Remote sensing image feature segmentation method based on denoising diffusion probability model

Abstract

The application discloses a remote sensing image ground object segmentation method based on a denoising diffusion probability model, and relates to the technical field of remote sensing image processing.The method comprises the following steps: performing grid division and interference feature recognition on a target remote sensing image, and performing grid confidence evaluation based on the recognition result; fitting a plurality of image sub-regions according to a plurality of interference feature types and a plurality of grid confidences; calling a sample denoising diffusion probability model library, and performing sample model matching and integration according to the regional position environment feature, the high-frequency interference feature type and the regional image confidence, so as to construct a plurality of adaptive image denoising plug-ins; performing image denoising on the plurality of image sub-regions respectively, obtaining a plurality of denoising regional images, and performing image splicing to obtain a denoising remote sensing image; performing ground object segmentation on the denoising remote sensing image by using a convolutional neural network, and outputting a ground object classification map.The application solves the technical problem that the ground object segmentation result is inaccurate due to poor denoising effect of the remote sensing image in the prior art.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image processing technology, and specifically to a method for land cover segmentation in remote sensing images based on a denoising diffusion probability model. Background Technology

[0002] With the rapid development of remote sensing technology, remote sensing image land cover segmentation plays an increasingly important role in fields such as land resource surveys, environmental monitoring, and urban planning. However, in practical applications, remote sensing images are often affected by various factors, such as cloud cover, atmospheric haze, sensor noise, and changes in illumination, leading to a decline in image quality and severely impacting the accuracy of land cover segmentation. Traditional land cover segmentation methods typically treat the entire image with a uniform processing approach, lacking targeted processing for local interference features. This results in poor performance in complex and ever-changing real-world environments, often requiring significant manual intervention and parameter adjustments, leading to low efficiency and difficulty in meeting the needs of large-scale remote sensing image processing. Especially when dealing with remote sensing images acquired under new environments or special weather conditions, the generalization ability of existing methods is insufficient, resulting in significant deviations in segmentation results. These shortcomings not only limit the accuracy and efficiency of remote sensing image land cover segmentation but also affect the reliability of subsequent applications. Summary of the Invention

[0003] This application provides a method for land cover segmentation of remote sensing images based on a denoising diffusion probability model, which is used to address the technical problem that inaccurate land cover segmentation results are caused by poor denoising effects of remote sensing images in existing technologies.

[0004] In view of the above problems, this application provides a method for land cover segmentation of remote sensing images based on a denoising diffusion probability model, the method comprising:

[0005] The target remote sensing image is divided into grids and interference features are identified. Based on the identification results, the grid confidence is evaluated, and several grid confidence scores are obtained.

[0006] Several image sub-regions are obtained by fitting several interference feature types and several grid confidence scores, wherein each image sub-region is identified by a high-frequency interference feature type and a region image confidence score.

[0007] The sample denoising diffusion probability model library is called, and sample model matching and model integration are performed based on the regional location environment characteristics, the high-frequency interference feature type, and the regional image confidence level to construct several adapted image denoising plugins.

[0008] The image denoising plugins are used to denoise the image sub-regions respectively, and the images of the denoised regions are obtained. The images are then stitched together to obtain the denoised remote sensing image.

[0009] A convolutional neural network is used to segment ground features in the denoised remote sensing image, and a ground feature classification map is output.

[0010] One or more technical solutions provided in this application have at least the following technical effects or advantages:

[0011] This application proposes a remote sensing image land cover segmentation method based on a denoising diffusion probability model. By performing fine-grained grid division and interference feature identification on the target remote sensing image, and then performing denoising processing based on the identification results before land cover segmentation, the accuracy and adaptability of remote sensing image land cover segmentation are significantly improved. Compared with traditional methods, the technical solution provided in this application significantly improves the ability to process low-quality remote sensing images, can effectively cope with interference factors in various complex environments, achieves adaptive processing of different regional features, avoids the performance degradation problem of traditional methods when processing remote sensing images with multiple interferences, and achieves the technical effect of improving the accuracy of remote sensing image land cover segmentation and improving processing efficiency. Attached Figure Description

[0012] To more clearly illustrate the technical solutions in the embodiments of the present invention, the accompanying drawings used in the description of the embodiments will be briefly introduced below. Obviously, the accompanying drawings described below are only some embodiments of the present invention. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0013] Figure 1 This is a flowchart illustrating the remote sensing image feature segmentation method based on a denoising diffusion probability model provided in an embodiment of this application.

