Farmer cultivated land boundary extraction interaction method based on satellite remote sensing and artificial intelligence

Abstract

The invention discloses a farmer cultivated land boundary extraction interaction method based on satellite remote sensing and artificial intelligence, and belongs to the technical field of intelligent agriculture. The method comprises the following steps: a basic data preparation step: receiving and preprocessing satellite remote sensing image data, digital elevation model data and soil type data; a core interaction step: displaying an image on a front-end interface and receiving a positive or negative feedback click operation of a user; a core algorithm processing step: inputting the interactive operation and the multi-modal data into an agricultural segmentation arbitrary model optimized by agricultural field data, generating a candidate cultivated land boundary through multi-modal feature fusion, and outputting an optimal cultivated land boundary based on an area consistency loss function and a result selector; and a real-time feedback and iterative optimization step: feeding back a boundary result to a front end in real time. According to the method, professional interior work is converted into simple field work operation, and through a special model and a closed-loop interaction mechanism, the operation cost and the technical threshold are greatly reduced while the result precision is ensured.

Description

Technical Field

[0001] This invention belongs to the field of interactive image segmentation technology based on satellite remote sensing and artificial intelligence, specifically relating to an interactive method for extracting farmland boundaries based on satellite remote sensing and artificial intelligence. Background Technology

[0002] With the development of spatial information technology, society is paying increasing attention to spatial information. In the agricultural sector, the state attaches great importance to the confirmation, registration and certification of rural collective land rights. Among the most important aspects is the determination of the spatial location and boundaries of land plots under land contract management rights. Nationwide rural cadastral surveys typically take four to five years or even longer, consuming a large amount of human and material resources. The timeliness of the data cannot meet the diverse needs of specific agricultural operations.

[0003] Existing solutions typically rely on real-world application scenarios, such as land management rights transfers and agricultural insurance purchases, where third-party organizations are used to map the latest spatial information and boundaries of farmers' farmland management rights.

[0004] Existing technical solutions can be divided into three steps: basic data preparation, spatial location determination, and farmland boundary extraction. Basic data typically uses near-ground UAV remote sensing and high-altitude satellite remote sensing as base map data; spatial location determination usually involves field surveys, using identification by farmers or relevant personnel, and mobile phone positioning assistance to determine spatial locations; farmland boundary extraction is typically done manually in the office, using specialized spatial data processing software to draw farmland boundaries and associate them with relevant farmer attribute information, completing the identification and extraction of spatial information and boundaries of farmer farmland plots. Commonly used devices are mobile phones and PCs.

[0005] How to improve efficiency and reduce costs through innovation in technologies such as artificial intelligence and information systems is a key challenge that various technical solutions urgently need to overcome. Currently, there are related attempts using technologies including artificial intelligence and human-computer interaction, but a complete solution has not yet been formed.

[0006] The patent with publication number CN117853511A, entitled "An Interactive Extraction Method for Typical Land Features in High-Resolution Remote Sensing Images," describes a method that involves preprocessing and multi-scale segmentation of high-resolution remote sensing images, followed by positive sample labeling of typical land features to be extracted from the images through manual interaction. Based on differences in spectral, shape, and texture features, a fully connected conditional random field is used to achieve the initial extraction of labeled typical land features. Finally, multiple manual contour optimizations are performed to extract the boundaries of the land features.

[0007] The patent "Method and Apparatus for Extracting Farmland Boundary Lines" with publication number CN118485921A describes the process of acquiring near-ground remote sensing images of farmland areas; outputting segmented images after classification through a farmland plot boundary line prediction model; then processing contiguous plots to finally determine the farmland boundary lines.

[0008] The patent with publication number CN117057936A, titled "An Agricultural Insurance Underwriting System," describes how a mobile terminal supporting real-time positioning is embedded in a GIS system. The target, i.e., the location information of farmland, is drawn in the GIS system, and the target information is imported into a geographic information system map to generate target information with spatial data, thus completing the underwriting operation.

[0009] The patent with publication number CN107067326A, titled "An Agricultural Insurance Underwriting System and Its Implementation Method," describes how staff use a handheld tablet to record latitude and longitude coordinates at the intersection of the edges of the land to be insured, thereby obtaining the precise geographical location, area information, plant species, and actual growth status of the land, which are then uniformly transmitted back to the target storage module.

[0010] The patent application CN112734579A, titled "A Precise Underwriting Method for Crop Insurance Based on Satellite Remote Sensing Technology," describes a process where, on-site, an insurance coordinator identifies the planting location of a large-scale farmer using high-resolution satellite imagery of the target area. Based on the coordinator's description and identification, a salesperson maps the farmer's insured plots. A human-computer interaction platform (APP client) displays and records the plot location and area in real time, and provides feedback on the plot area to guide the coordinator in correcting the plot location and area. When the planting location and map area provided by the coordinator and salesperson are substantially consistent, the farmer's insured plots are exported and saved in .shp format. The plot data is generated using artificial intelligence technology or is ownership confirmation data provided by the agricultural sector.

[0011] Existing technical solutions can be divided into three key steps: basic data preparation, spatial location determination, and farmland boundary extraction. Although existing solutions have completed the extraction of farmland boundary data for farmers to a certain extent, they lack advancement and cannot achieve breakthroughs in work efficiency and result quality. They also cannot provide data support in larger areas and more refined business scenarios. The following key issues can be summarized in the existing technologies.

