A method for fine field extraction based on multidimensional feature fusion and semantically guided watershed.

Abstract

This invention belongs to the field of agricultural remote sensing technology and relates to a method for fine-grained field extraction based on multi-dimensional feature fusion and semantically guided watershed. The method includes: multi-dimensional feature extraction, constructing feature subsets corresponding to each dimension; concatenating all feature subsets along the channel dimension; constructing a bi-branch multi-scale convolutional module to perform semantic segmentation on the multi-channel feature tensor; using the field boundary probability map as an energy topographic map to generate a constraint mask; extracting instance seed points; generating a marker map, combining the energy topographic map and the constraint mask to perform watershed transformation; and overlaying to obtain the farmland field segmentation result. This invention constructs a multi-channel input, improving the model's perception and analysis capabilities; significantly improves the instance segmentation accuracy and robustness of farmland fields through the bi-branch multi-scale convolutional module; uses the field boundary probability map as the energy topographic map of the watershed algorithm to find a unique instance seed for each independent field; and effectively avoids leakage of segmented regions into the background.

Description

Technical Field

[0001] This invention belongs to the field of agricultural remote sensing technology, specifically, it relates to a method for fine extraction of field plots based on multidimensional feature fusion and semantically guided watersheds. Background Technology

[0002] As the basic unit of agricultural production, accurate spatial boundary information of farmland plots is the core data foundation for achieving precision agricultural management and decision-making. High-resolution remote sensing imagery has made large-scale and automated extraction of farmland plots possible, but existing technologies still face significant challenges when dealing with specific complex scenarios.

[0003] In many regions, due to historical, geographical, or farming practices, farmland plots often exhibit a unique morphology of narrow strips and highly dense arrangement. This morphology results in highly similar spectral and textural features between plots, with weak and blurred plot boundaries, posing challenges to automated plot extraction due to instance adhesion and inaccurate boundary positioning.

[0004] Traditional edge- or region-based image processing methods are susceptible to noise interference, often resulting in severe oversegmentation or undersegmentation, and have poor robustness. While semantic segmentation methods based on deep learning, such as the UNet network, have emerged in recent years and can effectively identify farmland areas, they are essentially pixel-level classifications and cannot distinguish closely adjacent instances of the same type of land at the semantic level, causing all land parcels to appear as clustered together in the prediction results.

[0005] Furthermore, most existing studies only utilize the RGB three-channel information of images, failing to fully explore the diverse information contained in remote sensing data, such as near-infrared, color space, and gradient texture, thus limiting the performance of models in complex land cover discrimination. Therefore, there is an urgent need for a new method that can effectively integrate diverse features and solve the problem of dense target instance adhesion, in order to achieve accurate and automated extraction of narrow strips of farmland. Summary of the Invention

[0006] To address the aforementioned technical problems, this invention provides a method for fine field extraction based on multidimensional feature fusion and semantically guided watersheds, comprising:

[0007] Multi-dimensional feature extraction is performed on remote sensing images to construct feature subsets corresponding to each dimension; the dimensions include several original spectral bands, HSV color space and gradient;

[0008] All feature subsets are concatenated along the channel dimension to form a multi-channel feature tensor;

[0009] In the encoder of the UNet network, a dual-branch multi-scale convolutional module is constructed to perform semantic segmentation on the multi-channel feature tensor, resulting in a background probability map, a plot interior probability map, and a plot boundary probability map. The background probability map is used to represent the probability that each pixel belongs to the background; the plot interior probability map is used to represent the probability that each pixel belongs to the plot interior; and the plot boundary probability map is used to represent the probability that each pixel belongs to the plot boundary.

[0010] Using the probability map of the land parcel boundary as an energy topography map, the probability within the land parcel is analyzed. Figure 2 Value-based generation of constraint masks;

[0011] On the probability map inside the plot, the local peak maximization algorithm is used to extract instance seed points that satisfy the threshold and local optimum conditions;

[0012] Each seed point is assigned a unique ID to generate a label map. The energy topography map and the constraint mask are combined to perform watershed transformation, and the instance label map is output to obtain all independent farmland instances as the farmland segmentation result.