[0014] Figure 2 This is a flowchart illustrating the method for constructing a denoising diffusion probability model library provided in an embodiment of this application. Detailed Implementation

[0015] This application provides a method for land cover segmentation in remote sensing images based on a denoising diffusion probability model, which is intended to address the technical problem of inaccurate land cover segmentation results caused by poor denoising effects in existing remote sensing images.

[0016] The technical solutions of the embodiments of this application will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of this application, and not all of them. All other embodiments obtained by those skilled in the art based on the embodiments of this application without creative effort are within the scope of protection of this application.

[0017] It should be noted that the terms "comprising" and "having" are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or server that includes a series of steps or units is not necessarily limited to those steps or units that are explicitly listed, but may include other steps or modules that are not explicitly listed or that are inherent to these processes, methods, products, or devices.

[0018] Example 1, as Figure 1 As shown, this application provides a method for remote sensing image land cover segmentation based on a denoising diffusion probability model, wherein the method includes:

[0019] S10: Perform grid division and interference feature identification on the target remote sensing image, evaluate the grid confidence based on the identification results, and obtain several grid confidence scores.

[0020] Existing remote sensing image processing methods typically treat the entire image as a whole, lacking the ability to perform detailed analysis of local areas within the image. This inability to accurately identify and locate various interference features in different regions of the image prevents the adoption of differentiated processing strategies based on these interference features during subsequent processing, thus affecting the accuracy and adaptability of the overall processing results.

[0021] Step S10 in the method provided in this application embodiment includes:

[0022] The target remote sensing image is divided into regions according to a preset grid size to obtain multiple remote sensing grid images;

[0023] The interference feature types and interference intensity of the multiple remote sensing grid images are identified sequentially to obtain multiple interference identification results;

[0024] The types of interference features include clouds, cloud shadows, haze, localized uneven lighting, strip noise, random noise, building edges, road edges, water reflections, ripples, image registration errors, terrain shadows, and texture anomalies.

[0025] Based on the interference feature type and interference intensity, grid confidence is evaluated on the multiple interference identification results respectively, and several grid confidence scores are output.

[0026] In this embodiment, the target remote sensing image is divided into regions according to a preset grid size to obtain multiple remote sensing grid images. For example, the preset grid size is set to 256×256 pixels to facilitate accurate identification.

[0027] Interference feature identification and interference intensity identification are performed sequentially on multiple remote sensing grid images. Specifically, feature extraction is performed on each remote sensing grid image. For example, brightness statistical features such as brightness mean and variance are extracted to detect clouds, cloud shadows, haze, local uneven illumination, image registration errors, terrain shadows, and texture anomalies. High-frequency texture features such as gradients and Laplacian filtering are extracted to detect building edges, road edges, water reflections, and ripples. Fourier transform is used to extract strip noise and random noise from the remote sensing images.

[0028] The extracted features are input into a classification model such as a convolutional neural network to classify the type of interference. For example, the features are input into a random forest model for classification and identification, which contains 100 decision trees, each with a maximum depth of 20. The Gini coefficient is used as the splitting criterion to classify and identify the interference features, and the probability of the classification output is output as the interference intensity.

[0029] The types of interference features include clouds, cloud shadows, haze, localized uneven lighting, strip noise, random noise, building edges, road edges, water reflections, ripples, image registration errors, terrain shadows, and texture anomalies.

[0030] Based on the interference feature type and intensity, grid confidence scores are evaluated for multiple interference identification results, outputting several grid confidence scores. For example, grid confidence score = 1 - (interference intensity × type weight), where the type weight is a pre-set weight based on the influence degree of each interference type; the smaller the interference, the smaller the type weight. For instance, assigning a type weight of 0.3 to clouds, when the cloud interference intensity is 0.75, the grid confidence score = 1 - (0.75 × 0.3) = 0.775. A higher grid confidence score indicates less interference and stronger reliability of the remote sensing grid image.

[0031] By dividing the target remote sensing image into fine grid units and performing independent interference feature identification and confidence evaluation on each grid, localized and quantitative assessment of image quality is achieved, providing a detailed data foundation for subsequent differentiated processing.

