and the improved target separability results. These results are used to verify the actual contribution of polarization information to UAV rainy-day remote sensing image enhancement, providing a methodological foundation for subsequent research on grassland wet surface target characterization, rainy-day robust perception, and polarization-constrained image restoration.

[0135] Step 1: Construction of Rainy Day Enhanced Input Unit

[0136] The goal of this step is to organize the polarization remote sensing data acquired by low-altitude UAVs under rainfall conditions into standardized augmentation input units, providing a unified, stable, and physically meaningful data foundation for subsequent polarization prior construction and image enhancement. This step does not use raw RGB images as direct input, nor does it directly enhance uncalibrated or unnormalized polarization angle images. The input data must first undergo polarization response correction, attitude normalization, and rain quality screening to ensure that polarization parameters under different flight paths, attitudes, and observation angles are within a unified reference coordinate system.

[0137] For the k-th frame of the UAV polarization remote sensing image in rainy weather, the basic input includes the total light intensity image S under a unified reference coordinate system. 0,k Linear polarization component S 1,k and S 2,k The linear polarization degree DoLPk and polarization angle AoPk calculated from it are also input. Simultaneously, the structure sharpness index Q obtained from the preceding rainy day quality assessment method is input. S,k Rain strip shading index Q L,k Rain curtain scattering index Q M,k Polarization reliability index Q P,k and the overall availability index Q U,k The observed geometric information includes the UAV's flight altitude h. k Pixel observation angle ϑ k And the effective observation mask after attitude normalization. The above data together constitute the normalized polarization remote sensing data source before enhancement.

[0138] To mitigate the impact of light intensity and exposure differences on the polarization component, a normalized linear polarization component q is first constructed. k and u k :

[0139]

[0140] Where (x, y) represents the image pixel position, and ε is a very small positive number to avoid the denominator approaching zero. k and u kThis represents the normalized linear polarization component under a unified reference coordinate system, which is subsequently used to construct polarization priors and target structure preservation constraints.

[0141] Total light intensity image S 0,k Robust normalization is performed before inputting into the enhancement model. The intensity quantile within the effective observation region is used as the normalization benchmark to obtain the enhanced input intensity map I. k :

[0142]

[0143] Among them, P1(S 0,k ) and P 99 (S 0,k ) represent S within the effective observation area respectively 0,k The 1st and 99th percentiles. This processing suppresses the effects of local overexposure, dark areas, and extreme reflectance values ​​on the enhanced input scale, ensuring a consistent input range for rainy images from different flights and under different exposure conditions.

[0144] Polarization angle AoP k Due to its π-periodicity, it cannot be directly input into the enhancement model as a regular angle plot. To avoid angle jumps at the periodic boundaries, AoP... k Represented as a double angle in sine and cosine form:

[0145]

[0146] Among them, A c,k and A s,k This is a continuous encoding format for representing the polarization direction. This representation preserves the polarization angle direction information while avoiding the discontinuity caused by directly using angle values.

[0147] Because rainfall, low light levels, overexposure, and low linear polarization regions can lead to unstable polarization parameters, this step constructs two types of masks. The first type is an enhanced effective mask M. E,k The first type is used to mark pixels with basic imaging quality; the second type is a polarization-reliable mask M. P,k Used to mark regions where polarization information participates in enhanced constraints:

[0148]

[0149]

[0150] Among them, M 0,k τ represents the basic effective mask for the output of pre-processing polarization calibration and attitude normalization. U τ represents the overall availability threshold. P τ represents the polarization confidence threshold. DThis represents the effective threshold for linear polarization. M E,k M controls the effective area for image enhancement P,k Controlling polarization priors participates in the effective region of enhanced constraints.

[0151] Subsequently, polarization intensity, polarization direction, and quality assessment information are organized into normalized input channels for the enhancement model. The polarization-related channels are in M... P,k Under constraints, polarization information in low-confidence regions is not used as an enhancement basis when entering the model.

[0152]

[0153] The polarization direction channel also applies a reliable mask constraint:

[0154]

[0155] This processing ensures that the enhancement model uses polarization intensity and polarization direction information only in regions where polarization parameters are stable, avoiding the misleading effects of low-confidence polarization angles, low signal-to-noise ratio polarization components, and severely degraded regions in rainy weather on the enhancement results.

[0156] Rain degradation index serves as a degradation warning in enhanced input. Structural sharpness index Q. S,k The rain stripe occlusion index Q represents the degree of image structure preservation. L,k The rain curtain scattering index Q represents the intensity of directional rain streak interference. M,k The polarization confidence index Q characterizes the degree of low-frequency graying and contrast compression. P,k The comprehensive availability index Q characterizes the reliability of polarization information. U,k These indices characterize the overall reliability of an image entering the enhancement process. They are not used as the final enhancement result, but rather as inputs for subsequent polarization prior construction and enhancement intensity adjustment.

[0157] Finally, the rain enhancement input unit for the k-th frame image is defined as:

[0158]

[0159] Among them, I k To normalize the total light intensity input, and DoLPk represents the normalized linear polarization component within the reliable region, and is the degree of linear polarization. For polarization direction bi-angle encoding, Q S, k Q L,k Q M,k Q P,k and Q U,k M is the degradation quality index for rainy days. E,k and M P,kThese are the enhanced effective mask and the polarization reliable mask, respectively. k and ϑ k This represents geometric information from low-altitude observations.