[0013] After binarizing the background probability map, it is superimposed with the farmland plot segmentation results obtained by watershed transformation to obtain the final farmland plot extraction results.

[0014] Based on the above technical solution, the present invention can be further improved as follows.

[0015] Furthermore, the original spectral bands are extracted, including the bands corresponding to red, green, blue and near-infrared, and feature subsets corresponding to the original spectral bands are constructed.

[0016] The RGB portion of high-resolution remote sensing images is converted to the HSV color space, and the features corresponding to hue, saturation, and brightness are extracted to construct a feature subset corresponding to the HSV color space.

[0017] The high-resolution remote sensing image is converted to grayscale and the gradient magnitude map is calculated by applying an edge extraction operator.

[0018] Furthermore, all feature subsets are concatenated along the channel dimension to form a multi-channel feature tensor, including:

[0019] Let the feature subset corresponding to the original spectral band be . The feature subset corresponding to the HSV color space is The gradient magnitude diagram is as follows The multi-channel feature tensor is , If we represent the concatenation operation, then the multichannel feature tensor is represented as: .

[0020] Furthermore, a dual-branch multi-scale convolutional module is constructed in the encoder of the UNet network to perform semantic segmentation on the multi-channel feature tensor, obtaining a background probability map, a plot interior probability map, and a plot boundary probability map, including:

[0021] The same feature map is input into two independent convolutional branches with different receptive fields for parallel feature extraction. The feature maps output by the two convolutional branches are concatenated along the channel dimension. The concatenated features are integrated and dimensionality reduced through convolutional layers to obtain an output feature map containing multi-scale information.

[0022] The multi-channel feature tensor is input into the UNet network for forward propagation, and the output layer of the UNet network generates a probability map containing three channels through an activation function.

[0023] Furthermore, the input feature map is , for Convolution kernel, for Convolution kernel, for Convolution kernel, For the concatenation operation, the output feature map is: The output feature map is then represented as:

[0024] .

[0025] Furthermore, set Indicates vertical resolution. Indicates horizontal resolution. This indicates the category each pixel belongs to: background, plot interior, and plot boundary. This represents the pixel position; after normalization, the pixel is at position... Category The probability is expressed as , This is the original probability map output by the model. For summation, Indicates the position of the pixel The sum of the probabilities of a pixel belonging to the background, the interior of a plot, and the boundary of a plot is the normalized pixel's position. Category probability Represented as:

[0026] .

[0027] Furthermore, the parameters of the UNet network are optimized using a weighted cross-entropy loss function.

[0028] Furthermore, the probability within the plot Figure 2Value-based generation of constraint masks includes:

[0029] The probability map within each plot is binarized by setting a confidence threshold, generating a global constraint mask. Indicates vertical resolution. Indicates horizontal resolution. Indicates pixel location, with a confidence threshold of . The probability diagram inside the plot is as follows The global constraint mask is Then the global constraint mask is represented as:

[0030] .

[0031] Furthermore, on the probability map inside the plot, the local peak maximization algorithm is used to extract instance seed points that meet the threshold and local optimum conditions. This includes: extracting pixels on the probability map inside the plot whose probability value is greater than or equal to a preset absolute confidence threshold and whose probability value is greater than or equal to the probability value of all pixels in the set neighborhood around the pixel, thus obtaining instance seed points that meet the threshold and local optimum conditions.

[0032] Furthermore, a unique ID is assigned to each seed point to generate a labeled map. This map is then combined with the energy topography map and constraint mask to perform a watershed transformation, outputting an instance label map. This yields all independent farmland parcel instances as the farmland parcel segmentation result, including:

[0033] For all instance seed points, a unique integer ID is assigned to each independent instance seed point or a cluster of instance seed points consisting of several adjacent instance seed points using a connected component analysis algorithm, thereby generating a labeled graph.

