A remote sensing image patch interpretation optimization method based on multi-branch network and multi-stage judgment

Abstract

The application discloses a remote sensing image spot interpretation optimization method based on a multi-branch network and multi-stage judgment, and relates to the technical field of remote sensing image intelligent interpretation. The method overcomes the technical bottleneck of the existing remote sensing image intelligent interpretation technology in processing complex scenes, especially the problems of irregular spot boundary, geometric distortion and low automation degree caused by a traditional end-to-end segmentation model. The method specifically comprises the following steps: S1, multi-branch collaborative feature extraction; S2, two-class reliable screening based on spatial priori; S3, edge fine reconstruction guided by a detection frame; S4, geometric rule constraint spot morphological purification; and S5, result fusion and vectorization output.

Description

Technical Field

[0001] This invention relates to the field of intelligent interpretation technology of remote sensing images, specifically to an optimization method for remote sensing patch interpretation based on multi-branch networks and multi-stage analysis. Background Technology

[0002] In the field of intelligent interpretation of remote sensing images, the most mainstream technical paradigm in industry and academia is currently an end-to-end semantic segmentation architecture based on fully convolutional networks or visual Transformers, with typical examples including the U-Net series, DeepLab series, and SegNet. The core logic of this type of technology is to construct a deep neural network with an "encoder-decoder" structure. After inputting the remote sensing image into the model, it directly regresses the pixel-level category labels and outputs a preliminary segmentation mask.

[0003] Taking the classic U-Net and its variants as examples, the mainstream technical architecture and operating logic typically include the following steps: First, the encoder (downsampling path) extracts high-level semantic features of the image layer by layer to capture the category information of the target ground features; then, the decoder (upsampling path) gradually restores the spatial resolution, and shallow detailed texture information is fused with the help of skip connections; finally, a probability map or binary mask is output through the classification head. In actual business deployment, the large image is usually divided into fixed-size tiles to input the model for inference, and then the output tile results are stitched together to restore the global patch.

[0004] Although the aforementioned end-to-end segmentation model achieves high pixel accuracy on specific datasets, its inherent architectural limitations make it difficult to meet business-level "usability" standards when applied to complex real-world scenarios (such as multi-scale ground cover coexistence, shadow occlusion, and inter-class spectral confusion). These shortcomings manifest in the following three dimensions: 1. The limited sensitivity of single-scale feature extraction leads to insufficient multi-scale sensitivity: Traditional models typically rely on a single backbone network for feature extraction, resulting in a relatively fixed receptive field. When dealing with features in remote sensing imagery that vary significantly in scale and orientation, single-scale feature representation struggles to capture both deep semantics and shallow details. This leads to the model easily missing small targets, having blurred boundaries when dealing with large targets, and failing to effectively capture the contours of features with specific rotation angles. The output patches often exhibit jagged edges, internal holes, or missegmentation due to background noise.

[0005] 2. "Black-box" segmentation leads to geometric distortion: Due to quantization errors introduced by multiple downsampling and upsampling operations in the network structure, and the excessive emphasis on pixel accuracy in the loss function, the segmentation results output by existing models often lack constraints on the geometric topology of the patches. The interpreted patch boundaries usually exhibit irregular jagged or stepped shapes, poor shape compactness, and adjacent features are prone to sticking together. This geometric irregularity makes the original results unsuitable for direct use as standard vector data in GIS systems, requiring tedious manual smoothing and trimming operations.

