Farmland parcel extraction method and system based on large model fine tuning

Abstract

The invention discloses a farmland plot extraction method and system based on large model fine tuning, and the method comprises the steps: obtaining remote sensing image data of a target region, and carrying out the preprocessing of the remote sensing image data; field sample vectors are sketched and labeled for the remote sensing image data preprocessed in the step 1, and a semantic segmentation data set is made; constructing a farmland plot semantic segmentation model FARMSeg model, and training the FARMSeg by adopting a semantic segmentation data set; in consideration of spatial neighborhood information, performing block enhancement processing on a farmland remote sensing image to be extracted to obtain an image set, respectively inputting the image set after block enhancement into the trained FARMSeg for prediction, and performing fusion splicing on various prediction results of different blocks to obtain a final segmentation result image; and post-processing the final segmentation result map to generate a farmland plot result map meeting the resolution requirement. According to the method, high-precision farmland segmentation can be realized under the condition that only a small number of samples are needed, and the dependence on large-scale labeled data is remarkably reduced.

Description

Technical Field

[0001] This invention belongs to the technical field of agriculture and remote sensing, specifically relating to a method and system for extracting farmland plots based on large model fine-tuning. Background Technology

[0002] Farmland plots, as the basic spatial units of agricultural production and management, require precise extraction and dynamic monitoring to achieve digital, refined, and intelligent management of agricultural production. With the rapid development of high-resolution remote sensing technology and the continuous advancement of agricultural digital transformation, plot-scale precision agriculture management has become an important technological path to improve agricultural production efficiency and ensure food security. However, influenced by factors such as diversified farming methods and complex crop planting structures, farmland plots exhibit significant spatial heterogeneity. This spatial heterogeneity presents a major challenge to conducting high-precision, automated farmland plot extraction over large areas.

[0003] With the rise of deep learning, end-to-end semantic segmentation methods have gradually become an important technical approach for farmland parcel extraction. These methods fully utilize multi-dimensional information such as spectrum, shape, and texture in sample data, performing pixel-level classification through convolutional neural networks or visual Transformer architectures, greatly improving the accuracy and robustness of farmland parcel segmentation. Numerous studies have validated the effectiveness of deep learning segmentation models on publicly available remote sensing image datasets. However, these methods rely on a large number of high-quality, finely labeled samples. Farmland parcels, due to their complex shapes and fine boundaries, require extremely high manual annotation costs, resulting in a very limited number of publicly available farmland parcel datasets. Directly training deep neural networks under small sample conditions often fails to achieve ideal generalization performance, thus limiting their practical application in the automated extraction of large-scale farmland parcels.

[0004] In recent years, a new paradigm has emerged in the field of computer vision, represented by large-scale parameter pre-trained models. These models typically have hundreds of millions to billions of parameters and, through training on large-scale natural image datasets, have achieved powerful feature extraction capabilities and cross-task generalization abilities. This technological advancement has largely alleviated the bottleneck of lacking large-scale labeled data and high-computing resources in downstream tasks, providing a new solution for image segmentation tasks.

[0005] However, most existing pre-trained models are trained on datasets primarily composed of natural scene images, which differ significantly from remote sensing imagery in terms of spatial resolution, spectral characteristics, and observation angle. This distributional difference often results in suboptimal segmentation performance when directly transferring pre-trained models from natural images to remote sensing scenes. Therefore, combining large models with the remote sensing domain requires further fine-tuning to suit the characteristics of remote sensing imagery, in order to fully leverage the potential of large models in the automated extraction of farmland plots. Summary of the Invention

[0006] The purpose of this invention is to address the shortcomings of existing technologies by providing a farmland plot extraction method based on large model fine-tuning. This method can achieve high-precision farmland segmentation with only a small number of samples, significantly reducing the dependence on large-scale labeled data.