[0012] First, the existing basic data preparation mainly relies on drone imagery, which has a higher resolution than satellite imagery, and the textures of farmland and crops in the images are more easily recognized by the human eye. Therefore, drone imagery is used as the basis to meet the needs of subsequent manual delineation of boundaries. However, drone imagery is less efficient and more labor-intensive than satellite imagery, and without timely flight operations, it is impossible to obtain images of crops during critical time periods. Considering the overall cost, currently, drone imagery is usually only taken in areas of large-scale growers, lacking the collection of basic image data from all growers, and also unable to carry out subsequent farmland boundary collection work.

[0013] Secondly, the existing methods for determining the spatial location of farmland typically involve staff using professional positioning instruments or mobile phones, guided by the farmer to the farmland site to collect location data; or using remote sensing images where the farmer describes and identifies the plots on-site, which staff then record. Both methods require on-site communication and confirmation with the farmer before the data is processed in the office. Furthermore, the information on the spatial location and boundaries of the farmland confirmed on-site must be clear and unambiguous. This information plays a crucial role in the subsequent extraction of farmland boundaries in the office. If the boundary drawing is unqualified due to unclear information, a second on-site confirmation with the farmer is required.

[0014] Third, existing methods for extracting farmland boundaries typically involve geographic information professionals using specialized vector drawing software for indoor mapping. This work requires extensive prior agricultural knowledge and mapping experience, taking into account factors such as imagery, farmer information, total cultivated land area, and the spatial location of farmland identified by farmers to complete the boundary extraction. Due to its lengthy timeframe and the specific needs of the personnel involved, this work is usually not feasible in the field when costs are manageable. Existing technical solutions have attempted to provide interactive apps and on-site data collection apps, but the effectiveness and efficiency of field mapping are low, and the quality of the results cannot meet the high demands of business operations.

[0015] In summary, due to the unique nature of boundary extraction, mapping and on-site verification cannot be carried out simultaneously at low cost and high efficiency. On-site verification requires a significant amount of fieldwork, and the professional skills of the office operators largely determine the quality of the results. Existing technical solutions also include methods for pre-extracting the boundaries of all cultivated land patches, but this requires substantial time and cost upfront, and the pre-extracted boundaries often differ significantly from those identified by farmers, necessitating considerable manual post-processing. Summary of the Invention

[0016] The purpose of this invention is to provide an interactive method for extracting farmland boundaries for farmers based on satellite remote sensing and artificial intelligence, which fundamentally solves the core technical problems of long operation cycles, high communication costs, and high technical barriers in traditional farmland boundary extraction methods.

[0017] To solve the above-mentioned technical problems, the technical solution adopted by the present invention is as follows: An interactive method for extracting farmland boundaries based on satellite remote sensing and artificial intelligence includes the following steps: Basic data preparation steps: Receive satellite remote sensing image data, digital elevation model data, and soil type data; perform orthorectification, image fusion, and atmospheric correction preprocessing on the satellite remote sensing image data. Core interactive steps: The preprocessed satellite remote sensing image is displayed on the front-end visual operation interface, and the user's interactive click operation on the satellite remote sensing image is received. The interactive click operation includes positive feedback clicks and negative feedback clicks. The core algorithm processing steps are as follows: The interactive click operation, satellite remote sensing image data, digital elevation model data, and soil type data are input into the arbitrary agricultural segmentation model. Candidate farmland boundaries are generated through multimodal feature fusion. The optimal farmland boundary is output based on the area consistency loss function and the result selector. Real-time feedback step: The optimal farmland boundary is fed back to the front-end visual operation interface for display in real time; Iterative optimization steps: Repeat the core interaction steps and core algorithm processing steps until the optimal farmland boundary matches the actual farmland boundary identified by the farmer.

[0018] In one embodiment of the present invention, the core interaction step includes: The positive feedback click is executed outside the existing farmland boundary range and is used to indicate the expansion of the farmland boundary range; The negative feedback click is executed in areas that are not actually cultivated land within the existing cultivated land boundary range, and is used to indicate the reduction of the cultivated land boundary range.

[0019] In one embodiment of the present invention, the arbitrary agricultural segmentation model processes data through the following steps: Multimodal feature fusion steps: Weighted fusion of satellite remote sensing image features, digital elevation model features, and soil type features to generate a fused feature map; Mask decoding steps: Input the fused feature map and the prompt encoding of the user interaction click into the mask decoder to generate candidate segmentation results; Result selection steps: Based on the confidence level of the candidate segmentation results and their consistency with the cultivated land area provided by the growers, select the optimal cultivated land boundary.

[0020] In one embodiment of the present invention, the fusion function used in the multimodal feature fusion step is: ; Among them, the output It is a comprehensive feature map that integrates multi-dimensional information such as spectrum, topography, and phenology. It will replace single image features and be input into the subsequent mask decoder to provide richer and more robust contextual information for segmentation. , , These are image, elevation, and soil features, respectively. As learnable weights, their parameter values ​​will be automatically learned and adjusted during the training process through optimization algorithms; To activate the function, a nonlinear transformation is introduced, enabling the model to learn and represent more complex feature relationships, rather than a simple linear weighting. This is a bias term used to adjust the baseline of the fusion results, making the model fit better.

