A prior feature assisted directional coordinate attention remote sensing road extraction method

Abstract

The application discloses a prior feature auxiliary directional coordinate attention remote sensing road extraction method, and relates to the technical field of remote sensing image processing.The method comprises the following steps: extracting multi-source prior features representing road physical characteristics and optimizing and fusing weight to obtain a fused prior feature map; processing the RGB remote sensing image and the fused prior feature cooperatively to form a fused feature containing multi-scale semantic information and physical prior; predicting a pixel main direction angle in the feature through a directional coordinate attention mechanism, pooling the feature along the main direction and the orthogonal direction to generate an attention map and re-calibrate the original feature; performing step-by-step up-sampling on the decoder and introducing a deep supervision mechanism; and training the model by using a composite loss function containing a Dice loss, a Focal loss and a boundary perception loss.The application fully utilizes road physical characteristic prior knowledge, combines the directional coordinate attention mechanism to capture road geometric characteristics, and effectively improves the accuracy and integrity of remote sensing road extraction in a complex scene.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, specifically to a priori feature-assisted directional coordinate attention remote sensing road extraction method. Background Technology

[0002] With the rapid development of remote sensing technology, high-resolution remote sensing imagery has become an important data source for acquiring surface information. Roads, as critical surface infrastructure, have wide applications in automated extraction technologies for urban planning, traffic management, and emergency rescue. Currently, remote sensing image road extraction methods mainly include traditional image processing-based methods and deep learning-based methods. Traditional methods rely primarily on manually designed features and rules for road extraction, while deep learning methods automatically learn road feature representations through neural networks, exhibiting stronger adaptability and robustness in complex scenarios. In deep learning-based methods, the introduction of attention mechanisms significantly improves the model's ability to capture road features, and multi-scale feature fusion strategies have also addressed the disconnection problem in road extraction to some extent. Furthermore, some methods effectively alleviate fragmented road extraction and category confusion by combining directional and semantic features. In addition, methods combining bidirectional multi-level dynamic road feature fusion and dual-context dynamic extraction enhance the model's understanding of complex scenes and multi-layered road structures.

[0003] Despite significant progress in road extraction from remote sensing images, several technical challenges remain. First, existing methods lack sufficient integration of prior knowledge about road physical characteristics (such as color, brightness, structure, and texture), resulting in inadequate accuracy in representing road features against complex backgrounds. Second, while some methods consider directional features, they lack precise pixel-level direction prediction and direction-based feature enhancement mechanisms, making it difficult to effectively handle directional changes and intersections in complex road networks. Third, existing methods are insufficient in multi-scale feature extraction and fusion, particularly when dealing with roads of varying widths and complexities. The collaborative representation of features at different scales needs improvement, making it difficult to simultaneously consider global semantic information and local details. Finally, existing loss function designs fail to adequately address the specific characteristics of road extraction tasks, such as class imbalance and high boundary accuracy requirements, leading to weak robustness against complex background interference. To address these issues, we propose a prior feature-assisted directional coordinate attention remote sensing road extraction method. Summary of the Invention

[0004] To address the technical problems of existing road extraction methods, such as insufficient accuracy in explicit prior feature injection, limited multi-scale feature learning ability, and weak robustness to complex background interference, and to achieve the technical effects of fully integrating multi-source physical prior information, accurately restoring the road trunk and boundary structure, and end-to-end automated remote sensing road extraction, this invention provides a prior feature-assisted directional coordinate attention remote sensing road extraction method.

[0005] To achieve the above objectives, the technical solution adopted by the present invention is as follows:

[0006] A priori feature-assisted directional coordinate attention remote sensing road extraction method includes the following steps:

[0007] S1. For parallel extraction of multi-source prior features representing the physical characteristics of roads from remote sensing images, optimize the fusion weights of each feature based on the validation set to obtain a fused prior feature map. The multi-source prior features include color features, brightness and saturation joint features, structural features, directional texture features, and gradient consistency features.