[0034] Invert the energy topography map input;

[0035] Starting from each labeled instance seed point in the label map, within the range of the constraint mask, a simulated mask is performed, and the instance label map is output. All sets of pixels with the same non-zero integer ID constitute an independent farmland plot instance, and all independent farmland plot instances are obtained as the farmland plot segmentation result.

[0036] The beneficial effects of this invention are:

[0037] (1) Compared with the traditional method that only uses RGB three channels, the present invention innovatively constructs a multi-channel input, which can integrate different types of information. Through deep feature engineering, the standard remote sensing image is transformed into a multi-dimensional feature tensor with richer information dimensions and stronger discrimination ability, thereby improving the model's perception and analysis capabilities.

[0038] (2) The present invention adopts an innovatively structured and optimized UNet network. In the encoder path of the UNet network, a dual-branch multi-scale convolutional module is designed to replace the traditional single-path convolutional block, which enhances the network’s comprehensive perception of land features at different scales, performs end-to-end semantic segmentation on the constructed multi-dimensional feature tensor, and generates a probability graph rich in topological information that can directly guide subsequent instance separation, which significantly improves the instance segmentation accuracy and robustness of narrow strips and densely arranged farmland plots.

[0039] (3) This invention uses the land parcel boundary probability map directly as the energy terrain map of the watershed algorithm, which can prevent different land parcel instances from merging during the segmentation. The local peak maximization algorithm is applied on the probability map inside the land parcel to identify the core of the instance, find a unique instance seed for each independent land parcel, and achieve automation and high robustness.

[0040] (4) The present invention relates to the probability within the land parcel. Figure 2 Value-based generation of constraint masks limits the watershed transformation to areas where the network is highly certain to be farmland, effectively preventing leakage of segmented regions into the background. Attached Figure Description

[0041] Figure 1 A schematic diagram illustrating the principle of the field refinement method based on multidimensional feature fusion and semantically guided watershed provided by this invention;

[0042] Figure 2 A flowchart illustrating the field refinement method based on multidimensional feature fusion and semantically guided watershed provided by this invention;

[0043] Figure 3 This is a schematic diagram of the structure of a dual-branch multi-scale convolutional module.

[0044] Figure 4 This is a simulation diagram comparing the performance of the method used in this invention with that of a method using only the traditional UNet network for segmentation on two typical samples. Detailed Implementation

[0045] To make the objectives, technical solutions, and advantages of the embodiments of the present invention clearer, the technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the accompanying drawings. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. The components of the embodiments of the present invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations.

[0046] As an example, see the attached document. Figure 1 As shown, to solve the above-mentioned technical problems, this embodiment provides a method for fine field extraction based on multi-dimensional feature fusion and semantically guided watersheds, including:

[0047] Multi-dimensional feature extraction is performed on remote sensing images to construct feature subsets corresponding to each dimension; the dimensions include several original spectral bands, HSV color space and gradient;

[0048] All feature subsets are concatenated along the channel dimension to form a multi-channel feature tensor;

[0049] In the encoder of the UNet network, a dual-branch multi-scale convolutional module is constructed to perform semantic segmentation on the multi-channel feature tensor, resulting in a background probability map, a plot interior probability map, and a plot boundary probability map. The background probability map is used to represent the probability that each pixel belongs to the background; the plot interior probability map is used to represent the probability that each pixel belongs to the plot interior; and the plot boundary probability map is used to represent the probability that each pixel belongs to the plot boundary.

[0050] Using the probability map of the land parcel boundary as an energy topography map, the probability within the land parcel is analyzed. Figure 2 Value-based generation of constraint masks;

[0051] On the probability map inside the plot, the local peak maximization algorithm is used to extract instance seed points that satisfy the threshold and local optimum conditions;

[0052] Each seed point is assigned a unique ID to generate a label map. The energy topography map and the constraint mask are combined to perform watershed transformation, and the instance label map is output to obtain all independent farmland instances as the farmland segmentation result.

