A method for calibration and intelligent segmentation change detection of UAV remote sensing images

Abstract

This invention discloses a method for UAV remote sensing image calibration and intelligent segmentation change detection, belonging to the field of UAV remote sensing image processing technology. The method includes: acquiring the original remote sensing image sequence and performing atmospheric correction based on a ground object radiative transfer model to generate a radiometrically calibrated image set; employing a spatiotemporal feature encoding network based on dynamic graph convolution to jointly embed the spectral and spatial geometric features of each image in the radiometrically calibrated image set, constructing a multi-scale feature tensor pyramid; generating a heatmap of candidate change regions through a learnable region proposal module, and performing pixel-by-pixel differential excitation between the heatmap and the temporal images in the radiometrically calibrated image set to obtain a change intensity prior map; guided by the change intensity prior map, performing iterative semantic segmentation using an edge-guided conditional random field to output a change pixel probability matrix; and determining the final change detection binary map based on the change pixel probability matrix using an adaptive threshold selection strategy.

Description

Technical Field

[0001] This invention relates to the field of UAV remote sensing image processing technology, specifically a method for UAV remote sensing image calibration and intelligent segmentation change detection. Background Technology

[0002] In existing UAV remote sensing image change detection technologies, traditional methods such as the difference method, ratio method, or principal component analysis directly perform algebraic operations on two temporal images. The detection results are severely affected by atmospheric conditions, sensor noise, and differences in ground illumination, making it difficult to distinguish between genuine surface changes and pseudo-changes. Deep learning methods, such as fully convolutional networks or U-Net structures based on convolutional neural networks, can automatically extract features, but their training requires a large amount of pixel-level labeled remote sensing change detection datasets, and their generalization ability to unlabeled areas is weak. Conditional random fields (CRFs) are often used as post-processing modules to optimize segmentation edges, but the binary potential energy of conventional CRFs relies solely on pixel spatial distance and color differences, lacking active guidance on ground edge information, leading to over-smoothing or edge confusion at the boundary between changed and unchanged areas. Existing technologies struggle to simultaneously solve the problems of edge preservation in changed areas and accurate modeling of spectral differences without requiring extensive labeling. Summary of the Invention

[0003] This invention aims to provide a method for calibration and intelligent segmentation of UAV remote sensing images to detect changes, thereby solving the problems of existing technologies where change detection accuracy is greatly affected by the atmosphere, edge preservation capability is weak, and reliance on a large amount of labeled data.

[0004] The objective of this invention can be achieved through the following technical solutions: This invention provides a method for calibration and intelligent segmentation change detection of UAV remote sensing images. The method includes: acquiring an original remote sensing image sequence using a multispectral sensor mounted on the UAV, and performing atmospheric correction based on a ground-object radiative transfer model on the original remote sensing image sequence to generate a radiometrically calibrated image set. As a preferred technical solution, the specific steps of atmospheric correction are as follows: extracting the digital quantization value of each pixel in the original remote sensing image sequence, and combining it with the pre-calibrated spectral response function of the multispectral sensor mounted on the UAV, converting the digital quantization value into the corresponding atmospheric top-layer radiance value; inputting the atmospheric top-layer radiance value into the ground-object radiative transfer model, which inverts atmospheric upward radiation, atmospheric downward radiation, and solar direct radiation transmittance based on synchronously acquired atmospheric aerosol optical thickness and water vapor content; using the atmospheric upward radiation, atmospheric downward radiation, and solar direct radiation transmittance, performing surface reflectance inversion on the atmospheric top-layer radiance value to obtain the radiometrically calibrated image set. This step eliminates the influence of atmospheric scattering and absorption on surface reflectivity through precise atmospheric correction, enabling subsequent change detection to be based on real physical quantities, significantly improving the comparability and accuracy of the detection results. As a further preferred option, in the atmospheric correction based on the ground-object radiative transfer model, the atmospheric aerosol optical thickness and water vapor content are simultaneously acquired and collected in real time by the meteorological detection module of the multispectral sensor mounted on the UAV, thus eliminating reliance on external meteorological data and enhancing the autonomy and real-time performance of the method. Furthermore, before performing the atmospheric correction based on the ground-object radiative transfer model, a step of geometric fine correction of the original remote sensing image sequence is included: extracting feature points from the original remote sensing image sequence, matching the feature points with a pre-stored reference digital orthophoto image to obtain a set of corresponding point pairs; fitting the geometric deformation field of the original remote sensing image sequence relative to the reference digital orthophoto image using a thin-plate spline function based on the set of corresponding point pairs; applying the geometric deformation field to each pixel coordinate of the original remote sensing image sequence, and obtaining a geometrically registered remote sensing image sequence through bicubic interpolation resampling. Geometric correction eliminates geometric distortions caused by factors such as unstable UAV flight attitude and terrain undulations, ensuring pixel-level spatial correspondence between images from different time phases and providing a foundation for subsequent accurate pixel-level change detection.

[0005] Preferably, the method employs a spatiotemporal feature coding network based on dynamic graph convolution to jointly embed the spectral features and spatial geometric features of each image in the radiometrically calibrated image set, constructing a multi-scale feature tensor pyramid. The specific steps are as follows: For each image in the radiometrically calibrated image set, a local graph structure is constructed according to spatial adjacency relationships. The nodes of the local graph structure are pixels in the image, and the edges of the local graph structure are determined by the spatial distance and spectral angle difference between pixels. In the dynamic graph convolution, the edge connections of the local graph structure are dynamically reconnected based on the cosine similarity of the feature vectors of the nodes in the feature graph output of the current convolutional layer, generating a dynamically updated graph adjacency matrix. Based on the dynamically updated graph adjacency matrix, graph convolution operations are performed layer by layer. The output of each graph convolution layer is fused with the output of the previous layer through cross-layer residual fusion and then subjected to spatial pooling to form the multi-scale feature tensor pyramid. This dynamic graph convolutional network can adaptively adjust the graph structure according to the similarity of the feature space, capture irregular ground object boundaries and spatial context relationships, while the multi-scale feature pyramid retains the spectral-geometric joint features at different spatial resolutions, providing rich multi-level information for subsequent region proposal and semantic segmentation.