[0021] In one embodiment of the present invention, the training process of the arbitrary agricultural segmentation model employs a total loss function that includes an area consistency loss function, wherein the area consistency loss function is: ; in, To predict the boundary area of ​​arable land, This represents the actual boundary area of ​​arable land. The total loss function is: ; in, is the total training loss, which is the objective that the model ultimately needs to minimize; is the intersection-union loss; directly optimizes the overlap area between the predicted boundary and the true boundary, guiding the model to learn how to generate accurate shapes and boundaries; This is due to the loss of area consistency. The purpose of regularization loss is to prevent the model from overfitting, that is, the model performs too well on the training data and loses its ability to generalize to new data. , These are the weight hyperparameters for the area loss term and the regularization term, respectively, used to adjust their importance in the total loss.

[0022] In one embodiment of the present invention, an incremental learning step is also included: Collect farmland boundary data that has been manually verified and confirmed by users during the interaction process as new training samples; The parameters of the arbitrary agricultural segmentation model are updated using an elastic weight fixation strategy, and the loss function used is: ; in, The total loss is the overall objective function that the model needs to minimize during the incremental learning process. The loss is the standard loss on new tasks, driving the model to learn new knowledge from new data; This is a real number greater than 0, set manually, to control the proportion of "retaining old knowledge" in the total loss. The larger the value, the stronger the model protects old knowledge, and the slower and more cautious it is in learning new knowledge. This is the Fisher information matrix, used to protect important parameters; The larger the value, the more important parameter i is in the old task, and the more it needs to be protected when learning new tasks to avoid major changes. The current parameter values ​​of the model; This represents the optimal value after training on the old task.

[0023] In one embodiment of the present invention, the basic data preparation step further includes a step of data collection by growers: Receive grower login information via mobile terminal; The mobile terminal receives on-site photos and location data taken by farmers on the cultivated land plots. The on-site photos and location data are associated with and stored with the grower's information.

[0024] Furthermore, the result selector selects the optimal farmland boundary based on the following criteria: Calculate the confidence score for each candidate farmland boundary; Calculate the difference between the boundary area of ​​each candidate arable land and the reference value of arable land area provided by the grower; The candidate farmland boundary with the highest confidence score and the smallest area difference is selected as the optimal farmland boundary.

[0025] In one embodiment of the present invention, after the iterative optimization step, a manual correction step is further included: Manual vector correction is performed on local areas in the optimal cultivated land boundary where there is edge overlap, non-collinear edges of patches, or uneven boundaries.

[0026] In one embodiment of the present invention, in the core interaction step, when there is no self-collected location data from the growers, the grower's plot is located in the following way: By combining visual judgment of map annotations in satellite imagery, search results for place names, and on-site identification by farmers, the spatial location of the farmers' plots can be determined.

[0027] Compared with the prior art, the present invention has the following beneficial effects: This invention introduces artificial intelligence algorithms, changing the existing traditional work mode of drawing farmland boundaries in the office. Ordinary staff can achieve the same production quality as professionals in the traditional mode through quick and simple interactive operations in the field, solving the cost and efficiency problems caused by outdated technology in farmland boundary drawing.

[0028] This invention significantly improves the ease of operation and efficiency, while ensuring that the quality of the results is not reduced. It transforms the original highly specialized indoor drawing work into a field that can be completed efficiently. The goal is to complete the entire boundary extraction work at the same time as confirming the spatial location and farmland boundaries with the growers on site, and quickly output the corresponding extraction results. This allows growers to complete the farmland boundary identification work in one go, greatly reducing communication and labor costs.

[0029] This invention allows farmers to easily collect data on their mobile devices. Farmers (or relevant grassroots staff) can locate and take photos of the farmer's farmland to obtain the corresponding spatial information, which greatly facilitates the subsequent identification of the spatial location.

[0030] By introducing artificial intelligence algorithms, this invention reduces the excessive demand for image clarity required by human visual recognition. The original drone imagery data can be replaced by satellite imagery, reducing the cost of acquiring basic data and expanding the coverage of work at the same cost.

[0031] This invention takes artificial intelligence algorithms as its starting point and quality, efficiency, and cost as its core factors. Without compromising the quality of the results, it reduces the professional and technical requirements of the work, fundamentally changes the working mode of internal boundaries, simplifies the workflow, and thus achieves cost reduction and efficiency improvement in data collection, operation process, and communication with growers. Attached Figure Description

[0032] To more clearly illustrate the technical solutions of the embodiments of the present invention, the accompanying drawings used in the embodiments will be briefly introduced below. It should be understood that the following drawings only show some embodiments of the present invention and should not be regarded as a limitation of the scope. For those skilled in the art, other related drawings can be obtained from these drawings without creative effort.

[0033] Figure 1 This is a flowchart of the interactive method of the present invention.

[0034] Figure 2 This is a flowchart of the core algorithm of the present invention.

[0035] Figure 3 This is a flowchart of the incremental learning and model self-upgrading mechanism of the present invention.

[0036] Figure 4 This is an example screenshot of the boundary extraction results of the present invention.

[0037] Figure 5 The screenshots shown are enlarged details of the results of this invention.