[0006] 3. The interpretation and judgment logic are disconnected, lacking self-assessment of credibility: Existing technology is essentially a purely data-driven classifier, whose output is only a pixel probability map, lacking the ability to "understand" and "judge" the high-level semantics of the interpretation results. The model cannot automatically assess whether a segmented patch conforms to the physical laws of the real world. Therefore, the existing process must rely on manually setting empirical thresholds for post-processing (such as area filtering, morphological opening and closing operations). This loosely coupled "model segmentation + manual cleaning" mode is highly subjective and has weak generalization ability, failing to achieve a fully automated closed loop from image input to high-quality patch output, severely restricting the real-time performance and standardization level of remote sensing information services. Summary of the Invention

[0007] The purpose of this invention is to overcome the technical bottlenecks of existing intelligent remote sensing image interpretation technologies when handling complex scenes, particularly addressing the problems of irregular patch boundaries, geometric distortion, and low automation caused by traditional end-to-end segmentation models. This invention proposes an optimized remote sensing patch interpretation method based on multi-branch networks and multi-stage analysis. By introducing a multi-branch feature extraction mechanism and a multi-stage collaborative reasoning architecture of "coarse screening-fine screening-fusion," this method aims to solve the core pain points of existing technologies, such as missed detection of multi-scale targets due to the single feature extraction method, jagged boundaries due to the lack of geometric constraints, and high false alarm rates due to the separation of interpretation and analysis logic. This method strives to achieve fully automated production of high-quality vector patches from raw images, improving the geometric accuracy and business usability of the interpretation results.

[0008] The method is specifically as follows: S1. Multi-branch collaborative feature extraction: Receive the pre-processed high-resolution remote sensing image, input the image into the multi-branch network for interpretation, and generate a binary initial segmentation mask containing candidate patch regions, which is then input into step S2 for processing. S2. Reliable binary classification screening based on spatial prior: For false positives in the automatic interpretation results, a binary classification model that integrates spatial prior information is constructed to distinguish between true and false positives of candidate patches. S3. Edge Refinement Guided by Detection Box: To address the issues of jagged edges, breaks, and inaccurate positioning of patches after initial screening, the FastSAM model is introduced for sub-pixel-level repair. S4. Geometry-constrained morphological cleanup of patches: Using the binary instance mask output from step S3 as input, geometric noise is eliminated from the mask to improve the topological rationality and business compliance of the patches corresponding to the mask. S5. Result Fusion and Vectorization Output: The multi-stage processed map features are vectorized to generate the final interpretation result.

[0009] Furthermore, the multi-branch network uses Transformer and its variants as the backbone network, combined with two specially designed auxiliary branches in parallel. The backbone network is responsible for capturing long-distance dependencies and global feature associations. The first auxiliary branch enhances detail resolution through high-resolution paths and local convolutional filters to ensure that small features are not lost. The second auxiliary branch optimizes boundary extraction through specific magnification layers and edge-sensitive mechanisms. The features of the three branches are deeply fused through a weighted fusion strategy to generate an initial segmentation result that combines global consistency with local details.

[0010] Furthermore, the backbone network passes through the following units sequentially from input to output: image segmentation unit, initial Swin stage (this unit consists of a Linear Embedding layer and two Swin Transformer Blocks), three downsampling Swin modules with identical structures (each unit consists of a Patch Merging layer and N Swin Transformer Blocks, where the number of layers N is 2, 6 and 2 respectively), pyramid pooling module, feature aggregation module, and global-local module; The first auxiliary branch passes through the following layers from input to output: downsampling layer; first convolutional layer; high-resolution residual block; second convolutional layer; and upsampling and fusion layer. The second auxiliary branch passes through the following layers from input to output: feature input layer, edge response extraction layer, multi-scale dilated convolutional layer, feature splicing layer, dimensionality reduction convolutional layer, and addition fusion layer.

[0011] Furthermore, step S2 specifically includes: Step S2.1: Four-channel input construction; First, perform morphological closing operation and connected component analysis on the binary mask output from step S1 to extract the minimum bounding rectangle of each candidate region; then, generate a spatial indicator mask with the same size as the original image based on the rectangle; finally, stitch the spatial mask channel with the three channels of the original remote sensing image in the channel dimension to construct a four-channel input tensor. Step S2.2: Model Inference and Filtering; The four-channel tensor is input into the improved EfficientNet binary classification model. Based on the ImageNet pre-trained weights, the first layer convolutional kernel is expanded from 3 channels to 4 channels and fine-tuned through transfer learning to determine whether the input patch is a real target. The improved EfficientNet binary classification model outputs candidate regions that are determined to be real target patches to enter the next stage. At the same time, the minimum bounding rectangle corresponding to each candidate region that is determined to be a real target patch is output as a high-confidence detection box for use in step S3.