[0006] Based on the multi-scale feature tensor pyramid, a heatmap of candidate change regions is generated through a learnable region proposal module, and the heatmap is subjected to pixel-by-pixel differential excitation with the temporal images in the radiometrically calibrated image set to obtain a prior map of change intensity. The specific steps are as follows: Adaptive feature alignment is performed on feature tensors at the same spatial resolution but different scales in the multi-scale feature tensor pyramid. After alignment, the alignment is input into the learnable region proposal module. The learnable region proposal module consists of two parallel convolutional sub-networks. One convolutional sub-network is used to predict the change confidence of the anchor box, and the other convolutional network is used to regress the position offset of the anchor box. Based on the change confidence and the position offset, a heatmap of the candidate change regions is generated. The value of each pixel in the heatmap represents the initial confidence probability that the pixel belongs to the change region. Two remote sensing images from different time phases are selected from the radiometrically calibrated image set. A band-by-band spectral ratio calculation is performed on the two remote sensing images from different time phases. Specifically, for the pixel values ​​of each corresponding band in the two remote sensing images from different time phases, the spectral ratio feature is calculated according to the following formula: in, Indicates in band Medium and spatial location Spectral ratio at that location These represent the first and second time-phase remote sensing images in the bands, respectively. ,Location The surface reflectance value after radiometric calibration at that location. This is a preset regularization constant used to avoid zero denominators; its value range is [value range missing]. The result of the spectral ratio calculation is multiplied by the heatmap using a Hadamard product to obtain the change intensity prior map. This step utilizes the spectral ratio calculation to highlight the magnitude of the change in ground object spectra over time. Combined with the heatmap generated by the learnable region proposal network, the resulting change intensity prior map simultaneously incorporates semantic priors from deep feature learning and numerical priors from physical spectral differences, effectively suppressing the influence of spurious changes and noise, and providing high-quality guiding information for subsequent fine segmentation. As a preferred embodiment, the two parallel convolutional sub-networks in the learnable region proposal module share the feature tensor output by the multi-scale feature tensor pyramid as input, thus reducing the number of parameters and enhancing feature reuse. Furthermore, in the learnable region proposal module, the step of jointly training the two parallel convolutional sub-networks is as follows: The feature tensors of different scales in the multi-scale feature tensor pyramid are weighted and fused to generate a fused feature map, wherein the weight coefficients of the weighted fusion are adaptively generated by a spatial attention mechanism; the fused feature map is input to the convolutional sub-network predicting the change confidence of the anchor box and the convolutional sub-network regressing the position offset of the anchor box, respectively, wherein the convolutional sub-network predicting the change confidence of the anchor box outputs the change probability score of each anchor box, and the convolutional sub-network regressing the position offset of the anchor box outputs the center coordinates and adjustment amounts of the anchor box width and height; a multi-task loss function is used to simultaneously supervise the change probability score and the adjustment amounts, wherein the classification loss in the multi-task loss function adopts a focusing loss function, and the modulation factor in the focusing loss function is dynamically adjusted according to the spatial distribution of the change intensity prior map. This training strategy integrates multi-scale features through spatial attention, enabling the region proposal module to focus on regions with significant changes. The modulation factor of the focusing loss is dynamically adjusted according to the prior map of the change intensity, so that pixels that are more difficult to classify by the model receive larger gradient weights, thereby improving the detection capability of difficult samples.

[0007] Guided by the prior map of intensity variation, an edge-guided conditional random field is used to perform iterative semantic segmentation on the radiometrically calibrated image set, outputting a probability matrix of variation pixels. The specific steps are as follows: the prior map of intensity variation is used as the initial weight of a univariate potential, which is obtained by mapping the spectral feature vectors of each pixel in the radiometrically calibrated image set through a fully connected network. The output of the univariate potential is the probability distribution of each pixel belonging to the variation class or the non-variation class. A binary potential of the edge-guided conditional random field is constructed, comprising a Gaussian kernel function based on pixel spatial location and an edge intensity-guided kernel function calculated based on the radiometrically calibrated image set. The specific form of the edge intensity-guided kernel function is as follows: in, They represent the first The pixel and the A spatial coordinate vector of pixels, Let be the Euclidean distance between the two. pixels and pixels gradient magnitude at that point The spatial distance bandwidth parameter and gradient difference bandwidth parameter are positive real numbers, respectively. The edge intensity guiding kernel function modulates the penalty term between adjacent pixels using the gradient magnitude extracted by the Sobel operator. The univariate potential energy is added to the binary potential energy to form the total energy function of the edge-guided conditional random field. The total energy function is iteratively optimized using mean-field variational inference. After each iteration, the change probability of each pixel is updated until convergence, and the change pixel probability matrix is ​​output. This edge-guided conditional random field introduces image gradient information and applies edge intensity-related constraints between spatially adjacent pixels, enabling the segmentation boundary to accurately fit the edge of the ground object. This effectively avoids the boundary blurring and oversegmentation problems that are easily caused by traditional conditional random fields. At the same time, the change intensity prior map serves as the initial weight of the univariate potential energy, providing reliable prior knowledge for segmentation, accelerating convergence and improving segmentation accuracy. As a preferred solution, during the iterative optimization process of the mean-field variational inference, the global difference between the change pixel probability matrices of two adjacent iterations is calculated after each iteration. When the global difference is less than a preset convergence threshold, the iteration is terminated early, thereby reducing computational overhead while ensuring accuracy.

[0008] Based on the changed pixel probability matrix, an adaptive threshold selection strategy is used to determine the final change detection binary image. The specific steps are as follows: Histogram statistics are performed on all probability values ​​in the changed pixel probability matrix to obtain a probability distribution curve. Two local peaks are found on the probability distribution curve, corresponding to the probability concentration regions of non-changed pixels and changed pixels, respectively. The lowest valley point of the probability distribution curve is searched between the two local peaks, and the probability value corresponding to the lowest valley point is used as the adaptive segmentation threshold. The probability value of each pixel in the changed pixel probability matrix is ​​compared with the adaptive segmentation threshold. When the probability value is greater than or equal to the adaptive segmentation threshold, the pixel is marked as a changed pixel; when the probability value is less than the adaptive segmentation threshold, the pixel is marked as a non-changed pixel, thus generating the change detection binary image. This adaptive threshold selection strategy automatically determines the optimal threshold based on the statistical characteristics of the probability distribution, eliminating the need for manual setting. It can adapt to the probability distribution characteristics under different scenarios and different change intensities, significantly improving the robustness and accuracy of the binarization results. Preferably, the two local peaks on the probability distribution curve are determined by finding the position where the first derivative of the probability density function crosses zero and the second derivative is negative. This method is simple to calculate and can stably locate the peak points.

[0009] Preferably, the ground-based radiative transfer model further includes a reflectance verification step based on synchronous ground-based measurement points: While the UAV is flying, multiple standard gray boards with known reflectance are deployed on the ground, and the actual reflectance of the standard gray boards is measured using a ground-based spectroradiometer to obtain a measured reflectance dataset; the inverted reflectance at the locations of the standard gray boards in the radiometric calibration image set is compared with the measured reflectance dataset using linear regression analysis to calculate the root mean square error and coefficient of determination of the inverted reflectance; based on the root mean square error and coefficient of determination, the atmospheric correction parameters in the ground-based radiative transfer model are corrected using the least squares method. This verification and compensation step can correct the systematic errors of the atmospheric correction parameters in real time, further improving the accuracy of radiometric calibration, thereby providing a more accurate surface reflectance input for change detection.