[0038] Figure 6 This is an example diagram illustrating the operation of the present invention.

[0039] Figure 7 This is a vector result diagram returned by the algorithm after the first feedback (positive feedback) of this invention.

[0040] Figure 8 This is the vector result returned by the algorithm after the second feedback (positive feedback) of this invention.

[0041] Figure 9 This is a vector result diagram returned by the algorithm after the third feedback (negative feedback) of this invention.

[0042] Figure 10 This is an example diagram of the final interactive extraction result of the present invention. Detailed Implementation

[0043] In the following description, only certain exemplary embodiments are briefly described. As those skilled in the art will recognize, the described embodiments can be modified in various ways without departing from the spirit or scope of the embodiments of the invention. Therefore, the drawings and description are considered to be exemplary in nature and not restrictive.

[0044] The embodiments of the present invention will now be described in detail with reference to the accompanying drawings.

[0045] Example 1: This embodiment discloses an interactive method for extracting farmland boundaries for farmers based on satellite remote sensing and artificial intelligence, including the following steps: Basic data preparation steps: Receive satellite remote sensing image data, digital elevation model data, and soil type data; perform orthorectification, image fusion, and atmospheric correction preprocessing on the satellite remote sensing image data. Core interactive steps: The preprocessed satellite remote sensing image is displayed on the front-end visual operation interface, and the user's interactive click operation on the satellite remote sensing image is received. The interactive click operation includes positive feedback clicks and negative feedback clicks. The core algorithm processing steps are as follows: The interactive click operation, satellite remote sensing image data, digital elevation model data, and soil type data are input into the arbitrary agricultural segmentation model. Candidate farmland boundaries are generated through multimodal feature fusion. The optimal farmland boundary is output based on the area consistency loss function and the result selector. Real-time feedback step: The optimal farmland boundary is fed back to the front-end visual operation interface for display in real time; Iterative optimization steps: Repeat the core interaction steps and core algorithm processing steps until the optimal farmland boundary matches the actual farmland boundary identified by the farmer.

[0046] In one embodiment of the present invention, the core interaction step includes: The positive feedback click is executed outside the existing farmland boundary range and is used to indicate the expansion of the farmland boundary range; The negative feedback click is executed in areas that are not actually cultivated land within the existing cultivated land boundary range, and is used to indicate the reduction of the cultivated land boundary range.

[0047] In one embodiment of the present invention, the arbitrary agricultural segmentation model processes data through the following steps: Multimodal feature fusion steps: Weighted fusion of satellite remote sensing image features, digital elevation model features, and soil type features to generate a fused feature map; Mask decoding steps: Input the fused feature map and the prompt encoding of the user interaction click into the mask decoder to generate candidate segmentation results; Result selection steps: Based on the confidence level of the candidate segmentation results and their consistency with the cultivated land area provided by the growers, select the optimal cultivated land boundary.

[0048] In one embodiment of the present invention, the fusion function used in the multimodal feature fusion step is: ; Among them, the output It is a comprehensive feature map that integrates multi-dimensional information such as spectrum, topography, and phenology. It will replace single image features and be input into the subsequent mask decoder to provide richer and more robust contextual information for segmentation. , , These are image, elevation, and soil features, respectively. As learnable weights, their parameter values ​​will be automatically learned and adjusted during the training process through optimization algorithms; To activate the ReLU function, a nonlinear transformation is introduced, enabling the model to learn and represent more complex feature relationships, rather than a simple linear weighting. This is a bias term used to adjust the baseline of the fusion results, making the model fit better.

[0049] In one embodiment of the present invention, the training process of the arbitrary agricultural segmentation model employs a total loss function that includes an area consistency loss function, wherein the area consistency loss function is: ; in, To predict the boundary area of ​​arable land, This represents the actual boundary area of ​​arable land. The total loss function is: ; in, is the total training loss, which is the objective that the model ultimately needs to minimize; is the intersection-union loss; directly optimizes the overlap area between the predicted boundary and the true boundary, guiding the model to learn how to generate accurate shapes and boundaries; This is due to the loss of area consistency. The purpose of regularization loss is to prevent the model from overfitting, that is, the model performs too well on the training data and loses its ability to generalize to new data. , These are the weight hyperparameters for the area loss term and the regularization term, respectively, used to adjust their importance in the total loss.

[0050] In one embodiment of the present invention, an incremental learning step is also included: Collect data on farmland boundaries that users confirm are correct during the interaction process as new training samples; The parameters of the arbitrary agricultural segmentation model are updated using an elastic weight fixation strategy, and the loss function used is: ; in, The total loss is the overall objective function that the model needs to minimize during the incremental learning process. The loss is the standard loss on new tasks, driving the model to learn new knowledge from new data; This is a real number greater than 0, set manually, to control the proportion of "retaining old knowledge" in the total loss. The larger the value, the stronger the model protects old knowledge, and the slower and more cautious it is in learning new knowledge. This is the Fisher information matrix, used to protect important parameters; The larger the value, the more important parameter i is in the old task, and the more it needs to be protected when learning new tasks to avoid major changes. The current parameter values ​​of the model; This represents the optimal value after training on the old task.

[0051] In one embodiment of the present invention, the basic data preparation step further includes a step of data collection by growers: Receive grower login information via mobile terminal; The mobile terminal receives on-site photos and location data taken by farmers on the cultivated land plots. The on-site photos and location data are associated with and stored with the grower's information.