[0007] To solve the above-mentioned technical problems, the present invention adopts the following technical solution: A method for extracting farmland plots based on large model fine-tuning includes the following steps: Step 1: Acquire remote sensing image data of the target area and preprocess it; Step 2: Delineate and label the field sample vectors of the preprocessed remote sensing image data in Step 1 to create a semantic segmentation dataset; Step 3: Construct the FARMSeg model, a semantic segmentation model for farmland plots, and train the FARMSeg model using the semantic segmentation dataset from Step 2. Step 4: Taking into account spatial neighborhood information, the remote sensing images of farmland to be extracted are segmented and enhanced to obtain an image set. The segmented and enhanced image sets are then input into the trained FARMSeg model for prediction. The multiple prediction results of different segments are fused and stitched together to obtain the final segmentation result image. Step 5: Post-process the final segmentation result map to generate a farmland plot result map that meets the resolution requirements.

[0008] Furthermore, the data preprocessing method in step 1 includes: performing band filtering, resampling and cropping preprocessing operations on the remote sensing image, filtering the RGB band data of the image, and unifying its spatial resolution and pixel size.

[0009] Furthermore, the structure of the farmland parcel semantic segmentation model FARMSeg in step 3 includes: The pre-trained DINOv2 visual base model was used as the backbone network model to extract high-level semantic features from remote sensing images. Using Mask2Former as a decoder, it is used to perform multi-scale fusion and prediction of high-level semantic features of remote sensing images, and generate segmentation prediction results at the farmland instance level. Lightweight Rein modules are embedded between the layers of the backbone network model to keep the backbone network model parameters frozen and to enhance and refine the features of each layer layer by layer in a farmland instance-aware manner. Furthermore, the Rein module employs methods for progressive enhancement and refinement, including: A learnable token mechanism and an attention-inspired mechanism are introduced. The learnable token mechanism is used to adaptively optimize features for different farmland areas, and the attention-inspired mechanism is used to perform fine-grained feature optimization for different farmland areas in a single remote sensing image. The feature maps extracted by the backbone network model are also calculated. The dot product between the token and the feature map is generated. Similarity graph with token Based on this, the change in each feature map layer is calculated. ; Then calculate the change in the feature map obtained through the token. The features are fused with the original features, and the fused features are used as input to the next layer, allowing the features to be continuously optimized across multiple levels. , , , in, This represents the refined feature map. It is the input image. This represents the patch embedding layer in the model. Represent a The two-dimensional feature matrix space, Indicates the number of patches. express Dimensions Indicates the total number of floors; Finally, storage requirements are reduced by sharing parameters across layers, and redundant computations are reduced by using a low-rank token design.

[0010] Furthermore, the method for performing block-based enhancement processing on the remote sensing image in step 4 includes: Set the reserved window size when inputting remote sensing images. The above is divided into several sub-blocks using the sliding window method. Its center coordinates are defined as:

[0011] in, This indicates the input farmland remote sensing image. The tensor space dimension in which this remote sensing image exists contains One pixel, Indicates altitude, Indicates the width, each cell is composed of 3D spectral vector description, These correspond to the step size in the horizontal and vertical directions, respectively. These correspond to the width and height of the sliding window, respectively. Then for each sub-block Applying multiple enhancement operators The resulting enhanced image set:

[0012] in, Including geometric transformations and radiative transformations, This indicates the total number of data augmentation methods.

[0013] Furthermore, the method for fusing and stitching together the multiple prediction results from different blocks in step 4 includes: For each remote sensing image, only the central area within the window is retained in the prediction results. Prediction results:

[0014] in, Represents pixels The probability of belonging to a land parcel category predicted by the FARMSeg model. Represents element-wise product. For the indicator function of the central region, This indicates the number of pixels to retain in the width direction of the window. This indicates the number of pixels to retain in the height direction of the window; The prediction results of multiple remote sensing images of the same location are fused using a weighted average or voting strategy. Methods include:

[0015] In the formula, Indicates the first k The weight of each enhancement result in the final fusion k Indicates the first k There are K types of data augmentation methods. Concatenate the prediction results of all sub-blocks A weighted fusion strategy is adopted in the overlapping areas. The final segmentation result image is obtained: .

[0016] Furthermore, the post-processing methods in step 5 include boundary tracing, area filtering, and polygon simplification.

