Optical remote sensing image ground target extraction system and method based on deep learning

Abstract

The invention discloses an optical remote sensing image ground target extraction system and method based on deep learning in the technical field of remote sensing image processing. The method comprises the following steps: acquiring an original optical remote sensing image; performing radiation correction, geometric correction, denoising processing, resolution normalization and ground interference suppression processing on the original optical remote sensing image to obtain a standard image with an enhanced ground target; and inputting the ground target enhanced standardized image into an improved convolutional neural network, the improved convolutional neural network comprising a ground feature extraction skeleton, a multi-scale feature fusion module and a hierarchical attention enhancement module, and extracting ground target feature maps of different hierarchies. According to the invention, through ground interference suppression, hierarchical attention enhancement and multi-type target adaptation optimization, accurate and efficient extraction of the ground target in a complex scene is realized, and the defects of a traditional method and an existing deep learning method in anti-interference capability, target adaptability and extraction precision are overcome.

Description

Technical Field

[0001] This invention relates to a deep learning-based system and method for extracting ground targets from optical remote sensing images, belonging to the field of remote sensing image processing technology. Background Technology

[0002] Optical remote sensing imagery, with its high resolution, rich spectral information, and wide coverage, has become a core data source for ground target monitoring. Ground target extraction, a crucial task in remote sensing image interpretation, aims to separate ground entities of interest, such as buildings, roads, and water bodies, from complex backgrounds, providing foundational data for subsequent spatial analysis and decision-making. Traditional ground target extraction methods mainly include threshold segmentation, edge detection, region growing, and traditional machine learning (support vector machines, random forests). However, these methods have significant limitations: firstly, they are sensitive to cloud cover, shadow interference, and terrain undulations, making it difficult to distinguish between ground targets and non-ground interference; secondly, ground targets vary greatly in scale (e.g., large building complexes versus small houses, wide highways versus rural paths), making it difficult to adapt to a single feature extraction model; and thirdly, they rely on manually designed ground features (e.g., texture, spectral indices), resulting in insufficient generalization ability and an inability to adapt to diverse remote sensing imagery and ground target types.

[0003] Existing deep learning-based methods (such as U-Net and FCN) achieve end-to-end target extraction through an encoder-decoder structure, but they lack specific design for ground targets. On the one hand, they do not specifically enhance the ability to distinguish ground targets from non-ground interference (clouds, high-altitude noise), leading to problems such as cloud residue and false detection of shadows. On the other hand, they lack differentiated feature optimization for different types of ground targets (hard targets such as buildings / roads, soft targets such as water bodies / vegetation), limiting extraction accuracy. In addition, fixed threshold segmentation cannot adapt to the differences in grayscale distribution of different ground targets, further affecting the extraction effect. Therefore, there is an urgent need for an optical remote sensing image extraction technology that focuses on the characteristics of ground targets, enhances anti-interference capabilities, and adapts to multiple types of ground targets to improve the accuracy and robustness of ground target extraction. Summary of the Invention

[0004] The purpose of this invention is to overcome the shortcomings of the prior art and provide a ground target extraction system and method based on deep learning for optical remote sensing images. This system focuses on the characteristics of ground targets, enhances anti-interference capabilities, and adapts to optical remote sensing image extraction technology for various types of ground targets, thereby improving the accuracy and robustness of ground target extraction.

[0005] To achieve the above objectives, the present invention is implemented using the following technical solution:

[0006] In a first aspect, the present invention provides a method for extracting ground targets from optical remote sensing images based on deep learning, comprising:

[0007] Acquire raw optical remote sensing images;

[0008] Radiometric correction, geometric correction, denoising, resolution normalization, and ground interference suppression are performed on the original optical remote sensing images to obtain a standardized image with enhanced ground targets.

[0009] The normalized image with enhanced ground targets is input into an improved convolutional neural network, which includes a ground feature extraction skeleton, a multi-scale feature fusion module, and a hierarchical attention enhancement module to extract ground target feature maps at different levels.

[0010] The multi-scale feature map is weighted by the hierarchical attention enhancement module, which enhances non-ground interference such as cloud and high-altitude noise, including ground target area features. Then, the shallow detail features and deep semantic features are fused by the feature fusion module to obtain the ground target fused feature map.