[0052] Furthermore, the result selector selects the optimal farmland boundary based on the following criteria: Calculate the confidence score for each candidate farmland boundary based on the Softmax probability value output by the model; Calculate the difference between the boundary area of ​​each candidate arable land and the reference value of arable land area provided by the grower; The candidate farmland boundary with the highest confidence score and the smallest area difference is selected as the optimal farmland boundary.

[0053] In one embodiment of the present invention, after the iterative optimization step, a manual correction step is further included: Manual vector correction is performed on local areas in the optimal cultivated land boundary where there is edge overlap, non-collinear edges of patches, or uneven boundaries.

[0054] In one embodiment of the present invention, in the core interaction step, when there is no self-collected location data from the growers, the grower's plot is located in the following way: By combining visual judgment of map annotations in satellite imagery, search results for place names, and on-site identification by farmers, the spatial location of the farmers' plots can be determined.

[0055] To facilitate a better understanding of the present invention by those skilled in the art, the present invention will be further described below in conjunction with specific embodiments.

[0056] See Figures 1-10 An interactive method for extracting farmland boundaries for farmers, based on satellite remote sensing and artificial intelligence, is proposed. The entire extraction and interaction method consists of three steps: basic data preparation, core interaction steps, and core algorithm processing. Basic data preparation: This includes receiving, processing, and displaying remote sensing image data, basic data, spatial data, and data collected by farmers themselves.

[0057] It includes the following steps: Data reception: Receives raw remote sensing image data from satellites and drones, as well as images and location data acquired from mobile devices; receives information such as personal information of growers, quantity and area of ​​cultivated land, and types of crops planted, entered by users or imported in batches, and ensures the integrity and accuracy of the data.

[0058] Data preprocessing: Preprocessing of raw image data includes orthorectification, image fusion, atmospheric correction, and image mosaicking to improve the quality and usability of the image data. Radiometric correction primarily aims to eliminate image distortion caused by radiometric errors during image capture; geometric correction aims to eliminate geometric parameter distortions caused by various factors throughout the production process; and band fusion aims to merge the original grayscale images into RGB color images that are easily visible to the human eye.

[0059] Farmer self-collection of data: The data collection process for farmers adopts the simplest possible workflow. Farmers only need to log in on their mobile devices and take a photo to complete the collection of plot location information. The aim is to set the lowest possible barrier to entry to promote wider adoption among users, providing location information to assist in subsequent boundary identification and extraction work, and improving the efficiency of related tasks. The specific collection process includes the following steps: First, log in using the farmer's name and mobile phone number; after confirming that the login information is correct, go to the farmland plot to take photos. Usually, the photos only need to show the area of ​​the plot and the crops. The farmland location information will be automatically obtained and sent back; repeat the photo taking steps until all farmland has photos, and the spatial location of the farmer's farmland will be collected.

[0060] It should be added that, considering the actual technical level of farmers, the entire process can also be carried out by third-party business personnel (such as village officials, agricultural insurance coordinators, etc.) on behalf of the farmers.

[0061] Data services: Responsible for publishing various types of data, slicing data on the server side to ensure visualization efficiency and quality, and facilitating subsequent display and analysis.

[0062] Data visualization: Provides rendering and display capabilities for various data and services such as map services, image services, and vector data, allowing users to dynamically view various data and information on the map.

[0063] Core interaction steps are as follows Figure 1 As shown, the core interaction is usually based on the front-end visual operation interface. The staff interacts with the back-end core algorithm module through click operations, and the back-end feeds back the boundary vector results generated by the algorithm to the front-end interface for display in real time.

[0064] During the boundary extraction process, staff can click multiple times to continuously optimize the boundary based on communication with farmers on-site, maximizing the match with the farmers' identification, and ultimately completing the extraction of the plot boundary. It should be noted that the core interactive module is the core module of the boundary extraction work and can be conducted in the field. Typically, farmers in the area need to be gathered at village committees, town governments, or similar locations to cooperate with staff in conducting on-site identification.

[0065] The specific operation includes the following steps: 1. Staff quickly locate the farmers' plots based on existing spatial positioning data collected by farmers; if there is no positioning data, they conduct a comprehensive visual inspection on-site based on satellite imagery, map annotations, and place name searches, combined with the farmers' on-site visual identification to determine and locate the plots. Second, staff members interact and click on the corresponding spatial locations in the image. After calculation by the backend algorithm, the corresponding boundary extraction results are returned in real time. Staff members and farmers confirm on-site whether the results match the actual cultivated land area and boundary conditions. The clicking operation is repeated until the requirements are met, and the boundary extraction is completed.

[0066] Thirdly, it should be noted that the front-end interaction method is divided into positive and negative feedback, using the left and right mouse buttons as an example: After the plot boundary is fed back to the interface by the algorithm, if the currently extracted cultivated land area is smaller than the actual cultivated land (i.e., the boundary range is smaller than the actual boundary range), then clicking the left button outside the existing boundary range provides positive feedback, and the algorithm will expand the boundary range and return the extraction result again. If the currently extracted cultivated land area is larger than the actual cultivated land (i.e., the boundary range is larger than the actual boundary range), then clicking the right button in the non-actual cultivated land area within the existing boundary range provides negative feedback, and the algorithm will reduce the boundary range of the corresponding area and return the extraction result again. This process requires staff to communicate and interact with farmers on-site, ultimately achieving a match between the pre-defined boundary and the boundary indicated by the farmer. Figure 6 .