[0032] S20: Several image sub-regions are obtained by fitting several interference feature types and several grid confidence levels, wherein each image sub-region is identified by a high-frequency interference feature type and a regional image confidence level.

[0033] After obtaining grid-level recognition results, existing technologies lack effective region merging strategies, making it difficult to reasonably aggregate adjacent grids with similar characteristics into meaningful processing units. If processing is done simply on a grid-by-grid basis, the processing units will be too fragmented, increasing processing complexity; if the entire image is simply used as the processing unit, the differences between regions cannot be reflected.

[0034] Step S20 in the method provided in this application embodiment includes:

[0035] Q starting points are uniformly selected in the edge region of the target remote sensing image to determine Q starting grid images, where Q is an integer greater than or equal to 4;

[0036] Starting with the Q initial grid images, based on the interference feature type and grid confidence, the grid similarity between the initial grid image and the adjacent grid images is calculated respectively. If the grid similarity meets the preset grid similarity threshold, the adjacent grid images are merged into the initial grid image to obtain the updated grid image. Starting with the updated grid image, grid fitting is continued. The interference feature type of the updated grid image is the union of the merged grid image and the initial grid image, and the grid confidence is the mean of the merged grid image and the initial grid image.

[0037] If the grid similarity does not meet the preset grid similarity threshold, a dividing line is set between the starting grid image and the adjacent grid images, and grid fitting continues from the adjacent grid images as the starting point until the grid fitting is completed, resulting in several image sub-regions. Each image sub-region is identified by a high-frequency interference feature type and a region image confidence level.

[0038] In this embodiment, Q starting points are uniformly selected in the edge region of the target remote sensing image to determine Q starting grid images. Here, Q is an integer greater than or equal to 4. For example, if the target remote sensing image is rectangular, then four starting points at the four corners of the rectangle are selected to determine four starting grid images.

[0039] Starting with Q initial grid images, the grid similarity between the initial grid image and adjacent grid images is calculated based on the interference feature type and grid confidence level. For example, grid similarity = feature type similarity weight × type matching degree + confidence similarity weight × (1 - |confidence difference|). Wherein, if the interference feature types of two grids are exactly the same, the type matching degree is 1; otherwise, the number of features with the same interference type is divided by the number of the union of the interference type features of the two grids, which is used as the type matching degree. The confidence similarity weight and feature type similarity weight are each set to 0.5. For example, if the initial grid feature type is cloud and haze with a confidence level of 0.8, and the adjacent grid feature type is cloud with a confidence level of 0.75, then the similarity = 0.5 × 0.5 + 0.5 × (1 - |0.8 - 0.75|) = 0.25 + 0.5 × 0.95 = 0.725.

[0040] If the grid similarity meets a preset grid similarity threshold, for example, setting the grid similarity threshold to 0.7, then adjacent grid images are merged into the starting grid image to obtain an updated grid image. Starting from this updated grid image, grid fitting continues. Each time an image is merged, the interference feature type and grid confidence are updated. Specifically, the interference feature type of the updated grid image is the union of the merged grid image and the starting grid image, and the grid confidence is the mean of the merged grid image and the starting grid image.

[0041] If the grid similarity does not meet the preset grid similarity threshold, a dividing line is set between the starting grid image and the adjacent grid images, and grid fitting continues from the adjacent grid images until the grid fitting is completed, resulting in several image sub-regions. Each image sub-region is identified with a high-frequency interference feature type and a region image confidence level.

[0042] By fitting regions based on interference feature types and grid confidence, adjacent grids with similar features are aggregated into several image sub-regions, thus achieving optimized division of processing units. This maintains the rationality of processing units while ensuring feature consistency within each sub-region, providing crucial data support for achieving differentiated processing across regions.

[0043] S30: Call the sample denoising diffusion probability model library, perform sample model matching and model integration based on the regional location environment characteristics, the high-frequency interference feature type, and the regional image confidence, and construct several adapted image denoising plugins.

[0044] Existing denoising methods typically employ fixed, universal denoising models, which cannot adapt to the diverse characteristics and processing requirements of different regions. This approach struggles to simultaneously meet the removal requirements of various interference features under complex environments, resulting in unsatisfactory denoising performance.