[0012] Furthermore, step S3 specifically includes: Step S3.1: Spatial prior guidance mechanism; The high-confidence detection box generated in step S2 is used as a spatial prior cue input to the cue encoder of the FastSAM model; Step S3.2: Local instance segmentation; The FastSAM model activates feature responses within the local region defined by the detection box to generate a binary instance mask.

[0013] Furthermore, step S4 specifically includes: Step S4.1: Connectivity analysis and fragment removal; perform neighborhood connectivity analysis on the reconstructed mask, remove isolated fragments with an area smaller than a preset threshold, and eliminate the influence of salt-and-pepper noise; Step S4.2: Shape index rule filtering; introduce mathematical morphology indicators for logical filtering, calculate the shape compactness index and circumscribed rectangle fill rate of the patch, set a threshold range to automatically exclude non-target misjudgments caused by the curvature of natural water body shorelines, and retain regular patches that conform to the characteristics of artificial land features.

[0014] The beneficial effects of the method described in this invention are as follows: In the multi-branch feature extraction stage, this invention abandons the traditional single-backbone network feature extraction mode and constructs a multi-branch network structure. This structure can capture feature information of different scales and orientations in parallel, effectively enhancing the model's ability to perceive ground targets (such as buildings and structures) with significant size differences and arbitrary orientations in remote sensing images, providing a rich and robust feature foundation for subsequent accurate segmentation.

[0015] In the initial target screening and classification stage, a lightweight binary classification model is used to quickly scan the entire image, extract scene-level prior features, and achieve global screening of suspected target areas, thereby significantly reducing the redundancy of subsequent high-precision calculations.

[0016] In the local resegmentation stage guided by the detection box, a precise bounding box is generated for the suspected target regions output in the coarse screening stage using the target detection model. This bounding box serves as strong prior information, guiding the segmentation model to perform high-resolution local resegmentation only on the regions within the bounding box. This effectively avoids interference from background noise and solves the edge blurring and adhesion problems caused by contextual redundancy in traditional global segmentation.

[0017] In the geometric feature constraint and patch fusion optimization stage, morphological processing and geometric feature analysis are performed on the pixel-level mask output by resegmentation. By calculating geometric parameters such as patch compactness, fill rate, and aspect ratio, an automatic judgment rule base is established to intelligently remove pseudo-targets that do not conform to real physical characteristics (such as noise that is too fragmented or has discrete shapes). Finally, the intermediate results of the multi-stage output are fused and optimized using the "intersection and then union" strategy to output a final patch result with smooth boundaries, clear semantics, and regular geometry. Attached Figure Description

[0018] Figure 1 This is a flowchart of a remote sensing patch interpretation optimization method based on multi-branch network and multi-stage analysis as described in an embodiment of the present invention; Figure 2 This is a diagram of a multi-branch network structure in an embodiment of the present invention; Figure 3 This is a diagram showing the effect of refined edge reconstruction guided by the detection box in an embodiment of the present invention; Figure 4 This is a diagram illustrating the effect of the building experimental interpretation optimization method in an embodiment of the present invention. Detailed Implementation

[0019] The technical solution of the present invention will now be clearly and completely described with reference to the accompanying drawings. Obviously, the described embodiments are only some, not all, of the embodiments of the present invention. Based on the embodiments of the present invention, all other embodiments obtained by those skilled in the art without creative effort are within the protection scope of the present invention.