[0067] IV. After extraction, due to unusual farmland shapes, tree obstructions, building obstructions, etc., the extracted farmland boundaries may exhibit some non-standard features, such as edge overlap, non-collinearity of patch edges, and insufficient smoothness. These minor boundary issues can be easily corrected manually, for example (…). Figure 4 , Figure 5 ).

[0068] The core extraction algorithm is as follows: The A-SAM (Agricultural-Segment Anything Model) model proposed in this application is a novel farmland boundary extraction model optimized based on the basic SAM model through agricultural sample fine-tuning, multimodal data fusion, incremental learning mechanisms, and result selectors. Its overall architecture is as follows: Figure 2 As shown, the model supports feature-level fusion of multimodal data such as satellite imagery, digital elevation models (DEM), and soil type maps. User interactions (positive / negative samples) are fed into the encoder via prompts and, together with image features, generate candidate segmentation results in the mask decoder. The result selector comprehensively calculates the confidence of the candidate results and their consistency with the cultivated land area (weak label) of the farmers, and finally outputs the optimal boundary. The model also integrates an incremental learning module, which can continuously optimize model parameters using newly generated labeled data to achieve self-iterative upgrades.

[0069] 3.1 Detailed Explanation of Model Structure: Multimodal fusion module: Supports multi-source data input such as satellite imagery, elevation data, and soil type, enhancing the model's ability to perceive farmland boundaries through feature-level fusion. The fusion function can be expressed as: ; Among them, the output It is a comprehensive feature map that integrates multi-dimensional information such as spectrum, topography, and phenology. It will replace the single image feature in the original SAM model and be input into the subsequent mask decoder to provide richer and more robust contextual information for segmentation. , , These are image, elevation, and soil features, respectively. As learnable weights, their parameter values ​​will be automatically learned and adjusted during the training process through optimization algorithms; To activate the function, a nonlinear transformation is introduced, enabling the model to learn and represent more complex feature relationships, rather than a simple linear weighting. This is a bias term used to adjust the baseline of the fusion results, making the model fit better.

[0070] Incremental Learning Module: The model supports an online learning mechanism, continuously optimizing model parameters with new samples to avoid catastrophic forgetting. It employs an Elastic Weight Fixation (EWC) strategy, with the following loss function: ; in, The total loss is the overall objective function that the model needs to minimize during the incremental learning process. The loss is the standard loss on new tasks, driving the model to learn new knowledge from new data; This is a real number greater than 0, set manually, to control the proportion of "retaining old knowledge" in the total loss. The larger the value, the stronger the model protects old knowledge, and the slower and more cautious it is in learning new knowledge. This is the Fisher information matrix, used to protect important parameters. The larger the value, the more important parameter i is in the old task, and the more it needs to be "protected" when learning new tasks to avoid major changes. The current parameter values ​​of the model; This represents the optimal value after training on the old task.

[0071] The EWC strategy aims to ensure that, when learning a new task, the model retains the current values ​​of parameters that are important for older tasks. Do not deviate from the previous optimal value. Too far.

[0072] 3.2 Training and Fine-tuning Strategies: Fine-tuning was performed using over 100,000 farmland images and 1 million boundary-annotated samples, covering various types including plains, hills, and terraces. The model was trained using the Adam optimizer with an initial learning rate of 0.001, a batch size of 32, and 100 training epochs. An area consistency loss was introduced during training to ensure that the predicted boundary area matched the actual area.

[0073] The total loss function is:

[0074] in, is the total training loss, which is the objective that the model ultimately needs to minimize; is the intersection-union loss. It directly optimizes the overlap area between the predicted boundary and the true boundary, guiding the model to learn how to generate accurate shapes and boundaries; This refers to the area consistency loss defined above; The purpose of regularization loss is to prevent the model from overfitting, that is, the model performs too well on the training data and loses its ability to generalize to new data. , These are the weight hyperparameters for the area loss term and the regularization term, respectively, used to adjust their importance in the total loss.

[0075] 3.3 Comparative Experiments and Performance Evaluation: To verify the effectiveness of the A-SAM model (Agricultural Segmentation Arbitrary Model), we conducted comparative experiments on multiple test sets, and the results are shown in Table 1 below: Table 1:

[0076] Test area: We selected an area of approximately 25 square kilometers in the southwest region of China, with a total of 12,450 patches; an IoU > 0.75 is considered qualified. Evaluation metrics: Recall rate: The proportion of true patches that are correctly extracted; [[ID=ll]]Precision rate: The proportion of the extracted patches that are verified to be true arable land; F1-Score: The harmonic mean of recall rate and precision rate; Average number of feedbacks: The average number of interactive clicks required to extract a satisfactory plot.