[0160] This step has clearly defined technical boundaries: it standardizes and organizes the polarization remote sensing data from UAVs in rainy weather, constrains the reliable polarization region, introduces degradation quality information, and constructs enhanced input units; it does not perform image enhancement, generate polarization prior maps, retrieve rain field scattering parameters, or perform target identification. The output X of this step is... k It is the sole input for subsequent polarization prior construction, degradation type-guided enhancement, and target structure preservation constraints.

[0161] Step 2: Construction of Polarization Prior Map

[0162] The goal of this step is to generate a prior map for image enhancement from the standardized rain enhancement input unit. The prior map does not represent the exact inversion of the rain field scattering components, nor does it represent raindrop size, rain field optical thickness, or wet surface reflection parameters. This step only estimates three spatial cues directly relevant to the enhancement task: rainfall degradation intensity distribution, reliable regions of polarization information, and reliable regions of target structure. These three types of priors are used to control the enhancement intensity, adjust the polarization information weights, and constrain target edge texture preservation, respectively.

[0163] The input to the polarization prior map comes from the enhanced input unit constructed in the first step, including the normalized total intensity image I. k Normalized linear polarization components and Linear polarization degree DoLPk, polarization angle bi-angle encoding Structural clarity index Q S, k Rain strip shading index Q L,k Rain curtain scattering index Q M,k Polarization reliability index Q P,k Overall Availability Index Q U,k Enhance the effectiveness of the mask M E,k and polarization reliable mask M P,k The four-directional corrected polarization intensities retained during the preprocessing stage are used to calculate the polarization channel difference map as auxiliary information for prior construction, and are not directly input into the enhancement model as ordinary raw images.

[0164] First, construct the rain-day degradation prior map P. D,k This prior is used to describe the spatial distribution of rain streak occlusion, rain veil graying, and structure degradation in the image. Rain streak occlusion index Q L,k For directional linear interference, the rain curtain scattering index Q M,k Corresponding to low-frequency graying and contrast compression, the structural sharpness index Q S,kThe inverse vector corresponds to the edge and texture loss. The rain degradation prior is defined as:

[0165]

[0166] Where, ω L ω M and ω S The weights are non-negative, and the sum of the three is 1. P D,k The larger the value of (x, y), the stronger the degradation effect of rainfall on the pixel or local area. This prior is used to adjust the degradation intensity in the subsequent enhancement process: strengthening the suppression of linear interference in rain streak-dominated areas, strengthening the restoration of low-frequency contrast in rain curtain-dominated areas, and strengthening the restoration of edges and textures in structurally degraded areas.

[0167] Then, a polarization confidence prior map P is constructed. C,k This prior is used to determine whether polarization information can reliably participate in strengthening constraints. Polarization reliability is determined by four factors: the validity of linear polarization, the local consistency of polarization angles, the stability of the normalized Stokes component, and the anomalies in the structural responses of different polarization channels. Low DoLP, low quality, discrete polarization angles, and regions with abnormal channel differences are not considered as criteria for strong polarization constraints.

[0168] Polarization angle consistency is achieved using bi-angular circle statistics within a local window. Internal calculation:

[0169]

[0170] Among them, R A,k The closer (x, y) is to 1, the more stable the local polarization direction; R A,k The lower the (x, y) value, the more significantly the polarization angle is affected by low polarization degree, noise, rain scattering, or local mixed pixels.

[0171] Normalized Stokes stability is determined by The degree of local discretization is given as follows:

[0172]

[0173] Among them, R S,k The larger (x, y) is, the more stable the normalized linear polarization component is in the local region; σ S This is the stability scale parameter.

[0174] The polarization channel difference diagram is denoted as D. C,k It is calculated from the local structural response differences of the four-directional corrected polarization images. D C,kThe larger the value, the stronger the difference in structural response between different polarization channels in that region. This difference has structural cue value at the target edge, but it reduces the reliability of polarization constraint in areas with rain streaks or low signal-to-noise ratio. Therefore, the polarization confidence prior uses mass, linear polarization degree, polarization direction consistency, and Stokes stability as positive factors, and anomalous channel differences as suppression factors.

[0175]

[0176] Where, α Q α D α A α S and α C The weights are non-negative. P C,k The larger the value of P, the more suitable the polarization information in that region is as an enhancement prior; C,k The lower the value, the more likely the polarization information in that region is used as a weak reference or does not participate in the constraint.

[0177] Finally, the target structure is constructed while preserving the prior graph P. T,k This prior describes the target edges, textures, and local structural regions that should be prioritized for preservation during enhancement. The goal of UAV-based rainy-weather remote sensing image enhancement is not simply to increase brightness, but to preserve grassland-bare soil boundaries, waterlogged edges, degraded patch boundaries, sandy patch textures, and small-scale target contours. The structure-preserving prior is composed of both the intensity-structure response and the polarization-structure response, and is modulated by rain streak occlusion and polarization confidence.

[0178] The intensity-structured response is given by the normalized total intensity gradient, and the polarization-structured response is given by the normalized Stokes component and the linear polarization gradient.

