Remote sensing image enhancement method based on radiation credible partition and surface feature spectrum sequence keeping

Abstract

The present application belongs to the technical field of remote sensing image enhancement, and particularly relates to a remote sensing image enhancement method based on radiation credible partition and ground object spectrum sequence maintenance, steps of which comprise: acquiring a multi-channel low-light remote sensing image and performing normalization, constructing an over-dark weight map, a radiation credible weight map and a high-light protection weight map; dividing a ground object area to obtain an anchor point broadcast map; obtaining an original gain parameter based on a radiation correction model and acquiring a shared gain field acting on each band of the same pixel; acquiring an initial enhanced image and obtaining an enhanced remote sensing image; constructing a combined constraint function by using a pair of low-light remote sensing images and normal-illumination reference images to train the radiation correction model; and after the training is completed, outputting the enhanced remote sensing image based on an actual low-light remote sensing image. The present application can adjust the enhancement intensity according to the radiation credibility of the current image itself, suppress over-enhancement in the high-light area, and maintain the stable band sequence inside the ground object area while improving the brightness of the dark part.

Description

Technical Field

[0001] This invention belongs to the field of remote sensing image enhancement technology, specifically relating to a remote sensing image enhancement method based on radiometric reliability partitioning and ground feature spectral order preservation. Background Technology

[0002] Remote sensing imaging is often affected by factors such as solar altitude angle, cloud shadows and mountain shadows, imaging time, and payload exposure conditions, resulting in insufficient incident radiation in some image areas. This manifests as reduced visibility in dark areas, insufficient local contrast, and an increased relative noise level. Existing low-light image enhancement methods can be broadly classified into three categories: histogram- or curve-based brightness mapping methods, which directly redistribute the dynamic range; imaging decomposition-based methods, which decompose the image into components such as illumination and reflection, process them separately, and then reconstruct them; and end-to-end enhancement methods based on deep learning, which utilize data to learn a nonlinear mapping from low-light images to normal illumination images.

[0003] The existing methods described above have several shortcomings when applied to remote sensing images. First, a remote sensing image often contains dark shadows, normally exposed features, and highly reflective features simultaneously. Uniform mapping or end-to-end regression makes it difficult to distinguish the radiometric reliability of different areas, easily leading to insufficient or excessive enhancement of dark areas while simultaneously amplifying bright areas and amplifying noise in dark areas. Second, existing methods mostly aim to improve visual brightness, typically mapping or regressing each band independently, lacking constraints on cross-band relative response relationships. Furthermore, the relative band order and normalized band ratio of the same type of feature within local homogeneous regions are usually stable. Independent band processing may reverse band size relationships, adversely affecting subsequent remote sensing processing workflows such as feature classification, index calculation, and change analysis. In addition, low-light remote sensing images often contain some normally bright pixels with relatively reliable radiometrics. Most existing methods treat the entire low-light image as input of the same quality level, failing to utilize these relatively reliable local observations as a reference for enhancing dark pixels in the same area. Summary of the Invention

[0004] In view of the shortcomings of the prior art, the purpose of this invention is to provide a remote sensing image enhancement method based on radiometric reliability partitioning and ground object spectral order preservation, which can adjust the enhancement intensity according to the radiometric reliability of the current image itself, suppress over-enhancement in bright areas, and maintain a stable band order within the ground object area while improving the brightness of dark areas.

[0005] To achieve the above objectives, this invention provides a remote sensing image enhancement method based on radiometrically reliable partitioning and ground feature spectral order preservation, comprising the following steps: S1. Acquire multi-channel low-light remote sensing images and normalize them to obtain normalized observation values ​​for each pixel and each band. Calculate normalized brightness, local fluctuation degree, and normalized band ratio vector based on the normalized observation values. Combine the dark area threshold and bright area threshold of the current image statistics to construct an over-dark weight map, a radiation confidence weight map, and a bright area protection weight map. S2. Divide the land cover regions according to the structural boundary of normalized brightness and the normalized band ratio vector; select pixels with radiation confidence weight not less than the anchor point confidence threshold in each land cover region to form a set of confidence pixels, and calculate the regional median response and regional normalized band ratio of each band to obtain the spectral order anchor point; assign the spectral order anchor point of each land cover region to the pixels in that land cover region to obtain the anchor point broadcast map. S3. Input the low-light remote sensing image, the dark weight map, the radiometric confidence weight map, the bright protection weight map, and the anchor point broadcast map into the radiometric correction model to obtain the original gain parameters; perform bounded mapping on the original gain parameters and modulate them through the dark weight map and the bright protection weight map to obtain the shared gain field that affects all bands of the same pixel. S4. Pixel-by-pixel radiometric mapping is performed on the low-light remote sensing image using the shared gain field to obtain the initial enhanced image. For pixels in the initial enhanced image whose band order is inconsistent with the spectral order anchor point of the corresponding ground object area, a weighted monotonic projection is performed under the constraints of regional band order and output dynamic range, with the goal of minimizing the weighted change relative to the initial enhanced image, to obtain the enhanced remote sensing image. S5. Construct a combined constraint function using paired low-light remote sensing images and normal illumination reference images, and train the radiometric correction model. After training, execute steps S1 to S4 on the actual low-light remote sensing images to output enhanced remote sensing images.

[0006] As a preferred embodiment of the present invention, in S1, the method for calculating the normalized brightness, local fluctuation degree, and normalized band ratio vector is as follows: The normalized brightness at pixel p is obtained by weighted summation of the normalized observations for each band. : ; In the formula, C is the number of bands in the low-light remote sensing image, and c is the band index, with each channel corresponding to one band; Let a be the normalized observation value of pixel p in the c-th band; c Calculate the weight for the brightness of the c-th band, a c ≥0 and ; Within a local neighborhood centered on pixel p, the median of the absolute deviation of the normalized brightness from the local median value characterizes the degree of local fluctuation at pixel p. : ; In the formula, The local neighborhood centered at pixel p; med indicates taking the median; , Local neighborhood Normalized brightness at inner pixels q and r; The proportion of each band's normalized observation value at pixel p to the sum of all band's normalized observation values ​​is taken as the component of the corresponding band. The normalized band proportion vector at pixel p is obtained by combining the components of each band. The c-th component of the normalized band proportion vector at pixel p... for: ; In the formula, Let be the normalized observation value of pixel p in the j-th band, where j is the band index used for summation traversal; The zero constant is greater than zero.

