A method and system for optimizing remote sensing images of farmland plots

Abstract

A method and system for optimizing remote sensing images of farmland plots, relating to the field of remote sensing image optimization, alleviates problems such as unclear plot edges in existing technologies. The optimization method includes: a data acquisition stage: collecting data on farmland plots to be optimized; a preliminary screening stage: obtaining preliminary crop types through threshold screening; a multimodal optimization stage: obtaining segmented farmland plot remote sensing images and image recognition attribute information through a multimodal fusion segmentation model; a historical fusion stage: obtaining fused plot remote sensing images and fused attribute information through spatial matching and fusion; a coordinate transformation stage: obtaining corrected plot remote sensing images through a coordinate transformation model; a verification stage: if a four-layer verification system is met, proceeding to the optimization stage; otherwise, proceeding to the multimodal optimization stage; and an optimization stage: labeling the fused attribute information and preliminary crop types onto the corrected plot remote sensing images. This invention is applicable to scenarios such as the maturity period of northern crops and mixed crop planting.

Description

Technical Field

[0001] This invention relates to the field of remote sensing image optimization, specifically to the field of farmland monitoring. Background Technology

[0002] Traditional methods for optimizing remote sensing images of farmland plots rely on manual annotation, which is inefficient, costly, and fails to optimize for the characteristics of crop maturity periods in northern regions. Major northern crops (corn, soybeans, wheat, potatoes, etc.) typically mature from late summer to autumn, with ample sunlight, dense crop canopies, and stable high NDVI / EVI / SAVI indices, resulting in flattened spectral curves. Different varieties of corn and legumes of the same family exhibit extremely high spectral similarity, easily leading to confusion. Plots are often strip fields or sloping land with "discontinuous, multi-segmented, and zigzag" boundaries, significantly different from the continuous, smooth boundaries of paddy fields in the south and the large fluctuations in vegetation indices during crop immaturity.

[0003] The following problems exist in the existing technology: First, the boundaries of farmland plots are blurred and jagged, with indistinct edges and numerous jagged or blurred transition zones. Second, the plots are scattered, with many cavities or mottling within each plot, and inconsistent classifications within the same plot, including pinholes and noise points. Third, crops with similar spectral maturity periods in northern regions (such as different varieties of maize) are frequently misclassified. Fourth, scattered small plots are often missed or merged, resulting in the loss of small plots and the incorrect merging of adjacent plots with different attributes. Fifth, the optimized farmland plot remote sensing images exhibit positional shifts, shape distortions, and significant geometric deviations in the output layers.

[0004] In the prior art, Chinese patent document CN118941582A discloses "An Optimized Segmentation Method for Geological Survey Remote Sensing Images." This technical solution calculates the direction angle of each cell based on the gradient direction of all pixels in each cell of a connected domain, and obtains the direction angle of each edge curve through a weighted average method, thereby improving the accuracy of calculating the direction angle of the edge curve and thus improving the accuracy of subsequent ridgeline determination. This method optimizes the segmentation of remote sensing images based on pixels because the pixels of the ridgeline have strong, consistent, and continuous gradient directions, and the gradient changes within the ridge region are smooth with little noise. However, the gradient directions of farmland plots are chaotic, and the direction angles of the edge curves are not suitable for the "discontinuous and multi-segmented" boundaries of farmland plots. Therefore, the optimization method used for ridgelines is not suitable for farmland plots. If used directly, it will lead to problems such as blurred plot edges and difficulty in integrating scattered plots.

[0005] In summary, existing farmland plot remote sensing image optimization technologies suffer from problems such as unclear plot edges, difficulty in integrating scattered plots, low accuracy in optimizing and identifying images of crops in the northern crop maturity period, and location and shape distortion. Summary of the Invention

[0006] This invention alleviates the problems of unclear plot edges, difficulty in integrating scattered plots, low accuracy in optimizing and identifying images of northern crops at maturity, and positional and shape distortion in existing farmland plot remote sensing image optimization technologies.

[0007] This invention provides the following solution: Option 1: A method for optimizing remote sensing images of farmland plots, comprising the following stages: Data collection phase: Collect remote sensing images of farmland plots to be optimized, remote sensing images of several farmland plots from the previous year, and corresponding attribute information from the previous year; Preliminary screening stage: Threshold screening is performed on the remote sensing images of the farmland plots to be optimized to obtain preliminary crop types; Multimodal optimization stage: The remote sensing images of the farmland plots to be optimized are used as optimization objects. The multimodal fusion segmentation model is used to perform fusion segmentation to obtain several segmented farmland plot remote sensing images and corresponding image recognition attribute information after multimodal fusion segmentation. Historical fusion stage: Spatial matching is performed on the remote sensing images of the several segmented farmland plots and the remote sensing images of several farmland plots from the previous year to obtain several related plot remote sensing images, which are then combined into a fused plot remote sensing image; The image identification attribute information and the attribute information from the previous year are fused to obtain fused attribute information; Coordinate transformation stage: The fused land parcel remote sensing images are transformed using a coordinate transformation model to obtain the corresponding corrected land parcel remote sensing images; Verification phase: If the corrected land parcel remote sensing image, the fused attribute information, and the preliminary crop type satisfy the four-layer verification system, then proceed to the optimization phase; otherwise, use the corrected land parcel remote sensing image as the optimization object for the multimodal optimization phase and proceed to the multimodal optimization phase. Optimization phase: The fused attribute information and the preliminary crop type are labeled onto the corrected remote sensing image of the plot to obtain the optimized remote sensing image of the farmland plot.