[0020] Example 1 This embodiment provides a remote sensing patch interpretation optimization method based on multi-branch networks and multi-stage analysis, combining... Figure 1 The method will be described in detail.

[0021] This solution addresses the pain points of traditional single-model interpretation, such as discontinuous edges, high false detection rate, and inability to distinguish between new and existing changes. It constructs a five-stage closed-loop optimization architecture of "multi-branch extraction—credible screening—edge reconstruction—rule purification—change determination," with the specific steps as follows: Step 1: Feature extraction through multi-branch collaboration This step, serving as the "front-end perception" stage of the method, aims to generate high-quality initial feature patches. Leveraging big data analysis and a multi-sample training mechanism, this step constructs an innovative multi-branch fusion network architecture, overcoming the limitations of traditional single-branch feature extraction.

[0022] The input data for step 1 is a preprocessed high-resolution remote sensing image (RGB three channels). This image is input into a multi-branch network for interpretation, generating a binary initial segmentation mask containing candidate patch regions, which is then input into step 2 for processing.

[0023] Step 1.1: Multi-branch Feature Extraction and Fusion. This step employs a Transformer and its variants as the backbone network, combined with two specially designed auxiliary branches in parallel. The backbone network is responsible for capturing long-range dependencies and global feature associations; the first auxiliary branch enhances detail resolution through high-resolution paths and local convolutional filters, ensuring that small features are not lost; the second auxiliary branch optimizes boundary extraction through specific magnification layers and edge-sensitive mechanisms. The features from the three branches are deeply fused using a weighted fusion strategy to generate an initial segmentation result that combines global consistency with local details.

[0024] Multi-branch network structure such as Figure 2 As shown, the backbone network passes through the following units from input to output: The Patch Parition unit divides the input high-resolution remote sensing image into fixed-size non-overlapping image patches. Each image patch is flattened into a one-dimensional vector, which serves as the input sequence element for the Transformer encoder. This aims to reduce sequence length, preserve local spatial structure information, and provide a basis for subsequent linear embedding. The Initial Swin Stage consists of a Linear Embedding layer and two Swin Transformer Blocks. The function of this stage is as follows: the Linear Embedding layer maps the vector of each image patch to the hidden feature dimension; the two Swin Transformer Blocks employ a shift-window multi-head self-attention mechanism to calculate attention within a local window and initially capture the global context through the shift interaction between windows, generating an initial feature map containing shallow texture and edge information. Three identical downsampling Swin stages are used, each consisting of a Patch Merging layer and N Swin Transformer Blocks, with the number of layers N being 2, 6, and 2 respectively. The three downsampling Swin stages function as follows: the Patch Merging layer concatenates and linearly transforms the feature maps of adjacent 2×2 regions, halving the spatial resolution and doubling the number of channels; subsequently, the N Swin Transformer Blocks further extract multi-scale semantic features. These three stages sequentially refine mid-level semantic features, high-level abstract features, and global contextual features, forming a feature pyramid from fine to coarse. PPM (Pyramid Pooling Module) performs average pooling on feature maps at multiple scales (such as 1×1, 2×2, 3×3, 6×6) to capture contextual information from different receptive fields. It then upsamples the pooled feature maps to the original feature map size and stitches them together along the channel dimension, thereby enhancing the model's ability to perceive multi-scale features and effectively mitigating the problem of missed detections caused by scale differences. The Bag (Feature Aggregation Block) is responsible for summing or concatenating features from different levels of the backbone network by aligning channels through adaptive average pooling and 1×1 convolution, thereby achieving the fusion of multi-level features and enhancing the richness and robustness of feature representation. The Global-Local Block serves to process the global semantic branch and the local detail branch in parallel. The global branch uses global average pooling and fully connected layers to extract scene-level contextual information, while the local branch uses depthwise separable convolution to preserve high-resolution spatial details. The outputs of the two branches are added element-wise, so that the final feature has both global context awareness and local edge preservation capabilities. The first auxiliary branch (high-resolution detail enhancement branch) proceeds sequentially from input to output as follows: Downsampling layer: The input high-resolution remote sensing image is downsampled to half its original size through bilinear interpolation to reduce subsequent computation while preserving sufficient spatial details; First convolutional layer: 3×3 convolution, stride 1, padding 1, output channels 32, followed by batch normalization and ReLU activation function to keep the feature map size unchanged, used to extract shallow detailed features; High-resolution residual blocks consist of two 3×3 convolutions (both with a stride of 1, padding of 1, and 32 output channels). Each convolution is followed by batch normalization and ReLU, and residual connections are introduced to add the input and output element by element. This is used to enhance feature representation and avoid gradient vanishing while maintaining high spatial resolution. The second convolutional layer is a 3×3 convolution with a stride of 2 and padding of 1. It has 48 output channels and is followed by batch normalization and ReLU to halve the feature map size and increase the number of channels, thereby extracting higher-level detailed semantics. Upsampling and fusion layer: The feature map is upsampled to the original image size through bilinear interpolation, and then the number of channels is adjusted by 1×1 convolution to match the features of the corresponding layer of the backbone network. Finally, it is added element by element with the features of the intermediate layer of the backbone network to output the fused features with enhanced details.