[0007] As a preferred embodiment of the present invention, the process of constructing the dark weight map, the radiation confidence weight map, and the highlight protection weight map in S1 is as follows: Based on the normalized brightness distribution of all pixels in the current image, a preset low quantile ratio is taken respectively. The corresponding quantile is used as the dark area threshold t d Take the preset high quantile ratio The corresponding quantile is used as the highlight threshold t h ; The sigmoid function is used to calculate the under-darkness weight at pixel p. and highlight protection weight The radiative confidence weight at pixel p is formed by the non-darkness level, non-brightness level, and local fluctuation level. : ; ; ; In the formula, This represents the Sigmoid function; s d s h These are the dark area transition scale and the highlight transition scale, respectively. is the local fluctuation suppression coefficient; e is the natural constant; The under-darkness weight map, radiation confidence weight map, and highlight protection weight map are respectively constructed from the under-darkness weight, radiation confidence weight, and highlight protection weight of all pixels.

[0008] As a preferred embodiment of the present invention, the process of dividing the land feature area in S2 is as follows: A normalized brightness map is constructed from the normalized brightness of all pixels. The first-order difference between adjacent pixels is calculated on the normalized brightness map to obtain the structural boundary intensity. The structural boundary is constructed from the structural boundary intensity of all pixels. Pixels with structural boundary intensity lower than a preset boundary threshold and radiation confidence weight higher than a seed confidence threshold are selected as region growth seeds. For each region growing seed, examine the candidate neighbor pixels of the growing seed in that region. If a candidate neighbor pixel meets the following conditions simultaneously, it is merged into the current region: Condition 1: The difference between the average normalized brightness of the candidate neighboring pixels and the current region is not greater than the preset brightness difference threshold. Condition 2: The distance between the candidate neighboring pixel and the normalized band ratio vector of the current region is not greater than the preset ratio difference threshold; Condition 3: The candidate neighboring pixels do not cross structural boundaries with the current region; After completing the region growth, regions with an area smaller than the preset minimum area are merged with adjacent regions with the lowest boundary cost to obtain several feature regions.

[0009] As a preferred embodiment of the present invention, the process of obtaining the spectral sequence anchor point and the anchor point broadcast diagram in S2 is as follows: For the k-th land feature region R k The radiation confidence weight shall be no less than the anchor point confidence threshold. The pixels constitute R k The set of trustworthy pixels T k ; For T k Calculate the mid-range response of each band within the region: ; In the formula, For R k The regional midpoint response of the c-th band within the band; R k Arrange the regional median responses of each band within the band from largest to smallest to obtain R. k Band arrangement ; For T k The median of the c-th component of the normalized band scaling vector for each pixel is taken to obtain R. k Region normalized band ratio: ; In the formula, For R k The proportion of the region-normalized bands in the c-th band represents R. k The typical proportion of the response of the c-th band at the inner reliable pixel to the total response of all bands; and Together they constitute the land feature region Rk Spectral sequence anchor points; Assign spectral sequence anchor points to the pixels within each land cover region to obtain an anchor point broadcast map with the same spatial dimensions as the low-light remote sensing image. when When the difference between the regional median responses of two adjacent bands is less than a preset stability threshold, the adjacent band pair is marked as a weakly constrained band pair, and the order of the adjacent band pair is not forced in the weighted monotonic projection. When T k When the number of pixels in the array is less than a preset threshold, R is not forced to be used. k The spectral sequence anchor point does not perform weighted monotonic projection on the pixels in the area of ​​the ground feature, and directly retains the values ​​of each band of the initial enhanced image.

[0010] In a preferred embodiment of the present invention, in step S3, the radiation correction model outputs the original gain parameters of a single channel to obtain the shared gain field. ; In the formula, The shared gain at pixel p acts on all bands of the same pixel p; g max The maximum allowable gain; This represents the original gain parameter output by the radiometric correction model at pixel p. The shared gain across all pixels constitutes the shared gain field.

[0011] As a preferred embodiment of the present invention, in step S4, the method for obtaining the initial enhanced image is as follows: ; In the formula, The value of the c-th band at pixel p in the initial enhanced image; This represents a saturation function whose values ​​are restricted to the effective dynamic range of the output. Composed of all pixels, all bands This forms the initial enhanced image.

[0012] As a preferred embodiment of the present invention, the process of obtaining the enhanced remote sensing image in S4 is as follows: For those belonging to R k For pixel p, if the band order of the initial enhanced image at pixel p satisfies all conditions outside the weakly constrained band pair... If the values ​​of the initial enhanced image are used directly as the output, then if there is a band order violation, the following weighted monotonic projection is solved: ; The constraints are: ; In the formula, Let be the spectral vector at pixel p after spectral order correction; z is the spectral vector to be solved. c Let z be the component of the c-th band; The projection weight of the c-th band is based on T. k The degree of dispersion of the c-th band value is set; for The c-th order band index; For z in The component values ​​in the c-th sequential band; The weighted monotonic projection is solved using the adjacent disordered merging method: according to Arrange the values ​​of each band at pixel p in the initial enhanced image and form initial blocks band by band. When the weighted mean of two adjacent blocks violates the monotonic relationship, merge the two adjacent blocks and recalculate the weighted mean, replacing each element in the block with the merged weighted mean; repeat the above merging process until all adjacent blocks satisfy the monotonic relationship; finally, for The c-th position in the sequence is assigned its value to the z-th position. Each component is used to obtain the solved spectral vector; The obtained spectral vector is used as the spectral vector after spectral order correction at pixel p. ,by The band components replace the band values ​​at pixel p in the initial enhanced image; for pixels where the band order is not violated, the band values ​​of the initial enhanced image are retained; the enhanced remote sensing image is composed of the final values ​​of all pixels.

[0013] As a preferred embodiment of the present invention, in S5, the combined constraint function L includes the reference reconstruction loss L. rec , regional spectral ratio loss L rat Neighborhood brightness order loss L rank and boundary-aware gain smoothing loss L gain : ; In the formula, , , , L respectively rec L rat L rank L gain Loss weights; L rec The pixel differences between the constrained augmented remote sensing image and the corresponding normal illumination reference image are as follows: ; In the formula, H and W represent the height and width of the low-light remote sensing image, respectively; To enhance the value of the c-th band at pixel p in the remote sensing image; The value of the c-th band at pixel p is the reference image under normal illumination. L rat The constrained augmented remote sensing image is defined as the region whose normalized band proportion vector within each land cover area is close to the corresponding spectral sequence anchor point. Specifically: ; In the formula, N is the total number of pixel band terms participating in the statistics; To enhance the c-th component of the normalized band scaling vector at pixel p in the remote sensing image; L rank The relative brightness order of trusted adjacent pixels within the same ground feature area is constrained to remain consistent before and after enhancement, specifically as follows: ; In the formula, P is the set of adjacent pixel pairs in the same land feature area where the difference in normalized brightness exceeds the preset brightness difference stability threshold and the radiation confidence weights of the two pixels are both higher than the preset order confidence threshold. s is the number of pixel pairs in set P; pq The sign for the difference in normalized brightness between pixel p and pixel q in a low-light remote sensing image; , These represent the normalized brightness of the remotely sensed image at pixels p and q, respectively. The preset sequence interval; L gain Suppress abrupt changes in the shared gain field within the same land feature area and reduce the smoothing weight at structural boundaries, specifically: ; In the formula, E is the set of adjacent pixel pairs; k is the number of pixel pairs in set E. e This is the boundary attenuation coefficient; This represents the normalized brightness of the low-light remote sensing image at pixel q. This represents the shared gain at pixel q.