[0179]

[0180] Among them, G I,k G represents the light intensity structural response. P,k Representing the polarization structure response, the target structure retains the prior definition as:

[0181]

[0182] in, β represents the normalization operator based on the effective region. I and β P The weights are non-negative, and their sum is 1. Multiplier 1 − Q L,k (x, y) is used to suppress the misidentification of rain streak directional structures as target edges; P C,k (x, y) is used to ensure that the polarization structure only participates in target structure preservation within the polarization confidence region. T,kThe larger the value, the more structural details should be preserved in that area during the enhancement process.

[0183] Finally, the set of polarization prior maps is defined as:

[0184]

[0185] Among them, P D,k For the rainy day degradation prior diagram, P C,k For the polarization confidence prior diagram, P T,k The prior map is preserved for the target structure. These three elements respectively serve to control degradation intensity, adjust polarization prior weights, and constrain the target structure in the subsequent enhancement model.

[0186] The technical boundaries of this step are defined: it only generates the spatial priors required for enhancement, does not output the enhanced image, does not estimate raindrop size, does not invert the rain field scattering components, and does not separate the target reflected light from the background scattered light. (Through P) D,k P C,k and P T,k The augmentation model obtains interpretable input cues, thus avoiding the problems of polarization confidence and target structure preservation that are ignored by ordinary rainy day image augmentation methods.

[0187] Step 3: Polarization Prior Constraint-Based UAV Rainy Weather Remote Sensing Image Enhancement

[0188] The goal of this step is to enhance low-altitude UAV remote sensing images under rainfall conditions, constrained by the polarization prior map generated in the second step. This enhances the contrast, edge sharpness, and texture visibility of the target area, while suppressing rain streak occlusion, rain curtain graying, and false enhancements in low-confidence areas. This step employs a lightweight polarization prior-guided enhancement network as the determining factor, without using a complete rainfield physical inversion model, and does not estimate raindrop size, rainfield optical thickness, or wet surface reflection parameters.

[0189] The enhancement target is the normalized total light intensity image I obtained in the first step. k The second step yielded the rain-induced degradation prior map P. D,k Polarization confidence prior diagram P C,k Maintain the prior graph P with the target structure T,k As the input for the augmentation constraints, the input to the augmentation network for the k-th frame image is defined as:

[0190]

[0191] Among them, I k Provide information on total solar intensity during rainy weather that needs improvement, P D,k P indicates the spatial distribution of rain streaks, rain curtains, and structural degradation. C,k Indicating the reliable region of polarization information, P T,kIndicates the target edges and texture structure that should be preserved, M E,k Indicates the effective area for enhancement. The network output is the enhanced total light intensity image. .

[0192] The enhancement network consists of a degradation-guided branch, a polarization confidence-guided branch, a structure-preserving branch, and a fusion reconstruction module. The degradation-guided branch uses I... k and P D,k As input, extract features such as rain streak occlusion, rain curtain graying, and local low-contrast degradation; polarization confidence guides the branch with I k and P C,k As input, enhance auxiliary features are extracted from reliable polarization prior regions; structure-preserving branches are used with I k and P T,k As input, it extracts target boundary, texture, and local structural features. All three branches employ lightweight convolutional residual structures, avoiding the introduction of large-scale semantic recognition networks and thus preventing the image enhancement problem from being extended into a full-fledged object detection or classification task.

[0193] The characteristics of the three branches are represented as follows:

[0194]

[0195] in, This indicates the degradation-guided feature extraction module. This indicates a polarization confidence-guided feature extraction module. This represents the structure-preserving feature extraction module. F D, k Primarily used to guide enhancement of areas where rain stripes and rain curtains degrade, F C, k The weights F used to control the participation of polarization priors in enhancement. T, k Used to constrain the preservation of target structural information.

[0196] The fusion and reconstruction module fuses the features of the three branches according to prior weights. The degenerate region is determined by P. D,k Adjusting the enhancement intensity, the reliable polarization region is determined by P C,k Adjusting the polarization prior contribution, the target structure region is determined by P T,k Adjust edge and texture weights. The fusion feature is defined as:

[0197]

[0198] Where ⊙ represents pixel-by-pixel weighting, This represents the feature fusion module. This design ensures that the network does not indiscriminately enhance the entire image, but rather performs spatially adaptive enhancement based on rainfall degradation intensity, polarization confidence, and target structure reliability.

[0199] The network outputs an enhanced image using residual enhancement. The fused features are processed by the reconstruction module to generate an intensity residual ∆I. k The final enhancement result is defined as:

[0200]

[0201] The clip(·) function is used to limit the enhancement result to the effective intensity range. E,k The enhancement is ensured to be applied only to the effective imaging area; the invalid area retains the original input and continues to be marked in the output, thus avoiding the network from incorrectly enhancing low-quality areas.

[0202] The training objectives of the augmentation network consist of intensity recovery, structure preservation, fidelity preservation of low-degradation regions, and polarization prior consistency. During the training phase, reference clear images are generated using data from an artificial rainfall platform and simulated rainfall samples. The overall loss function is defined as:

[0203]

[0204] in, For strength recovery loss, To maintain loss for the target structure, To preserve fidelity in low-degradation areas, λ represents the consistency loss within the polarization confidence region. T , λ I and λ C The weights are non-negative.