[0011] Based on the pixel distribution characteristics of the fused feature map of ground targets, an adaptive threshold algorithm is used to determine the optimal segmentation threshold, and the fused feature map is binarized to obtain candidate regions of ground targets.

[0012] Morphological processing, hole filling, edge smoothing, and non-ground false target removal are performed on the candidate ground target regions to remove false targets including cloud remnants and shadows that are falsely detected, thus obtaining the final ground target extraction results.

[0013] Furthermore, the ground interference suppression processing includes cloud detection and masking, and shadow correction. Cloud detection uses a threshold method combined with texture features to identify cloud areas and generate a mask. Shadow correction uses a radiation compensation algorithm to restore the true spectral information of the ground target.

[0014] Furthermore, the improved convolutional neural network is based on the ResNet-50 network as its backbone. It first extracts shallow detail features, mid-level morphological features, and deep semantic features of ground targets through multiple stages, and then achieves multi-scale fusion through upsampling and channel concatenation.

[0015] Furthermore, the hierarchical attention enhancement module includes a target type attention submodule and a spatial attention submodule. The target type attention submodule adaptively adjusts the feature channel weights through a squeeze-excitation operation based on the spectral characteristics of different ground targets. The spatial attention submodule extracts spatial features through 3×3 convolution to generate a spatial attention map, which enhances the features of the ground area and suppresses the features of the high-altitude area.

[0016] Furthermore, the formula for calculating the channel weight is as follows:

[0017]

[0018] In the formula: W c Here, H represents the attention weight for the c-th feature channel; H and W represent the spatial height and width of the feature map; F represents the attention weight for the c-th feature channel. c (i,j) represents the pixel value of the c-th feature map at position (i,j); Z is the learnable parameter matrix; σ is the Sigmoid activation function.

[0019] Furthermore, the adaptive threshold algorithm is based on the maximum inter-class variance method, and dynamically adjusts the threshold by combining the gray-level histogram of the ground target, optimizing the threshold calculation logic for different ground target types.

[0020] Furthermore, the removal of non-ground false targets is achieved through ground elevation constraints and morphological feature screening, including: filtering areas higher than the preset ground elevation based on digital elevation model data, and removing non-ground irregular false targets through morphological features, including area threshold and aspect ratio.

[0021] Secondly, the present invention provides a deep learning-based ground target extraction system for optical remote sensing images, comprising:

[0022] Data acquisition module: used to acquire raw optical remote sensing images;

[0023] Data processing module: used to perform radiometric correction, geometric correction, noise reduction, resolution normalization and ground interference suppression on the original optical remote sensing images to obtain standardized images with enhanced ground targets;

[0024] Feature extraction module: This module is used to input the normalized image of the enhanced ground target into an improved convolutional neural network. The improved convolutional neural network includes a ground feature extraction skeleton, a multi-scale feature fusion module, and a hierarchical attention enhancement module to extract ground target feature maps at different levels.

[0025] The allocation and fusion module is used to assign weights to the multi-scale feature map through the hierarchical attention enhancement module, enhance non-ground interference such as cloud and high-altitude noise, including ground target area features, and then fuse shallow detail features and deep semantic features through the feature fusion module to obtain the ground target fusion feature map.

[0026] The segmentation processing module is used to determine the optimal segmentation threshold based on the pixel distribution characteristics of the fused feature map of ground targets, and to perform binarization processing on the fused feature map to obtain candidate regions of ground targets.

[0027] The target extraction module is used to perform morphological processing, hole filling, edge smoothing, and non-ground false target removal on ground target candidate regions. It removes false targets including cloud remnants and shadows that are falsely detected, and obtains the final ground target extraction result.

[0028] Thirdly, the present invention provides a ground target extraction device based on deep learning for optical remote sensing images, including a processor and a storage medium;

[0029] The storage medium is used to store instructions;

[0030] The processor is configured to operate according to the instructions to perform the steps of the method according to any of the foregoing.

[0031] Fourthly, the present invention provides a computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of any of the methods described above.