[0045] Step S30 in the method provided in this application embodiment includes:

[0046] The method for constructing the sample denoising diffusion probability model library is as follows: Figure 2 As shown, it includes:

[0047] Multi-source remote sensing images were collected as training samples and divided according to regional location and environmental characteristics and interference characteristics to obtain several sample image sets;

[0048] A first sample image set is randomly selected. With the goal of minimizing the denoising error, the denoising diffusion probability model is trained by forward diffusion and backward diffusion respectively to obtain the first sample denoising diffusion probability model. The first sample denoising diffusion probability model is then labeled with features based on the location and environmental features of the first region and the first type of interference features.

[0049] Several sample denoising and diffusion probability models are obtained by sequentially training the several sample image sets, and a sample denoising and diffusion probability model library is constructed.

[0050] Randomly select the first high-frequency interference feature type and the first region image confidence level of the first image sub-region;

[0051] Based on the regional location and environmental features and the first high-frequency interference feature type, feature similarity comparison is performed on several feature tags in the sample denoising and diffusion probability model library, and several feature similarity coefficients are output.

[0052] If the number of similarity coefficients among the plurality of features that are greater than the first similarity coefficient threshold is not zero, then the sample denoising diffusion probability model corresponding to the largest feature similarity coefficient is selected as the first matching model.

[0053] Calculate the ratio of the confidence level of the preset standard image to the confidence level of the first region image, and round it down in combination with the number of preset standard models to obtain the number of first adaptation models;

[0054] Based on the first number of adaptation models, multiple first matching models are selected, and the models are integrated with the image fitting layer to construct a first adaptation image denoising plugin, which is then added to the plurality of adaptation image denoising plugins.

[0055] If the number of similarity coefficients greater than the first similarity coefficient threshold is 0 and the number of similarity coefficients greater than the second similarity coefficient threshold is not 0, the similarity coefficients that meet the conditions are set as selected similarity coefficients, wherein the second similarity coefficient threshold is less than the first similarity coefficient threshold.

[0056] Select the sample denoising diffusion probability model corresponding to the selected feature similarity coefficient, and construct the second matching model set;

[0057] The product of the first number of adapted models and the similarity coefficient of the selected features is rounded down to the second number of adapted models, thus obtaining the second number of adapted models.

[0058] Based on the second number of adaptation models and the second matching model set, the first adaptation image denoising plugin is constructed by integrating the image fitting layer and performing model integration.

[0059] If the number of similarity coefficients among the aforementioned features that are greater than the second similarity coefficient threshold is 0, then the third matching model set is constructed by selecting all samples from the denoising diffusion probability model.

[0060] The product of the first number of adaptive models and the feature similarity coefficient is rounded down to the third number of adaptive models, thus obtaining the third number of adaptive models.

[0061] Based on the third matching model set and the number of third adaptation models, the first adaptation image denoising plugin is constructed by integrating the model with the image fitting layer.

[0062] In this embodiment, a sample denoising diffusion probability model library is constructed. Specifically, multi-source remote sensing images are collected as training samples and divided according to regional location environmental characteristics and interference feature types to obtain several sample image sets.

[0063] Based on the U-Net architecture, a diffusion probability model is constructed, consisting of four layers each for the encoder and decoder. Each layer uses a 3×3 convolutional kernel and employs a skip connection structure. A first sample image set is randomly selected, and training is performed with the goal of minimizing the denoising error. The training process is divided into two stages: forward diffusion and backward diffusion. In the forward diffusion stage, Gaussian noise is gradually added to the clear image, causing it to degenerate into pure noise. In the backward diffusion stage, the network is trained to reconstruct the clear image from the noise. The model parameters are optimized by minimizing the mean square error between the predicted noise and the actual added noise until convergence, for example, when the denoising error is less than or equal to 0.1, resulting in the first sample denoising diffusion probability model. The first sample denoising diffusion probability model is then feature-labeled according to the location and environmental features of the first region and the type of the first interference feature. For example, if the location and environmental feature of the first region is a city and the type of interference is clouds, then the first sample denoising diffusion probability model is labeled with the content "city" and "clouds".

[0064] Several sample denoising diffusion probability models are obtained by training several sample image sets in sequence, and a sample denoising diffusion probability model library is constructed.

[0065] The first high-frequency interference feature type and the first region image confidence level of the first image sub-region are randomly selected as the current processing objects.