[0010] Preferably, after generating the change detection binary image, the method further includes a post-processing filtering step: performing connected component analysis on the change detection binary image to extract connected regions composed of all changed pixels, and calculating the total number of pixels contained in each connected region; marking connected regions with a total number of pixels less than a preset area threshold as noise regions, and flipping all pixels in the noise region from changed pixels to non-changed pixels; performing morphological closing operations on the noise-removed change detection binary image, wherein the morphological closing operation first performs a dilation operation on the changed region to fill holes, and then performs an erosion operation to restore the original edges, outputting the final change detection result image. This post-processing step removes noise isolated points and small patches of salt-and-pepper noise through area filtering, fills small holes inside the changed region through morphological closing operations, and maintains the original shape of the edges, making the final output detection result image more accurate and consistent with the spatial continuity of the actual ground feature change area.

[0011] The beneficial effects of this invention are: A learningable region proposal module combined with spectral ratio differential excitation is employed. Specifically, the learningable region proposal module predicts the change confidence and positional offset of anchor boxes in parallel from a multi-scale feature tensor pyramid, generating a heatmap of candidate change regions. Spectral ratio calculation calculates ratio features for each corresponding band pixel value of two temporal remote sensing images. By introducing a preset regularization constant to avoid zero denominators, the ratio result is multiplied with the heatmap using a Hadamard product to obtain a change intensity precursor map. This approach differs from the conventional method of directly using depth features to predict change probabilities. It utilizes the ratio response at the physical level of multispectral bands to filter and enhance the confidence of region proposals. This ensures that high-response regions in the heatmap possess both deep semantic features and spectral physical consistency, effectively suppressing spurious change responses caused by seasonal fluctuations in illumination, shadows, or vegetation, thus improving the accuracy and reliability of change region localization.

[0012] An iterative semantic segmentation technique using edge-guided conditional random fields (CRFs) for radiometrically calibrated image sets is proposed. The binary potential energy of this CRF includes a Gaussian kernel function based on pixel spatial location and an edge intensity-guided kernel function based on gradient magnitude. The edge intensity-guided kernel function modulates the penalty term between adjacent pixels using gradient magnitudes extracted by the Sobel operator. Pixel pairs with large gradient magnitude differences are penalized with smaller penalty terms to preserve edges, while pixel pairs with small gradient magnitude differences are penalized with larger penalty terms to promote smoothing of homogeneous regions. The unary potential energy uses a leading map of varying intensity as initial weights and obtains the probability distribution by mapping the spectral features of each pixel through a fully connected network. The total energy function is iteratively optimized through mean-field variational inference until convergence, outputting a probability matrix of varying pixels. This scheme differs from conventional conditional random fields that rely solely on color and spatial distance. It proactively introduces ground feature edge information as a potential energy modulation factor, maintaining the clarity of the boundaries between changing and non-changing regions during iterative inference, thus avoiding over-segmentation or edge blurring. Simultaneously, it utilizes a lead graph to constrain the univariate potential energy, ensuring that the segmentation results are consistent with spectral physical differences at the semantic level. This allows for precise edge localization and reliable estimation of changing pixel probabilities without requiring a large amount of labeled data. Attached Figure Description

[0013] The invention will now be further described with reference to the accompanying drawings.

[0014] Figure 1 This is a flowchart of the method for calibration and intelligent segmentation change detection of UAV remote sensing images; Figure 2 This is a flowchart of edge-guided conditional random field iterative semantic segmentation; Figure 3 This is a flowchart of the adaptive threshold change detection binary image generation process; Figure 4 This is a flowchart of an atmospheric correction and change detection method based on a ground-based radiative transfer model. Detailed Implementation

[0015] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the scope of protection of the present invention.

[0016] See appendix Figure 1This invention provides a method for calibration and intelligent segmentation of UAV remote sensing images for change detection, comprising: acquiring an original remote sensing image sequence using a multispectral sensor mounted on a UAV, and performing atmospheric correction based on a ground object radiative transfer model on the original remote sensing image sequence to generate a radiometrically calibrated image set; employing a spatiotemporal feature coding network based on dynamic graph convolution to jointly embed the spectral features and spatial geometric features of each image in the radiometrically calibrated image set to construct a multi-scale feature tensor pyramid; generating a heatmap of candidate change regions based on the multi-scale feature tensor pyramid through a learnable region proposal module, and performing pixel-by-pixel differential excitation between the heatmap and the temporal images in the radiometrically calibrated image set to obtain a change intensity prior map; guided by the change intensity prior map, performing iterative semantic segmentation of the radiometrically calibrated image set using an edge-guided conditional random field to output a change pixel probability matrix; and determining the final change detection binary map based on the change pixel probability matrix using an adaptive threshold selection strategy. Example