[0032] Compared with the prior art, the beneficial effects achieved by the present invention are as follows:

[0033] This solution addresses the morphological and spectral characteristics of ground targets such as buildings, roads, water bodies, and vegetation. It constructs a deep learning network that integrates ground feature enhancement, combining a hierarchical attention mechanism and an adaptive threshold optimization strategy to achieve accurate extraction of ground targets from optical remote sensing images in complex scenarios. First, the remote sensing image undergoes ground target-oriented preprocessing to suppress non-ground interference such as clouds and shadows. Then, an improved convolutional neural network extracts multi-scale ground features, and a hierarchical attention module enhances ground target features and suppresses non-ground interference. Finally, adaptive threshold segmentation and post-processing optimization of ground targets output high-precision ground target extraction results. This solution solves the problems of low extraction accuracy and weak anti-interference ability of traditional methods in scenarios with cloud cover, terrain undulations, and large differences in ground target scale. It is applicable to multiple fields such as land spatial planning, ecological environment monitoring, disaster emergency response, and traffic network surveys, and has broad application prospects. Attached Figure Description

[0034] The accompanying drawings, which form part of this specification, are used to provide a further understanding of the invention. The illustrative embodiments of the invention and their descriptions are used to explain the invention and do not constitute an undue limitation of the invention. In the drawings:

[0035] Figure 1 This is a flowchart illustrating the deep learning-based ground target extraction method for optical remote sensing images provided in Embodiment 1 of the present invention. Detailed Implementation

[0036] The present invention will now be described in detail with reference to the accompanying drawings and embodiments. It should be noted that, unless otherwise specified, the embodiments and features described herein can be combined with each other.

[0037] The following detailed description is exemplary and intended to provide further detailed explanation of the invention. Unless otherwise specified, all technical terms used in this invention have the same meaning as commonly understood by one of ordinary skill in the art. The terminology used in this invention is for describing particular embodiments only and is not intended to limit the scope of exemplary embodiments according to the invention.

[0038] Example 1:

[0039] Please see Figure 1 This embodiment proposes a deep learning-based method for extracting ground targets from optical remote sensing images, employing the following technical solution:

[0040] I. Remote Sensing Image Preprocessing

[0041] First, acquire raw optical remote sensing images and their accompanying digital elevation models (DEMs). Image sources include satellite and aerial remote sensing data. The preprocessing steps add ground interference suppression to the traditional workflow. By performing radiometric correction, geometric correction, denoising, resolution normalization, and ground interference suppression on the raw optical remote sensing images, a standardized image with enhanced ground targets is obtained. The specific processing steps include:

[0042] Radiation correction: The FLAASH atmospheric correction model is used to eliminate atmospheric scattering and sensor errors, restoring the true radiation information of the ground target;

[0043] Geometric correction: Based on ground control points and DEM data, a rational function model is used to correct geometric distortions caused by terrain undulations, ensuring the accuracy of the spatial location of ground targets;

[0044] Noise reduction: A combination of Gaussian filtering (to remove Gaussian noise) and median filtering (to remove salt-and-pepper noise) is used to preserve the edge details of ground targets;

[0045] Resolution normalization: Unifies the scaling of images of different resolutions to a preset resolution (1m, 2m or 30m) to adapt to the scale of ground targets;

[0046] Ground interference suppression: ① Cloud detection and masking: Cloud areas are identified by using a threshold method (setting a reflectivity threshold) combined with texture features (uniformity of cloud texture) to generate cloud masks to shield non-ground interference; ② Shadow correction: A radiation compensation algorithm is used to restore the true brightness of ground targets in shadow areas based on the spectral information of adjacent non-shadow areas, avoiding missed detections caused by shadows.

[0047] II. Design of Multi-Scale Ground Feature Extraction Network

[0048] The normalized imagery with enhanced ground targets is input into an improved convolutional neural network. This improved convolutional neural network includes a ground feature extraction skeleton, a multi-scale feature fusion module, and a hierarchical attention enhancement module to extract ground target feature maps at different levels. The deep learning network designed in this scheme uses the ResNet-50 network as its basic skeleton and optimizes the network structure for the characteristics of ground targets, specifically including:

[0049] Ground feature extraction skeleton: The first four stages of the ResNet-50 network serve as the basic feature extraction part. The first and second stages extract shallow detail features of ground targets (such as building edges, road textures, and water boundaries), the third stage extracts mid-level morphological features (such as building outlines and road directions), and the fourth stage extracts deep semantic features (such as ground target category information).