[0008] Preferably, the color feature extraction method specifically involves: converting the original image from RGB space to HSV space; utilizing the distribution differences of road and non-road pixels in the H, S, and V channels; employing the Otsu adaptive thresholding algorithm to segment each channel separately to obtain initial segmentation results for each channel; applying median filtering to the initial segmentation results to suppress isolated noise points generated during the segmentation process; and performing morphological processing on the filtered results to remove residual small noise areas and highlight the color dimension differences between the road and the background.

[0009] Preferably, the method for extracting the joint features of brightness and saturation specifically involves: using a sliding window to calculate the local mean of the V channel pixel values ​​in the HSV space and the local standard deviation of the S channel pixel values; adaptively setting a screening threshold based on the statistical characteristics of the local mean and local standard deviation; and filtering out non-road pixels with high saturation by comparing the local mean and local standard deviation of the pixels with the threshold, while retaining pixels whose brightness and saturation conform to the characteristics of road areas.

[0010] Preferably, the structural feature extraction method specifically involves: using a multi-scale edge detection operator to calculate the gradients in the horizontal and vertical directions of the remote sensing image, extracting linear features from the image based on the gradient results using a linear structure extraction algorithm; performing joint processing on the multi-scale edge responses, connecting broken line segments using morphological closing operations, and then removing isolated noise line segments using morphological opening operations.

[0011] Preferably, the directional texture feature extraction method specifically involves: performing convolution processing on the remote sensing image using a multi-scale, multi-directional filter to obtain filter response maps at each scale and in each direction; calculating the local energy in a local region centered on the current pixel for each filter response map, and obtaining the local energy by calculating the square and square root of the filter response values ​​within the region; statistically analyzing the mean and variance of the local energy, and constructing a texture saliency map using the mean and variance.

[0012] Preferably, the optimization of the fusion weights of each feature to obtain the fusion prior feature map specifically involves: setting a corresponding feature mask and fusion weight for each type of prior feature; the fusion prior feature map is obtained by multiplying each type of feature mask by its corresponding weight and then summing the results; the optimization of the fusion weights is achieved by minimizing the loss on the validation set, with the average loss between the fusion prior feature map and the true label of all samples in the validation set as the optimization objective, wherein the sample loss is calculated based on the fusion prior feature map and the true road region label of the sample; and the contribution of different prior features is dynamically adjusted through an adaptive weight optimization mechanism.

[0013] S2. Input feature collaborative modeling and multi-scale processing: RGB remote sensing images and fusion prior features are collaboratively processed. Multi-scale feature extraction is performed on the RGB images to form a feature pyramid. The fusion prior features are transformed into a high-dimensional representation. The two types of features are spliced ​​together in the channel dimension to obtain the fusion features.

[0014] Preferably, the input feature collaborative modeling and multi-scale processing specifically involves: performing multi-scale convolution processing on the RGB remote sensing image, retaining the original resolution, performing 2x downsampling and 4x downsampling operations to extract semantic features at different scales, upsampling the downsampled features to the original resolution using an interpolation method to form a multi-scale feature pyramid; converting the fused prior features into a high-dimensional representation using a feature fusion module, with a size equal to the original image height × width × number of fused channels; and concatenating the RGB multi-scale features and the high-dimensional representation of the prior features along the channel dimension to form a fused feature containing multi-scale semantic information and physical priors.

[0015] S3. Oriented coordinate attention mechanism processing: predict the main direction angle of pixels in the feature through oriented convolution, pool the feature along the main direction and orthogonal direction, generate attention map and recalibrate the original feature;

[0016] Preferably, the directional coordinate attention mechanism specifically involves: predicting the principal direction angle of each pixel in the input features through directional convolution operations, where the angle reflects the extension direction of the road at that pixel location; performing one-dimensional global average pooling along the principal direction to capture the feature dependency of the road extension direction, and performing one-dimensional global average pooling along a direction orthogonal to the principal direction to capture the feature dependency of the road width direction; concatenating the two types of pooled features in the channel dimension, processing them through dimensionality reduction convolution and activation function, and then performing dimensionality increase convolution to generate two sets of attention maps through the activation function; recalibrating the original input features, and summing the original features by multiplying them by the elements of the two sets of attention maps respectively to obtain the output features.