[0017] Another object of the present invention is to provide a system for implementing the above-described method for extracting farmland plots based on large model fine-tuning, comprising: The data acquisition module is used to acquire remote sensing image data of the target area and preprocess it; The dataset creation module is used to delineate and label field sample vectors from preprocessed remote sensing image data, creating a semantic segmentation dataset. The model building and training module is used to build the farmland plot semantic segmentation model FARMSeg and train the FARMSeg model using a semantic segmentation dataset. The fusion and stitching processing module takes into account spatial neighborhood information, performs block enhancement processing on the farmland remote sensing image to be extracted to obtain an image set, inputs the block-enhanced image set into the trained FARMSeg model for prediction, and fuses and stitches the multiple prediction results of different blocks to obtain the final segmentation result image. The post-processing module is used to post-process the final segmented result image to generate a farmland plot result image that meets the resolution requirements.

[0018] An electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above method.

[0019] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0020] Compared with the prior art, the beneficial effects of the present invention are as follows: (1) The plot-level farmland segmentation model FARMSeg proposed in this invention still has a strong segmentation capability under the condition of scarce samples. The model uses DINOv2 as the backbone network, combined with the lightweight Mask2Former decoder as the segmentation task head, and uses the Rein module for efficient fine-tuning. It can achieve high-precision farmland segmentation with only a small number of samples, which significantly reduces the dependence on large-scale labeled data. (2) The present invention has outstanding cross-domain generalization performance. Without retraining, the method can adapt to remote sensing images of different regions and different sensors, stably realize the automated extraction of large-scale farmland plots, and show broad application potential. (3) The present invention also proposes a large-scale farmland collaborative extraction method that integrates context awareness and reasoning enhancement. By designing a dual-window mechanism and data enhancement strategy, only the results of the central area of ​​the prediction window are retained to avoid boundary effects. The final segmentation map is generated by integrating multiple enhanced prediction results. This effectively solves the problems of plot boundary breakage and missed detection of small farmland, and significantly improves the accuracy and completeness of large-scale extraction. Attached Figure Description

[0021] Figure 1 This is an overall flowchart of the farmland plot extraction method based on large model fine-tuning in an embodiment of the present invention; Figure 2 This is a detailed flowchart of the farmland plot extraction method based on large model fine-tuning in an embodiment of the present invention. Detailed Implementation

[0022] The technical solutions of the embodiments of the present invention will be clearly and completely described below with reference to the embodiments of the present invention. Obviously, the described embodiments are only some embodiments of the present invention, and not all embodiments. Based on the embodiments of the present invention, all other embodiments obtained by those of ordinary skill in the art without creative effort are within the scope of protection of the present invention.

[0023] It should be noted that, unless otherwise specified, the embodiments and features described in the present invention can be combined with each other.

[0024] The present invention will be further described below with reference to specific embodiments, but these are not intended to limit the scope of the invention.

[0025] like Figure 1 and Figure 2 As shown, this invention discloses a method for farmland plot extraction based on large-scale model fine-tuning. It aims to fully utilize sub-meter resolution Google Earth Satellite imagery to construct a plot-level farmland segmentation model, FARMSeg, with cross-domain generalization capabilities. Furthermore, it applies a large-scale farmland collaborative extraction method that integrates block-based enhanced inference and boundary optimization to complete an application example of large-scale farmland plot extraction within a province. Specifically, it includes the following steps: Step 1: Acquire remote sensing image data of the target area and preprocess it; In this embodiment, high spatial resolution remote sensing image data of typical agricultural areas of the province under study are acquired, and the remote sensing images are preprocessed by band filtering, resampling and cropping to filter the RGB band data of the images and make them have uniform spatial resolution and pixel size.

[0026] Step 2: For the preprocessed remote sensing image data in Step 1, delineate and label the field sample vectors, and create a semantic segmentation dataset. In this embodiment, through manual interaction, based on the texture, color, shape, and boundary features of farmland in the image, visual interpretation is performed to finely delineate the boundaries of farmland plots and generate vector-format labels. The vector-format labels then undergo quality checks and corrections, followed by random sampling to further verify the annotation quality. Once the quality requirements are met, an image-label pair dataset is generated for subsequent model training.