[0050] Multi-scale feature fusion module: The third-stage feature map is upsampled by 2 times and then stitched with the second-stage feature map channel; the fourth-stage feature map is upsampled by 4 times and then stitched with the above-mentioned stitched feature map again, realizing the fusion of shallow, medium and deep features, taking into account the extraction needs of small ground targets (such as small houses) and large targets (such as urban building complexes);

[0051] Layered Attention Enhancement Module: Designed specifically for ground targets, this module weights multi-scale feature maps to enhance ground target region features (edges, textures, contours) and suppress non-ground interference such as clouds and high-altitude noise. Then, a feature fusion module fuses shallow detail features with deep semantic features to obtain a fused ground target feature map. The layered attention enhancement module includes a target type attention submodule and a spatial attention submodule, wherein:

[0052] The target type attention submodule: Based on the spectral characteristics of different ground targets (e.g., low reflectivity of water, medium reflectivity of buildings, and high near-infrared reflectivity of vegetation), it adaptively adjusts the feature channel weights through a squeeze-excitation operation (global average pooling + fully connected layer + sigmoid activation) to enhance the feature response corresponding to the target type. It should be noted that the formula for calculating the channel weights is:

[0053]

[0054] In the formula: W c Here, H represents the attention weight for the c-th feature channel; H and W represent the spatial height and width of the feature map; F represents the attention weight for the c-th feature channel. c (i,j) represents the pixel value at position (i,j) of the c-th feature map; Z is the learnable parameter matrix; σ is the Sigmoid activation function. Weights W cThis is used to perform channel weighting on the feature map, enhance the feature response related to ground targets (such as water bodies and buildings), and suppress irrelevant channels. The weighted feature map is then input into the spatial attention submodule for further optimization.

[0055] Spatial Attention Submodule: Combining ground elevation information provided by DEM data, spatial features are extracted through 3×3 convolution to generate a spatial attention map, which enhances features in ground areas (within a reasonable elevation range) and suppresses features in high-altitude areas (clouds, high-altitude noise).

[0056] III. Adaptive Threshold Segmentation

[0057] After feature extraction and optimization, based on the pixel distribution characteristics of the fused feature map of ground targets, an adaptive thresholding algorithm is used to determine the optimal segmentation threshold. The fused feature map is then binarized to obtain candidate regions for ground targets. To address the differences in grayscale distribution among different ground targets, an adaptive thresholding segmentation based on the Otsu's algorithm (maximum inter-class variance method) is employed. The optimization process is as follows:

[0058] 1) Statistically analyze the gray-level histogram of the fused feature map to determine the range of gray-level values;

[0059] 2) Adjust the threshold calculation logic for different ground target types: For hard targets (buildings, roads), focus on distinguishing the gray-level abrupt changes between the target and the background; for soft targets (water bodies, vegetation), focus on adapting to gray-level gradation characteristics.

[0060] 3) Iterate through all possible thresholds, calculate the inter-class variance between the target class and the background class, and take the threshold corresponding to the maximum value of the inter-class variance as the optimal segmentation threshold. In this embodiment, the expression for the optimal segmentation threshold is:

[0061]

[0062] In the formula: T adaptive ω is the optimal segmentation threshold; T is the threshold; hard For hard targets, the weighting coefficient is used. ω represents the inter-class variance of the hard target at threshold T; soft For soft targets, the weighting coefficients are used. The inter-class variance of the soft target at threshold T; This represents the threshold T corresponding to the maximum inter-class variance. The optimal segmentation threshold T. adaptive Used to binarize the fused feature map and generate candidate regions for ground targets. The weight coefficients are dynamically adjusted based on the output of the target type attention module; ω is increased when hard targets account for a high proportion. hard Weighting increases ω when soft targets have a high proportion. soft Weights.

[0063] 4) Binarize the fused feature map based on the optimal segmentation threshold to obtain ground target candidate regions (pixel value of 1) and background regions (pixel value of 0).

[0064] IV. Post-processing optimization of ground targets

[0065] To further improve the accuracy of ground target extraction, a post-processing step is added to remove non-ground false targets. Morphological processing, hole filling, edge smoothing, and non-ground false target removal are performed on candidate ground target regions to remove false targets such as cloud remnants and shadow false detections, resulting in the final ground target extraction result. Specifically, this includes:

[0066] Morphological processing: Using a 3×3 rectangular structural element, erosion followed by dilation is performed to remove small-area pseudo-targets (such as cloud residues and noise points).