[0077] In addition, we also conducted a breakdown test for different farmland types, and the results are shown in Table 2 below: Table 2:

[0078] 3.4 Multimodal data support: A-SAM supports the input of multimodal data, such as: optical satellite images (RGB, multispectral), digital elevation models (DEM), soil type distribution maps, historical arable land boundary vectors. Through multimodal feature fusion, compared with previous data inputs, the model can still maintain a high extraction accuracy in complex scenarios such as occlusion, shadow, and crop type changes. The specific processing flow includes: Data registration and preprocessing: Ensure the precise spatial registration of data such as optical images and DEM, and process them to the same resolution; Feature extraction: Specialized encoders are used for deep feature extraction of each modal data, rather than using the original pixel values; Feature fusion and normalization: The extracted features will be normalized to eliminate the dimensional difference, and then weighted fusion will be performed.

[0079] 3.5 Incremental learning and self-upgrading mechanism: A-SAM has the ability of online learning and can perform incremental training through newly collected samples to continuously improve the model's performance in specific regions or for new crop types. For example Figure 3As shown, after the model is deployed, the system continuously collects high-quality labeled data (i.e., the boundaries ultimately confirmed as correct) generated by users during interactions and passed quality checks. When new data accumulates to a certain scale or reaches a predetermined update time, the system automatically triggers the incremental learning training process. This process employs strategies such as Elastic Weight Consolidation (EWC) to avoid forgetting previously learned knowledge while acquiring new knowledge. After training, a new model version with optimized performance (such as A-SAM v1) is generated, and after testing, it is seamlessly deployed back to the application system, forming a closed-loop continuous optimization ecosystem that enables the model to continuously adapt to new regions, crops, and terrains.

[0080] It should be noted that multimodal data registration: spatial registration of image, DEM and soil data is achieved through the GDAL library, and the registration error is controlled within 1 pixel.

[0081] Model inference optimization: The ONNX runtime is used for model deployment, and the time for a single inference is controlled within 3 seconds.

[0082] Area consistency verification: The area of ​​the polygon is calculated using the Shapely library and compared in real time with the area reference value provided by the user.

[0083] This invention encapsulates professional capabilities within a backend agricultural segmentation model, transforming what was previously a high-skilled task requiring professional expertise indoors into a standardized operation that can be performed by ordinary staff in the field. This paradigm shift directly eliminates the significant communication costs and time delays caused by the separation between field identification and indoor mapping in traditional solutions, enabling integrated operations of simultaneous communication, confirmation, and mapping. Actual test data shows that the average number of interactions required to extract the boundary of a plot has decreased from 6.2 in the basic SAM model to 3.5, improving efficiency and significantly reducing reliance on professional GIS technicians, directly saving labor costs.

[0084] This invention lowers the technical threshold while ensuring the stability and high accuracy of the results. It doesn't simply replace manual labor with automation; instead, it constructs a robust intelligent auxiliary core through multimodal data fusion, an area consistency loss function, and a result selector mechanism. This core effectively integrates multidimensional information such as spectrum, topography, and soil, generating high-precision candidate boundaries even in complex scenarios with occlusion or shadows in the imagery. Simultaneously, area consistency constraints ensure a high degree of match between the output results and the area information provided by the growers. This results in an average IoU of 0.87 and an F1-Score of 0.90 for results obtained by non-professional operators, approaching the level of professional manual mapping (IoU 0.92) in key indicators, achieving an optimal balance between quality and efficiency.

[0085] The incremental learning mechanism introduced in this invention enables arbitrary agricultural segmentation models to continuously optimize themselves using newly labeled data generated in practical applications. This overcomes the inherent limitations of single static models in adapting to different regions, crop types, and time-varying landforms, endowing the system with the robustness and generalization capabilities required for long-term service, fundamentally enhancing the lifecycle and application value of the technical solution.

[0086] This invention successfully replaces traditional UAV imagery as the primary data source with satellite remote sensing imagery, which offers lower cost, wider coverage, and more timely updates, significantly reducing the cost and time required to acquire basic data. Simultaneously, the simplified data self-collection process designed for farmers further lowers the initial barrier to business startup, laying the foundation for large-scale promotion and application.

[0087] Although preferred embodiments of the invention have been described, those skilled in the art, upon learning the basic inventive concept, can make other changes and modifications to these embodiments. Therefore, the appended claims are intended to be interpreted as including both the preferred embodiments and all changes and modifications falling within the scope of the invention.

[0088] The above description is merely a preferred embodiment of the present invention and is not intended to limit the present invention. It should be noted that any modifications, equivalent substitutions, and improvements made within the spirit and principles of the present invention should be included within the protection scope of the present invention.

Claims

1. An interactive method for extracting farmland boundaries based on satellite remote sensing and artificial intelligence, characterized in that, Includes the following steps: Basic data preparation steps: Receive satellite remote sensing image data, digital elevation model data, and soil type data; perform orthorectification, image fusion, and atmospheric correction preprocessing on the satellite remote sensing image data. Core interactive steps: The preprocessed satellite remote sensing image is displayed on the front-end visual operation interface, and the user's interactive click operation on the satellite remote sensing image is received. The interactive click operation includes positive feedback clicks and negative feedback clicks. The core algorithm processing steps are as follows: The interactive click operation, satellite remote sensing image data, digital elevation model data, and soil type data are input into the arbitrary agricultural segmentation model. Candidate farmland boundaries are generated through multimodal feature fusion. The optimal farmland boundary is output based on the area consistency loss function and the result selector. Real-time feedback step: The optimal farmland boundary is fed back to the front-end visual operation interface for display in real time; Iterative optimization steps: Repeat the core interaction steps and core algorithm processing steps until the optimal farmland boundary matches the actual farmland boundary identified by the farmer.