[0053] After binarizing the background probability map, it is superimposed with the farmland plot segmentation results obtained by watershed transformation to obtain the final farmland plot extraction results.

[0054] As attached Figure 2 The diagram shows the flowchart of the method. This invention aims to significantly improve the accuracy and robustness of instance segmentation for narrow strips and densely arranged farmland plots by deeply fusing multi-dimensional feature information from remote sensing images and organically combining the semantic perception capabilities of deep learning with the advantages of topological segmentation in traditional image processing.

[0055] This invention utilizes deep feature engineering to transform standard remote sensing imagery into a multidimensional feature tensor with richer information dimensions and stronger discriminative capabilities, thereby enhancing the perception and analysis capabilities of subsequent deep learning models. Compared to traditional methods that only use three channels of RGB (Red, Green, Blue, RGB color mode), this invention innovatively constructs an 8-channel input, integrating complementary information of different types to address the challenges of complex scenarios.

[0056] Optionally, the original spectral bands are extracted, including the bands corresponding to red, green, blue and near-infrared, and a feature subset corresponding to the original spectral bands is constructed.

[0057] The RGB portion of high-resolution remote sensing images is converted to the HSV color space, and the features corresponding to hue, saturation, and brightness are extracted to construct a feature subset corresponding to the HSV (Hue; Saturation; Value) color space.

[0058] The high-resolution remote sensing image is converted to grayscale and the gradient magnitude map is calculated by applying an edge extraction operator.

[0059] The feature subset corresponding to the original spectral bands retains the most basic and direct physical reflectance information of ground features. Unlike the RGB space, where color and brightness are highly coupled, the HSV color space decouples hue, saturation, and brightness attributes. In other words, it splits the attributes of the color space into three completely independent channels that do not interfere with each other. Adjusting one channel only changes the corresponding attribute and does not affect the other two. This invention introduces image gradients as an important geometric feature. By calculating gradient magnitude maps, the network can explicitly and actively focus on the edge structure of ground features.

[0060] Optionally, all feature subsets are concatenated along the channel dimension to form a multi-channel feature tensor, including:

[0061] Let the feature subset corresponding to the original spectral band be . The feature subset corresponding to the HSV color space is The gradient magnitude diagram is as follows The multi-channel feature tensor is , If we represent the concatenation operation, then the multichannel feature tensor is represented as: .

[0062] set up This represents the rate of change of brightness of a pixel in the horizontal direction. This represents the rate of change of pixel brightness in the vertical direction. The quantization value representing the edge intensity at a pixel location is:

[0063] .

[0064] Gradient magnitude maps can highlight areas where pixel intensity changes drastically, i.e., the outlines and boundaries of ground features.

[0065] The extracted feature subsets are stacked along the channel dimension to form the final 8-channel input feature tensor. This multi-dimensional feature construction strategy enables the input data to contain multiple information such as spectrum, color and geometry, providing richer information input for subsequent network segmentation tasks that are refined and highly robust.

[0066] This invention employs an innovatively structured and optimized UNet network to perform end-to-end semantic segmentation on the constructed multidimensional feature tensor, with the goal of generating a probabilistic graph rich in topological information that can directly guide subsequent instance separation.

[0067] Optionally, a dual-branch multi-scale convolutional module is constructed in the encoder of the UNet network to perform semantic segmentation on the multi-channel feature tensor, obtaining a background probability map, a plot interior probability map, and a plot boundary probability map, including:

[0068] The same feature map is input into two independent convolutional branches with different receptive fields for parallel feature extraction. The feature maps output by the two convolutional branches are concatenated along the channel dimension. The concatenated features are integrated and dimensionality reduced through convolutional layers to obtain an output feature map containing multi-scale information.

[0069] The multi-channel feature tensor is input into the UNet network for forward propagation, and the output layer of the UNet network generates a probability map containing three channels through an activation function.

