A hierarchical saliency enhancement and residual topology reconstruction collaborative farmland regional plot geometric shape identification method and system

Abstract

The application discloses a layered saliency enhancement and residual topology reconstruction collaborative farmland regional plot geometric morphology identification method and system, and belongs to the technical field of remote sensing image processing and intelligent agricultural informatization. The application is based on a static modal knowledge injection mechanism, uses a preset high-dimensional representation encoder to perform feature stream extraction, sharpens plot boundary details through a spatial topology sensitivity strengthening mechanism, and uses an intra-class structure consistency anchoring mechanism to ensure global attribute collaboration. In the decoding stage, an attention residual skip connection reconstruction mechanism is adopted to suppress non-target noise and realize pixel-level fidelity reconstruction. Finally, a multi-dimensional geometric constraint collaborative supervision framework is adopted to realize the dual consistency of geometric boundaries and internal connectivity. The application solves the problems of plot extraction boundary blur and internal semantic fragmentation in the prior art under complex environment, significantly improves the recognition accuracy and topology integrity of farmland plots while maintaining extremely low training load.

Description

Technical Field

[0001] This invention relates to the fields of intelligent processing of remote sensing images and agricultural information technology, specifically to a scheme for accurate extraction of farmland boundaries based on fine-tuning of deep perception models and synergy of multi-dimensional features. Background Technology

[0002] When using high-resolution remote sensing imagery for agricultural area analysis, extracting the geometric shape of plots is a core step. However, existing visual segmentation methods face a dual challenge when dealing with agricultural plots: on the one hand, although large-scale pre-trained models have generalized perception capabilities, it is difficult to achieve a balance between computational efficiency and task-specific representation when directly transferred to specific agricultural scenarios; on the other hand, agricultural plots often have complex topological structures and significant intra-class heterogeneity. Traditional feature fusion methods are prone to losing subtle boundary signals during downsampling or ignoring semantic consistency within plots during global perception, resulting in extraction results with edge blunting, internal holes, and impaired spatial connectivity. Summary of the Invention

[0003] The purpose of this invention is to provide a method for identifying the geometric shape of farmland plots based on hierarchical saliency enhancement and residual topology reconstruction, so as to solve the technical problems mentioned in the background art.

[0004] To achieve the above objectives, the present invention provides the following technical solution: a method for recognizing the geometric morphology of farmland plots based on hierarchical saliency enhancement and residual topology reconstruction, characterized by the following steps: based on a static modal knowledge injection mechanism, a pre-set high-dimensional representation encoder is used to perform feature streaming extraction on remote sensing images to obtain original feature sequences of different perceptual dimensions; for low-order spatial signals in the original feature sequences, a spatial topology sensitivity enhancement mechanism is adopted to achieve sharpening of plot boundary edge details by capturing the geometric gradient correlation between pixels; for high-order semantic signals in the original feature sequences, an intra-class structural consistency anchoring mechanism is adopted to eliminate homogeneity interference in complex backgrounds by establishing attribute collaboration of global context; for the enhanced heterogeneous representation signals, an attention residual skip-connection reconstruction mechanism is adopted to gradually restore spatial resolution while suppressing non-target noise through adaptive adjustment of saliency weights to complete the faithful reconstruction of the plot mask; for the output plot morphology, a multi-dimensional geometric constraint collaborative supervision framework is used, combined with distance field evolution and contour topology guidance, to ensure the dual consistency of the recognition results in geometric boundaries and internal connectivity.

[0005] Furthermore, the static modal knowledge injection mechanism uses a parameter-frozen generalized visual kernel as the feature extraction basis, and through non-training weight solidification, it seamlessly transfers the generalized perception capability in the large-scale pre-trained model to the farmland extraction task.

[0006] Furthermore, the spatial topology sensitivity enhancement mechanism extracts minute geometric breakpoints hidden in shallow networks by constructing a spatial activation map within the local receptive field, and uses structural operators to perform directional enhancement of the plot edges.

[0007] Furthermore, the intra-class structural consistency anchoring mechanism involves global dependency modeling of high-dimensional feature maps, and by calculating the cosine affinity between feature vectors, it enhances the sense of feature belonging within the same plot and ensures the continuity of attributes within large-scale plots.