[0205] The strength recovery loss constraint enhancement result is close to the reference clear image:

[0206]

[0207] Structure preservation loss in the target structure preservation prior map P T,k The calculation is performed under constraints to protect the target edges and textures.

[0208]

[0209] Low-degradation region fidelity loss is used to suppress unnecessary and spurious enhancements. The low-degradation region consists of 1 − P. D,k express:

[0210]

[0211] This requirement mandates that low-degradation areas maintain their original brightness and texture, preventing the network from creating false edges or excessive contrast in areas where degradation is not obvious in rainy weather.

[0212] The polarization confidence region consistency loss is used to ensure that the enhanced structural changes are consistent with the target structural cues in the polarization confidence region. This term only applies to P. C,k and P T,k At the same time, higher-level areas are affected:

[0213]

[0214] The purpose of this constraint is not to keep the enhancement process completely unchanged, but rather to suppress non-physical shifts in the edge orientation and position of the polarization-confidential target structure region. Combined with... Subsequently, the network restored the degradation caused by rain while maintaining the spatial consistency of grassland target boundaries, water accumulation edges, bare soil textures, and degraded patch outlines.

[0215] During the inference phase, no clear reference image is used; only the Zk and prior graph sets are input. The network is based on P. D, k Controlling degradation and enhancing strength, based on P C,k Controlling the degree of involvement of polarization priors, based on P T,k The system maintains the target structure. Stronger linear interference suppression is applied to areas with strong rain streak occlusion, stronger low-frequency contrast recovery is applied to areas with strong rain curtain graying, polarization prior weights are reduced in areas with low polarization confidence, and edge and texture preservation is enhanced in structurally reliable areas.

[0216] The output of this step is an enhanced image. And retain the corresponding effective mask M E, k Polarization confidence prior P C,k Target structure maintains prior P T,k and degenerate prior P D,k The output proceeds to the next step, target information preservation evaluation. The technical boundaries of this step are clearly defined: it performs image enhancement under polarization prior constraints, does not perform complete rain field physical inversion, does not separate target light from scattered light, does not output target recognition results, and does not estimate grassland ecological parameters.

[0217] Step 4: Evaluation of Target Information Maintenance

[0218] The goal of this step is to evaluate whether the enhanced image truly preserves and improves target information in UAV rainy remote sensing images. The evaluation object is not simply visual sharpness or target detection accuracy, but rather the changes in target structure, texture details, target-background separability, and polarization prior consistency before and after enhancement. This step does not output a target recognition category, generate a classification map, or claim to have completed automatic grassland target recognition; its purpose is to verify whether polarization prior constraint enhancement is beneficial to subsequent target representation and separability analysis.

[0219] Input includes an enhanced image of total light intensity before rain (I). kEnhanced image Enhance the effectiveness of the mask M E, k Rainy day degradation priori diagram P D,k Polarization confidence prior diagram P C,k Target structure preserves prior graph P T,k and typical targets and background samples of grassland Preferably, it is constructed from the related patent filed by the applicant on the same day, which describes a method for extracting UAV polarization remote sensing features and assessing the separability of targets and backgrounds for wet grassland surfaces. Evaluations are only available on M. E,k Evaluation is conducted within the marked effective area; severely invalid areas, low-quality areas, and areas not involved in enhancement are excluded from the target information preservation evaluation.

[0220] First, the edge preservation and edge enhancement effects are calculated. Edge evaluation of an enhanced image cannot solely focus on increasing the gradient, as over-enhancement can produce false edges. Therefore, this method simultaneously evaluates edge intensity enhancement and edge orientation consistency. The edge intensity gain within the target structural region is defined as:

[0221]

[0222] Among them, G E,k This indicates the degree of improvement in the edge strength of the target structural region after enhancement compared to before enhancement. Only G... E,k This is insufficient to demonstrate that the edges remain reliable, therefore we further calculate the consistency of gradient directions:

[0223]

[0224] C E,k The higher the value, the closer the enhanced edge location and orientation are to the unenhanced observable structure; G E,k With C E,k A joint evaluation is conducted to determine whether the enhancement improves edge strength while avoiding edge drift and pseudo-structure generation.

[0225] Next, the texture preservation index is calculated. Grassland wet surfaces, bare soil, sandy patches, and degraded patches are often characterized by fine-grained texture differences; rain-induced graying and excessive smoothing weaken this information. This method uses bandpass texture response to evaluate the degree of texture preservation before and after enhancement. Let... If the fixed bandpass filter operator is used, then the texture preservation index is defined as:

[0226]

[0227] G T,k This metric describes the degree of preservation and enhancement of texture response within a target structural region. It is only used in P... T,kThe target structure is constrained and the reliable region is calculated to avoid mistaking rain streaks or noise for real textures.