[0017] S4. Decoding and deep supervision: The decoder samples and splices the corresponding layer features of the encoder level by level. After processing by the directional coordinate attention mechanism, an auxiliary supervision branch is introduced at the decoder level.

[0018] Preferably, the decoding and deep supervision mechanism specifically involves: the decoder restoring the spatial resolution of features through progressive upsampling operations; concatenating the output features of each decoder level with the output features of the corresponding level encoder in the channel dimension; and processing the concatenated features using an oriented coordinate attention mechanism. Auxiliary supervision branches are introduced at multiple levels of the decoder, with each auxiliary supervision branch's loss being a composite loss including Dice loss, Focal loss, and boundary-aware loss. The total loss objective function consists of the main output loss and the loss terms of each auxiliary branch, with each auxiliary branch loss term including the correlation calculation between the composite loss of that branch and the corresponding uncertainty weighting coefficient. Through multi-level auxiliary supervision, the model's learning ability for multi-scale features and training convergence speed are improved, while suppressing gradient vanishing.

[0019] S5. A composite loss function including Dice loss, Focal loss and boundary-aware loss is adopted to balance region consistency, class imbalance and boundary accuracy.

[0020] Preferably, the composite loss function is designed as follows: the composite loss function is composed of Dice loss, Focal loss and boundary-aware loss combined with preset weights. Dice loss is used to measure the overlap between the segmentation result and the real label region, which is achieved by calculating the intersection-union ratio (IU) of the predicted region and the real region. Focal loss is used to balance the ratio of positive and negative samples and focus on difficult samples, which is achieved by assigning low weights to easy samples and high weights to difficult samples. Boundary-aware loss is used to improve the segmentation accuracy of road boundaries, which is based on edge detection operators to calculate pixel-level weights and assign higher weights to boundary pixels.

[0021] S6. Train the model by inputting the original image and fused prior features into the trained model to obtain the road extraction results.

[0022] Compared with the prior art, the beneficial effects of the present invention are as follows:

[0023] The proposed directional coordinate attention remote sensing road extraction method extracts road features from multiple sources in parallel and adaptively fuses these features, including color features, joint brightness and saturation features, structural features, directional texture features, and gradient consistency features. This fully integrates multi-source physical prior information, effectively improving the saliency and structural integrity of road targets in remote sensing images. It enhances the ability to identify complex scenes and areas with weak texture, significantly outperforming methods that rely solely on single features or are purely data-driven. Through a directional coordinate attention mechanism, targeting the directional and linear continuity features of roads, it uses directional convolution to predict the principal direction angle of pixels, performs one-dimensional global average pooling along tangents and normals, and generates an attention map to weight and label the original features. The combined deep supervised optimization strategy can accurately restore the main road and boundary structure, effectively suppress the interference of background noise on the segmentation results, and improve the segmentation accuracy and robustness of the model. It is particularly suitable for various terrains, different image resolutions and complex road morphologies. By designing a composite loss function that includes Dice loss, Focal loss and boundary-aware loss, and comprehensively considering regional consistency, class imbalance and boundary accuracy, an end-to-end automated remote sensing road extraction process is realized. This greatly reduces the workload of manual annotation and feature design, improves data processing and production efficiency, and can be quickly adapted and promoted to automatic extraction tasks of various linear features such as power lines and rivers, which is conducive to intelligent inspection and refined management in related industries. Attached Figure Description

[0024] Figure 1 This is an overall flowchart used in the embodiments of the present invention;

[0025] Figure 2 This is the road extraction result of the present invention on two datasets. Detailed Implementation

[0026] The following description is intended to disclose the invention and enable those skilled in the art to implement it. The preferred embodiments described below are merely examples, and other obvious variations will occur to those skilled in the art.

[0027] Example 1

[0028] Reference Figure 1 and Figure 2 As shown, a priori feature-assisted directional coordinate attention remote sensing road extraction method includes the following steps:

[0029] S1: For parallel extraction of multi-source prior features representing the physical characteristics of roads from remote sensing images, the fusion weights of each feature are optimized based on the validation set to obtain a fused prior feature map.