[0017] In specific implementation, the method for performing atmospheric correction based on the ground object radiative transfer model on the original remote sensing image sequence to generate a radiocalibrated image set includes the following steps: First, extract the digital quantization value of each pixel in the original remote sensing image sequence. The original remote sensing image sequence is acquired by a multispectral sensor mounted on a UAV, and the digital quantization value of each pixel is the original recorded value of that pixel in each band. Using the pre-calibrated spectral response function of the multispectral sensor mounted on the UAV, the digital quantization value is converted into the corresponding band's top-atmosphere radiance value. The spectral response function describes the response characteristics of each band of the sensor to radiant energy and is obtained through laboratory calibration or field calibration. Top-atmosphere radiance value Indicates band Medium and spatial location The radiance received by the sensor, in units of Next, the atmospheric top-layer radiance value is input into the ground-object radiative transfer model. The ground-object radiative transfer model employs classic atmospheric radiative transfer models such as MODTRAN or 6SV. The model inverts atmospheric upward radiation based on synchronously acquired atmospheric aerosol optical thickness and water vapor content. Atmospheric downward radiation and solar direct radiation transmittance Atmospheric aerosol optical thickness and water vapor content were collected in real time by the meteorological detection module of the multispectral sensor carried by the UAV. The atmospheric upward radiation... This refers to the upward radiance emitted by the atmosphere itself; the downward radiance of the atmosphere. The solar transmittance represents the radiance emitted downwards from the atmosphere to the ground; This represents the transmittance of direct solar radiation after atmospheric attenuation. Then, the atmospheric upward radiation is used. Atmospheric downward radiation and solar direct radiation transmittance The surface reflectance is inverted from the radiance value of the upper atmosphere. The inversion formula is: in, Indicates band Medium and spatial location Surface reflectance at that location; The solar irradiance at the top of the atmosphere is provided in the sensor calibration file, and the unit is 1. The solar zenith angle is calculated from the UAV flight time and geographic coordinates. Through the above inversion, the radiometric calibration image set is obtained, which contains multiple surface reflectance images from different time phases. In a specific implementation, a spatiotemporal feature coding network based on dynamic graph convolution is used to jointly embed the spectral features and spatial geometric features of each image in the radiometric calibration image set, constructing a multi-scale feature tensor pyramid. The specific implementation method includes the following steps: First, for each image in the radiometric calibration image set, a local graph structure is constructed according to spatial adjacency relationships. The nodes of the local graph structure are pixels in the image, with each pixel serving as a node. The edges of the local graph structure are determined by the spatial distance and spectral angle differences between pixels. Specifically, for any pixel... Its spatial coordinates are Spectral eigenvectors are (Dimension equals number of bands), then with pixels Pixels with edge connections Two conditions must be met simultaneously: spatial Euclidean distance Smaller than the preset spatial radius (Values ​​are taken from 3 pixels), and the spectral angle Less than the preset spectral angle threshold (Value is 0.1 radians). An undirected edge is established between pixels that satisfy the above two conditions. Next, in the dynamic graph convolution, the edge connections of the local graph structure are dynamically reconnected. The basis for dynamic reconnection is the cosine similarity of the feature vectors of the nodes in the feature map output by the current convolutional layer. Let the th... The feature map output by the layer graph convolution is in The total number of nodes. For the first The feature dimension of a layer. For any two nodes... Calculate its eigenvectors cosine similarity For each node The k neighbor nodes with the highest cosine similarity (k takes the value of 8) are retained as the neighbors after reconnection, and the remaining edges are removed, thus generating a dynamically updated graph adjacency matrix. Then, based on the dynamically updated graph adjacency matrix... Perform graph convolution operations layer by layer. The graph convolution operation is defined as follows: in, It is the identity matrix; It is a diagonal matrix, with diagonal elements For the first The learnable weight matrix of the layer;

[0018] Activation function. Output of each graph convolution layer. All are the same as the output of the previous layer Perform cross-layer residual fusion, the fusion method is as follows The fused feature tensors are then subjected to spatial pooling. Max pooling is used, with a pooling window size of [value missing]. With a stride of 2, the spatial resolution is reduced by half. After multi-layer graph convolution and spatial pooling (4 layers, with output feature dimensions of 64, 128, 256, and 512 respectively), the multi-scale feature tensor pyramid is formed. This multi-scale feature tensor pyramid contains feature tensors with different spatial resolutions, each corresponding to one output layer. The spatial resolution gradually decreases from the original resolution of the input image, while the number of feature channels gradually increases. Optionally, before executing the spatiotemporal feature encoding network based on dynamic graph convolution, each image in the radiometrically calibrated image set is normalized, linearly scaling the surface reflectance value of each band to a different scale. The interval is used to accelerate network training convergence. It can be understood that the input of the spatiotemporal feature encoding network of the dynamic graph convolution is a single image from the radiometrically calibrated image set (or a three-dimensional tensor formed by stacking temporal images), and the output is the multi-scale feature tensor pyramid. Each layer of the multi-scale feature tensor pyramid retains the spectral and spatial geometric features of the original image, and adaptive modeling of irregular spatial adjacency relationships is achieved through dynamic graph convolution. Example

[0019] In specific implementation, based on the multi-scale feature tensor pyramid, a heatmap of candidate change regions is generated through a learnable region proposal module. The heatmap is then subjected to pixel-by-pixel differential excitation with the temporal images in the radiometrically calibrated image set to obtain a priori map of change intensity. The specific implementation method includes the following steps: The multi-scale feature tensor pyramid adaptively aligns feature tensors at the same spatial resolution but different scales. The multi-scale feature tensor pyramid contains multiple scale layers, each corresponding to a spatial resolution. For feature tensors at the same spatial resolution (e.g., the original image spatial resolution or a resolution after pooling) but from different scale layers (e.g., the outputs of different graph convolutional layers), bilinear interpolation is used to upsample the lower spatial resolution feature tensor to the higher spatial resolution size, ensuring that all aligned feature tensors have the same spatial size and number of channels. The aligned feature tensors are then concatenated along the channel dimension to form a unified feature tensor.

[0020] Optionally, in the adaptive feature alignment process, the high spatial resolution feature tensor output from the bottom layer of the multi-scale feature tensor pyramid is fused with the low spatial resolution, high semantic feature tensor output from the top layer to preserve spatial details and semantic information.

[0021] The aligned, unified feature tensor is input to the learnable region proposal module. The learnable region proposal module consists of two parallel convolutional sub-networks. One convolutional sub-network predicts the change confidence of the anchor boxes, and the other convolutional network regresses the positional offset of the anchor boxes. For each spatial location of the anchor box on the feature map, multiple rectangular boxes of different scales and aspect ratios are predefined, and the center of each anchor box is aligned with the pixel center on the feature map. The convolutional sub-network used to predict the change confidence of the anchor boxes employs a... A convolutional layer followed by a Convolutional layer, with the number of output channels being in Number of anchor points predefined for each spatial location 2 represents the confidence score for each anchor box to belong to the changed or unchanged class. The convolutional subnetwork used to regress the positional offset of the anchor boxes also uses a... A convolutional layer followed by a Convolutional layer, with the number of output channels being 4 indicates the offset of the anchor point frame center coordinates. and width and height scaling

[0022] Based on the change confidence level and the position offset, each anchor point frame is corrected: the center coordinates of the corrected anchor point frame are... The corrected width is The corrected height is in The coordinates, width, and height of the original anchor boxes are defined as follows: Anchor boxes with a change confidence score higher than a preset confidence threshold (set to 0.5) are selected as candidate change regions. All candidate change regions are spatially mapped back to the original image size, generating a heatmap of these candidate change regions. The value of each pixel in the heatmap represents the initial confidence probability that the pixel belongs to a change region. The spatial resolution of the heatmap is consistent with the original image. For pixels belonging to multiple overlapping candidate change regions, the highest change confidence score among all candidate change regions covering that pixel is taken as the initial confidence probability of that pixel.