[0008] Furthermore, the attention residual jump-connection reconstruction mechanism introduces residual learning logic in the decoding stage, and uses a saliency gating unit to nonlinearly couple deep semantic information with shallow anatomical structure to compensate for high-frequency pixel information lost during downsampling.

[0009] Furthermore, the multidimensional geometric constraint collaborative supervision framework includes a mask branch, an Euclidean distance transformation branch, and an edge morphology analysis branch; the loss functions of the three are calculated in parallel and joint gradient optimization is implemented.

[0010] Compared with the prior art, the beneficial effects of the present invention are as follows: the present invention reduces the computational load by distributing the frozen basic encoder; it resolves the contradiction between "boundary precision" and "semantic connectivity" by using a hierarchical enhancement mechanism; it achieves high-fidelity reconstruction of pixel-level high-frequency information by introducing attention residual logic, thus eliminating semantically ambiguous areas; and it achieves high-fidelity extraction of farmland plots in highly dynamic and complex backgrounds while maintaining extremely low training costs. Attached Figure Description

[0011] Figure 1 This is a schematic diagram of the overall algorithm flow of the method of the present invention.

[0012] Figure 2 This is a detailed algorithm flowchart of the method of the present invention. Detailed Implementation

[0013] The technical solution of the present invention will be described in detail below with reference to the accompanying drawings and embodiments.

[0014] Example 1: As Figure 1 and Figure 2As shown, this invention discloses a method for farmland plot extraction based on a hierarchical attention-enhanced residual network (HARNet). The aim is to fully utilize sub-meter level remote sensing images from Sentinel-2 and Gaofen series satellites to construct a HARNet model with synergistic effects of spatial structure and semantic consistency. An application example of large-scale farmland plot extraction in typical agricultural areas of Sichuan and Shandong provinces is demonstrated by applying feature flow extraction using the FastSAM (Freeze-of-Visual-Based Model) and a multi-dimensional geometric constraint collaborative supervision mechanism. Specifically, the method includes the following steps: Step 1: Feature flow extraction based on static modality knowledge injection; In this embodiment, the four-level feature map sequence is first extracted using a parameter-frozen visual base model encoder. (where k=1,2,3,4), and the corresponding step sizes are respectively Subsequently, a lightweight adapter is used to perform dimensionality compression and resampling on the features at each level, generating an original feature sequence with a uniform feature distribution. ,in For the adapted channel dimensions, and These represent the height and width pixel values ​​of the input remote sensing image in spatial dimensions, respectively.

[0015] Step 2: Implementation of spatial topology sensitivity enhancement; targeting the low-order feature map output from Step 1. (k=1,2) In this embodiment, a dedicated spatial detail excitation characterization unit is constructed. This unit captures the geometric gradient correlation between pixels and amplifies the activation intensity of the plot edges using residual mapping logic. First, channel descriptors are extracted using global average pooling (GAP). Secondly, attention weight vectors are generated through activation functions. ,in For the Sigmoid function, The network is mapped; finally, the sharpened features are output: ,in This represents element-wise multiplication. Through this mechanism, we can effectively extract the subtle boundaries hidden in the original features and prevent the boundary information from being buried by background noise.

[0016] Step 3: Implementation of intra-class structural consistency anchoring; for higher-order semantic features (k=1,2), this embodiment introduces a global attribute attribution awareness module, utilizing features Generate spatial structure guidance diagram and higher-order semantic features Modulation is performed to generate a modulated signal: Then introduce channel gating attention. The enhanced modulation signal is obtained by increasing the global attribute affinity: The final fusion output is: Where ReLU is the nonlinear activation operator and BN is the numerical normalization operator. To integrate the weight matrix, this step ensures the integrity and internal continuity of the extracted results for super-large plots by enhancing the attribute affinity within the same topological unit.

[0017] Step 4: Attention Residual Skip-Connection Reconstruction Mechanism; During the decoder feature upsampling and integration process, bilinear upsampling is performed on the decoded features to obtain... And concatenate it with the enhanced jump connection feature to obtain the composite feature. ;Utilizing adaptive convolution kernels Calculate efficient channel attention weights, where the kernel size k has a logarithmic relationship with the number of channels C; introduce a residual optimization path into the fused output to maintain training stability. ,in This is the mapping function used to learn the predicted residuals.