[0030] In this method, the multi-source prior features include color features, combined brightness and saturation features, structural features, and directional texture features.

[0031] The color feature extraction method is as follows: the original image is converted from RGB space to HSV space, and the distribution differences of road and non-road pixels in the H, S and V channels are used to segment each channel using the Otsu adaptive thresholding algorithm to obtain the initial segmentation results for each channel; median filtering is applied to the initial segmentation results to suppress isolated noise points generated during the segmentation process; morphological processing is performed on the filtered results to remove residual small noise areas and highlight the differences between the road and the background in the color dimension.

[0032] The joint feature extraction method for brightness and saturation is as follows: a sliding window is used to calculate the local mean of the V channel pixel values ​​in the HSV space. Calculate the local standard deviation for the S-channel pixel values. Based on the statistical characteristics of local mean and local standard deviation, the filtering threshold is adaptively set. By comparing the local mean and local standard deviation of the pixels with the threshold, high-saturation non-road pixels are filtered out, while pixels whose brightness and saturation conform to the characteristics of road areas are retained.

[0033] The structural feature extraction method is as follows: multi-scale edge detection operators are used to calculate the gradients in the horizontal and vertical directions of the remote sensing image, and linear features in the image are extracted based on the gradient results using a linear structure extraction algorithm; multi-scale edge responses are jointly processed, broken line segments are connected by morphological closing operations, and isolated noise line segments are removed by morphological opening operations.

[0034] The directional texture feature extraction method specifically involves: convolving the remote sensing image with multi-scale (frequency 0.1, 0.2, 0.4) and multi-directional (0°, 45°, 90°, 135°) filters to obtain filtered response maps at each scale and direction; for each filtered response map, calculating the local energy within a local region centered on the current pixel; and obtaining the local energy by calculating the square and square root of the filtered response values ​​within the region. The calculation is as follows:

[0035]

[0036] in, Indicates frequency as , direction is The filtered response at the pixel The value at that location, Represents a local region centered on the current pixel; calculates the mean of the local energy. and variance We construct a texture saliency map using the mean and variance.

[0037] Optimizing the fusion weights of each feature to obtain the fused prior feature map involves setting a corresponding feature mask for each class of prior features. (in (Feature category index, value range 1-4) and fusion weights , fusing prior feature maps The expression is obtained by multiplying each feature mask by its corresponding weight and then summing the results:

[0038]

[0039] The optimization of the fusion weights is achieved by minimizing the loss on the validation set. The optimization objective is the average loss between the fused prior feature maps and the true labels of all samples in the validation set. The objective function is:

[0040]

[0041] in The total number of validation set samples (index of total sample count) ), For loss function, Indicates the first Fusion prior feature maps of individual image samples Indicates the first The true labels of each image; through an adaptive weight optimization mechanism, the contribution of different prior features is dynamically adjusted.

[0042] S2: Input feature collaborative modeling and multi-scale processing, which collaboratively processes RGB remote sensing images and fused prior features, extracts multi-scale features from RGB images to form a feature pyramid, performs high-dimensional representation transformation on fused prior features, and concatenates the two types of features in the channel dimension to obtain fused features.

[0043] Input feature co-modeling and multi-scale processing specifically involves: processing RGB remote sensing images ( Multi-scale convolution processing is performed, retaining the original resolution and then performing 2x and 4x downsampling operations to extract semantic features at different scales. The downsampled features are then upsampled to the original resolution using interpolation methods to form a multi-scale feature pyramid. The fusion prior features are converted into a high-dimensional representation by a feature fusion module, with dimensions equal to the original image height × width × number of fusion channels. The RGB multi-scale features and the high-dimensional representation of prior features are concatenated along the channel dimension to form a fused feature that includes multi-scale semantic information and physical priors. The size is (in (This indicates the final number of channels after splicing).

[0044] S3: Oriented coordinate attention mechanism processing, predicts the main direction angle of pixels in the feature through directional convolution, pools the feature along the main direction and orthogonal direction, generates attention map and recalibrates the original feature.