[0025] This branch maintains the feature map size at least 1 / 4 of the original image throughout the process, and is specifically designed to capture small features and boundary details, ensuring that small targets are not lost.

[0026] The second auxiliary branch (edge ​​enhancement branch) proceeds sequentially from input to output as follows: Feature Input Layer: Receives the feature map output by the first downsampled Swin module (i.e., the Swin Stage with N=2 layers) in the backbone network. This feature map retains relatively rich spatial structure and edge information. Edge response extraction layer: A set of learnable edge convolution kernels (3×3 convolution, stride of 1, padding of 1, output channel number of 16) is used to simulate the function of the Sobel operator to extract the edge response map in the feature map, followed by batch normalization and ReLU activation function; Multi-scale dilated convolutional layer: It consists of three parallel dilated convolutional branches with dilation rates of 2, 4 and 6 respectively. Each branch uses 3×3 convolution with a stride of 1 and padding with the corresponding dilation rate to keep the size constant. The number of output channels is 32. Batch normalization and ReLU are then applied to capture edge context information at different scales. Feature concatenation layer: The outputs of the three dilated convolution branches are concatenated along the channel dimension to obtain a 96-channel feature map; Dimensionality reduction convolutional layer: 1×1 convolution with a stride of 1, the number of output channels is set to be the same as the number of output channels of the last Swin module of the backbone network (such as 256 or 384), used to align the feature dimensions, followed by batch normalization. Additive fusion layer: The dimension-reduced edge enhancement features are added element-wise to the output of the last Swin module of the backbone network to output the final feature with optimized boundaries.

[0027] Step 1.2: Scale and Orientation Adaptive Adjustment. During feature extraction, a Scale Self-Attention (VSA) module and a Rotational Attention (RVSA) mechanism are introduced. The VSA module enhances adaptability to multi-scale targets by dynamically adjusting the size and position of the attention window; the RVSA module further predicts the rotation angle of the window, enabling the attention mechanism to accurately focus on tilted or rotated ground features, thus solving the problem of incomplete feature extraction caused by traditional horizontally aligned windows.

[0028] Step 2: Reliable screening based on spatial priors in binary classification This step addresses the false alarms (such as those caused by shadows or spectral confusion) in the automatic interpretation results by constructing a binary classification model that integrates spatial prior information to distinguish between true and false positives for candidate patches.

[0029] Step 2.1: Four-channel input construction. First, morphological closing operations and connected component analysis are performed on the binary mask output from Step 1 to extract the minimum bounding rectangle of each candidate region. Then, a spatial indicator mask with the same size as the original image (1 inside, 0 outside) is generated based on this rectangle. Finally, the channels of this spatial mask are concatenated with the three channels (RGB) of the original remote sensing image in the channel dimension to construct a four-channel input tensor of "RGB+Mask".