[0070] To enhance the network's comprehensive perception of features at different scales and address the challenges of large scale variations in narrow strip land parcels and the need to consider both macroscopic shapes and locate microscopic boundaries, this invention designs a dual-branch multi-scale convolutional module in the encoder path of the UNet network to replace the traditional single-path convolutional block, as shown in the attached figure. Figure 3 As shown, the core idea of ​​this dual-branch multi-scale convolutional module is to enable the network to examine and process features in parallel at each layer with different receptive fields. Specifically, for any feature map input to this dual-branch multi-scale convolutional module, it is simultaneously fed into two independent convolutional branches with different receptive fields for parallel feature extraction: one branch uses a small 3×3 convolutional kernel, focusing on extracting high-frequency detail information, which is crucial for accurately capturing and locating plot boundaries that are only one or two pixels wide; the other branch uses a large 7×7 convolutional kernel, which can capture spatial context information over a wider range, thereby better understanding the overall shape, orientation, and topological relationship of the plot with surrounding features, helping to avoid misclassifying noise with similar local textures as plots. After the two branches complete feature extraction, the output feature maps are concatenated along the channel dimension. This operation allows the fused feature map to simultaneously possess local detail information and global contextual information. Finally, a 1×1 convolutional layer efficiently integrates and reduces the dimensionality of the concatenated features, resulting in the module's final output feature map, which is rich in multi-scale information.

[0071] Optionally, the input feature map is , for Convolution kernel, for Convolution kernel, for Convolution kernel, For the concatenation operation, the output feature map is: The output feature map is then represented as:

[0072] .

[0073] The constructed 8-channel feature tensor is input into the dual-branch multi-scale UNet network for forward propagation. This invention innovatively defines the segmentation task as a three-class classification problem instead of a two-class classification problem, fundamentally solving the instance sticking problem. Let... For the height of remote sensing images, To determine the width of the remote sensing image, the network's output layer uses the Softmax activation function to ultimately generate a... The probability diagram of the three channels.

[0074] Optional, set Indicates vertical resolution. Indicates horizontal resolution. This indicates the category each pixel belongs to: background, plot interior, and plot boundary. This represents the pixel position; after normalization, the pixel is at position... Category The probability is expressed as , This is the original probability map output by the model. For summation, Indicates the position of the pixel The sum of the probabilities of a pixel belonging to the background, the interior of a plot, and the boundary of a plot is the normalized pixel's position. Category probability Represented as:

[0075] .

[0076] This indicates that the pixel belongs to the background. This indicates that the pixel belongs to the area inside the plot. This indicates that a pixel belongs to a plot boundary. The three channels of this probability map represent the probability that each pixel belongs to the background, the plot interior, and the plot boundary, respectively, providing a perfect topological prior for the subsequent watershed algorithm.

[0077] Optionally, the parameters of the UNet network can be optimized using a weighted cross-entropy loss function.

[0078] As linear features, plot boundaries have far fewer pixels than the background and the interior of the plots. Therefore, this invention employs a weighted cross-entropy loss function during model training to address the severe class imbalance problem. By assigning a significantly higher weight to the plot boundary class and lower weights to other classes, this loss function effectively amplifies the penalty for misclassification of boundary pixels, thereby driving the UNet network to invest more resources in learning and accurately depicting the subtle features crucial for instance separation.

[0079] Specifically, let's set The weighted cross-entropy loss function is... Represents the learnable parameters of the model. As a category, This serves as the row index when traversing remotely sensed images. For column indexing when traversing remote sensing images, For the height of remote sensing images, The width of the remote sensing image. This represents the total number of pixels. For pixels In category The real labels on For category If the weights are equal, then:

[0080] .

[0081] To accurately transform the output semantically rich probabilistic map into discrete farmland instance label maps with unique identifiers, semantic terrain segmentation, constraint region construction, automated instance seed extraction, label assignment, and watershed transformation under constraints are performed.

[0082] This invention interfaces the output of deep learning with the input of traditional algorithms. Specifically, it directly uses the land parcel boundary probability map as the energy terrain map for the watershed algorithm. In the energy terrain map, regions predicted by the UNet network to have higher boundary probabilities have higher energy values, forming more difficult-to-cross energy ridges. This provides a solid physical basis for preventing the merging of different land parcel instances during segmentation. At the same time, to ensure the accuracy of the segmentation process, the probability map inside the land parcel is binarized by setting a confidence threshold to generate a global constraint mask.