[0014] As a preferred embodiment of the present invention, the process of training the radiation correction model in S5 is as follows: The training samples are composed of paired low-light remote sensing images and normal illumination reference images, and the training set and validation set are divided. Steps S1 to S4 are executed sequentially for each training sample to obtain the enhanced remote sensing image. The combined constraint function is calculated based on the enhanced remote sensing image and the corresponding normal illumination reference image. The learnable parameters of the radiometric correction model are updated by gradient backpropagation. When the combined constraint function value on the validation set no longer decreases within a preset number of consecutive training rounds, the parameters of the radiation correction model are saved, and training is complete.

[0015] The beneficial effects of this invention are: This invention links the enhancement intensity to the radiation reliability of the current image itself through radiation reliability partitioning. Dark areas obtain a larger usable gain range, and the gain of the highlight protection area automatically approaches 1. Dark area brightening and highlight protection are achieved by the same gain modulation structure, without the need for additional semantic tags. At the same time, the spectral order anchor point is taken from reliable pixels within the same ground feature area. Stable observations within the ground feature area provide a band relationship reference from the current image itself for dark pixel enhancement, avoiding the forced application of global color statistics of the entire image to different ground features.

[0016] This invention restricts brightness recovery to a shared gain affecting all bands of the same pixel, ensuring that the band proportions of the input pixel are maintained during the initial enhancement stage without saturation truncation. For local band order anomalies caused by dark area noise, weighted monotonic projection is performed only on pixels inconsistent with the regional spectral sequence anchor point, and the result with the smallest weighted change is selected from those satisfying the band order constraint, resulting in minimal disturbance to the original radiometric observation. Furthermore, this invention does not perform illumination-reflection decomposition, does not generate virtual exposure sequences, and does not rely on frequency domain branching or semantic segmentation results, allowing for implementation using the same process on remote sensing inputs with different numbers of bands. Attached Figure Description

[0017] Figure 1 This is a flowchart illustrating the principle of this invention; Figure 2 This is a schematic diagram illustrating the process of enhancing remote sensing images based on actual low-light remote sensing images. Detailed Implementation

[0018] The embodiments of the present invention will be further described below with reference to the accompanying drawings: Example 1: As Figure 1 As shown, the remote sensing image enhancement method based on radiometric reliability partitioning and ground feature spectral order preservation includes the following steps: S1. Acquire multi-channel low-light remote sensing images and normalize them to obtain normalized observation values ​​for each pixel and each band. Calculate normalized brightness, local fluctuation degree, and normalized band ratio vector based on the normalized observation values. Combine the dark area threshold and bright area threshold of the current image statistics to construct an over-dark weight map, a radiation confidence weight map, and a bright area protection weight map. S2. Divide the land cover regions according to the structural boundary of normalized brightness and the normalized band ratio vector; select pixels with radiation confidence weight not less than the anchor point confidence threshold in each land cover region to form a set of confidence pixels, and calculate the regional median response and regional normalized band ratio of each band to obtain the spectral order anchor point; assign the spectral order anchor point of each land cover region to the pixels in that land cover region to obtain the anchor point broadcast map. S3. Input the low-light remote sensing image, the dark weight map, the radiometric confidence weight map, the bright protection weight map, and the anchor point broadcast map into the radiometric correction model to obtain the original gain parameters; perform bounded mapping on the original gain parameters and modulate them through the dark weight map and the bright protection weight map to obtain the shared gain field that affects all bands of the same pixel. S4. Pixel-by-pixel radiometric mapping is performed on the low-light remote sensing image using the shared gain field to obtain the initial enhanced image. For pixels in the initial enhanced image whose band order is inconsistent with the spectral order anchor point of the corresponding ground object area, a weighted monotonic projection is performed under the constraints of regional band order and output dynamic range, with the goal of minimizing the weighted change relative to the initial enhanced image, to obtain the enhanced remote sensing image. S5. Construct a combined constraint function using paired low-light remote sensing images and normal illumination reference images, and train the radiometric correction model. After training, execute steps S1 to S4 on the actual low-light remote sensing images to output enhanced remote sensing images.

[0019] In this embodiment, a multi-channel low-light remote sensing image of H×W×C is used as input, where H and W represent the height and width of the low-light remote sensing image, respectively, and C represents the number of bands, with each channel corresponding to one band. The input can be a three-channel visible light remote sensing image or a multispectral image with more discrete bands.

[0020] Linear normalization is performed on each band according to the effective quantization range of the sensor (i.e., remote sensing sensor, hereinafter the same), so that the normalized observation value of each pixel in each band is within the range of 0 to 1. If the sensor data contains invalid values, saturation values, or missing measurement markers, the corresponding pixels are first excluded from subsequent statistical calculations based on the product quality label. The remaining pixels after exclusion are the valid pixels. For example, when reading the quality label bands associated with the image product, pixels marked as cloud, cloud shadow, saturation, or missing measurement status are set as invalid pixels, and the remaining pixels are used as valid pixels. For data without quality label bands, pixels with a value of 0 in any band, a value equal to the upper limit of the quantization range (saturation value), or a value that is always zero in all bands are determined as invalid pixels, and the remaining pixels are used as valid pixels. In this embodiment, pixels in the subsequent process all refer to valid pixels.

[0021] In S1, the normalized brightness, local fluctuation level, and normalized band scale vector are calculated as follows: The normalized brightness at pixel p is obtained by weighted summation of the normalized observations for each band. : ; In the formula, C is the number of bands in the low-light remote sensing image, and c is the band index; Let a be the normalized observation value of pixel p in the c-th band; c Calculate the weight for the brightness of the c-th band, a c ≥0 and ; a c The acquisition method is as follows: for three-channel visible light remote sensing images, fixed weights can be used, for example, the weights of the three bands can be set to 0.299, 0.587, and 0.114 respectively; for general multispectral input, each band can be processed with equal weights, i.e., a c =1 / C, or normalized after weighting each band according to the sensor calibration coefficient. The normalized brightness map of the low-light remote sensing image is composed of the normalized brightness of all pixels.

[0022] Within a local neighborhood centered on pixel p, the median of the absolute deviation of the normalized brightness from the local median value characterizes the degree of local fluctuation at pixel p. : ; In the formula, The local neighborhood centered at pixel p can be a 5×5 or 7×7 square window; med indicates taking the median. , Local neighborhood Normalized brightness at inner pixels q and r; The reason for using the median absolute bias form is that it is insensitive to noise and outliers, and can distinguish between true dark radiation regions and local texture variations, as well as texture-rich areas. Larger, homogeneous dark areas The size is small, thus avoiding misclassification of textured areas as reliable areas.