[0045] The directional coordinate attention mechanism specifically predicts input features through directional convolution operations. ( The principal orientation angle of each pixel in the feature map tensor is represented by the tensor. ( (This is a directional parameter), where the angle reflects the direction of road extension at that pixel location; one-dimensional global average pooling is performed on the input features along the main direction to capture the feature dependency of the road extension direction, and one-dimensional global average pooling is performed along a direction orthogonal to the main direction to capture the feature dependency of the road width direction, yielding:

[0046]

[0047]

[0048] in This represents the features after pooling along the principal direction. This represents the features after pooling along orthogonal directions; the two types of pooled features are concatenated along the channel dimension, processed by dimensionality reduction convolution and activation function, and then subjected to dimensionality increase convolution, generating two sets of attention maps through the activation function. and (in and (These are two sets of directional attention weights respectively).

[0049]

[0050] The original input features are recalibrated, and the output features are obtained by multiplying the original features by the elements of the two attention maps respectively and then summing the results. :

[0051]

[0052] in This indicates element-wise multiplication.

[0053] S4: Decoding and deep supervision. The decoder samples and splices the corresponding layer features from the encoder level, and after processing by the directional coordinate attention mechanism, an auxiliary supervision branch is introduced at the decoder level.

[0054] The decoding and deep supervision mechanism works as follows: the decoder restores the spatial resolution of the features through progressive upsampling operations; the output features of each decoder level are concatenated with the output features of the corresponding level encoder in the channel dimension; the concatenated features are then processed by the directional coordinate attention mechanism; auxiliary supervision branches are introduced at each of the three levels of the decoder, and the loss of each auxiliary supervision branch is a composite loss including Dice loss, Focal loss, and boundary-aware loss. (in (Indicates the auxiliary branch layer number); the total loss objective function consists of the main output loss and the loss terms of each auxiliary branch:

[0055]

[0056] in This indicates principal component loss. Indicates the total number of auxiliary branches. The uncertainty-adaptive weighting coefficients are used; through multi-level auxiliary supervision, the model's ability to learn multi-scale features and the training convergence speed are improved, while suppressing the gradient vanishing problem.

[0057] S5: Employ a composite loss function that includes Dice loss, Focal loss, and boundary-aware loss to balance region consistency, class imbalance, and boundary accuracy.

[0058] The composite loss function is designed as follows: it is composed of Dice loss, Focal loss, and boundary-aware loss combined with preset weights.

[0059]

[0060] The Dice loss is used to measure the overlap between the segmentation result and the true label region, and its calculation formula is:

[0061]

[0062] In the formula The total number of pixels in the segmented region ( (pixel index) For the first The network prediction value for each pixel. For real labels ( ), For smoothing terms;

[0063] Focal loss is used to balance the ratio of positive to negative samples and focus on hard-to-classify samples. The calculation formula is as follows:

[0064]

[0065] in This is the positive and negative sample balance coefficient. For focusing parameters;

[0066] Boundary-aware loss is used to improve the segmentation accuracy of road boundaries. It calculates pixel-level weights based on edge detection operators, assigning higher weights to boundary pixels. The calculation formula is as follows:

[0067]

[0068] in These are pixel-level weights based on the Sobel operator.

[0069] S6: Train the model by inputting the original image and fused prior features into the trained model to obtain the road extraction results.

[0070] During training, the remote sensing image dataset and corresponding road labeling data are first prepared, and the dataset is divided into training, validation, and test sets. For each remote sensing image in the training set, multi-source prior features are extracted and fused according to step S1, while the original RGB image is preprocessed. The original RGB image and the fused prior features are used as input to train the model using a deep learning framework.

[0071] During the training phase, the model first extracts and fuses features through feature co-modeling and multi-scale processing in step S2. Then, it enhances road directionality features through the directional coordinate attention mechanism in step S3. Finally, it generates road segmentation results through decoding and deep supervision mechanisms in step S4. The training process uses the composite loss function designed in step S5 to calculate the loss and update the model parameters.