[0027] Step 3: Construct the FARMSeg model, a semantic segmentation model for farmland plots, and train the FARMSeg model using the semantically segmented dataset from Step 2. In this step, the constructed FARMSeg model structure is as follows: The pre-trained DINOv2 visual base model is used as the backbone network model to extract high-level semantic features of remote sensing images. Mask2Former is used as a decoder to perform multi-scale fusion and prediction of high-level semantic features of remote sensing images, generating segmentation prediction results at the farmland instance level. In addition, lightweight Rein modules are embedded between the layers of the backbone network model to keep the backbone network model parameters frozen, and the features of each layer are enhanced and refined layer by layer in a farmland instance perception manner.

[0028] It should be noted that the Rein module in this embodiment enhances and refines each layer of features layer by layer as follows: First, a learnable token mechanism and an attention-inspired mechanism are introduced. Specifically, the learnable token mechanism optimizes adaptive features for different farmland regions. The Rein module constructs a series of randomly initialized learnable tokens and freezes the backbone network during fine-tuning of the large model. New farmland knowledge is learned from the fine-tuning dataset and embedded into the tokens to bridge the domain gap between pre-trained natural images and farmland remote sensing images. The attention-inspired mechanism performs fine-grained feature optimization on different farmland instances in a single image, bridging the task differences between the pre-training task (masked image modeling) and the fine-tuning task (farmland semantics) segmentation. To measure the correlation between each vector in the feature map and the token—that is, which locations in the image are semantically closest to the target—the dot product between the feature map extracted by the backbone network model and the token is calculated, generating a similarity map between the feature vector and the token.

[0029] in, Representing the backbone network model i Feature maps extracted from layers, This represents a similarity graph generated using the dot product operation. Indicates the first The token sequence of the layer, express The number of tokens in the middle, n represents the number of tokens in the middle. The number of pixels in the layer feature map This represents a two-dimensional matrix space consisting of the similarity between each pixel and each learnable token; Secondly, utilize Quantitatively evaluate the relationship between different tokens and feature vectors, that is, to Softmax normalization is used to measure the association strength; a higher value indicates that the location is semantically close to a field. Based on this, the change in each layer of the feature map is calculated. ;

[0030] for Attention weights obtained by softmax normalization It is a learnable linear transformation matrix.

[0031] Inter-layer feature enhancement strategy design; feature change calculated using tokens The features are fused with the original features, and the fused features are used as input to the next layer, allowing the features to be continuously optimized across multiple levels. , , , in, This represents the refined feature map. Indicates the first The nonlinear characteristic transformation function of the layer It is the input image. This represents the patch embedding layer in the model. Represent a The two-dimensional feature matrix space, Indicates the number of patches. express Dimensions Indicates the total number of floors.

[0032] Finally, an efficient parameter sharing mechanism is implemented. Through cross-layer parameter sharing, specifically the Rein module, different layers of the backbone network share the same MLP (Multilayer Perceptron) weights, thus avoiding the need to store MLP parameters independently for each layer and significantly reducing storage requirements. Furthermore, a low-rank token design reduces redundant token computation, improving model inference efficiency. The formula for reducing token computation redundancy through low-rank matrix factorization is as follows:

[0033] in, This represents the low-rank coefficient matrix of tokens, where each row corresponds to the coefficients of a token in the low-rank space. This represents a low-rank basis matrix, mapping the low-rank coefficients back to the original token space dimension. Much smaller than the original token dimension It can reduce computational complexity without significantly reducing the ability to express information. Let the representation matrix of the token in the low-rank eigenspace be: Each token is used for [number] tokens. Represented using low-rank coefficients. Let represent the linear mapping matrix from the low-rank space to the original high-dimensional token space.

[0034] By sharing MLP weights across layers and using a low-rank token design, Rein maintains segmentation performance close to complete fine-tuning while reducing the amount of training parameter optimization by more than 99%.

[0035] Then, the small sample semantic segmentation dataset created in step 2 was used to train the plot-level farmland segmentation model FARMSeg, resulting in the trained FARMSeg model.