[0023] The proportion of each band's normalized observation value at pixel p to the sum of all band's normalized observation values ​​is taken as the component of the corresponding band. The normalized band proportion vector at pixel p is obtained by combining the components of each band. The c-th component of the normalized band scaling vector at pixel p for: ; In the formula, Let be the normalized observation value of pixel p in the j-th band, where j is the band index used for summation traversal; For a zero-valued constant, 10 can be taken.-6 The magnitude is used to avoid division-zero singularities when all band observations at dark pixels are close to zero simultaneously.

[0024] The sum of all components of the normalized band scale vector is always equal to 1. The value of this vector reflects the relative compositional relationship between the responses of each band. When the observed values ​​of each band of a pixel change proportionally as a whole (e.g., only the brightness increases or decreases as a whole), this vector remains unchanged. The same type of land cover usually has similar normalized band scale vectors in local homogeneous areas. Therefore, this vector can be used to distinguish different land cover areas and also to provide a reference for the band relationship of dark area pixels.

[0025] The process of constructing the dark weight map, the radiation confidence weight map, and the highlight protection weight map is as follows: Based on the normalized brightness distribution of all pixels in the current image, a preset low quantile ratio is taken respectively. The corresponding quantile is used as the dark area threshold t d Take the preset high quantile ratio The corresponding quantile is used as the highlight threshold t h ; A value between 0.20 and 0.35 is acceptable, with 0.25 being preferred; Values ​​between 0.90 and 0.98 are acceptable, with 0.95 being preferred. Since the dark area threshold and highlight threshold are obtained statistically from the brightness distribution of the current image itself, they are adaptive to the current image at different exposure levels, eliminating the need to manually set the dark area threshold and highlight threshold for each image.

[0026] The sigmoid function is used to calculate the under-darkness weight at pixel p. and highlight protection weight The radiative confidence weight at pixel p is formed by the non-darkness level, non-brightness level, and local fluctuation level. : ; ; ; In the formula, This represents the Sigmoid function, whose values ​​range from 0 to 1 and are monotonically continuous. It uses continuous mapping instead of a hard threshold to avoid abrupt changes in weights at threshold boundaries; s d s h These are the dark transition scale and the bright transition scale, respectively, both greater than 0, controlling the steepness of the weighted transition, which can be taken as 0.02 to 0.10 times the normalized brightness dynamic range; This is the local fluctuation suppression coefficient, which can be a value between 0.5 and 5. The larger the value, the stronger the suppression effect of local fluctuations on the radiation confidence weight; e is the natural constant; Larger areas are overly dark and allow for stronger brightening; Larger areas are high-brightness protection zones and should be kept at their original values. Larger location radiation observations are more reliable and can be used for subsequent regional radiation statistics.

[0027] The under-darkness weight map, radiation confidence weight map, and highlight protection weight map are respectively constructed from the under-darkness weight, radiation confidence weight, and highlight protection weight of all pixels. The three weight maps are continuously evaluated weights, rather than mutually exclusive hard partition labels.

[0028] In S2, the process of dividing the land feature regions is as follows: The first-order difference between adjacent pixels is calculated on the normalized brightness map to obtain the structural boundary intensity. The structural boundary is composed of the structural boundary intensities of all pixels. The structural boundary intensity at pixel p can be the maximum or average of the absolute values ​​of the differences between the normalized brightness of the pixel and its four or eight neighboring pixels. The preset boundary threshold can be the upper quantile of the structural boundary intensity of all pixels, such as the 80th quantile. The structural boundary is defined between adjacent pixels whose structural boundary intensity is higher than the preset boundary threshold.

[0029] Pixels with structural boundary strength lower than a preset boundary threshold and radiation confidence weight higher than a seed confidence threshold are selected as region growth seeds. The seed confidence threshold can be a value between 0.5 and 0.7. Its purpose is to ensure that region growth starts from pixels with reliable radiation observation and no boundary, and to avoid the seed falling on a location with strong noise or blurred boundary.

[0030] For each region growing seed, examine the candidate neighbor pixels of the growing seed in that region. If a candidate neighbor pixel meets the following conditions simultaneously, it is merged into the current region: Condition 1: The difference between the average normalized brightness of the candidate neighboring pixels and the current region is not greater than the preset brightness difference threshold; the preset brightness difference threshold can be a value between 0.05 and 0.15 (normalized brightness range 0 to 1). Condition 2: The distance between the candidate neighboring pixel and the normalized band ratio vector of the current region is not greater than the preset ratio difference threshold; the distance can be Euclidean distance, and the preset ratio difference threshold can be a value between 0.05 and 0.20; Condition 3: The candidate neighboring pixels do not cross structural boundaries with the current region; After region growing is completed, regions with areas smaller than the preset minimum area are merged with adjacent regions having the lowest boundary cost, resulting in several feature regions. The preset minimum area is A. minThe boundary cost is set to tens to hundreds of pixels based on the image spatial resolution. The boundary cost between two regions is measured by the sum or average of the structural boundary strengths of the pixels on their common boundary. The smaller the boundary cost, the less significant the radiation difference between the two regions, and the more suitable they are for merging. This step only determines the statistical range of local homogeneity and does not perform semantic recognition of land cover categories, therefore no semantic annotation is required.

[0031] The process of obtaining the spectral order anchor points and anchor point broadcast diagram is as follows: For the k-th land feature region R k The radiation confidence weight shall be no less than the anchor point confidence threshold. The pixels constitute R k The set of trustworthy pixels T k ; A value between 0.45 and 0.65 can be used to filter reliable pixels that participate in regional radiation statistics.

[0032] For T k Calculate the mid-range response of each band within the region: ; In the formula, For R k The regional midpoint response of the c-th band within the band; R k Arrange the regional median responses of each band within the band from largest to smallest to obtain R. k Band arrangement This is the band index sequence obtained by arranging all bands in descending order of their median response. The median is used instead of the mean because the median is more robust to residual noise and anomalous pixels in dark areas.

[0033] For T k The median of the c-th component of the normalized band scaling vector for each pixel is taken to obtain R. k Region normalized band ratio: ; In the formula, For R k The proportion of the region-normalized bands in the c-th band represents R. k The typical proportion of the response of the c-th band at the inner reliable pixel to the total response of all bands; and Together they constitute the land feature region R kThe spectral sequence anchor point represents the stable relative order and typical band proportions of radiatively reliable pixels within the same ground feature area. It can be used as a reference for the band relationships after dark pixel enhancement within that area. Because the spectral sequence anchor point is taken from reliable observations within the same ground feature area, rather than global statistics of the entire image, it will not incorrectly apply the band relationships of one type of ground feature to another.