[0067] Void filling: Filling holes (such as cavities inside buildings or shadows of islands in water) in the candidate area of ​​ground targets to ensure the integrity of the targets;

[0068] Edge smoothing: Gaussian filtering is used to smooth the target edges, reducing jagged edges;

[0069] Non-ground false target removal: ① Elevation constraint: Combine DEM data to filter areas above the preset ground elevation (e.g., 50m, which can be customized) to remove residual high-altitude interference; ② Morphological screening: Remove irregular non-ground false targets by using morphological features such as area threshold (to remove excessively small areas) and aspect ratio (e.g., the aspect ratio of roads is greater than the preset value).

[0070] Compared with existing technologies, this solution effectively filters interference from clouds, high-altitude noise, and shadows through ground interference suppression processing and a hierarchical attention module, solving the problem of weak anti-interference capability of traditional methods and specifically enhancing the ability to distinguish between ground targets and non-ground interference. Through the target type attention submodule and multi-scale feature fusion, it takes into account the extraction needs of hard / soft ground targets and ground targets of different sizes, with stronger generalization ability and adaptability to multiple types and scales of ground targets. The newly added target type configuration module supports users to customize target types and fine-tune datasets, adapting to diverse application scenarios such as land monitoring and traffic census, making it more flexible, convenient and practical.

[0071] Example 2:

[0072] The deep learning-based optical remote sensing image ground target extraction system can implement the deep learning-based optical remote sensing image ground target extraction method described in Example 1, including:

[0073] Data acquisition module: used to acquire raw optical remote sensing images;

[0074] Data processing module: used to perform radiometric correction, geometric correction, noise reduction, resolution normalization and ground interference suppression on the original optical remote sensing images to obtain standardized images with enhanced ground targets;

[0075] Feature extraction module: This module is used to input the normalized image of the enhanced ground target into an improved convolutional neural network. The improved convolutional neural network includes a ground feature extraction skeleton, a multi-scale feature fusion module, and a hierarchical attention enhancement module to extract ground target feature maps at different levels.

[0076] The allocation and fusion module is used to assign weights to the multi-scale feature map through the hierarchical attention enhancement module, enhance non-ground interference such as cloud and high-altitude noise, including ground target area features, and then fuse shallow detail features and deep semantic features through the feature fusion module to obtain the ground target fusion feature map.

[0077] The segmentation processing module is used to determine the optimal segmentation threshold based on the pixel distribution characteristics of the fused feature map of ground targets, and to perform binarization processing on the fused feature map to obtain candidate regions of ground targets.

[0078] The target extraction module is used to perform morphological processing, hole filling, edge smoothing, and non-ground false target removal on ground target candidate regions. It removes false targets including cloud remnants and shadows that are falsely detected, and obtains the final ground target extraction result.

[0079] Example 3:

[0080] This invention also provides a deep learning-based ground target extraction device for optical remote sensing images, which can realize the deep learning-based ground target extraction method for optical remote sensing images described in Embodiment 1, including a processor and a storage medium;

[0081] The storage medium is used to store instructions;

[0082] The processor is configured to operate according to the instructions to perform the steps of the following method:

[0083] Acquire raw optical remote sensing images;

[0084] Radiometric correction, geometric correction, denoising, resolution normalization, and ground interference suppression are performed on the original optical remote sensing images to obtain a standardized image with enhanced ground targets.

[0085] The normalized image with enhanced ground targets is input into an improved convolutional neural network, which includes a ground feature extraction skeleton, a multi-scale feature fusion module, and a hierarchical attention enhancement module to extract ground target feature maps at different levels.

[0086] The multi-scale feature map is weighted by the hierarchical attention enhancement module, which enhances non-ground interference such as cloud and high-altitude noise, including ground target area features. Then, the shallow detail features and deep semantic features are fused by the feature fusion module to obtain the ground target fused feature map.

[0087] Based on the pixel distribution characteristics of the fused feature map of ground targets, an adaptive threshold algorithm is used to determine the optimal segmentation threshold, and the fused feature map is binarized to obtain candidate regions of ground targets.

[0088] Morphological processing, hole filling, edge smoothing, and non-ground false target removal are performed on the candidate ground target regions to remove false targets including cloud remnants and shadows that are falsely detected, thus obtaining the final ground target extraction results.

[0089] Example 4:

[0090] This invention also provides a computer-readable storage medium that implements the deep learning-based ground target extraction method for optical remote sensing images described in Embodiment 1. The medium stores a computer program that, when executed by a processor, performs the steps of the following method:

[0091] Acquire raw optical remote sensing images;

[0092] Radiometric correction, geometric correction, denoising, resolution normalization, and ground interference suppression are performed on the original optical remote sensing images to obtain a standardized image with enhanced ground targets.