[0023] In specific implementation, the two parallel convolutional sub-networks in the learnable region proposal module share the feature tensor output by the multi-scale feature tensor pyramid as input. Specifically, both convolutional sub-networks use the aligned and concatenated unified feature tensor as input, without using different feature sources individually. Optionally, the method for jointly training the two parallel convolutional sub-networks in the learnable region proposal module includes the following steps: First, the feature tensors of different scales in the multi-scale feature tensor pyramid are weighted and fused to generate a fused feature map. The weighted fusion process employs a spatial attention mechanism: for each layer of feature tensor in the multi-scale feature tensor pyramid... Firstly, through a Convolutional layers unify the number of channels to a fixed dimension. Obtain the adjusted feature tensor Then, the adjusted feature tensors of all layers are upsampled to the maximum spatial resolution using bilinear interpolation. The upsampled feature tensors are concatenated along the spatial dimensions to form a structure with dimension 1. A stacked tensor, where L is the number of pyramid levels (L is set to 4). This stacked tensor is input to a spatial attention module, which consists of a... The convolutional layer consists of a sigmoid activation function and outputs a spatial attention weight map. Then, the adjusted feature tensors of each layer Element-wise multiplication by spatial attention weight map The corresponding position values ​​are used to obtain the weighted feature tensors of each layer. All weighted feature tensors of each layer are summed along the channel dimension to generate a fused feature map. Next, the fused feature map The convolutional subnetworks for predicting the change confidence of the anchor boxes and regressing the position offset of the anchor boxes are respectively input into the convolutional subnetworks. The convolutional subnetwork for predicting the change confidence of the anchor boxes outputs a change probability score for each anchor box. The convolutional subnetwork that calculates the position offset of the regression anchor frame outputs the center coordinates and the adjustment amounts of the width and height of the anchor frame. Then, a multi-task loss function is used to simultaneously monitor the change probability score. and the adjustment amount The multi-task loss function Defined as classification loss Regression loss Weighted sum: in Set as The regression loss is set to 2. The SmoothL1 loss function is used. The classification loss... A focusing loss function is employed, in which the modulation factor is dynamically adjusted based on the spatial distribution of the prior map of intensity changes. The focusing loss function is defined as follows: in, This represents the total number of anchor point frames; This represents the probability that the anchor box predicted by the model belongs to the true category. If the true label of the anchor box is the variation class, then... For the probability score of change, if the true label is the invariant class, then for Set the category balancing weight to 0.25; For the dynamic modulation factor, for the first An anchor box with dynamic modulation factor Defined as: in, The reference modulation factor is set to 2; The intensity value of the pixel at the center of the corresponding anchor box in the prior image of the change intensity (the value ranges from 0 to 1). For the first The center coordinates of each anchor box are mapped to the original image size. Through dynamic adjustment, anchor boxes in low-intensity regions (i.e., regions with low probability of change) in the prior image of change intensity receive larger modulation factors, making the model pay more attention to difficult samples; anchor boxes in high-intensity regions in the prior image of change intensity receive smaller modulation factors, reducing attention to easily classified samples.

[0024] In specific implementation, two remote sensing images from different time phases are selected from the radiometrically calibrated image set, and band-by-band spectral ratio calculations are performed on the two remote sensing images from different time phases. Let the first time phase remote sensing image be... The second phase remote sensing image is Both have the same number of bands and spatial resolution. For each band Regarding spatial location The spectral ratio feature of the pixel at that location is calculated according to the following formula. in, This indicates the first phase of remote sensing imagery in the band. ,Location The surface reflectance values ​​after radiometric calibration at the location range from 0 to... . This indicates that the second-phase remote sensing image is in the band. ,Location The surface reflectance value after radiometric calibration at the location ranges from 0 to 1. This is a preset regularization constant used to avoid the denominator being zero. The range of values ​​is Specifically, it is set to 0.001.

[0025] The result of the spectral ratio calculation is then multiplied by the heatmap using a Hadamard product operation. The result of the spectral ratio calculation is a multiband tensor. in The total number of bands, and The image height and width are specified in the heatmap. This is a single-channel tensor. The Hadamard product operation is defined as: converting the heatmap... Copy along channel dimension Next, make its spatial dimensions consistent with the spectral ratio tensor, and then multiply it element by element: in, The prior diagram representing the intensity of change in the band Location The intensity value at that location. Finally, the prior map of the changing intensity is obtained. The prior map of change intensity integrates the spatial confidence and spectral ratio information of candidate change regions. The value of each pixel and each band reflects the prior information of the change intensity at that location in that band. Example

[0026] For specific implementation details, please refer to the appendix. Figure 2 Guided by the prior map of intensity variation, the radiometrically calibrated image set is iteratively segmented using an edge-guided conditional random field to output the probability matrix of varying pixels. The specific implementation method includes the following steps: The prior map of the change intensity The initial weights for the unary potential energy are used. This unary potential energy is obtained by mapping the spectral feature vectors of each pixel in the radiometrically calibrated image set through a fully connected network. Each pixel in each image of the radiometrically calibrated image set has a spectral feature vector, which is composed of the surface reflectance values ​​of that pixel across all bands, with a dimension of [missing information]. The fully connected network consists of two fully connected layers. The first fully connected layer transforms the input spectral feature vector from one dimension to the other. Mapped to dimension 128, using The activation function; the second fully connected layer maps dimension 128 to dimension 2, outputting two values ​​representing the log odds of belonging to the changed class and the unchanged class, respectively. The output of the fully connected network is then processed... The function yields the probability distribution of each pixel belonging to either the changed or unchanged class. The univariate potential energy... Defined as:

[0027] in, Indicates the first The spectral feature vector of each pixel; Indicates the first Category labels per pixel, 0 represents a non-mutable class, and 1 represents a mutable class; The first prediction of a fully connected network Each pixel belongs to the category The probability value; For the first The initial weights for each pixel are taken from the prior map of change intensity. In the corresponding spatial location The multi-band average intensity value at a given location is calculated as follows:

[0028] The output of the univariate potential energy is the probability distribution of each pixel belonging to the changing or non-changing class.