2. The interactive method for extracting farmland boundaries based on satellite remote sensing and artificial intelligence according to claim 1, characterized in that, In the core interaction steps: The positive feedback click is executed outside the existing farmland boundary range and is used to indicate the expansion of the farmland boundary range; The negative feedback click is executed in areas that are not actually cultivated land within the existing cultivated land boundary range, and is used to indicate the reduction of the cultivated land boundary range.

3. The interactive method for extracting farmland boundaries based on satellite remote sensing and artificial intelligence according to claim 1, characterized in that, The arbitrary agricultural segmentation model processes data through the following steps: Multimodal feature fusion steps: Weighted fusion of satellite remote sensing image features, digital elevation model features, and soil type features to generate a fused feature map; Mask decoding steps: Input the fused feature map and the prompt encoding of the user interaction click into the mask decoder to generate candidate segmentation results; Result selection steps: Based on the confidence level of the candidate segmentation results and their consistency with the cultivated land area provided by the growers, select the optimal cultivated land boundary.

4. The interactive method for extracting farmland boundaries based on satellite remote sensing and artificial intelligence according to claim 3, characterized in that, The fusion function used in the multimodal feature fusion step is: ； Among them, the output It is a comprehensive feature map that integrates multi-dimensional information such as spectrum, topography, and phenology. It will replace single image features and be input into the subsequent mask decoder to provide richer and more robust contextual information for segmentation. , , These are image, elevation, and soil features, respectively. As learnable weights, their parameter values ​​will be automatically learned and adjusted during the training process through optimization algorithms; To activate the function, a nonlinear transformation is introduced, enabling the model to learn and represent more complex feature relationships, rather than a simple linear weighting. This is a bias term used to adjust the baseline of the fusion results, making the model fit better.

5. The interactive method for extracting farmland boundaries based on satellite remote sensing and artificial intelligence according to claim 1, characterized in that, The training process of the arbitrary agricultural segmentation model employs a total loss function that includes an area consistency loss function, which is: ； in, To predict the boundary area of ​​arable land, This represents the actual boundary area of ​​arable land. The total loss function is: ； in, is the total training loss, which is the objective that the model ultimately needs to minimize; is the intersection-union loss; directly optimizes the overlap area between the predicted boundary and the true boundary, guiding the model to learn how to generate accurate shapes and boundaries; This is due to the loss of area consistency. The purpose of regularization loss is to prevent the model from overfitting, that is, the model performs too well on the training data and loses its ability to generalize to new data. , These are the weight hyperparameters for the area loss term and the regularization term, respectively, used to adjust their importance in the total loss.

6. The interactive method for extracting farmland boundaries based on satellite remote sensing and artificial intelligence according to claim 1, characterized in that, It also includes incremental learning steps: Collect data on farmland boundaries that users confirm are correct during the interaction process as new training samples; The parameters of the arbitrary agricultural segmentation model are updated using an elastic weight fixation strategy, and the loss function used is: ； in, The total loss is the overall objective function that the model needs to minimize during the incremental learning process. The loss is the standard loss on new tasks, driving the model to learn new knowledge from new data; This is a real number greater than 0, set manually, to control the proportion of "retaining old knowledge" in the total loss. The larger the value, the stronger the model protects old knowledge, and the slower and more cautious it is in learning new knowledge. This is the Fisher information matrix, used to protect important parameters; The larger the value, the more important parameter i is in the old task, and the more it needs to be protected when learning new tasks to avoid major changes. The current parameter values ​​of the model; This represents the optimal value after training on the old task.

7. The interactive method for extracting farmland boundaries based on satellite remote sensing and artificial intelligence according to claim 1, characterized in that, The basic data preparation steps also include a step of data collection by growers themselves: Receive grower login information via mobile terminal; The mobile terminal receives on-site photos and location data taken by farmers on the cultivated land plots. The on-site photos and location data are associated with and stored with the grower's information.

8. The interactive method for extracting farmland boundaries based on satellite remote sensing and artificial intelligence according to claim 1, characterized in that, The result selector selects the optimal farmland boundary based on the following criteria: Calculate the confidence score for each candidate farmland boundary; Calculate the difference between the boundary area of ​​each candidate arable land and the reference value of arable land area provided by the grower; The candidate farmland boundary with the highest confidence score and the smallest area difference is selected as the optimal farmland boundary.

9. The interactive method for extracting farmland boundaries based on satellite remote sensing and artificial intelligence according to claim 1, characterized in that, Following the iterative optimization step, a manual correction step is also included: Manual vector correction is performed on local areas in the optimal cultivated land boundary where there is edge overlap, non-collinear edges of patches, or uneven boundaries.

10. The interactive method for extracting farmland boundaries based on satellite remote sensing and artificial intelligence according to claim 1, characterized in that, In the core interaction steps, when there is no self-collected location data from the growers, the grower's plot is located using the following method: By combining visual judgment of map annotations in satellite imagery, search results for place names, and on-site identification by farmers, the spatial location of the farmers' plots can be determined.

Images