[0093] The normalized image with enhanced ground targets is input into an improved convolutional neural network, which includes a ground feature extraction skeleton, a multi-scale feature fusion module, and a hierarchical attention enhancement module to extract ground target feature maps at different levels.

[0094] The multi-scale feature map is weighted by the hierarchical attention enhancement module, which enhances non-ground interference such as cloud and high-altitude noise, including ground target area features. Then, the shallow detail features and deep semantic features are fused by the feature fusion module to obtain the ground target fused feature map.

[0095] Based on the pixel distribution characteristics of the fused feature map of ground targets, an adaptive threshold algorithm is used to determine the optimal segmentation threshold, and the fused feature map is binarized to obtain candidate regions of ground targets.

[0096] Morphological processing, hole filling, edge smoothing, and non-ground false target removal are performed on the candidate ground target regions to remove false targets including cloud remnants and shadows that are falsely detected, thus obtaining the final ground target extraction results.

[0097] As is known from common technical knowledge, this invention can be implemented through other embodiments that do not depart from its spirit or essential characteristics. Therefore, the disclosed embodiments described above are merely illustrative and not exhaustive. All modifications within the scope of this invention or its equivalents are included in this invention.

[0098] Those skilled in the art will understand that embodiments of the present invention can be provided as methods, systems, or computer program products. Therefore, the present invention can take the form of a completely hardware embodiment, a completely software embodiment, or an embodiment combining software and hardware aspects. Furthermore, the present invention can take the form of a computer program product embodied on one or more computer-usable storage media (including, but not limited to, disk storage, CD-ROM, optical storage, etc.) containing computer-usable program code.

[0099] This invention is described with reference to flowchart illustrations and / or block diagrams of methods, apparatus (systems), and computer program products according to embodiments of the invention. It will be understood that each block of the flowchart illustrations and / or block diagrams, and combinations of blocks in the flowchart illustrations and / or block diagrams, can be implemented by computer program instructions. These computer program instructions can be provided to a processor of a general-purpose computer, special-purpose computer, embedded processor, or other programmable data processing apparatus to produce a machine, such that the instructions, which execute via the processor of the computer or other programmable data processing apparatus, generate instructions for implementing the flowchart illustrations and / or block diagrams. Figure 1 One or more processes and / or boxes Figure 1 A device that provides the functions specified in one or more boxes.

[0100] These computer program instructions may also be stored in a computer-readable storage medium that can direct a computer or other programmable data processing device to function in a particular manner, such that the instructions stored in the computer-readable storage medium produce an article of manufacture including instruction means, which are implemented in a process Figure 1 One or more processes and / or boxes Figure 1 The function specified in one or more boxes.

[0101] These computer program instructions may also be loaded onto a computer or other programmable data processing equipment to cause a series of operational steps to be performed on the computer or other programmable equipment to produce a computer-implemented process, thereby providing instructions that execute on the computer or other programmable equipment for implementing the process. Figure 1 One or more processes and / or boxes Figure 1 The steps of the function specified in one or more boxes.

[0102] Finally, it should be noted that the above embodiments are only used to illustrate the technical solutions of the present invention and not to limit it. Although the present invention has been described in detail with reference to the above embodiments, those skilled in the art should understand that modifications or equivalent substitutions can still be made to the specific implementation of the present invention. Any modifications or equivalent substitutions that do not depart from the spirit and scope of the present invention should be covered within the scope of protection of the claims of the present invention.

Claims

1. A deep learning-based method for extracting ground targets from optical remote sensing images, characterized by: include: Acquire raw optical remote sensing images; Radiometric correction, geometric correction, denoising, resolution normalization, and ground interference suppression are performed on the original optical remote sensing images to obtain a standardized image with enhanced ground targets. The normalized image with enhanced ground targets is input into an improved convolutional neural network, which includes a ground feature extraction skeleton, a multi-scale feature fusion module, and a hierarchical attention enhancement module to extract ground target feature maps at different levels. The multi-scale feature map is weighted by the hierarchical attention enhancement module, which enhances non-ground interference such as cloud and high-altitude noise, including ground target area features. Then, the shallow detail features and deep semantic features are fused by the feature fusion module to obtain the ground target fused feature map. Based on the pixel distribution characteristics of the fused feature map of ground targets, an adaptive threshold algorithm is used to determine the optimal segmentation threshold, and the fused feature map is binarized to obtain candidate regions of ground targets. Morphological processing, hole filling, edge smoothing, and non-ground false target removal are performed on the candidate ground target regions to remove false targets including cloud remnants and shadows that are falsely detected, thus obtaining the final ground target extraction results.