[0034] The spectral sequence anchor point of each land cover region is assigned to the pixels within that region, resulting in an anchor point broadcast map with the same spatial size as the low-light remote sensing image. The anchor point broadcast map is a multi-channel feature map. At each pixel location, the band arrangement information of the land cover region to which the pixel belongs and the regional normalized band ratio value of each band are filled in, so that the subsequent radiometric correction model can know the typical band relationship of the land cover region to which each pixel belongs.

[0035] when The difference in the regional median response between two adjacent bands is less than a preset stability threshold. When this occurs, it indicates that the two bands are difficult to distinguish in terms of size in the reliable statistics of the ground feature area. The adjacent band pair is marked as a weakly constrained band pair, and the order of the adjacent band pair is not forced in the weighted monotonic projection, so as to avoid creating artificial differences for bands that are difficult to distinguish. The normalized observation dynamic range can be taken as 0.01 to 0.05 times.

[0036] When T k When the number of pixels in the data is less than a preset threshold (e.g., 10 to 50 pixels), it indicates a lack of sufficient reliable observations in the area, and R is not mandatory in this case. k The spectral sequence anchor point does not perform weighted monotonic projection on the pixels in the area of ​​the ground feature, and directly retains the values ​​of each band of the initial enhanced image.

[0037] Alternatively, the spectral sequence anchor point of the adjacent land cover region with the closest boundary statistical features can be selected as the fallback anchor point. Boundary statistical features are the radiometric statistics of several pixels inside the boundary of the land cover region (e.g., a strip 3 to 5 pixels wide inside the boundary), including the median of normalized brightness and the median of each component of the normalized band scale vector. "Closest boundary statistical features" means that the distance between the boundary statistical features of the adjacent land cover region and the boundary statistical features of the local land cover region is minimized (e.g., the weighted sum of the differences between the statistics). The fallback anchor point is a spectral sequence anchor point from an adjacent, similar land cover region used to replace the local land cover region's spectral sequence anchor point when a reliable spectral sequence anchor point cannot be established. Using a fallback anchor point can preserve the spectral sequence constraint capability while avoiding imposing anchor points from land cover regions with excessively different relationships onto the local land cover region.

[0038] In S3, the radiometric correction model is a convolutional neural network that employs a multi-scale encoder-decoder structure: The input consists of C bands of the low-light remote sensing image, an overly dark weight map, a radiometric confidence weight map, a highlight protection weight map, and an anchor point broadcast map, concatenated along the channel dimension. The network itself comprises several convolutional layers and activation functions, progressively downsampling at two or three spatial scales to extract features to expand the receptive field, and then progressively upsampling to restore the original spatial dimensions. The output is a single-channel output layer that generates the original gain parameter map consistent with the spatial dimensions of the low-light remote sensing image, i.e., the original gain parameters for each pixel. The number of convolutional layers, the number of channels in each layer, and the type of activation function can be adjusted based on computational resources.

[0039] The radiation correction model outputs the raw gain parameters of a single channel, from which the shared gain field is obtained: ; In the formula, The shared gain at pixel p acts on all bands of the same pixel p; g max For the maximum allowable gain, and g max >1, which can be a value between 2 and 8 depending on the dynamic range of the training data; The original gain parameters are output by the radiometric correction model at pixel p; the Sigmoid function is used to convert the gain parameters to the original gain parameters. Compress to the interval between 0 and 1 to achieve a bounded mapping.

[0040] The function of this gain modulation structure is to adjust the brightness of pixel p when it is in the highlight protection zone. It approaches 1, regardless of the output of the radiation correction model. , All values ​​approach 1, meaning that the highlight area automatically maintains near-unit gain to avoid over-enhancement; when pixel p is in an overly dark area and not within the highlight protection zone... The larger the g, the larger the usable gain range of the radiation correction model; max Explicitly limit the maximum gain to prevent extremely dark noise from being amplified indefinitely. Because... As a single-channel scalar, it acts on all bands of the same pixel, and the proportional relationship between the bands is not artificially changed during the gain mapping stage. The shared gain at all pixels constitutes the shared gain field.

[0041] The learnable parameters of the radiometric correction model are all the convolutional kernel weights and bias parameters within the network; if the network contains normalization layers, the scaling and translation parameters of the normalization layers are also included. The radiometric confidence partitioning in step S1 and the land cover region division and spectral sequence anchor construction in step S2 are calculated according to deterministic rules and do not include learnable parameters.

[0042] In S4, the initial enhanced image is obtained as follows: ; In the formula, The value of the c-th band at pixel p in the initial enhanced image; This indicates a saturation function whose values ​​are restricted to the effective dynamic range of the output (0 to 1), that is, values ​​less than 0 are set to 0 and values ​​greater than 1 are set to 1; Composed of all pixels, all bands This forms the initial enhanced image. Because the same pixel shares bands... When no saturation truncation occurs, the normalized band ratio vector of each pixel in the initial enhanced image remains consistent with that of the low-light remote sensing image, meaning that the band ratio relationship of the input pixels is naturally maintained during the initial enhancement stage.

[0043] For dark regions, the original observations may be affected by noise, resulting in inconsistencies between the local band order and the reliable statistics of the region. Therefore, the process of continuing spectral order checking and correction to obtain the enhanced remote sensing image is as follows: For those belonging to R k For pixel p, if the band order of the initial enhanced image at pixel p satisfies all conditions outside the weakly constrained band pair... If so, the pixel does not require any color or spectral correction and is directly output using the value of the initial enhanced image; If there is a band order violation, then solve for the weighted monotonic projection, that is, to minimize the weighted change from the initial enhanced spectral vector while ensuring that the corrected spectral vector satisfies the regional band order constraint and the output dynamic range constraint; the weighted monotonic projection is: ; The regional band sequence constraint is: ; The output dynamic range constraint is: ; In the formula, Let be the spectral vector at pixel p after spectral order correction; z is the spectral vector to be solved. c Let z be the component of the c-th band; The projection weight of the c-th band is based on T. k The dispersion of the c-th band value is set; the smaller the dispersion, the larger the projection weight. The dispersion can be expressed as T. k The standard deviation or median absolute deviation of the normalized observations in the c-th band is a measure of, for example, taking... , where s c The degree of dispersion of the c-th band, It is a small constant that is greater than zero; for The c-th order band index; For z in The component values ​​in the c-th sequential band; The weighted monotonic projection is solved using the adjacent disordered merging method: according to Arrange the band values ​​at pixel p of the initial enhanced image and form initial blocks band by band, with each initial block containing only one band value; check the weighted mean of adjacent blocks from front to back. If the weighted mean of two adjacent blocks violates the monotonic relationship (the weighted mean of the previous block is less than the weighted mean of the next block), merge the two adjacent blocks and recalculate the weighted mean, replacing each element in the block with the merged weighted mean; repeat the above merging process until all adjacent blocks satisfy the monotonic relationship; finally, for The c-th position in the sequence is assigned its value to the z-th position. The solution is obtained by solving the spectral vector by considering each component. The solution obtained by this process is the optimal solution of the weighted monotonic projection problem. Its computational complexity is linearly related to the number of bands and is suitable for pixel-by-pixel execution.