[0083] Optionally, the probability within the plot can be... Figure 2 Value-based generation of constraint masks includes:

[0084] The probability map within each plot is binarized by setting a confidence threshold, generating a global constraint mask. Indicates vertical resolution. Indicates horizontal resolution. Indicates pixel location, with a confidence threshold of . The probability diagram inside the plot is as follows The global constraint mask is Then the global constraint mask is represented as:

[0085] .

[0086] This mask limits the watershed transformation to areas where the network is highly certain to be farmland, effectively preventing leakage of segmented regions into the background.

[0087] Finding a unique instance seed for each individual plot is crucial for the successful marking of the control watershed. To automate and ensure the robustness of this process, this invention applies a local peak maximization algorithm to the probability map within each plot to identify the core of the instance. This algorithm searches for all pixels that meet the high confidence requirement and have the highest confidence within their local neighborhood, thereby stably locating the central region of each plot.

[0088] Optionally, on the probability map inside the plot, instance seed points that satisfy the threshold and local optimum conditions are extracted by the local peak maximization algorithm, including: extracting pixels on the probability map inside the plot whose probability value is greater than or equal to a preset absolute confidence threshold and whose probability value is greater than or equal to the probability value of all pixels in the set neighborhood around the pixel, to obtain instance seed points that satisfy the threshold and local optimum conditions.

[0089] The condition for a pixel to be identified as a local peak can be formally defined as follows: the pixel must simultaneously satisfy two conditions: an absolute threshold and a local optimum. First, satisfying the absolute threshold means that the probability value of the pixel must be higher than a preset absolute confidence threshold to ensure that seed points are only generated in areas where the network is highly confident they are within a plot of land. Let the preset absolute confidence threshold be... The probability value of a pixel is ,but: Secondly, the probability value of this pixel must be greater than or equal to the probability values ​​of all points within a specific neighborhood around it. Let's assume... The neighborhood centered on is represented as , express Remove from the neighborhood Points other than those mentioned above. Optionally, a point can be used... Centered on the minimum spacing parameter Define a circular or square neighborhood. This condition ensures the minimum distance between seed points, effectively suppressing dense pseudo-peaks caused by small fluctuations within the same plot, thus avoiding over-segmentation. This condition is expressed as: , .

[0090] Optionally, a unique ID is assigned to each seed point to generate a labeled map. This map is then combined with the energy topography map and constraint mask to perform a watershed transformation, outputting an instance label map. This yields all independent farmland parcel instances as the farmland parcel segmentation result, including:

[0091] For all instance seed points, a unique integer ID is assigned to each independent instance seed point or a cluster of instance seed points consisting of several adjacent instance seed points using a connected component analysis algorithm, thereby generating a labeled graph.

[0092] Invert the energy topography map input;

[0093] Starting from each labeled instance seed point in the label map, within the range of the constraint mask, a simulated mask is performed, and the instance label map is output. All sets of pixels with the same non-zero integer ID constitute an independent farmland plot instance, and all independent farmland plot instances are obtained as the farmland plot segmentation result.

[0094] Specifically, for all extracted instance seed points, a connected component analysis algorithm is applied to assign a unique integer ID to each independent seed point or cluster of seed points, thereby generating a labeled graph. Finally, the watershed transformation function is called, and all semantic guidance information is integrated. Let the labeled graph be... The constraint mask is The output instance label image is The watershed transform function is The notation for the watershed transform function originates from... Constraint mask The region-restricted mask for the watershed transform function is expressed as follows: The internal probability map of a land parcel is an energy topography map based on the watershed transform function, represented as follows: ,but:

[0095] .