[0066] Based on regional location environmental features and the first high-frequency interference feature type, feature similarity comparisons are performed on several feature markers in the sample denoising diffusion probability model library, outputting several feature similarity coefficients. Specifically, similarity = environmental feature similarity weight × environmental matching degree + interference feature similarity weight × interference feature matching degree. The union of regional location environmental features and environmental features in the feature markers is calculated to obtain the environmental feature union. The number of identical environmental features in the regional location environmental features and the environmental feature union is divided by the total number of environmental features in the environmental feature union to obtain the environmental feature matching degree. Using the same method as the environmental feature matching degree, the interference feature matching degree is calculated using the first high-frequency interference features and the interference features in the feature markers. The sum of the environmental feature similarity weight and the interference feature similarity weight is 1. The calculated feature similarity coefficients represent the feature similarity between the first image sub-region and all models in the sample denoising diffusion probability model library, used to preliminarily characterize the adaptability of all models in the sample denoising diffusion probability model library and the first image sub-region.

[0067] If, among a number of feature similarity coefficients, the number exceeding the first similarity coefficient threshold is not zero, then the sample denoising diffusion probability model corresponding to the largest feature similarity coefficient is selected as the first matching model. For example, the first similarity coefficient threshold is set to 0.7 to filter out a more suitable model.

[0068] Calculate the ratio of the confidence level of the preset standard image to the confidence level of the first region image, and round it down to obtain the number of first adaptation models. For example, if the confidence level of the first region image is 0.8, the confidence level of the preset unit image is 0.2, and the number of preset standard models is 10, then the number of first adaptation models = confidence level of preset standard image ÷ confidence level of first region image × number of preset standard models = 0.2 ÷ 0.8 × 10 = 2.5, rounded up to 3, that is, the number of first adaptation models is 3.

[0069] Based on the number of first adaptation models, multiple first matching models are selected and integrated with the image fitting layer. For example, the image fitting layer includes two convolutional layers to adjust the output features in order to construct the first adaptation image denoising plugin. The constructed first adaptation image denoising plugin is then added to several adaptation image denoising plugins.

[0070] If, among a number of feature similarity coefficients, the number greater than the first similarity coefficient threshold is 0 and the number greater than the second similarity coefficient threshold is not 0, then the feature similarity coefficients that meet the conditions are set as the selected feature similarity coefficients. Here, the second similarity coefficient threshold is less than the first similarity coefficient threshold. For example, if the first similarity threshold is 0.7, then the second similarity threshold is set to 0.5. The feature similarity coefficients that meet the conditions are the maximum value among the similarity coefficients less than 0.7 and greater than or equal to 0.5, which are set as the selected feature similarity coefficients. For example, if the feature similarity coefficients that meet the conditions are 0.55, 0.63, and 0.66, then 0.66 is selected as the selected feature similarity coefficient.

[0071] Select the sample denoising diffusion probability model corresponding to "select feature similarity coefficient" to construct the second matching model set.

[0072] The product of the first number of adapted models and the selected feature similarity coefficients, rounded down, is taken as the second number of adapted models. The second number of adapted models = the first number of adapted models × the selected feature similarity coefficients.

[0073] Based on the second set of adaptation models and the second set of matching models, a first adaptation image denoising plugin is constructed by integrating the image fitting layer. The image fitting layer is constructed using three convolutional layers to adjust the output features.

[0074] If the number of similarity coefficients among several features that are greater than the second similarity coefficient threshold is 0, it indicates that the first image sub-region is difficult to identify. The third matching model set is constructed by selecting all samples to denoise the diffusion probability model.

[0075] The product of the number of first-fitting models and the feature similarity coefficient is rounded down to the number of third-fitting models, thus obtaining the number of third-fitting models.

[0076] Based on the third matching model set and the number of third adaptation models, the first adaptation image denoising plugin is constructed by integrating the image fitting layer.

[0077] Through intelligent matching and model integration, the most suitable denoising plugin is dynamically constructed for each image sub-region, achieving precise adaptation of the processing model. It can provide customized denoising solutions according to the specific characteristics and processing needs of each region, greatly improving the pertinence and effectiveness of denoising processing and providing reliable technical support for obtaining high-quality denoised remote sensing images.

[0078] S40: Use the several adapted image denoising plugins to denoise the several image sub-regions respectively, obtain several denoised region images, and stitch the images together to obtain denoised remote sensing images.