[0228] Then, the polarization consistency metric is calculated. Here, polarization consistency does not require the augmented image to change or restore polarization parameters, but rather evaluates whether the augmented intensity structure maintains spatial consistency with the believable polarization structure. The augmented output is an intensity image; polarization parameters are not included in the network output. Polarization consistency is only used to verify whether the augmented edges and textures follow the structural cues within the believable polarization region. Let the polarization structure response be G. P,k It consists of normalized Stokes components and the gradient of DoLP; the enhanced image structure response is The polarization structure consistency is defined as:

[0229]

[0230] C P,k A higher value indicates a closer alignment between the enhanced intensity structure and the reliable polarization structure. This metric is used to suppress spurious edges with inconsistent orientations generated by the enhancement network in the reliable polarization structure region.

[0231] Next, the target-background separability change is calculated. This evaluation assesses the target-background sample pairs. However, it only compares the changes in light intensity / texture separability in the images before and after enhancement, without outputting the recognition result. For the m-th target-background sample pair, the pre-enhancement data is based on... Extract light intensity / texture features and calculate the separability index. Enhanced based on Extract similar features and calculate the separability index The enhanced target-context separability is defined as...

[0232]

[0233] Where, ∆C m,k > 0 indicates that the separability of the target-background sample pair in the intensity / texture feature space is improved after enhancement; ∆C m,k = 0 indicates that the enhancement does not change the separability; ∆C m,k A value less than 0 indicates that the enhancement reduced the separability of the sample pairs. This index is used to determine whether the enhancement truly helps preserve the information of typical grassland targets.

[0234] To avoid the dominance of a single sample pair in the evaluation results, all target-background sample pairs are weighted and aggregated. The overall separability improvement of the k-th frame image is defined as:

[0235]

[0236] Among them, w m,k∆C represents the evaluation weight of the m-th sample pair, which is determined by the proportion of valid pixels in the sample, the quality confidence score, and the sample size. k Used to measure the improvement of image enhancement on the overall object-background separability.

[0237] Finally, a comprehensive score for object information preservation is constructed. This score is composed of edge intensity gain, edge orientation consistency, texture preservation, polarization structure consistency, and object-background separability enhancement.

[0238]

[0239] in, Represents the normalization function, γ E γ C γ T γ P and γ S The weights are non-negative, and the sum of the weights is 1. Q O,k The higher the value, the better the enhancement results are at maintaining and improving the target structure, texture, polarization consistency, and target-background separability.

[0240] The output of this step is:

[0241]

[0242] Among them, V k The result represents the target information preservation evaluation of the augmentation result in the k-th frame. This result is used to verify whether the polarization prior constraint augmentation method improves the target structure visibility and target-background separability in rainy UAV remote sensing images.

[0243] The technical boundaries of this step are clearly defined: it evaluates the degree to which the enhancement result preserves target information, without performing target identification, outputting target categories, calculating detection accuracy, or retrieving rain field physical parameters. Its purpose is to demonstrate that the enhancement method does not simply increase visual brightness or contrast, but rather preserves and enhances typical grassland target information at the levels of edge, texture, polarization structure consistency, and target-background separability.

[0244] This application provides a polarization prior constraint-based UAV rainy weather remote sensing image enhancement method. This method takes the UAV polarization remote sensing product after attitude normalization and quality screening as input, and constructs rainy weather degradation priors, polarization confidence priors, and target structure preservation priors using linear polarization degree, polarization angle consistency, normalized Stokes component, polarization channel difference, rain streak occlusion index, and rain curtain scattering index. During enhancement, the enhancement intensity is adjusted according to the degradation type, the polarization prior weights are adjusted according to the polarization confidence, and edges and textures are preserved based on the target structure prior. Finally, the enhancement effect is evaluated through edge preservation, texture preservation, polarization structure consistency, and target-background separability changes. This technology overcomes the shortcomings of existing ordinary rainy weather enhancement methods, such as the lack of polarization constraints, the incompatibility of existing polarization restoration methods with low-altitude UAV rainy weather remote sensing, and the lack of target information preservation verification in existing enhancement evaluations. Specifically:

[0245] First, this application provides a method for constructing a "standardized polarization enhancement input unit". This input unit is not the original RGB image, but rather consists of the total light intensity image after pose normalization and quality screening, normalized Stokes components, linear polarization degree, polarization angle biangular coding, polarization quality mask, rain degradation index, and observation geometry information. This is used to distinguish it from ordinary image deraining methods.

[0246] Second, this application provides a construction mechanism for "rain degradation prior, polarization reliability prior, and target structure preservation prior". The rain degradation prior is used to indicate the intensity of rain streak occlusion, rain curtain graying, and structure degradation; the polarization reliability prior is used to determine whether polarization information can reliably participate in enhancement; and the target structure preservation prior is used to indicate the areas where grassland feature edges, textures, and local structures need to be protected.

[0247] Third, this application provides a "polarization confidence adjustment mechanism." Instead of inputting all polarization parameters indiscriminately into the enhancement model, this application determines the confidence level of polarization information based on linear polarization degree, polarization angle consistency, Stokes stability, and polarization channel differences. The weight of polarization priors is reduced in low-confidence regions, while the effect of polarization priors is enhanced in high-confidence regions. This prevents low-quality polarization regions from misleading the enhancement results.