[0036] Step 4: Taking into account spatial neighborhood information, the remote sensing images of farmland to be extracted are segmented and enhanced to obtain an image set. The segmented and enhanced image sets are then input into the trained FARMSeg model for prediction. The multiple prediction results of different segments are fused and stitched together to obtain the final segmentation result image. In this embodiment, considering spatial neighborhood information, the provincial imagery is input into the FARMSeg model for prediction, organized by district / county administrative region. Before prediction, an appropriate window size is set; then, an optimized sliding window coordinate grid is constructed. Specifically, this is done when inputting remote sensing images. The above is divided into several sub-blocks using a sliding window method. Its center coordinates are defined as:

[0037] in, Indicates input remote sensing image The image contains the tensor space dimension. One pixel, Indicates altitude, Indicates the width, each cell is composed of 3D spectral vector description, These correspond to the step size in the horizontal and vertical directions, respectively. These correspond to the width and height of the sliding window, respectively. In this case... ; Then, various types of data augmentation transformations are applied to the aforementioned sub-blocks. Specifically, for each sub-block... Applying multiple enhancement operators The resulting enhanced image set is as follows:

[0038] in, This includes geometric transformations (rotation, translation, scaling, mirroring) and radiometric transformations (brightness, contrast, spectral perturbation). This indicates the total number of data augmentation methods.

[0039] The enhanced image set is input into the FARMSeg model for inference to obtain the prediction mask:

[0040] in, Represents pixels The probability of belonging to a land parcel category as predicted by the FARMSeg model; For each image, the prediction result of its central confidence region is extracted, that is, only the central region within the window is retained. To avoid edge redundancy and overlap:

[0041] in, Represents element-wise product. For the indicator function of the central region, This indicates the number of pixels to retain in the width direction of the window. This indicates the number of pixels to retain in the height direction of the window; in this case... .

[0042] Since the data augmentation process described above has resulted in multiple images of the same location, each with a corresponding prediction result, the multiple augmentation prediction results for the same location are fused here. Specifically, this embodiment uses a weighted average or voting strategy for fusion:

[0043] In the formula, This represents the weight of the k-th enhancement result in the final fusion. k Let K represent the k-th data augmentation method, and there are a total of K methods. Finally, seamlessly stitch together the merged result of all sub-blocks. A weighted fusion strategy is adopted in the overlapping areas. The final segmentation result image is obtained: .

[0044] Step 5: Post-process the final segmentation result map to generate a farmland plot result map that meets the resolution requirements; The post-processing methods in this step include boundary tracing, area filtering, and polygon simplification. The boundary tracing method includes an 8-neighborhood-based raster vectorization algorithm, which converts the original segmented raster results into vector polygon features with topological integrity through boundary tracing technology, with a designed area threshold of 100m². 2 An area threshold filtering mechanism is used to eliminate small noisy polygons and ensure data quality.

[0045] Area filtering is performed on the boundary-traced map. Specifically, the Douglas-Peucker algorithm is applied to simplify polygons with a tolerance of 1m. While maintaining the overall shape characteristics of the farmland plots, redundant fixed points are reduced by 30%-50%, thereby improving data storage and processing efficiency.

[0046] Next, polygon simplification is performed on the area-filtered graph, which involves a comprehensive topology check and repair, including eliminating dangling nodes, closing tiny gaps, and merging adjacent similar polygons. The threshold is set to 1m to ensure that the output data has strict topological correctness.

[0047] Finally, the GeoTIFF format raster farmland segmentation map is converted into vector data in the same format that conforms to the GIS industry standard, providing a reliable data foundation for subsequent farmland management, spatial analysis, and decision support.

[0048] To illustrate the effectiveness of this embodiment, a province in central my country is used as an example. As a major agricultural province in my country, this province features diverse farmland distribution and a complex planting structure. Its geographical location, climate, and topography provide a typical experimental scenario for farmland plot extraction research. The province's farmland plots exhibit distinct characteristics, including large-scale contiguous farmland in plains areas as well as small-scale fragmented farmland in mountainous and hilly regions. Major crops include wheat and corn, providing rich experimental data and research value for large-scale, high-precision farmland extraction research. Therefore, this embodiment uses this province as an example, applying the FARMSeg model—a plot-level farmland segmentation model with cross-domain generalization capabilities—and a large-scale collaborative farmland extraction method that integrates block-based enhanced inference and boundary optimization to complete an application example of large-scale farmland plot extraction. This embodiment extracted over 1.8 million plots in this province, covering a total farmland area of ​​102 million mu (approximately 6.67 million hectares), with an average plot size of 56.7 mu (approximately 3.45 hectares).