[0029] Next, the binary potential energy of the edge-guided conditional random field is constructed. This binary potential energy includes a Gaussian kernel function based on pixel spatial location and an edge intensity-guided kernel function calculated based on the radiometrically calibrated image set. The specific form of the edge intensity-guided kernel function is as follows:

[0030] in, Indicates the first A spatial coordinate vector of pixels, in the form of: Indicates the first A spatial coordinate vector of pixels, in the form of: Indicates the first The pixel and the Euclidean spatial distance between pixels, in pixels; Indicates the first The gradient magnitude at each pixel is calculated as follows: the first and second phase images in the radiometric calibration image set are converted into single-channel grayscale images (calculated by averaging the values ​​of each band). For each phase grayscale image, the Sobel operator is used to extract the horizontal and vertical gradients, and then the gradient magnitude is calculated. The maximum value of the two phase gradient magnitudes is taken as the gradient magnitude at each pixel. The final value; Indicates the first The gradient magnitude at each pixel is obtained in the same way as... same; The spatial distance bandwidth parameter is set to 3.0 pixels, determined based on the spatial resolution of the image (set to 0.1 meters / pixel) and the desired neighborhood range, so that the kernel function value decays to approximately 0.135 when the spatial distance exceeds 3 pixels; The gradient difference bandwidth parameter is set to 0.1, determined based on the normalized range of the gradient magnitude (normalized to [0,1]), so that the kernel function value decays to approximately 0.135 when the gradient difference exceeds 0.1. The edge intensity guiding kernel function modulates the penalty term between adjacent pixels using the gradient magnitude extracted by the Sobel operator: the kernel function value between adjacent pixels with large gradient magnitude differences is smaller, thus imposing a smaller penalty in the binary potential energy.

[0031] The binary potential energy Defined as:

[0032] in, For tag compatibility functions, use Model, i.e. when and The value is 1 if they cannot both be the same, and 0 if they are the same. Here, is the number of kernel functions. Set to 1; For the first The weights of each kernel function, Set to 1.0; This is the edge strength guiding kernel function.

[0033] Adding the univariate potential energy to the binary potential energy constitutes the total energy function of the edge-guided conditional random field. in, This represents the category label vector for all pixels. The total energy function is iteratively optimized using mean-field variational inference. The mean-field variational inference approximates the posterior probability distribution. Iteratively update the change probability of each pixel. The iterative update formula is:

[0034] in, Indicates the number of iterations; As a normalization constant, ensure Indicates the first After the nth iteration Each pixel belongs to the category The probability of the initial state; From the univariate potential energy The following steps are given. After each iteration, the probability value of each pixel is updated. The iteration continues until convergence, and the changed pixel probability matrix is ​​output. ,in Indicates the first The probability that a pixel belongs to the change class. This represents the probability of belonging to the non-changing class.

[0035] Optionally, in the iterative optimization process of the mean-field variational inference, the global difference between the changed pixel probability matrices of two adjacent iterations is calculated after each iteration. The global difference... Defined as: in, Indicates the first After the nth iteration The probability that a pixel belongs to the change class; Indicates the first After the nth iteration The probability that a pixel belongs to the change class. When the global difference... Less than the preset convergence threshold If the iteration is terminated early, the current probability matrix of the changed pixels is output. This is the final output. If the convergence threshold is not reached, the iteration continues until the maximum number of iterations is reached. Example

[0036] See appendix Figure 3 In specific implementation, the method for determining the final change detection binary image based on the changed pixel probability matrix using an adaptive threshold selection strategy includes the following steps: First, the changed pixel probability matrix Histogram analysis was performed on all probability values ​​to obtain the probability distribution curve. The changed pixel probability matrix... Each pixel in the image corresponds to two probability values: the probability of belonging to the change class. The probability of belonging to the non-changing class. Histogram statistics represent the probability value of each pixel belonging to the change class. Perform, all The probability values ​​are divided into numerical ranges. equidistant boxes The number of pixels in each box is counted to form the probability distribution curve. ,in This represents the probability value of the change class, and its value range is... , This indicates the number of pixels whose probability value falls within the corresponding box.

[0037] The probability distribution curve Two local peaks are identified on the probability distribution curve. These two local peaks correspond to the probability concentration regions of non-changing pixels and changing pixels, respectively. The two local peaks on the probability distribution curve are determined by finding the positions where the first derivative of the probability density function crosses zero and the second derivative is negative. Specifically, for the probability distribution curve... Smoothing is performed by convolving the histogram with a Gaussian filter, and the standard deviation of the Gaussian kernel is used. The box width was set to 2.0 to reduce spurious peaks caused by noise. The smoothed curve is denoted as... calculate first derivative and second derivative The first derivative is approximated using the central difference: in For the first The central probability value of each box. The second derivative is used. In satisfying and The location was determined as a local peak point. Obtain all local peak points within the range of 0 to 1. Select the local peak point with the lowest probability value as the peak point corresponding to the non-changing class. Select the local peak point with the highest probability value as the peak point corresponding to the change class. Set to 1.0.

[0038] In the two local peaks and The search range is from the lowest trough point of the probability distribution curve. The corresponding box index The corresponding box index is used to search within the specified range. The probability value corresponding to the minimum value is the lowest point of the probability distribution curve. The probability value corresponding to the lowest valley point As an adaptive segmentation threshold

[0039] The changed pixel probability matrix The probability value of the change class of each pixel With the adaptive segmentation threshold Comparison. When Greater than or equal to the adaptive segmentation threshold When, mark the pixel as a changed pixel; when Less than the adaptive segmentation threshold When this happens, mark the pixel as a non-changing pixel. Traverse all spatial locations. Generate the change detection binary map in Indicates the change in pixels. Indicates non-changing pixels.

[0040] Optionally, the two local peaks on the probability distribution curve are determined by finding the points where the first derivative of the probability density function crosses zero and the second derivative is negative; the specific implementation method is already included in the above description. When the probability distribution curve has multiple peaks, only the two peaks with the smallest and largest probability values ​​are selected as the two local peaks, and other intermediate peaks are ignored. Example

[0041] See appendix Figure 4 In specific implementation, during the atmospheric correction based on the ground-object radiative transfer model, the atmospheric aerosol optical thickness and water vapor content are simultaneously acquired and collected in real time by the meteorological detection module integrated into the multispectral sensor mounted on the UAV. The meteorological detection module, integrated on the UAV platform, includes an aerosol optical thickness sensor and a water vapor content sensor. The aerosol optical thickness sensor employs the principle of a multi-band solar photometer, continuously measuring the intensity of direct solar radiation during flight and retrieving the atmospheric aerosol optical thickness value through a built-in algorithm. The water vapor content sensor uses the near-infrared water vapor absorption band measurement principle, retrieving the atmospheric column water vapor content by measuring the attenuation of specific water vapor absorption bands. The acquired atmospheric aerosol optical thickness and water vapor content data are strictly aligned with the exposure times of the original remote sensing image sequence and are used as input parameters to the ground-object radiative transfer model for calculating atmospheric upward radiation, atmospheric downward radiation, and direct solar radiation transmittance.