[0248] Fourth, this application provides a "degradation type-aware enhancement network structure". The enhancement network includes a degradation guidance branch, a polarization confidence guidance branch, and a structure preservation branch. The degradation guidance branch handles rain streaks and rain curtain degradation, the polarization confidence guidance branch controls the degree of polarization prior involvement, and the structure preservation branch protects target edges, textures, and local structures. This structure differs from ordinary deraining networks with single input and single output.

[0249] Fifth, this application provides a "spatial adaptive enhancement mechanism based on prior images". This application adjusts the enhancement intensity based on the prior of rain degradation, adjusts the polarization feature weights based on the prior of polarization confidence, and adjusts the edge texture preservation intensity based on the prior of target structure preservation, so that different enhancement strategies are applied to different regions, rather than uniformly enhancing the entire image.

[0250] Sixth, this application provides a "target information preservation evaluation mechanism." This application evaluates not only the visual quality of the enhanced image, but also edge preservation, texture preservation, polarization structure consistency, and changes in target-background separability. This point serves to demonstrate that the enhancement results serve to preserve target information in UAV remote sensing, rather than simply improving subjective visual effects.

[0251] Compared with ordinary RGB rainy day image enhancement methods, this application introduces polarization prior constraints, which can provide additional physical information for the enhancement process by utilizing linear polarization degree, normalized Stokes component and polarization direction consistency, reducing spurious enhancement and edge misrecovery caused by relying solely on light intensity texture.

[0252] Compared to traditional rain removal methods, this application does not treat all rain-related degradation uniformly. Instead, it differentiates between rain streak occlusion, rain curtain graying, and structural degradation areas through rain-related degradation priors, thus creating spatial differences in enhancement intensity. Rain streak-dominant areas are targeted for suppressing linear interference, rain curtain graying areas are targeted for restoring low-frequency contrast, and target structural areas are targeted for protecting edges and textures.

[0253] Compared to existing polarization dehazing or underwater polarization restoration methods, this application does not require the establishment of a complete Stokes-Mueller physical inversion model, nor does it estimate raindrop size, optical thickness, or scattering components. Instead, it transforms polarization parameters into enhancement priors for preserving target information in UAV rainy remote sensing images. This approach reduces engineering complexity while retaining the constraining value of polarization information for the enhancement process.

[0254] Compared with conventional UAV rain enhancement or target detection preprocessing methods, this application adds a polarization confidence adjustment mechanism to avoid the polarization information of low DoLP, low signal-to-noise ratio, unstable polarization angle and strong rain strip occlusion areas from misleading the enhancement network, thereby reducing the risk of false textures, false edges and over-enhancement in low confidence areas.

Claims

1. A method for enhancing UAV remote sensing images in rainy weather with polarization prior constraints, characterized in that, The method includes: Step 1: Construct rain-enhanced input unit; the original polarization image first undergoes polarization response correction, attitude normalization and quality screening to obtain the total light intensity image, normalized Stokes component, linear polarization degree, polarization angle, polarization quality mask, rain degradation index and observation geometry information under a unified reference coordinate; Robust normalization is performed on the total light intensity image, bi-angle sine and cosine coding is used for the polarization angle, and mask constraint is applied to the low-confidence polarization region to form a standardized enhanced input unit. Step 2: Construct a polarization prior map; generate rain degradation priors based on rain stripe occlusion index, rain curtain scattering index, and structure sharpness index; generate polarization confidence priors based on linear polarization degree, polarization angle consistency, normalized Stokes stability, and polarization channel differences; generate target structure preservation priors based on light intensity gradient, polarization structure gradient, and polarization confidence; these three types of priors are used to control enhancement intensity, polarization participation weight, and target edge texture protection, respectively. Step 3: Perform image enhancement with polarization prior constraints; input the total light intensity image of rainy day, rain degradation prior, polarization confidence prior, and target structure preservation prior into a lightweight enhancement network; the network includes a degradation guidance branch, a polarization confidence guidance branch, a structure preservation branch, and a fusion reconstruction module; the degradation guidance branch extracts degradation features related to rain streaks and rain curtains; the polarization confidence guidance branch extracts reliable polarization auxiliary features; the structure preservation branch extracts target edge and texture features; the fusion reconstruction module outputs the enhancement residual, which is then superimposed with the input total light intensity image to generate the enhancement result; Step 4: Perform target information preservation evaluation; calculate edge intensity gain, edge direction consistency, texture preservation index, and polarization structure consistency for the images before and after enhancement; further utilize typical grassland target and background sample pairs to calculate the changes in target and background separability before and after enhancement; the evaluation results include enhanced image, edge preservation results, texture preservation results, polarization structure consistency results, and target and background separability improvement results; this evaluation is used to determine whether the enhancement results preserve and improve target information.

2. The method for enhancing UAV remote sensing images in rainy weather with polarization prior constraints according to claim 1, characterized in that: Step 1 includes: Constructing a standardized polarization remote sensing data source before enhancement: For the k-th frame of the UAV polarization remote sensing image in rainy weather, the basic input includes the total light intensity image S under a unified reference coordinate system. 0,k Linear polarization component S 1,k and S 2,k The linear polarization degree DoLPk and polarization angle AoPk calculated from it are also input; simultaneously, the structure sharpness index Q obtained from the preceding rainy day quality assessment method is input. S,k Rain strip shading index Q L,k Rain curtain scattering index Q M,k Polarization reliability index Q P,k and the overall availability index Q U,k The observed geometric information includes the UAV's flight altitude h. k Pixel observation angle ϑ k And an effective observation mask after attitude normalization; Constructing the normalized linear polarization component q k and u k : Where (x, y) represents the image pixel position, and ε is a very small positive number to avoid the denominator from approaching zero.