[0018] Step 5: Construction of the loss function for multi-dimensional geometric constraint collaborative supervision; To ensure that the final output plot shape conforms to the topological logic of geographic entities, this embodiment adopts a cross-task gradient-guided collaborative optimization strategy. In the design of the loss function, a joint loss function is used. ;in, This is a weighted sum of the binary cross-entropy loss and the Dice loss. For edge classification loss, This represents the mean square error loss of the Euclidean distance field. The weighting of the loss for different tasks is determined; through multi-dimensional joint supervision, the model is forced to output a geometric mask with clear boundaries and accurate topological relationships.

[0019] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. For those skilled in the art, the present invention can have various modifications and variations; any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. A method for identifying the geometric morphology of agricultural land parcels through the synergy of hierarchical saliency enhancement and residual topology reconstruction, characterized in that, Includes the following steps: 1) Based on the static modal knowledge injection mechanism, a pre-set high-dimensional representation encoder is used to extract features from remote sensing images in a streaming manner to obtain the original feature sequences of different perception dimensions. 2) For the low-order spatial signals in the original feature sequence, a spatial topology sensitivity enhancement mechanism is adopted to achieve sharpening of the details of the plot boundary edges by capturing the geometric gradient correlation between pixels. 3) For the high-order semantic signals in the original feature sequence, an intra-class structural consistency anchoring mechanism is adopted to eliminate homogeneity interference in complex backgrounds by establishing attribute collaboration of the global context. 4) For the enhanced heterogeneous representation signal, an attention residual skip-connection reconstruction mechanism is adopted to gradually restore the spatial resolution while suppressing non-target noise through adaptive adjustment of saliency weights, thus completing the faithful reconstruction of the plot mask. 5) For the output land parcel morphology, a multi-dimensional geometric constraint collaborative supervision framework is used, combined with distance field evolution and contour topology guidance, to ensure that the recognition results are consistent in both geometric boundaries and internal connectivity.

2. The method for identifying the geometric morphology of farmland parcels through the synergy of hierarchical saliency enhancement and residual topology reconstruction as described in claim 1, characterized in that, The static modal knowledge injection mechanism uses a parameter-frozen generalized visual kernel as the feature extraction basis. Through non-training weight solidification, it seamlessly transfers the generalized perception capability in the large-scale pre-trained model to the farmland extraction task to extract the original feature sequence.

3. The method for identifying the geometric morphology of farmland parcels through the synergy of hierarchical saliency enhancement and residual topology reconstruction as described in claim 2, characterized in that, The spatial topology sensitivity enhancement mechanism targets low-order spatial signals in the original feature sequence. It extracts minute geometric breakpoints hidden in the shallow network by constructing a spatial activation mapping within the local receptive field, and uses structural operators to perform directional enhancement of the plot edges.

4. The method for identifying the geometric morphology of farmland parcels through the synergy of hierarchical saliency enhancement and residual topology reconstruction as described in claim 2, characterized in that, The intra-class structural consistency anchoring mechanism targets the high-order semantic signals in the original feature sequence. It involves global dependency modeling of the high-dimensional feature map, and by calculating the cosine affinity between feature vectors, it enhances the sense of feature belonging within the same plot and ensures the continuity of attributes within large-scale plots.

5. The method for identifying the geometric morphology of farmland parcels through the synergy of hierarchical saliency enhancement and residual topology reconstruction as described in claim 3, characterized in that, The attention residual jump-connection reconstruction mechanism introduces residual learning logic in the decoding stage. It uses a saliency gating unit to nonlinearly couple the low-order spatial signal enhanced by the spatial topology sensitivity with the high-order semantic signal anchored by the intra-class structural consistency, thereby compensating for the high-frequency pixel information lost during downsampling.

6. The method for identifying the geometric morphology of farmland parcels through the synergy of hierarchical saliency enhancement and residual topology reconstruction as described in claim 1, characterized in that, The multidimensional geometric constraint collaborative supervision framework includes a mask branch, an Euclidean distance transformation branch, and an edge morphology analysis branch. By calculating the loss functions of the three branches in parallel and implementing joint gradient optimization, the model can automatically regress the center point orientation and boundary clarity of the plot during the optimization process.

Images