[0079] In this embodiment, several adapted image denoising plugins are used to denoise several image sub-regions, obtaining several denoised region images. Based on the fitted segmentation of these sub-regions in the target remote sensing image, the images are stitched together at their original positions to obtain the denoised remote sensing image. By using specially adapted denoising plugins to process each sub-region specifically, differentiated and precise denoising is achieved. This effectively removes various interferences while preserving image details, ensuring optimal denoising results for each region. The final image stitching ensures the overall consistency and integrity of the denoised remote sensing image, providing high-quality input data for subsequent ground feature segmentation.

[0080] S50: Use a convolutional neural network to segment the denoised remote sensing image and output a land cover classification map.

[0081] In this embodiment, a pre-trained convolutional neural network is used to segment ground features in denoised remote sensing images. The denoised remote sensing image is input into the pre-trained convolutional neural network, which outputs the ground feature type and probability for each region. The method provided in this application effectively utilizes the enhanced feature information of the denoised image to achieve more accurate and reliable ground feature classification and segmentation. The final output ground feature classification map has higher accuracy and better detail preservation capabilities, fully demonstrating the value of the preliminary fine processing and meeting the application needs of remote sensing image analysis.

[0082] In summary, the embodiments of this application have at least the following technical effects:

[0083] This application proposes a remote sensing image land cover segmentation method based on a denoising diffusion probability model. By identifying the target remote sensing image and then denoising it before land cover segmentation, the accuracy and adaptability of remote sensing image land cover segmentation are significantly improved. Specifically, the target remote sensing image is subjected to fine-grained grid division and interference feature identification. Based on the identification results, grid confidence is evaluated. Then, multiple image sub-regions with different characteristics are fitted according to the interference feature type and grid confidence. Next, based on the regional location environment characteristics and high-frequency interference feature types, the most suitable denoising model is matched and integrated from a sample denoising diffusion probability model library to construct an adapted image denoising plugin. Finally, these plugins are used to perform targeted denoising on each image sub-region and stitch them together to obtain a high-quality denoised image. Finally, a convolutional neural network is used to achieve accurate land cover segmentation, significantly improving the accuracy and adaptability of remote sensing image land cover segmentation. Compared with traditional methods, the technical solution provided in this application significantly improves the ability to process low-quality remote sensing images, effectively copes with interference factors in various complex environments, achieves adaptive processing of different regional features, avoids the performance degradation problem of traditional methods when processing remote sensing images with multiple interferences, and achieves the technical effects of improving the accuracy of remote sensing image land cover segmentation and improving processing efficiency.

[0084] It should be noted that the order of the embodiments described above is merely for descriptive purposes and does not represent the superiority or inferiority of the embodiments. Furthermore, the above description focuses on specific embodiments of this specification. Additionally, the processes depicted in the accompanying drawings do not necessarily require a specific or sequential order to achieve the desired results. In some implementations, multitasking and parallel processing are possible or may be advantageous.

[0085] The above description is only a preferred embodiment of this application and is not intended to limit this application. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of this application should be included within the protection scope of this application.

[0086] This specification and accompanying drawings are merely illustrative examples of this application and are intended to cover any and all modifications, variations, combinations, or equivalents within the scope of this application. Clearly, those skilled in the art can make various alterations and modifications to this application without departing from its scope. Therefore, if such modifications and modifications fall within the scope of this application and its equivalents, this application intends to include such modifications and modifications.