[0072] After the model training is complete, for new remote sensing images, the same prior feature extraction and fusion steps are performed to convert the original RGB images. and fusion of prior features Input the trained model This will yield high-precision road extraction results:

[0073]

[0074] Experiments show that this method can effectively address the challenges of road extraction from remote sensing images, such as complex backgrounds, varying road widths, and diverse directions, thereby improving the accuracy and completeness of road extraction.

[0075] The foregoing has shown and described the basic principles, main features, and advantages of the present invention. Those skilled in the art should understand that the present invention is not limited to the above embodiments. The embodiments and descriptions in the specification are merely principles of the invention. Various changes and modifications can be made to the invention without departing from its spirit and scope, and all such changes and modifications fall within the scope of the claimed invention. The scope of protection claimed by the appended claims and their equivalents is defined.

Claims

1. A priori feature-assisted directional coordinate attention remote sensing road extraction method, characterized in that, Includes the following steps: S1. For parallel extraction of multi-source prior features representing the physical characteristics of roads from remote sensing images, the fusion weights of each feature are optimized based on the validation set to obtain a fused prior feature map. The multi-source prior features include color features, brightness and saturation joint features, structural features and directional texture features. S2. Input feature collaborative modeling and multi-scale processing: RGB remote sensing images and fusion prior features are collaboratively processed. Multi-scale feature extraction is performed on the RGB images to form a feature pyramid. The fusion prior features are transformed into a high-dimensional representation. The two types of features are spliced ​​together in the channel dimension to obtain the fusion features. S3. The directional coordinate attention mechanism processes the data by predicting the principal direction angle of pixels in the features through directional convolution, pooling the features along the principal direction and orthogonal directions, generating an attention map, and recalibrating the original features. Specifically, the directional coordinate attention mechanism is as follows: Predict the principal orientation angle of each pixel in the input features using directional convolution operations. This angle reflects the direction in which the road extends at that pixel location. One-dimensional global average pooling is performed on the input features along the main direction to capture the feature dependencies in the road extension direction, and one-dimensional global average pooling is performed along the direction orthogonal to the main direction to capture the feature dependencies in the road width direction. The two types of pooling features are concatenated along the channel dimension, processed by dimensionality reduction convolution and activation function, and then dimensionality increase convolution is performed. Two sets of attention maps are generated by the activation function. The original input features are recalibrated, and the output features are obtained by multiplying the original features by the elements of the two attention maps respectively and summing the results. S4. Decoding and deep supervision: The decoder samples and splices the corresponding layer features of the encoder level by level. After processing by the directional coordinate attention mechanism, an auxiliary supervision branch is introduced at the decoder level. S5. A composite loss function including Dice loss, Focal loss and boundary-aware loss is adopted to balance region consistency, class imbalance and boundary accuracy. S6. Train the model by inputting the original image and fused prior features into the trained model to obtain the road extraction results.

2. The prior feature-assisted directional coordinate attention remote sensing road extraction method according to claim 1, characterized in that, In step S1, the color feature extraction method is specifically as follows: The original image was converted from RGB space to HSV space. Taking advantage of the difference in the distribution of road and non-road pixels in the H, S and V channels, the Otsu adaptive thresholding algorithm was used to segment each channel separately to obtain the initial segmentation results for each channel. Median filtering is applied to the initial segmentation results to suppress isolated noise points generated during the segmentation process; Morphological processing is performed on the filtered results to remove residual small noise areas and highlight the differences in color between the road and the background.

3. The prior feature-assisted directional coordinate attention remote sensing road extraction method according to claim 1, characterized in that, In step S1, the method for extracting the joint features of brightness and saturation is specifically as follows: A sliding window is used to calculate the local mean of the V channel pixel values ​​in the HSV space and the local standard deviation of the S channel pixel values. The filtering threshold is adaptively set based on the statistical characteristics of local mean and local standard deviation. By comparing the local mean and local standard deviation of pixels with the threshold, high-saturation non-road pixels are filtered out, while pixels whose brightness and saturation conform to the characteristics of road areas are retained.