3. The method for enhancing UAV remote sensing images in rainy weather with polarization prior constraints according to claim 2, characterized in that: In step 1, the method for forming the standardized enhanced input unit includes: Total light intensity image S 0,k Robust normalization is performed before entering the enhancement model; the intensity quantile within the effective observation area is used as the normalization benchmark to obtain the enhanced input intensity map I. k : Among them, P1(S 0,k ) and P 99 (S 0,k ) represent S within the effective observation area respectively 0,k The 1st percentile and the 99th percentile; Polarization angle AoP k Represented as a double angle in sine and cosine form: Among them, A c,k and A s,k A continuous encoding format representing the polarization direction; Two types of masks are constructed; the first type is the enhanced effective mask M. E,k The first type is used to mark pixels with basic imaging quality; the second type is a polarization-reliable mask M. P,k Used to mark regions where polarization information participates in enhanced constraints: Among them, M 0,k τ represents the basic effective mask for the output of pre-processing polarization calibration and attitude normalization. U τ represents the overall availability threshold. P τ represents the polarization confidence threshold. D Indicates the effective threshold for linear polarization; Subsequently, polarization intensity, polarization direction, and quality assessment information were organized into standardized input channels for the enhancement model; the polarization-related channels in M P,k Entering the model under constraints: The polarization direction channel also applies a reliable mask constraint: 。 4. The method for enhancing UAV remote sensing images in rainy weather with polarization prior constraints according to claim 1, characterized in that: Step 2 includes: Constructing a priori graph P of degradation in rainy weather D,k : Where, ω L ω M and ω S The weights are non-negative, and the sum of the three is 1; P D,k The larger (x, y) is, the stronger the impact of rainfall degradation on that pixel or local area; Constructing a priori graph of polarization confidence P C,k ; Polarization angle consistency is achieved using bi-angular circle statistics within a local window. Internal calculation: Among them, R A,k The closer (x, y) is to 1, the more stable the local polarization direction; R A,k The lower the (x, y) value, the more significantly the polarization angle is affected by low polarization degree, noise, rain scattering, or local mixed pixels. Normalized Stokes stability is determined by The degree of local discretization is given as follows: Among them, R S,k The larger (x, y) is, the more stable the normalized linear polarization component is in the local region; σ S For stability scale parameters; The polarization channel difference diagram is denoted as D. C,k It is calculated from the local structural response differences of the four-directional corrected polarization images; D C,k The larger the value, the stronger the difference in structural response between different polarization channels in this region; Where, α Q α D α A α S and α C The weights are non-negative; P C,k The larger the value of P, the more suitable the polarization information in that region is as an enhancement prior; C,k The lower the value, the more likely the polarization information in that region is used as a weak reference or does not participate in the constraint. Construct the target structure while preserving the prior graph P T,k ; The intensity-structured response is given by the normalized total intensity gradient, and the polarization-structured response is given by the normalized Stokes component and the linear polarization gradient. Among them, G I,k G represents the light intensity structural response. P,k The polarization structure response is represented by ∇, which represents the spatial gradient operator in the image plane, consisting of first-order differences in the horizontal and vertical directions, used to calculate the intensity of changes in the image or polarization parameters in the local neighborhood; the target structure preservation prior is defined as: in, β represents the normalization operator based on the effective region. I and β P The weights are non-negative, and their sum is 1; the multiplier is 1 − Q. L,k (x, y) is used to suppress the misidentification of rain streak directional structures as target edges; P C,k (x, y) is used to ensure that the polarization structure only participates in target structure preservation within the polarization confidence region; P T,k The larger the value, the more structural details should be preserved in that area during the enhancement process.

5. The method for enhancing UAV remote sensing images in rainy weather with polarization prior constraints according to claim 1, characterized in that: Step 3 includes: The augmentation network input for the k-th frame image is defined as: Among them, I k Provide information on total solar intensity during rainy weather that needs improvement, P D,k P indicates the spatial distribution of rain streaks, rain curtains, and structural degradation. C,k Indicating the reliable region of polarization information, P T,k Indicates the target edges and texture structure that should be preserved, M E,k Indicates the effective area for enhancement; the network output is the enhanced total light intensity image. ; The enhancement network consists of a degradation-guided branch, a polarization confidence-guided branch, a structure-preserving branch, and a fusion reconstruction module; the degradation-guided branch uses I... k and P D,k As input, extract features such as rain streak occlusion, rain curtain graying, and local low-contrast degradation; polarization confidence guides the branch with I k and P C,k As input, enhance auxiliary features are extracted from reliable polarization prior regions; structure-preserving branches are used with I k and P T,k As input, it extracts target boundary, texture, and local structural features; all three branches employ lightweight convolutional residual structures. The characteristics of the three branches are represented as follows: in, This indicates the degradation-guided feature extraction module. This indicates a polarization confidence-guided feature extraction module. F represents the structure-preserving feature extraction module; D, k Primarily used to guide enhancement of areas where rain stripes and rain curtains degrade, F C, k The weights F used to control the participation of polarization priors in enhancement. T, k Used to constrain and preserve target structural information; The fusion reconstruction module fuses the three types of branch features according to prior weights; the degenerate region is determined by P. D,k Adjusting the enhancement intensity, the reliable polarization region is determined by P C,k Adjusting the polarization prior contribution, the target structure region is determined by P T,k Adjusting edge and texture weights; the fusion feature is defined as: Where ⊙ represents pixel-by-pixel weighting, This indicates a feature fusion module; The network outputs an enhanced image using residual enhancement; the fused features are used by the reconstruction module to generate an intensity residual ∆I. k The final enhancement result is defined as: Among them, clip(·) is used to limit the enhancement result to the effective intensity range; M E,k The enhancement is ensured to be applied only to the effective imaging area; the invalid area retains the original input and continues to be marked in the output, thus avoiding the network from incorrectly enhancing low-quality areas.