[0044] The obtained spectral vector is used as the spectral vector after spectral order correction at pixel p. ,by The band components replace the band values ​​at pixel p in the initial enhanced image; for pixels where the band order is not violated, the band values ​​of the initial enhanced image are retained; the enhanced remote sensing image is composed of the final values ​​of all pixels.

[0045] Spectral order correction applies only to pixels that are indeed out of order, and selects the result with the smallest weighted distance to the initial enhancement value from all results that satisfy the constraints, thus minimizing additional perturbations to the original radiometric observations. For inputs with a large number of bands, to reduce computational load, monotonic constraints can be established only between adjacent sorted bands with high stability scores, or the bands can be divided into several adjacent band groups according to the sensor's spectral response and then projection can be performed separately.

[0046] In S5, the training samples consist of paired low-light remote sensing images and normal illumination reference images, divided into training and validation sets. Training samples are obtained by using low-light remote sensing images and normal illumination reference images of the same scene under different exposure conditions; or, while ensuring geometric registration, generating corresponding low-light remote sensing image samples using radiometrically corrected normal illumination reference images according to a preset exposure attenuation model. Training data should be paired within the same band definition and dynamic range; if the band response functions of different sensors are inconsistent, corresponding models should be established separately or radiometric normalization between sensors should be completed first.

[0047] Steps S1 to S4 are executed sequentially for each training sample to obtain an enhanced remote sensing image, and a combined constraint function is calculated based on the enhanced remote sensing image and the corresponding normal illumination reference image.

[0048] The combined constraint function L includes the reference reconstruction loss L rec , regional spectral ratio loss L rat Neighborhood brightness order loss L rank and boundary-aware gain smoothing loss L gain : ; In the formula, , , , L respectively rec L rat L rank L gain The loss weights can be taken separately. =1、 =0.1 to 1 =0.01 to 0.1 =0.01 to 0.1, the actual value is adjusted according to the effect of the validation set; L rec To enhance the average pixel-by-pixel absolute difference between the remotely sensed image and the corresponding normal illumination reference image, the overall radiometric accuracy of the enhancement results is constrained: ; In the formula, To enhance the value of the c-th band at pixel p in the remote sensing image; The value of the c-th band at pixel p is the reference image under normal illumination. L rat Constraint-enhanced remote sensing images assign greater weight to pixels that are too dark and not protected by highlight protection, by multiplying them by a factor of 1. The normalized band proportion vector of each land cover region in the constrained-enhanced remote sensing image is close to the normalized band proportion of the corresponding spectral sequence anchor point. As a pixel-by-pixel weight, it strengthens the spectral ratio constraint of dark areas and prevents bright areas from being forcibly pulled into regional statistics: ; In the formula, N is the total number of pixel band terms participating in the statistics; To enhance the c-th component of the normalized band scaling vector at pixel p in the remote sensing image; L rank To ensure that the relative brightness order of trusted adjacent pixels within the same feature area remains consistent before and after enhancement, for adjacent pixel pairs in set P... The loss is zero when the sign of the enhanced brightness difference matches the sign of the input brightness difference; a penalty is applied when the order is reversed. ; In the formula, P is the set of adjacent pixel pairs in the same land feature area where the difference in normalized brightness exceeds the preset brightness difference stability threshold and the radiation confidence weights of the two pixels are both higher than the preset order confidence threshold. The preset order confidence threshold can be the same value as the anchor point confidence threshold or slightly higher. s is the number of pixel pairs in set P; pq The sign for the difference in normalized brightness between pixel p and pixel q in a low-light remote sensing image; , These represent the normalized brightness of the remotely sensed image at pixels p and q, respectively. The preset sequence interval is used to ensure that the enhanced brightness difference maintains a positive safety interval, and can be 0.01 to 0.05. L gain Within the same feature area, abrupt changes in the shared gain field are suppressed, and the smoothing weight is reduced at structural boundaries. This is achieved when the normalized brightness difference between two adjacent pixels... When the exponential term is large (i.e., located near the structural boundary), As brightness decreases, the smoothing constraint weakens; within homogeneous regions, where brightness differences are small, the smoothing constraint strengthens. ; In the formula, E is the set of adjacent pixel pairs; k is the number of pixel pairs in set E. e This is the boundary attenuation coefficient, which can be a value between 5 and 50; This represents the normalized brightness of the low-light remote sensing image at pixel q. This represents the shared gain at pixel q.

[0049] The learnable parameters of the radiometric correction model are updated using gradient backpropagation. If weighted monotonic projection is used in training, it can be implemented as a piecewise differentiable operator. Alternatively, the spectral order relationship can be approximated during training using region spectral ratio loss and neighborhood brightness order loss, and then a strict weighted monotonic projection can be performed during inference. When the combined constraint function value on the validation set no longer decreases within a preset number of training rounds (e.g., 10 to 20 rounds), the parameters of the radiometric correction model are saved, and training is complete.

[0050] like Figure 2 As shown, after training is completed, steps S1 to S4 are performed on the actual low-light remote sensing image to output the enhanced remote sensing image. Figure 2 In the diagram, solid arrows represent the main data stream, while dashed arrows represent constraint or modulation relationships. Weight map modulation indicates that the overly dark weight map and the highlight protection weight map do not participate in image content transfer; instead, they are handled through... The formula modulates the gain amplitude; the spectral order anchor constraint indicates that the spectral order anchor is not used as the input of image content into the projection stage, but rather provides regional band order constraints and projection weights.

[0051] When the input is a three-channel visible light remote sensing image, C equals 3, and the spectral order anchor only involves the median response ranking of the three bands, resulting in a small scale of weighted monotonic projection. If the reliable pixel statistics for a certain area show that the response of band 1 is higher than that of band 2, and band 2 is higher than that of band 3, then after performing shared gain enhancement on the dark pixels in that area, only the corresponding band order reversal is checked. If no reversal occurs, no color correction is performed; if a reversal occurs, the conflicting bands are adjusted to the nearest position that satisfies the regional statistical order through adjacent disordered merging. This approach can both correct dark area anomalies using reliable regional color relationships and avoid performing uniform white balance or global color stretching on all pixels.

[0052] When inputting a multispectral image containing multiple narrow bands, the complete band arrangement can be obtained based on the regional mid-band response. Alternatively, only the difference in the median response of the regions can be retained. Stable ordered band pairs participate in the constraint. For relatively uniform areas such as large water bodies, vegetation, and bare land, reliable pixels can provide relatively stable regional band relationships. For land cover boundaries, mixed pixels, or locations with insufficient statistics within the region, the intensity of spectral order constraint is reduced or projection is not performed, thereby avoiding the incorrect application of the band relationship of one region to another land cover.