[0042] Optionally, before performing the atmospheric correction based on the ground-object radiative transfer model, a step of geometric fine correction of the original remote sensing image sequence is included. Feature points are extracted from the original remote sensing image sequence, and a scale-invariant feature transform algorithm is used to detect feature points in each image to generate a descriptor for each feature point. The feature points are matched with a pre-stored reference digital orthophoto image, which is an orthorectified image with known geographic coordinates. The matching process uses the nearest neighbor distance ratio method: for each feature point, the two feature points with the closest Euclidean distance are searched in the reference digital orthophoto image. If the ratio of the nearest distance to the second nearest distance is less than a preset matching threshold (the matching threshold is set to 0.8), the feature point corresponding to the nearest distance is taken as the matching point, forming a corresponding point pair. The above matching process is repeated for all images to obtain a set of corresponding point pairs. Each corresponding point pair contains the coordinates of a feature point in the original image and the corresponding geographic coordinates on the reference digital orthophoto image.

[0043] Based on the set of corresponding point pairs, a thin-plate spline function is used to fit the geometric deformation field of the original remote sensing image sequence relative to the reference digital orthophoto map. The basis functions of the thin-plate spline function are radial basis functions. in This represents the distance between control points. The fitting process solves for the thin-plate spline parameters, obtaining a mapping function from the original image coordinates to the reference geographic coordinates. and These represent the longitude and latitude values ​​after coordinate mapping, respectively. The geometric deformation field is applied to each pixel coordinate of the original remote sensing image sequence. For each pixel coordinate in the original image... Calculate its target coordinates in the reference coordinate system. By resampling using bicubic interpolation, pixel values ​​are calculated at the target coordinates to obtain a geometrically registered remote sensing image sequence. The geometrically registered remote sensing image sequence has similarity to the reference digital orthophoto image. Figure 1 Achieving consistent georeference and spatial resolution.

[0044] In its implementation, the ground-based radiative transfer model also includes a step of reflectivity verification based on synchronous ground-based measurement points. During the drone's flight, multiple standard gray boards with known reflectivities are deployed on the ground. The reflectivity values ​​of these standard gray boards are laboratory-calibrated, exhibiting a flat spectral response in the visible and near-infrared bands, with a range of 0.2 to 0.8. Five standard gray boards are evenly distributed within the drone's flight coverage area. The actual reflectivity of the standard gray boards is measured using a portable ground-based spectroradiometer, whose spectral range covers all bands of the multispectral sensor. During measurement, multiple measurements (10 times) are performed at each standard gray board location, and the average value is calculated to obtain the measured reflectivity dataset. in This indicates the band index, and k indicates the standard gray card number.

[0045] The inversion reflectance at the location corresponding to the standard gray card in the radiometrically calibrated image set is extracted. The radiometrically calibrated image set is an atmospherically corrected surface reflectance image. The corresponding pixel location on the image is found using the geographic coordinates of the standard gray card, and the inversion reflectance values ​​for each band of that pixel are obtained. The inverted reflectance With the measured reflectance dataset Perform linear regression analysis. The linear regression model is as follows: in For band Gain coefficient, For band The offset coefficient. The gain coefficient is solved using the least squares method. and offset coefficient The root mean square error of the inverted reflectance was calculated. in Standard gray board quantity Calculate the coefficient of determination The average of the measured reflectance of all standard gray boards.

[0046] Based on the root mean square error and the coefficient of determination, the atmospheric correction parameters in the ground-feature radiative transfer model are corrected using the least squares method for secondary compensation. Specifically, the gain coefficient is... and offset coefficient As a correction factor, the surface reflectance of all pixels in the radiometrically calibrated image set is compensated, and the compensation formula is as follows: Concentrated bands of the original radiometric calibration image Location Inversion reflectance at that location This is the compensated reflectance value. If the determining factor is... If the value is greater than 0.95, the compensation is deemed effective and applied; otherwise, the original inverted reflectance value remains unchanged.

[0047] In specific implementations, after generating the change detection binary image, the process further includes a post-processing filtering step on the change detection binary image. Connectivity analysis is performed, using the 8-neighborhood connectivity criterion to label all connected regions formed by changing pixels. Connectivity analysis involves scanning the change detection binary map; for each changing pixel with a value of 1, recursively searching for all pixels with a value of 1 within its 8-neighborhood and merging them into the same connected region. All connected regions are extracted, and each region is assigned a unique region identifier. The total number of pixels contained in each connected region is calculated. in For connected region indexes.

[0048] The total number of pixels Less than the preset area threshold Connected regions are marked as noise regions. The preset area threshold... The threshold is set to 100 pixels. This value is determined based on the spatial resolution of the UAV imagery (0.1 m / pixel) and the actual physical size of the smallest changing ground feature (the smallest changing target size corresponds to 10 square meters, or 1000 pixels; however, considering that noisy areas are usually composed of a single or a few pixels, 100 pixels is sufficient as a noise filtering threshold to remove isolated pixels while retaining meaningful small targets). All pixels in the noisy area are flipped from changed pixels to unchanged pixels, i.e., in the change detection binary image... In the middle, change the value at the corresponding position from 1 to 0.

[0049] A morphological closing operation is performed on the noise-removed binary image of the change detection. This morphological closing operation first dilates the changed region, then performs an erosion operation. The dilation operation uses... The rectangular structuring element is used to traverse the noise-removed binary map of change detection. For each pixel, if a changed pixel exists in its neighborhood, the pixel is marked as a changed pixel, and small holes inside the changed region are filled. The erosion operation uses the same... A rectangular structuring element is used to traverse the dilated binary image. For each pixel, the condition is met only if the pixel and its corresponding element are equal. A pixel is retained as a changed pixel only if all pixels in its neighborhood are changed pixels; otherwise, it is marked as a non-changed pixel to restore the original edge. The closing operation outputs the final change detection result image, which is the final change detection binary image.

[0050] The foregoing has provided a detailed description of one embodiment of the present invention, but this description is merely a preferred embodiment and should not be construed as limiting the scope of the invention. All equivalent variations and modifications made within the scope of the claims of this invention should still fall within the patent coverage of this invention.

Claims

1. A method for calibration and intelligent segmentation change detection of UAV remote sensing images, characterized in that, The method includes: The original remote sensing image sequence is acquired by a multispectral sensor carried by a drone, and atmospheric correction based on the ground object radiative transfer model is performed on the original remote sensing image sequence to generate a radiometrically calibrated image set. A spatiotemporal feature coding network based on dynamic graph convolution is used to jointly embed the spectral features and spatial geometric features of each image in the radiometrically calibrated image set to construct a multi-scale feature tensor pyramid. Based on the multi-scale feature tensor pyramid, a heat map of candidate change regions is generated through a learnable region proposal module, and the heat map is subjected to pixel-by-pixel differential excitation with the temporal images in the radiometrically calibrated image set to obtain a prior map of change intensity. Guided by the prior map of the changing intensity, the radiometrically calibrated image set is subjected to iterative semantic segmentation using an edge-guided conditional random field, and the changing pixel probability matrix is ​​output. Based on the changed pixel probability matrix, the final change detection binary image is determined through an adaptive threshold selection strategy.