6. The method for enhancing UAV remote sensing images in rainy weather with polarization prior constraints according to claim 1, characterized in that: The training objectives for the augmented network in step 3 include: During the training phase, data from artificial rainfall platforms and simulated rainfall samples are used to create a clear reference image. The overall loss function is defined as: in, For strength recovery loss, To maintain loss for the target structure, To preserve fidelity in low-degradation areas, For the consistency loss of the polarization confidence region; λ T , λ I and λ C Non-negative weights; The strength recovery loss constraint enhancement result is close to the reference clear image: Where ||1 represents the L1 norm, which is the sum of the absolute errors of all valid pixels within the parentheses; Structure preservation loss in the target structure preservation prior map P T,k The calculation is performed under constraints to protect the target edges and textures. Low-degradation region fidelity loss is used to suppress unnecessary and spurious enhancements; the low-degradation region consists of 1 − P D,k express: 。 7. The method for enhancing UAV remote sensing images in rainy weather with polarization prior constraints according to claim 1, characterized in that: Step 4 includes: Input includes an enhanced image of total light intensity before rain (I). k Enhanced image Enhance the effective mask M E, k Rainy day degradation priori diagram P D,k Polarization confidence prior diagram P C,k Target structure preserves prior graph P T,k and typical targets and background samples of grassland Evaluation is only available in M. E,k Perform within the marked valid area; The edge intensity gain within the target structural region is defined as: Among them, G E,k This indicates the degree of improvement in the edge strength of the target structural region after enhancement compared to before enhancement; Further calculation of gradient direction consistency: C E,k The higher the value, the closer the enhanced edge position and orientation are to the unenhanced observable structure; set up If the fixed bandpass filter operator is used, then the texture preservation index is defined as: G T,k This metric is used to describe the degree of preservation and enhancement of texture response within a target structural region; it is only used in P... T,k The reliable region of the constrained target structure is calculated to avoid mistaking rain streaks or noise for real textures. Let the polarization structure response be G. P,k It consists of normalized Stokes components and the gradient of DoLP; the enhanced image structure response is Polarization structure consistency is defined as: C P,k The higher the value, the more consistent the enhanced strength structure is with the reliable polarization structure; this index is used to suppress the generation of spurious edges with inconsistent orientations in the polarization reliable structure region of the enhancement network; For the m-th target and background sample pair, the enhancement is based on... Extract light intensity / texture features and calculate the separability index. Enhanced based on Extract similar features and calculate the separability index The enhancement of target-context separability is defined as follows: Where, ∆C m,k > 0 indicates that the separability of the target and background sample pairs in the intensity / texture feature space is improved after enhancement; ∆C m,k = 0 indicates that the enhancement does not change the separability; ∆C m,k < 0 indicates that the enhancement reduces the separability of the sample pairs; this index is used to determine whether the enhancement is truly beneficial to the preservation of typical grassland target information. The overall separability improvement of the k-th frame image is defined as: Among them, w m,k The evaluation weight of the m-th sample pair is determined by the proportion of valid pixels, the quality confidence level, and the sample size; ∆C k Used to measure the improvement in the overall separability of the target and background of an enhanced image; Finally, construct the target information to maintain a comprehensive score: in, Represents the normalization function, γ E γ C γ T γ P and γ S The weights are non-negative, and the sum of the weights is 1; Q O,k The higher the value, the better the enhancement results are at maintaining and improving the consistency of target structure, texture, polarization, and the separability between the target and the background.

8. A polarization prior-constrained UAV remote sensing image enhancement system for rainy weather, characterized in that: The system has a program module corresponding to the steps of any one of claims 1-7, and executes the steps in the UAV rainy weather remote sensing image enhancement method with polarization prior constraints when running.

9. A computer device, characterized in that: It includes a memory and a processor, wherein the memory stores a computer program, and when the processor runs the computer program stored in the memory, it performs the steps of the polarization prior constraint UAV rainy weather remote sensing image enhancement method according to any one of claims 1-7.

10. A computer-readable storage medium, characterized in that: The storage medium is used to store a computer program that executes a polarization prior constraint method for enhancing UAV remote sensing images in rainy weather, according to any one of claims 1-7.

Images