[0030] Step 2.2: Model Inference and Filtering. The four-channel tensor is input into the improved EfficientNet binary classification model. Based on the ImageNet pre-trained weights, this model adapts the first-layer convolutional kernel (expanding it from 3 channels to 4 channels) and fine-tunes it through transfer learning to determine whether the input image patches are real targets. Candidate regions identified as real target areas are sent to the next stage, effectively filtering background interference. Simultaneously, the minimum bounding rectangle corresponding to each candidate region identified as a real target area is output as a high-confidence bounding box for use in Step 3.

[0031] Step 3: Refined edge reconstruction guided by the detection box This step addresses the issues of jagged edges, breaks, and inaccurate positioning of patches after initial screening by introducing the FastSAM model for sub-pixel-level repair.

[0032] Step 3.1: Spatial Prior Guidance Mechanism. Abandoning FastSAM's original full-image blind search mode, the high-confidence bounding boxes generated in Step 2 are used as spatial prior prompts and input into FastSAM's prompt encoder.

[0033] Step 3.2: Local Instance Segmentation. The FastSAM model activates feature responses within the local region defined by the detection box, leveraging its powerful zero-shot transfer capability to generate a continuous, closed binary instance mask that fits the contours of ground features. This process compresses the computational domain to 5%-15% of the original image area, improving inference speed by tens of times compared to full-image SAM, achieving high-efficiency and high-precision boundary reconstruction.

[0034] The effect of fine-grained edge reconstruction guided by the detection box is shown in the figure below. Figure 3 As shown.

[0035] Step 4: Morphological purification of geometrically constrained patches This step uses the binary instance mask output from step 3 as input. The connected regions with a pixel value of 1 in the binary instance mask represent the pixel-level representation of the candidate patches, with each connected region corresponding to a patch to be optimized. This step aims to eliminate geometric noise from the mask, improving the topological rationality and business compliance of the patches corresponding to the mask. After subsequent vectorization processing, the purified mask can generate a final vector patch with smooth boundaries and regular shape.

[0036] Step 4.1: Connectivity Analysis and Fragment Removal. Perform neighborhood connectivity analysis on the reconstructed mask to remove isolated fragments with an area smaller than a preset threshold (e.g., 50m²) and eliminate the influence of salt-and-pepper noise.

[0037] Step 4.2: Shape Index Rule Filtering. Mathematical morphology indicators are introduced for logical filtering. The shape compactness index and circumscribed rectangle fill rate of the patches are calculated. Threshold ranges are set (e.g., compactness C > 0.55, fill rate F > 0.5) to automatically exclude non-target misjudgments caused by the curvature of natural water bodies' shorelines, retaining regular patches that conform to the characteristics of artificial features.

[0038] Step 5: Result Fusion and Vectorization Output The image patches, after undergoing the aforementioned multi-stage processing, are vectorized to generate the final interpretation result, achieving an automated closed loop from "pixel" to "decision." The effect diagram of the building experimental interpretation optimization method is shown below. Figure 4 As shown.

[0039] Example 2 This embodiment further defines Embodiment 1. The present invention will be described in detail with reference to the accompanying drawings and specific experimental data.

[0040] This embodiment selects a 0.5-meter resolution orthophoto of a certain area as the raw data. The data coverage includes various land types such as construction land, forest land, water bodies and unused land. The focus is on interpreting, optimizing and verifying the change patches of construction land.

[0041] In this embodiment, the original image is first preprocessed, including radiometric and orthorectification, and then divided into 512×512 pixel tiles for subsequent parallel processing. To verify the effectiveness of the invention, the interpretation results of this embodiment are compared with the results of manual on-site verification as the ground truth for accuracy evaluation.