[0053] If subsequent tasks require maintaining the original calibration units, the enhanced remote sensing image can be inversely transformed according to the quantization range of the original sensor after enhancement in the normalized space of 0 to 1, to obtain an output consistent with the input data type. At the same time, the shared gain field and the highlight protection weight map can be saved as a quality auxiliary layer to facilitate the determination of which areas have undergone strong radiometric adjustment in subsequent processing.

[0054] Example 2: A remote sensing image enhancement device based on radiometric reliability zoning and ground feature spectral order preservation, comprising: One or more processors; Memory, used to store one or more computer programs; When one or more computer programs are executed by one or more processors, the one or more processors perform the method in Embodiment 1.

[0055] Example 3: A computer-readable storage medium having executable instructions stored thereon, which, when executed by a processor, cause the processor to perform the method in Example 1.

[0056] The above description is merely a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any person skilled in the art can make equivalent substitutions or modifications based on the technical solution and concept of the present invention within the scope of the technology disclosed in the present invention, and such modifications should also be considered to fall within the scope of protection of the present invention.

Claims

1. A remote sensing image enhancement method based on radiometric reliability partitioning and ground feature spectral order preservation, characterized in that, Includes the following steps: S1. Acquire multi-channel low-light remote sensing images and normalize them to obtain normalized observation values ​​for each pixel and each band. Calculate normalized brightness, local fluctuation degree, and normalized band ratio vector based on the normalized observation values. Combine the dark area threshold and bright area threshold of the current image statistics to construct an over-dark weight map, a radiation confidence weight map, and a bright area protection weight map. S2. Divide the ground feature regions according to the structural boundary of normalized brightness and the normalized band ratio vector; select pixels with radiation confidence weight not less than the anchor point confidence threshold in each ground feature region to form a set of confidence pixels, and statistically analyze the regional median response and regional normalized band ratio of each band to obtain the spectral order anchor point. Assign spectral sequence anchor points to the pixels within each feature region to obtain the anchor point broadcast map; S3. Input the low-light remote sensing image, the dark weight map, the radiometric confidence weight map, the bright protection weight map, and the anchor point broadcast map into the radiometric correction model to obtain the original gain parameters; perform bounded mapping on the original gain parameters and modulate them through the dark weight map and the bright protection weight map to obtain the shared gain field that affects all bands of the same pixel. S4. Perform pixel-by-pixel radiometric mapping on the low-light remote sensing image using the shared gain field to obtain the initial enhanced image; For pixels in the initial enhanced image whose band order is inconsistent with the spectral sequence anchor point of the corresponding ground cover area, under the constraints of regional band order and output dynamic range, a weighted monotonic projection is performed with the goal of minimizing the weighted change relative to the initial enhanced image to obtain the enhanced remote sensing image. S5. Construct a combined constraint function using paired low-light remote sensing images and normal illumination reference images, and train the radiometric correction model. After training, execute steps S1 to S4 on the actual low-light remote sensing images to output enhanced remote sensing images.

2. The remote sensing image enhancement method based on radiometric reliability partitioning and ground feature spectral order preservation according to claim 1, characterized in that, In S1, the normalized brightness, local fluctuation degree, and normalized band ratio vector are calculated as follows: The normalized brightness at pixel p is obtained by weighted summation of the normalized observations for each band. : ； In the formula, C is the number of bands in the low-light remote sensing image, and c is the band index, with each channel corresponding to one band; Let a be the normalized observation value of pixel p in the c-th band; c Calculate the weight for the brightness of the c-th band, a c ≥0 and ; Within a local neighborhood centered on pixel p, the median of the absolute deviation of the normalized brightness from the local median value characterizes the degree of local fluctuation at pixel p. : ； In the formula, The local neighborhood centered at pixel p; med indicates taking the median; , Local neighborhood Normalized brightness at inner pixels q and r; The proportion of each band's normalized observation value at pixel p to the sum of all band's normalized observation values ​​is taken as the component of the corresponding band. The normalized band proportion vector at pixel p is obtained by combining the components of each band. The c-th component of the normalized band proportion vector at pixel p... for: ； In the formula, Let be the normalized observation value of pixel p in the j-th band, where j is the band index used for summation traversal; The zero constant is greater than zero.

3. The remote sensing image enhancement method based on radiometric reliability partitioning and ground feature spectral order preservation according to claim 1, characterized in that, In S1, the process of constructing the dark weight map, the radiation confidence weight map, and the highlight protection weight map is as follows: Based on the normalized brightness distribution of all pixels in the current image, a preset low quantile ratio is taken respectively. The corresponding quantile is used as the dark area threshold t d Take the preset high quantile ratio The corresponding quantile is used as the highlight threshold t h ; The sigmoid function is used to calculate the under-darkness weight at pixel p. and highlight protection weight The radiative confidence weight at pixel p is formed by the non-darkness level, non-brightness level, and local fluctuation level. : ； ； ； In the formula, This represents the Sigmoid function; s d s h These are the dark area transition scale and the highlight transition scale, respectively. This is the local fluctuation suppression coefficient; The normalized brightness at pixel p; The local fluctuation at pixel p; e is the natural constant; The under-darkness weight map, radiation confidence weight map, and highlight protection weight map are respectively constructed from the under-darkness weight, radiation confidence weight, and highlight protection weight of all pixels.

4. The remote sensing image enhancement method based on radiometric reliability partitioning and ground feature spectral order preservation according to claim 1, characterized in that, In S2, the process of dividing the land feature area is as follows: A normalized brightness map is constructed from the normalized brightness of all pixels. The first-order difference between adjacent pixels is calculated on the normalized brightness map to obtain the structural boundary intensity. The structural boundary is constructed from the structural boundary intensity of all pixels. Pixels with structural boundary intensity lower than a preset boundary threshold and radiation confidence weight higher than a seed confidence threshold are selected as region growth seeds. For each region growing seed, examine the candidate neighbor pixels of the growing seed in that region. If a candidate neighbor pixel meets the following conditions simultaneously, it is merged into the current region: Condition 1: The difference between the average normalized brightness of the candidate neighboring pixels and the current region is not greater than the preset brightness difference threshold. Condition 2: The distance between the candidate neighboring pixel and the normalized band ratio vector of the current region is not greater than the preset ratio difference threshold; Condition 3: The candidate neighboring pixels do not cross structural boundaries with the current region; After completing the region growth, regions with an area smaller than the preset minimum area are merged with adjacent regions with the lowest boundary cost to obtain several feature regions.