[0049] Farmland is widely distributed throughout the province, and its spatial distribution exhibits a variation closely related to topographical conditions. Overall, farmland is densely and continuously distributed in the low-lying eastern areas, gradually becoming sparser and more dispersed as altitude increases, forming a clear macro-pattern of "dense in plains and sparse in mountains." This east-west differentiation pattern highly coincides with the province's "flat in the east and mountainous in the west" topographical characteristics, reflecting the significant influence of topographical conditions on the spatial pattern of farmland.

[0050] To further illustrate the effectiveness of the method in this embodiment, a total of 4,000 samples from four cities in the province were selected as validation data and compared with various current mainstream farmland extraction methods (Method 1 is from A 30-mannual cropland dataset of China from 1986 to 2021; Method 2 is from Dynamic World, Near realtime global 10 m land use land cover mapping; Method 3 is from SinoLC-1: the first 1 m resolution national-scale land-cover map of China created with a deep learning framework and open-access data). The results show that this embodiment outperforms current mainstream farmland extraction methods in multiple accuracy evaluation indicators, as shown in Table 1. As can be seen from Table 1, the average intersection-union ratio (mIoU) reached 0.7114, the average F1 score (mF1) was 0.8115, the average precision (mPrecision) was 0.7882, the average recall (mRecall) was as high as 0.8749, and the overall accuracy (OA) reached 0.8984, indicating that the method in this embodiment has good cross-regional generalization ability and application potential.

[0051] Table 1. Comparison of accuracy of different farmland data products in this province (unit: %)

[0052] This invention also provides a system for implementing the above-described method for extracting farmland plots based on large model fine-tuning, comprising: The data acquisition module is used to acquire remote sensing image data of the target area and preprocess it; The dataset creation module is used to delineate and label field sample vectors from preprocessed remote sensing image data, creating a semantic segmentation dataset. The model building and training module is used to build the farmland plot semantic segmentation model FARMSeg and train the FARMSeg model using a semantic segmentation dataset. The fusion and stitching processing module takes into account spatial neighborhood information, performs block enhancement processing on the farmland remote sensing image to be extracted to obtain an image set, inputs the block-enhanced image set into the trained FARMSeg model for prediction, and fuses and stitches the multiple prediction results of different blocks to obtain the final segmentation result image. The post-processing module is used to post-process the final segmented result image to generate a farmland plot result image that meets the resolution requirements.

[0053] An electronic device includes a memory and a processor, the memory storing a computer program, the processor executing the computer program to implement the steps of the above method.

[0054] A computer-readable storage medium having a computer program stored thereon, which, when executed by a processor, implements the steps of the above-described method.

[0055] The above are merely preferred embodiments of the present invention and are not intended to limit the implementation methods and scope of protection of the present invention. Those skilled in the art should recognize that any equivalent substitutions and obvious changes made based on the content of this specification should be included within the scope of protection of the present invention.

Claims

1. A method for extracting farmland plots based on large model fine-tuning, characterized in that, Includes the following steps: Step 1: Acquire remote sensing image data of the target area and preprocess it; Step 2: Delineate and label the field sample vectors of the preprocessed remote sensing image data in Step 1 to create a semantic segmentation dataset; Step 3: Construct the FARMSeg model, a semantic segmentation model for farmland plots, and train the FARMSeg model using the semantic segmentation dataset from Step 2. Step 4: Taking into account spatial neighborhood information, the remote sensing images of farmland to be extracted are segmented and enhanced to obtain an image set. The segmented and enhanced image sets are then input into the trained FARMSeg model for prediction. The multiple prediction results of different segments are fused and stitched together to obtain the final segmentation result image. Step 5: Post-process the final segmentation result map to generate a farmland plot result map that meets the resolution requirements.

2. The method for extracting farmland plots based on large model fine-tuning according to claim 1, characterized in that, The data preprocessing methods in step 1 include: performing band filtering, resampling and cropping preprocessing operations on the remote sensing image, filtering the RGB band data of the image, and unifying its spatial resolution and pixel size.