2. The method for calibration and intelligent segmentation change detection of UAV remote sensing images according to claim 1, characterized in that, The specific steps for acquiring the original remote sensing image sequence using a multispectral sensor mounted on a UAV, and performing atmospheric correction based on a ground-object radiative transfer model on the original remote sensing image sequence to generate a radiometrically calibrated image set are as follows: The digital quantization value of each pixel is extracted from the original remote sensing image sequence, and combined with the pre-calibrated spectral response function of the multispectral sensor carried by the UAV, the digital quantization value is converted into the radiance value of the top atmospheric layer of the corresponding band. The atmospheric top radiance value is input into the ground object radiative transfer model, which inversely calculates atmospheric upward radiation, atmospheric downward radiation, and solar direct radiation transmittance based on the synchronously acquired atmospheric aerosol optical thickness and water vapor content. By using the atmospheric upward radiation, atmospheric downward radiation, and solar direct radiation transmittance, the surface reflectance inversion of the atmospheric top radiance value is performed to obtain the radiometric calibration image set.

3. The method for calibration and intelligent segmentation change detection of UAV remote sensing images according to claim 2, characterized in that, The specific steps for constructing a multi-scale feature tensor pyramid by jointly embedding the spectral and spatial geometric features of each image in the radiometrically calibrated image set using a spatiotemporal feature coding network based on dynamic graph convolution are as follows: For each image in the radiometrically calibrated image set, a local graph structure is constructed according to the spatial adjacency relationship. The nodes of the local graph structure are the pixels in the image, and the edges of the local graph structure are determined by the spatial distance and spectral angle difference between the pixels. In the dynamic graph convolution, the edge connections of the local graph structure are dynamically reconnected. The reconnection is based on the cosine similarity of the feature vectors of the nodes in the feature graph output by the current convolutional layer, generating a dynamically updated graph adjacency matrix. Based on the dynamically updated graph adjacency matrix, graph convolution operations are performed layer by layer. The output of each graph convolution layer is fused with the output of the previous layer through cross-layer residual fusion and then subjected to spatial pooling to form the multi-scale feature tensor pyramid.

4. The method for calibration and intelligent segmentation change detection of UAV remote sensing images according to claim 3, characterized in that, Based on the multi-scale feature tensor pyramid, a heatmap of candidate change regions is generated through a learnable region proposal module. The heatmap is then subjected to pixel-by-pixel differential excitation with the temporal images in the radiometrically calibrated image set to obtain a priori map of change intensity. The specific steps are as follows: The feature tensors in the multi-scale feature tensor pyramid that are at the same spatial resolution but different scales are adaptively aligned. After alignment, they are input into the learnable region proposal module. The learnable region proposal module consists of two parallel convolutional sub-networks. One convolutional sub-network is used to predict the change confidence of the anchor box, and the other convolutional network is used to regress the position offset of the anchor box. Based on the change confidence level and the position offset, a heatmap of the candidate change region is generated, wherein the value of each pixel in the heatmap represents the initial confidence probability that the pixel belongs to the change region; Two remote sensing images from different time phases are selected from the radiometrically calibrated image set. A band-by-band spectral ratio calculation is performed on the two remote sensing images from different time phases. The result of the spectral ratio calculation is then multiplied by the Hadamard product with the heat map to obtain the a priori map of the change in intensity.

5. The method for calibration and intelligent segmentation change detection of UAV remote sensing images according to claim 4, characterized in that, In the learnable region proposal module, two parallel convolutional sub-networks share the feature tensor output by the multi-scale feature tensor pyramid as input.

6. The method for calibration and intelligent segmentation change detection of UAV remote sensing images according to claim 4, characterized in that, The specific steps for performing iterative semantic segmentation on the radiometrically calibrated image set using an edge-guided conditional random field, guided by the prior map of the intensity variation, and outputting the probability matrix of the variation pixels are as follows: The prior map of the change intensity is used as the initial weight of the unary potential energy, which is obtained by mapping the spectral feature vectors of each pixel in the radiometrically calibrated image set through a fully connected network. The output of the unary potential energy is the probability distribution of each pixel belonging to the change class or the non-change class. The binary potential energy of the edge-guided conditional random field is constructed. The binary potential energy includes a Gaussian kernel function based on the pixel spatial location and an edge intensity-guided kernel function calculated based on the radiometrically calibrated image set. The edge intensity-guided kernel function modulates the penalty term between adjacent pixels by the gradient magnitude extracted by the Sobel operator. The univariate potential energy is added to the binary potential energy to form the total energy function of the edge-guided conditional random field. The total energy function is iteratively optimized using mean-field variational inference. The change probability of each pixel is updated after each iteration until convergence, and the change pixel probability matrix is ​​output.

7. The method for calibration and intelligent segmentation change detection of UAV remote sensing images according to claim 6, characterized in that, In the iterative optimization process of the mean-field variational inference, after each iteration, the global difference of the changed pixel probability matrix between two adjacent iterations is calculated, and the iteration is terminated in advance when the global difference is less than a preset convergence threshold.

8. The method for calibration and intelligent segmentation change detection of UAV remote sensing images according to claim 6, characterized in that, Based on the changed pixel probability matrix, the specific steps for determining the final change detection binary image using an adaptive threshold selection strategy are as follows: Histogram statistics are performed on all probability values ​​in the variable pixel probability matrix to obtain a probability distribution curve. Two local peaks are found on the probability distribution curve. The two local peaks correspond to the probability concentration areas of non-variable pixels and the probability concentration areas of variable pixels, respectively. Search for the lowest valley point of the probability distribution curve between the two local peaks, and use the probability value corresponding to the lowest valley point as the adaptive segmentation threshold; The probability value of each pixel in the changed pixel probability matrix is ​​compared with the adaptive segmentation threshold. When the probability value is greater than or equal to the adaptive segmentation threshold, the pixel is marked as a changed pixel. When the probability value is less than the adaptive segmentation threshold, the pixel is marked as a non-changed pixel, and the change detection binary map is generated.

9. The method for calibration and intelligent segmentation change detection of UAV remote sensing images according to claim 8, characterized in that, The two local peaks on the probability distribution curve are determined by finding the position where the first derivative of the probability density function crosses zero and the second derivative is negative.

10. The method for calibration and intelligent segmentation change detection of UAV remote sensing images according to claim 1, characterized in that, In the atmospheric correction based on the ground object radiative transfer model, the atmospheric aerosol optical thickness and water vapor content are simultaneously acquired and collected in real time by the meteorological detection module of the multispectral sensor carried by the UAV.

Images