5. The remote sensing image enhancement method based on radiometric reliability partitioning and ground feature spectral order preservation according to claim 1, characterized in that, In S2, the process of obtaining the spectral sequence anchor point and the anchor point broadcast diagram is as follows: For the k-th land feature region R k The radiation confidence weight shall be no less than the anchor point confidence threshold. The pixels constitute R k The set of trustworthy pixels T k ; For T k Calculate the mid-range response of each band within the region: ； In the formula, For R k The regional midpoint response of the c-th band within the band; is the normalized observation value of pixel p in the c-th band; med indicates taking the median. R k Arrange the regional median responses of each band within the band from largest to smallest to obtain R. k Band arrangement ; For T k The median of the c-th component of the normalized band scaling vector for each pixel is taken to obtain R. k Region normalized band ratio: ； In the formula, For R k The proportion of the region-normalized bands in the c-th band represents R. k The typical proportion of the response of the c-th band at the inner reliable pixel to the total response of all bands; is the c-th component of the normalized band scaling vector at pixel p; and Together they constitute the land feature region R k Spectral sequence anchor points; Assign spectral sequence anchor points to the pixels within each land cover region to obtain an anchor point broadcast map with the same spatial dimensions as the low-light remote sensing image. when When the difference between the regional median responses of two adjacent bands is less than a preset stability threshold, the adjacent band pair is marked as a weakly constrained band pair, and the order of the adjacent band pair is not forced in the weighted monotonic projection. When T k When the number of pixels in the array is less than a preset threshold, R is not forced to be used. k The spectral sequence anchor point does not perform weighted monotonic projection on the pixels in the area of ​​the ground feature, and directly retains the values ​​of each band of the initial enhanced image.

6. The remote sensing image enhancement method based on radiometric reliability partitioning and ground feature spectral order preservation according to claim 1, characterized in that, In S3, the radiation correction model outputs the original gain parameters of a single channel, thereby obtaining the shared gain field: ； In the formula, The shared gain at pixel p acts on all bands of the same pixel p; g max The maximum allowable gain; , These are the weights for underexposure and highlight protection at pixel p, respectively. Represents the Sigmoid function; This represents the original gain parameter output by the radiometric correction model at pixel p; The shared gain across all pixels constitutes the shared gain field.

7. The remote sensing image enhancement method based on radiometric reliability partitioning and ground feature spectral order preservation according to claim 1, characterized in that, In S4, the initial enhanced image is obtained in the following way: ； In the formula, The value of the c-th band at pixel p in the initial enhanced image; Let p be the normalized observation value of pixel p in the c-th band; This represents the shared gain at pixel p. This represents a saturation function whose values ​​are restricted to the effective dynamic range of the output. Composed of all pixels, all bands This forms the initial enhanced image.

8. The remote sensing image enhancement method based on radiometric reliability partitioning and ground feature spectral order preservation according to claim 5, characterized in that, In step S4, the process of obtaining the enhanced remote sensing image is as follows: For those belonging to R k For pixel p, if the band order of the initial enhanced image at pixel p satisfies all conditions outside the weakly constrained band pair... If the values ​​of the initial enhanced image are used directly as the output, then if there is a band order violation, the following weighted monotonic projection is solved: ； The constraints are: ； In the formula, Let be the spectral vector at pixel p after spectral order correction; z is the spectral vector to be solved. c Let z be the component of the c-th band; C is the number of bands in the low-light remote sensing image. The projection weight of the c-th band is based on T. k The degree of dispersion of the c-th band value is set; for The c-th order band index; For z in The component values ​​in the c-th sequential band; The value of the c-th band at pixel p in the initial enhanced image; The weighted monotonic projection is solved using the adjacent disordered merging method: according to Arrange the values ​​of each band at pixel p in the initial enhanced image and form initial blocks band by band. When the weighted mean of two adjacent blocks violates the monotonic relationship, merge the two adjacent blocks and recalculate the weighted mean, replacing each element in the block with the merged weighted mean; repeat the above merging process until all adjacent blocks satisfy the monotonic relationship; finally, for The c-th position in the sequence is assigned its value to the z-th position. Each component is used to obtain the solved spectral vector; The obtained spectral vector is used as the spectral order-corrected spectral vector at pixel p. ,by The band components replace the band values ​​at pixel p in the initial enhanced image; for pixels where the band order is not violated, the band values ​​of the initial enhanced image are retained; the enhanced remote sensing image is composed of the final values ​​of all pixels.

9. The remote sensing image enhancement method based on radiometric reliability partitioning and ground feature spectral order preservation according to claim 1, characterized in that, In S5, the combined constraint function L includes the reference reconstruction loss L. rec , regional spectral ratio loss L rat Neighborhood brightness order loss L rank and boundary-aware gain smoothing loss L gain : ； In the formula, , , , L respectively rec L rat L rank L gain Loss weights; L rec The pixel differences between the constrained augmented remote sensing image and the corresponding normal illumination reference image are as follows: ； In the formula, H and W are the height and width of the low-light remote sensing image, respectively; C is the number of bands in the low-light remote sensing image. To enhance the value of the c-th band at pixel p in the remote sensing image; The value of the c-th band at pixel p is the reference image under normal illumination. L rat The constrained augmented remote sensing image is defined as the region whose normalized band proportion vector within each land cover area is close to the region's normalized band proportion corresponding to the spectral sequence anchor point. Specifically: ； In the formula, N is the total number of pixel band terms participating in the statistics; R k This represents the k-th feature region; , These are the weights for underexposure and highlight protection at pixel p, respectively. To enhance the c-th component of the normalized band scaling vector at pixel p in the remote sensing image; For R k The region-normalized band ratio of the c-th band within the band; L rank The relative brightness order of trusted adjacent pixels within the same ground feature area is constrained to remain consistent before and after enhancement, specifically as follows: ； In the formula, P is the set of adjacent pixel pairs in the same land feature area where the difference in normalized brightness exceeds the preset brightness difference stability threshold and the radiation confidence weights of the two pixels are both higher than the preset order confidence threshold. s is the number of pixel pairs in set P; pq The sign for the difference in normalized brightness between pixel p and pixel q in a low-light remote sensing image; , These represent the normalized brightness of the remotely sensed image at pixels p and q, respectively. The preset sequence interval; L gain Suppressing abrupt changes in the shared gain field within the same land feature area and reducing smoothing weights at structural boundaries, specifically: ； In the formula, E is the set of adjacent pixel pairs; k is the number of pixel pairs in set E. e This is the boundary attenuation coefficient; , These represent the normalized brightness of the low-light remote sensing image at pixels p and q, respectively. , These are the shared gains at pixels p and q, respectively.

10. The remote sensing image enhancement method based on radiometric reliability partitioning and ground feature spectral order preservation according to claim 1, characterized in that, In S5, the process of training the radiation correction model is as follows: The training samples are composed of paired low-light remote sensing images and normal illumination reference images, and the training set and validation set are divided. Steps S1 to S4 are executed sequentially for each training sample to obtain an enhanced remote sensing image. The combined constraint function is calculated based on the enhanced remote sensing image and the corresponding normal illumination reference image. The learnable parameters of the radiometric correction model are updated using gradient backpropagation. When the combined constraint function value on the validation set no longer decreases within a preset number of consecutive training rounds, the parameters of the radiation correction model are saved, and training is complete.

Images