[0042] In the feature extraction stage, the multi-branch network configuration used in this embodiment is as follows: the backbone network uses Swin-Transformer, auxiliary branch 1 uses a high-resolution convolutional path with 48 channels, and auxiliary branch 2 uses an edge enhancement module with dilated convolution. During training, a VSA module is introduced to dynamically adjust the receptive field of the feature map, and the rotation angle prediction range of the RVSA module is set to [-45°, 45°]. After 50 rounds of iterative training, the network model converges, and the generated initial segmentation results can visually cover the target area well, but there are still some jagged effects and small holes at the boundaries.

[0043] Subsequently, the confidence screening stage begins. In this embodiment, the initial segmentation mask is stitched together with the original image channels to construct a four-channel input. A pre-trained EfficientNet-B5 model is fine-tuned, and the classification head is replaced with a Sigmoid activation function to adapt to the binary classification task. After screening, approximately 25% of low-confidence candidate regions are identified as false positives and removed. These removed regions are mainly shadow areas, bare soil areas, and interfering features with spectral characteristics similar to the background, significantly improving the purity of candidate patches.

[0044] For the selected high-confidence patches, this embodiment uses the FastSAM model for edge reconstruction. Unlike traditional full-image segmentation, this embodiment uses the detection boxes output in the previous step as input prompts to limit the computational range of FastSAM. Experimental data shows that this strategy reduces the edge reconstruction time of a single patch from an average of 3.2 seconds for full-image processing to 0.4 seconds, an efficiency improvement of approximately 8 times. The reconstructed patch boundaries are smooth and continuous, closely matching the contours of ground features, and sub-pixel-level details are restored.

[0045] In the post-processing stage, this embodiment sets the area threshold to 64 pixels (approximately 51.2 square meters) to remove residual small fragments. Simultaneously, it sets the shape compactness C > 0.5 and the bounding rectangle fill rate F > 0.45 to filter irregular edges of natural features. After the above purification process, the final generated patch shape is regular and conforms to the geometric characteristics of artificial features.

[0046] To quantitatively evaluate the performance of this embodiment, Intersection over Union (IoU), Completeness, and Correctness were selected as evaluation metrics. Experimental results show that the average IoU of the interpretation results in this embodiment reaches 0.85, the Completeness is 90.3%, and the Correctness is 93.6%. Compared with the traditional single U-Net model (IoU of 0.78), this invention significantly improves the accuracy and boundary quality of patch interpretation while maintaining a high Completeness, effectively solving the problems of edge discontinuity and high false detection rate.

[0047] The accuracy comparison of the final results of the method of this invention and the comparative method is shown in Table 1: Table 1:

Claims

1. A remote sensing patch interpretation optimization method based on multi-branch networks and multi-stage analysis, characterized in that, The method includes the following steps: S1. Multi-branch collaborative feature extraction: Receive the pre-processed high-resolution remote sensing image, input the image into the multi-branch network for interpretation, and generate a binary initial segmentation mask containing candidate patch regions, which is then input into step S2 for processing. S2. Reliable binary classification screening based on spatial prior: For false positives in the automatic interpretation results, a binary classification model that integrates spatial prior information is constructed to distinguish between true and false positives of candidate patches. S3. Edge Refinement Guided by Detection Box: To address the issues of jagged edges, breaks, and inaccurate positioning of patches after initial screening, the FastSAM model is introduced for sub-pixel-level repair. S4. Geometry-constrained morphological cleanup of patches: Using the binary instance mask output from step S3 as input, geometric noise is eliminated from the mask to improve the topological rationality and business compliance of the patches corresponding to the mask. S5. Result Fusion and Vectorization Output: The multi-stage processed map features are vectorized to generate the final interpretation result.