2. The method for extracting ground targets from optical remote sensing images based on deep learning according to claim 1, characterized in that, The ground interference suppression process includes cloud detection and masking, and shadow correction. Cloud detection uses a threshold method combined with texture features to identify cloud areas and generate a mask. Shadow correction uses a radiation compensation algorithm to restore the true spectral information of the ground target.

3. The method for extracting ground targets from optical remote sensing images based on deep learning according to claim 1, characterized in that, The improved convolutional neural network is based on the ResNet-50 network as its backbone. It first extracts shallow detail features, mid-level morphological features and deep semantic features of ground targets through multiple stages, and then achieves multi-scale fusion through upsampling and channel concatenation.

4. The method for extracting ground targets from optical remote sensing images based on deep learning according to claim 1, characterized in that, The hierarchical attention enhancement module includes a target type attention submodule and a spatial attention submodule. The target type attention submodule adaptively adjusts the feature channel weights through a squeeze-excitation operation based on the spectral characteristics of different ground targets. The spatial attention submodule extracts spatial features through 3×3 convolution, generates a spatial attention map, enhances features in the ground region, and suppresses features in the high-altitude region.

5. The method for extracting ground targets from optical remote sensing images based on deep learning according to claim 4, characterized in that, The formula for calculating the channel weight is: In the formula: W c Let be the attention weight for the c-th feature channel; H and W represent the spatial height and width of the feature map; F c (i,j) represents the pixel value of the c-th feature map at position (i,j); Z is the learnable parameter matrix; σ is the Sigmoid activation function.

6. The method for extracting ground targets from optical remote sensing images based on deep learning according to claim 1, characterized in that, The adaptive threshold algorithm is based on the maximum inter-class variance method and dynamically adjusts the threshold by combining the gray-level histogram of ground targets. The threshold calculation logic is optimized for different ground target types.

7. The method for extracting ground targets from optical remote sensing images based on deep learning according to claim 1, characterized in that, The removal of non-ground false targets is achieved through ground elevation constraints and morphological feature screening, including: filtering areas higher than the preset ground elevation based on digital elevation model data, and removing non-ground irregular false targets through morphological features, including area threshold and aspect ratio.

8. A deep learning-based ground target extraction system for optical remote sensing images, characterized by: include: Data acquisition module: used to acquire raw optical remote sensing images; Data processing module: used to perform radiometric correction, geometric correction, noise reduction, resolution normalization and ground interference suppression on the original optical remote sensing images to obtain standardized images with enhanced ground targets; Feature extraction module: This module is used to input the normalized image of the enhanced ground target into an improved convolutional neural network. The improved convolutional neural network includes a ground feature extraction skeleton, a multi-scale feature fusion module, and a hierarchical attention enhancement module to extract ground target feature maps at different levels. The allocation and fusion module is used to assign weights to the multi-scale feature map through the hierarchical attention enhancement module, enhance non-ground interference such as cloud and high-altitude noise, including ground target area features, and then fuse shallow detail features and deep semantic features through the feature fusion module to obtain the ground target fusion feature map. The segmentation processing module is used to determine the optimal segmentation threshold based on the pixel distribution characteristics of the fused feature map of ground targets, and to perform binarization processing on the fused feature map to obtain candidate regions of ground targets. The target extraction module is used to perform morphological processing, hole filling, edge smoothing, and non-ground false target removal on ground target candidate regions. It removes false targets including cloud remnants and shadows that are falsely detected, and obtains the final ground target extraction result.

9. A ground target extraction device based on deep learning for optical remote sensing images, characterized in that, Including processor and storage media; The storage medium is used to store instructions; The processor is configured to operate according to the instructions to perform the steps of the method according to any one of claims 1 to 7.

10. A computer-readable storage medium having a computer program stored thereon, characterized in that, When executed by a processor, the program implements the steps of the method according to any one of claims 1 to 7.

Images