[0096] Invert the energy topography map input so that the high-probability boundary corresponds to the high ridge of the watershed, from the labeled map Starting with each labeled seed point, a simulated mask is created on the energy topography map. The entire process is strictly constrained by the constraint mask. The range, and the final output instance label graph. In this context, the set of all pixels with the same non-zero integer ID constitutes a single, well-defined, and topologically correct individual farmland plot instance that has been successfully and precisely segmented.

[0097] In practical applications, if only the probability maps inside the fields and the probability maps at the field boundaries are used for watershed transformation, the result of the watershed transformation may spread to the regions that the UNet network judges as background. However, by superimposing the background probability map with the result of the watershed algorithm, it can be ensured that the regions that the UNet network judges as background by confidence score still belong to the background region in the final result, thereby ensuring the consistency between the output of the UNet network and the output of the watershed algorithm.

[0098] After the background probability map is binarized, it is superimposed with the farmland plot segmentation result obtained by the watershed transformation to obtain the final farmland plot extraction result, thereby ensuring the consistency between the UNet network and the watershed algorithm.

[0099] As attached Figure 4 As shown, the performance of the method used in this invention compared with that of a method using only the traditional UNet network for segmentation on two typical samples. The points in the probability graph represent continuous values ​​between 0 and 1. (Appendix) Figure 4 The left column shows two different input samples, both containing typical narrow strips of densely packed farmland; (Attached) Figure 4 The middle column shows the results of binary (farmland and background) semantic segmentation using only the traditional UNet network. It is evident that, due to the high similarity in spectral and texture features between plots, while the traditional UNet network can identify the approximate area of ​​farmland, the output is one or more large, connected white masks. Fine-grained boundary information between plots is almost completely lost. This result cannot be used for subsequent individual plot analysis, demonstrating the inherent limitations of single semantic segmentation methods when handling dense target instance segmentation tasks. (Appendix) Figure 4 The right column shows the final fine segmentation results obtained by using the complete technical solution of this invention, multi-dimensional feature input, a dual-branch multi-scale UNet network, and semantically guided watershed. In stark contrast to the middle column, the method used in this invention successfully separates contiguous regions into clearly defined and independent narrow strip plots. Each slender plot is precisely segmented, powerfully demonstrating that this invention, through its innovative three-category semantic learning strategy and semantically guided watershed post-processing, successfully overcomes the challenge of instance contiguousness, achieving a leap from coarse region-level segmentation to precise instance-level fine segmentation.

[0100] The above are merely preferred embodiments of the present invention and are not intended to limit the present invention. Various modifications and variations can be made to the present invention by those skilled in the art. Any modifications, equivalent substitutions, improvements, etc., made within the spirit and principles of the present invention should be included within the scope of protection of the present invention.

Claims

1. A method for fine field extraction based on multi-dimensional feature fusion and semantically guided watersheds, characterized in that, include: Multi-dimensional feature extraction is performed on remote sensing images to construct feature subsets corresponding to each dimension; the dimensions include several original spectral bands, HSV color space and gradient; All feature subsets are concatenated along the channel dimension to form a multi-channel feature tensor; A dual-branch multi-scale convolutional module is constructed in the encoder of the UNet network to perform semantic segmentation on the multi-channel feature tensor, obtaining a background probability map, a plot interior probability map, and a plot boundary probability map. This includes: inputting the same feature map into two independent convolutional branches with different receptive fields for parallel feature extraction; concatenating the feature maps output by the two convolutional branches along the channel dimension; integrating and reducing the dimensionality of the concatenated features through convolutional layers to obtain an output feature map containing multi-scale information; inputting the multi-channel feature tensor into the UNet network for forward propagation; and generating the background probability map, plot interior probability map, and plot boundary probability map through activation functions in the output layer of the UNet network. The background probability map represents the probability that each pixel belongs to the background; the plot interior probability map represents the probability that each pixel belongs to the plot interior; and the plot boundary probability map represents the probability that each pixel belongs to the plot boundary. The probability map of the land parcel boundary is used as an energy topographic map, and the probability map inside the land parcel is binarized to generate a constraint mask. On the probability map inside the plot, the local peak maximization algorithm is used to extract instance seed points that satisfy the threshold and local optimum conditions; Each seed point is assigned a unique ID to generate a labeled map. A watershed transformation is performed using the energy topography map and a constraint mask to output an instance label map. All independent farmland parcel instances are obtained as the farmland parcel segmentation result. This process includes: assigning a unique integer ID to each independent instance seed point or a cluster of instance seed points composed of several adjacent instance seed points using a connected component analysis algorithm, generating a labeled map; inverting the energy topography map as input; starting from each labeled instance seed point in the labeled map, performing a simulated mask within the constraint mask's range, outputting an instance label map; and ensuring that all pixel sets with the same non-zero integer ID constitute an independent farmland parcel instance, thus obtaining all independent farmland parcel instances as the farmland parcel segmentation result. After binarizing the background probability map, it is superimposed with the farmland plot segmentation results obtained by watershed transformation to obtain the final farmland plot extraction results.