3. The method for extracting farmland plots based on large model fine-tuning according to claim 1, characterized in that, The structure of the FARMSeg model in step 3 includes: The pre-trained DINOv2 visual base model was used as the backbone network model to extract high-level semantic features from remote sensing images. Using Mask2Former as a decoder, it is used to perform multi-scale fusion and prediction of high-level semantic features of remote sensing images, and generate segmentation prediction results at the farmland instance level. Lightweight Rein modules are embedded between the layers of the backbone network model to keep the backbone network model parameters frozen and to enhance and refine the features of each layer layer by layer in a farmland instance-aware manner.

4. The method for extracting farmland plots based on large model fine-tuning according to claim 3, characterized in that, The methods for enhancing and refining the Rein module layer by layer include: A learnable token mechanism and an attention-inspired mechanism are introduced. The learnable token mechanism is used to adaptively optimize features for different farmland areas, and the attention-inspired mechanism is used to perform fine-grained feature optimization for different farmland areas in a single remote sensing image. The feature maps extracted by the backbone network model are also calculated. The dot product between the token and the feature map is generated. Similarity graph with token Based on this, the change in each layer of feature maps is calculated. ; Then calculate the change in the feature map obtained through the token. The features are fused with the original features, and the fused features are used as input to the next layer, allowing the features to be continuously optimized across multiple levels. , , , in, This represents the refined feature map. It is the input image. This represents the patch embedding layer in the model. Represent a The two-dimensional feature matrix space, Indicates the number of patches. express Dimensions Indicates the total number of floors; Finally, storage requirements are reduced by sharing parameters across layers, and redundant computations are reduced by using a low-rank token design.

5. The method for extracting farmland plots based on large model fine-tuning according to claim 1, characterized in that, The methods for performing block-based enhancement processing on the remote sensing image in step 4 include: Set the reserved window size when inputting remote sensing images. The above is divided into several sub-blocks using the sliding window method. Its center coordinates are defined as: in, This indicates the input farmland remote sensing image. The tensor space dimension in which this remote sensing image exists contains One pixel, Indicates altitude, Indicates the width, each cell is composed of 3D spectral vector description, These correspond to the step size in the horizontal and vertical directions, respectively. These correspond to the width and height of the sliding window, respectively. Then for each sub-block Applying multiple enhancement operators The resulting enhanced image set: in, Including geometric transformations and radiative transformations, This indicates the total number of data augmentation methods.

6. The method for extracting farmland plots based on large model fine-tuning according to claim 5, characterized in that, The methods for fusing and stitching together multiple prediction results from different blocks in step 4 include: For each remote sensing image, only the central area within the window is retained in the prediction results. Prediction results: in, Represents pixels The probability of belonging to a land parcel category predicted by the FARMSeg model. Represents element-wise product. For the indicator function of the central region, This indicates the number of pixels to retain in the width direction of the window. This indicates the number of pixels to retain in the height direction of the window; The prediction results of multiple remote sensing images of the same location are fused using a weighted average or voting strategy. Methods include: In the formula, Indicates the first k The weight of each enhancement result in the final fusion k Indicates the first k There are K types of data augmentation methods. Concatenate the prediction results of all sub-blocks A weighted fusion strategy is adopted in the overlapping areas. The final segmentation result image is obtained: 。 7. The method for extracting farmland plots based on large model fine-tuning according to claim 1, characterized in that, The post-processing methods in step 5 include boundary tracing, area filtering, and polygon simplification.

8. A system for implementing the farmland plot extraction method based on large model fine-tuning as described in any one of claims 1-7, characterized in that, include: The data acquisition module is used to acquire remote sensing image data of the target area and preprocess it; The dataset creation module is used to delineate and label field sample vectors from preprocessed remote sensing image data, creating a semantic segmentation dataset. The model building and training module is used to build the farmland plot semantic segmentation model FARMSeg and train the FARMSeg model using a semantic segmentation dataset. The fusion and stitching processing module takes into account spatial neighborhood information, performs block enhancement processing on the farmland remote sensing image to be extracted to obtain an image set, inputs the block-enhanced image set into the trained FARMSeg model for prediction, and fuses and stitches the multiple prediction results of different blocks to obtain the final segmentation result image. The post-processing module is used to post-process the final segmented result image to generate a farmland plot result image that meets the resolution requirements.

9. 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 to 7.

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

Images