4. The prior feature-assisted directional coordinate attention remote sensing road extraction method according to claim 1, characterized in that, In step S1, the structural feature extraction method is specifically as follows: Multi-scale edge detection operators are used to calculate the gradients in the horizontal and vertical directions of remote sensing images, and linear features in the images are extracted based on the gradient results using a linear structure extraction algorithm. Multi-scale edge responses are jointly processed by using morphological closing operations to connect broken line segments and then using morphological opening operations to remove isolated noise line segments.

5. The prior feature-assisted directional coordinate attention remote sensing road extraction method according to claim 1, characterized in that, In step S1, the directional texture feature extraction method is specifically as follows: Multi-scale, multi-directional filters are used to convolve remote sensing images to obtain filter response maps at each scale and in each direction. For each filtered response map, the local energy is calculated in a local region centered on the current pixel. The local energy is obtained by calculating the square and square root of the filtered response values ​​in the region. The mean and variance of local energies are statistically analyzed, and a texture saliency map is constructed using the mean and variance.

6. The prior feature-assisted directional coordinate attention remote sensing road extraction method according to claim 1, characterized in that, The optimization of the fusion weights of each feature to obtain the fused prior feature map in step S1 is specifically as follows: For each type of prior feature, a corresponding feature mask and fusion weight are set. The fused prior feature map is obtained by multiplying each type of feature mask by its corresponding weight and then summing the results. The optimization of the fusion weights is achieved by minimizing the loss on the validation set. The optimization objective is the average loss of the fusion prior feature map and the true label of all samples in the validation set. The sample loss is calculated based on the fusion prior feature map and the true road region label of the sample. The contribution of different prior features is dynamically adjusted through an adaptive weight optimization mechanism.

7. The prior feature-assisted directional coordinate attention remote sensing road extraction method according to claim 1, characterized in that, In step S2, the input feature collaborative modeling and multi-scale processing specifically involve: Multi-scale convolution processing is performed on RGB remote sensing images to extract semantic features at different scales by retaining the original resolution, performing 2x downsampling and 4x downsampling operations respectively. The downsampled features are then upsampled to the original resolution using interpolation methods to form a multi-scale feature pyramid. The fusion prior features are converted into a high-dimensional representation by the feature fusion module, and its size is the original image height × width × number of fusion channels; By concatenating RGB multi-scale features with high-dimensional representations of prior features in the channel dimension, a fusion feature containing multi-scale semantic information and physical priors is formed.

8. The prior feature-assisted directional coordinate attention remote sensing road extraction method according to claim 1, characterized in that, In step S4, the decoding and deep supervision mechanism specifically includes: The decoder restores the spatial resolution of the features through progressive upsampling operations. The output features of each level decoder are concatenated with the output features of the corresponding level encoder in the channel dimension. The concatenated features are then processed by the directional coordinate attention mechanism. Auxiliary supervision branches are introduced at multiple levels of the decoder. The loss of each auxiliary supervision branch is a composite loss that includes Dice loss, Focal loss and boundary-aware loss. The total loss objective function consists of the main output loss and the loss terms of each auxiliary branch. Each auxiliary branch loss term includes the correlation calculation of the composite loss of that branch and the corresponding uncertainty weighting coefficient. By employing multi-level auxiliary supervision, the model's ability to learn multi-scale features and its training convergence speed are improved, while the gradient vanishing problem is suppressed.

9. The prior feature-assisted directional coordinate attention remote sensing road extraction method according to claim 1, characterized in that, In step S5, the design of the composite loss function is specifically as follows: The composite loss function is composed of Dice loss, Focal loss and boundary-aware loss combined with preset weights. Dice loss is used to measure the overlap between the segmentation result and the real label region, which is achieved by calculating the intersection-union ratio of the predicted region and the real region. Focal loss is used to balance the ratio of positive and negative samples and focus on difficult samples, which is achieved by assigning low weights to easy samples and high weights to difficult samples. Boundary-aware loss is used to improve the segmentation accuracy of road boundaries. It calculates pixel-level weights based on edge detection operators and assigns higher weights to boundary pixels.

Images