Claims

1. A method for land cover segmentation in remote sensing images based on a denoising diffusion probability model, characterized in that, The methods include: The target remote sensing image is divided into grids and interference features are identified. Based on the identification results, the grid confidence is evaluated, and several grid confidence scores are obtained, including: The target remote sensing image is divided into regions according to a preset grid size to obtain multiple remote sensing grid images; The interference feature types and interference intensity of the multiple remote sensing grid images are identified sequentially to obtain multiple interference identification results; Based on the interference feature type and interference intensity, grid confidence is evaluated on the multiple interference identification results respectively, and several grid confidence scores are output. Several image sub-regions are obtained by fitting several interference feature types and several grid confidence scores. Each image sub-region is identified by a high-frequency interference feature type and a region image confidence score, including: Q starting points are uniformly selected in the edge region of the target remote sensing image to determine Q starting grid images, where Q is an integer greater than or equal to 4; Starting with the Q initial grid images, based on the interference feature type and grid confidence, the grid similarity between the initial grid image and the adjacent grid images is calculated respectively. If the grid similarity meets the preset grid similarity threshold, the adjacent grid images are merged into the initial grid image to obtain the updated grid image. Starting with the updated grid image, grid fitting is continued. The interference feature type of the updated grid image is the union of the merged grid image and the initial grid image, and the grid confidence is the mean of the merged grid image and the initial grid image. If the grid similarity does not meet the preset grid similarity threshold, a dividing line is set between the starting grid image and the adjacent grid image, and grid fitting is continued from the adjacent grid image as the starting point until the grid fitting is completed, resulting in several image sub-regions. Each image sub-region is identified with a high-frequency interference feature type and a region image confidence level. The sample denoising diffusion probability model library is invoked. Based on the regional location and environmental characteristics, the type of high-frequency interference features, and the confidence level of the regional image, sample model matching and model integration are performed to construct several adapted image denoising plugins, including: Multi-source remote sensing images were collected as training samples and divided according to regional location and environmental characteristics and interference characteristics to obtain several sample image sets; A first sample image set is randomly selected. With the goal of minimizing the denoising error, the denoising diffusion probability model is trained by forward diffusion and backward diffusion respectively to obtain the first sample denoising diffusion probability model. The first sample denoising diffusion probability model is then labeled with features based on the location and environmental features of the first region and the first type of interference features. Several sample denoising and diffusion probability models are obtained by sequentially training the several sample image sets, and a sample denoising and diffusion probability model library is constructed. Randomly select the first high-frequency interference feature type and the first region image confidence level of the first image sub-region; Based on the regional location and environmental features and the first high-frequency interference feature type, feature similarity comparison is performed on several feature tags in the sample denoising and diffusion probability model library, and several feature similarity coefficients are output. If the number of similarity coefficients among the plurality of features that are greater than the first similarity coefficient threshold is not zero, then the sample denoising diffusion probability model corresponding to the largest feature similarity coefficient is selected as the first matching model. Calculate the ratio of the confidence level of the preset standard image to the confidence level of the first region image, and round it down in combination with the number of preset standard models to obtain the number of first adaptation models; Based on the first number of adaptation models, multiple first matching models are selected, and the models are integrated with the image fitting layer to construct a first adaptation image denoising plugin, which is then added to the plurality of adaptation image denoising plugins. The image denoising plugins are used to denoise the image sub-regions respectively, and the images of the denoised regions are obtained. The images are then stitched together to obtain the denoised remote sensing image. A convolutional neural network is used to segment ground features in the denoised remote sensing image, and a ground feature classification map is output.

2. The remote sensing image land cover segmentation method based on a denoising diffusion probability model according to claim 1, characterized in that, The types of interference features include clouds, cloud shadows, haze, localized uneven lighting, strip noise, random noise, building edges, road edges, water reflections, ripples, image registration errors, terrain shadows, and texture anomalies.

3. The remote sensing image land cover segmentation method based on a denoising diffusion probability model according to claim 1, characterized in that, If the number of similarity coefficients greater than the first similarity coefficient threshold is 0 and the number of similarity coefficients greater than the second similarity coefficient threshold is not 0, the similarity coefficients that meet the conditions are set as selected similarity coefficients, wherein the second similarity coefficient threshold is less than the first similarity coefficient threshold. Select the sample denoising diffusion probability model corresponding to the selected feature similarity coefficient, and construct the second matching model set; The product of the first number of adapted models and the similarity coefficient of the selected features is rounded down to the second number of adapted models, thus obtaining the second number of adapted models. Based on the second number of adaptation models and the second matching model set, the model is integrated with the image fitting layer to construct the first adaptation image denoising plugin.

4. The remote sensing image feature segmentation method based on a denoising diffusion probability model according to claim 1, characterized in that, If the number of similarity coefficients among the aforementioned features that are greater than the second similarity coefficient threshold is 0, then the third matching model set is constructed by selecting all samples from the denoising diffusion probability model. The product of the first number of adapted models and the feature similarity coefficient is rounded down to the third number of adapted models, thus obtaining the third number of adapted models. Based on the third matching model set and the number of third adaptation models, the first adaptation image denoising plugin is constructed by integrating the model with the image fitting layer.

Images