2. The method for fine field extraction based on multi-dimensional feature fusion and semantically guided watersheds according to claim 1, characterized in that, Extract the original spectral bands, including the bands corresponding to red, green, blue and near-infrared, and construct the feature subsets corresponding to the original spectral bands; The RGB portion of high-resolution remote sensing images is converted to the HSV color space, and the features corresponding to hue, saturation, and brightness are extracted to construct a feature subset corresponding to the HSV color space. The high-resolution remote sensing image is converted to grayscale and the gradient magnitude map is calculated by applying an edge extraction operator.

3. The method for fine field extraction based on multi-dimensional feature fusion and semantically guided watersheds according to claim 2, characterized in that, All feature subsets are concatenated along the channel dimension to form a multi-channel feature tensor, including: Let the feature subset corresponding to the original spectral band be . The feature subset corresponding to the HSV color space is The gradient magnitude diagram is as follows The multi-channel feature tensor is , If we represent the concatenation operation, then the multichannel feature tensor is represented as: .

4. The method for fine field extraction based on multi-dimensional feature fusion and semantically guided watersheds according to claim 1, characterized in that, The input feature map is , for Convolution kernel, for Convolution kernel, for Convolution kernel, For the concatenation operation, the output feature map is: The output feature map is then represented as: 。 5. The method for fine field extraction based on multidimensional feature fusion and semantically guided watersheds according to claim 1, characterized in that, set up Indicates vertical resolution. Indicates horizontal resolution. This indicates the category each pixel belongs to: background, plot interior, and plot boundary. This represents the pixel position; after normalization, the pixel is at position... Category The probability is expressed as , This is the original probability map output by the model. For summation, Indicates the position of the pixel The sum of the probabilities of a pixel belonging to the background, the interior of a plot, and the boundary of a plot is the normalized pixel's position. Category probability Represented as: 。 6. The method for fine field extraction based on multi-dimensional feature fusion and semantically guided watersheds according to claim 1, characterized in that, The parameters of the UNet network are optimized using a weighted cross-entropy loss function.

7. The method for fine field extraction based on multidimensional feature fusion and semantically guided watersheds according to claim 1, characterized in that, Binarizing the probability map within the land parcel to generate a constraint mask includes: The probability map within each plot is binarized by setting a confidence threshold, generating a global constraint mask. Indicates vertical resolution. Indicates horizontal resolution. Indicates pixel location, with a confidence threshold of . The probability diagram inside the plot is as follows The global constraint mask is Then the global constraint mask is represented as: 。 8. The method for fine field extraction based on multidimensional feature fusion and semantically guided watersheds according to claim 1, characterized in that, On the probability map inside the plot, the local peak maximization algorithm is used to extract instance seed points that meet the threshold and local optimum conditions. This includes: extracting pixels on the probability map inside the plot whose probability value is greater than or equal to the preset absolute confidence threshold and whose probability value is greater than or equal to the probability value of all pixels in the set neighborhood around the pixel, thus obtaining instance seed points that meet the threshold and local optimum conditions.

Images