2. The remote sensing patch interpretation optimization method based on multi-branch network and multi-stage analysis according to claim 1, characterized in that, The multi-branch network uses Transformer and its variants as the backbone network, combined with two specially designed auxiliary branches in parallel. The backbone network is responsible for capturing long-distance dependencies and global feature associations. The first auxiliary branch enhances detail resolution through high-resolution paths and local convolutional filters to ensure that small features are not lost. The second auxiliary branch optimizes boundary extraction through specific magnification layers and edge-sensitive mechanisms. The features of the three branches are deeply fused through a weighted fusion strategy to generate an initial segmentation result that combines global consistency with local details.

3. The remote sensing patch interpretation optimization method based on multi-branch network and multi-stage analysis according to claim 2, characterized in that, The backbone network passes through the following units from input to output: image segmentation unit, initial Swin stage (this unit consists of a Linear Embedding layer and two Swin Transformer Blocks), three downsampling Swin modules with the same structure (each unit consists of a Patch Merging layer and N Swin Transformer Blocks, where the number of layers N is 2, 6 and 2 respectively), pyramid pooling module, feature aggregation module and global-local module; The first auxiliary branch passes through the following layers from input to output: downsampling layer; first convolutional layer; high-resolution residual block; second convolutional layer; and upsampling and fusion layer. The second auxiliary branch passes through the following layers from input to output: feature input layer, edge response extraction layer, multi-scale dilated convolutional layer, feature splicing layer, dimensionality reduction convolutional layer, and addition fusion layer.

4. The remote sensing patch interpretation optimization method based on multi-branch network and multi-stage analysis according to claim 3, characterized in that, Step S2 is as follows: Step S2.1: Four-channel input construction; First, perform morphological closing operation and connected component analysis on the binary mask output from step S1 to extract the minimum bounding rectangle of each candidate region; then, generate a spatial indicator mask with the same size as the original image based on the rectangle; finally, stitch the spatial mask channel with the three channels of the original remote sensing image in the channel dimension to construct a four-channel input tensor. Step S2.2: Model inference and filtering; Input the four-channel tensor into the improved EfficientNet binary classification model. Based on the ImageNet pre-trained weights, this model expands the first layer convolutional kernel from 3 channels to 4 channels and fine-tunes it through transfer learning to determine whether the input image patch is a real target. The improved EfficientNet binary classification model outputs candidate regions that are determined to be true target patches and moves them to the next stage. At the same time, the minimum bounding rectangle corresponding to each candidate region determined to be a true target patch is output as a high-confidence detection box for use in step S3.

5. The remote sensing patch interpretation optimization method based on multi-branch network and multi-stage analysis according to claim 4, characterized in that, Step S3 is as follows: Step S3.1: Spatial prior guidance mechanism; The high-confidence detection box generated in step S2 is used as a spatial prior cue input to the cue encoder of the FastSAM model; Step S3.2: Local instance segmentation; The FastSAM model activates feature responses within the local region defined by the detection box to generate a binary instance mask.

6. The remote sensing patch interpretation optimization method based on multi-branch network and multi-stage analysis according to claim 5, characterized in that, Step S4 is as follows: Step S4.1: Connectivity analysis and fragment removal; perform neighborhood connectivity analysis on the reconstructed mask, remove isolated fragments with an area smaller than a preset threshold, and eliminate the influence of salt-and-pepper noise; Step S4.2: Shape index rule filtering; introduce mathematical morphology indicators for logical filtering, calculate the shape compactness index and circumscribed rectangle fill rate of the patch, set a threshold range to automatically exclude non-target misjudgments caused by the curvature of natural water body shorelines, and retain regular patches that conform to the characteristics of artificial land features.

7. An electronic device comprising a memory and a processor, wherein the memory stores a computer program, characterized in that, When the processor executes the computer program, it implements the steps of the method according to any one of claims 1-6.

8. A computer-readable storage medium for storing computer instructions, characterized in that, When the computer instructions are executed by the processor, they implement the steps of the method according to any one of claims 1-6.

Images