[0008] Furthermore, in one embodiment of the present invention, the threshold screening in the preliminary screening stage is as follows: The following processing is performed on each pixel in the remote sensing image of the farmland plot to be optimized to obtain several northern crop pixels as preliminary crop types: If the NDVI of the pixel is within the NDVI threshold range, the Enhanced Vegetation Index (EVI) is within the Enhanced Vegetation Index threshold range, and the Soil Adjusted Vegetation Index (SAVI) is within the Soil Adjusted Vegetation Index threshold range, then the pixel is a northern crop pixel; otherwise, it is a non-northern crop pixel.

[0009] Furthermore, in one embodiment of the present invention, the multimodal optimization stage further includes obtaining several smooth farmland remote sensing images by respectively optimizing the contours of the several segmented farmland remote sensing images, which are then used as several segmented farmland remote sensing images.

[0010] Furthermore, in one embodiment of the present invention, during the multimodal optimization stage, the multimodal fusion segmentation model is obtained through the following method: In the dataset construction phase, a dataset is constructed based on remote sensing images of all farmland plots from the previous year and their corresponding attribute information. During the training phase, based on the dataset, all remote sensing images of farmland plots from the previous year are used as input, and several segmented farmland plot remote sensing images and corresponding image recognition attribute information are used as output. The SAM3 model is trained using the joint loss function and the total loss function to obtain the trained SAM3 model as a multimodal fusion segmentation model. The joint loss function includes the overlap loss function, the boundary distance loss function, and the intersection-union ratio loss function; The total loss function includes a classification loss function and a smoothness constraint loss function.

[0011] Furthermore, in one embodiment of the present invention, during the verification stage, the four-layer verification system includes a boundary verification layer, a semantic verification layer, a fusion verification layer, and a coordinate verification layer; The boundary verification layer is used to verify the farmland boundaries in the corrected remote sensing image of the farmland plot. The verification conditions are: the overlap rate between the farmland boundary and the texture feature index abrupt change point of the remote sensing image of the farmland plot to be optimized is ≥95%, and the overlap rate between the farmland boundary and the farmland boundary of the remote sensing image of the farmland plot in the previous year is ≥92%. The semantic verification layer is used to verify the matching degree between the crop type in the fused attribute information and the preliminary crop type. If the matching degree is ≥85%, the verification passes. The fusion verification layer is used to verify the matching degree between the fused attribute information and the previous year's attribute information. If the matching degree is ≥90%, the verification passes. The coordinate verification layer is used to verify the error between the corrected land parcel remote sensing image and the fused land parcel remote sensing image. If the error is ≤0.2 meters, the verification is passed.

[0012] Option 2: A remote sensing image optimization system for farmland plots, comprising the following modules: Module 1 is used to collect remote sensing images of farmland plots to be optimized, remote sensing images of several farmland plots from the previous year, and corresponding attribute information from the previous year. Module 2 is used to perform threshold screening on the remote sensing images of the farmland plots to be optimized to obtain preliminary crop types; Module 3 is used to take the remote sensing image of the farmland plot to be optimized as the optimization object, and perform fusion segmentation through a multimodal fusion segmentation model to obtain several segmented farmland plot remote sensing images and corresponding image recognition attribute information after multimodal fusion segmentation. Module 4 is used to spatially match the remote sensing images of the several segmented farmland plots with the remote sensing images of several farmland plots from the previous year to obtain several related plot remote sensing images, and combine them into a fused plot remote sensing image; and to fuse the image recognition attribute information and the attribute information from the previous year to obtain fused attribute information. Module 5 is used to obtain the corresponding corrected remote sensing image of the land parcel through a coordinate transformation model; Module 6 is used to enter the optimization stage if the corrected land parcel remote sensing image, the fused attribute information, and the preliminary crop type satisfy the four-layer verification system; otherwise, the corrected land parcel remote sensing image is used as the optimization object in the multimodal optimization stage, and the multimodal optimization stage is entered. Module 7 is used to label the fused attribute information and the preliminary crop type onto the corrected plot remote sensing image to obtain an optimized farmland plot remote sensing image.

[0013] Option 3: An electronic device, comprising a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other via the communication bus; Memory, used to store computer programs; When a processor executes a computer program stored in memory, it implements the method described in Scheme 1.

[0014] Option 4: A computer-readable storage medium storing a computer program, wherein the computer program, when executed by a processor, implements the method described in Option 1.

[0015] The farmland plot remote sensing image optimization method and system described in this invention is based on the fusion of multimodal large models and image textures, effectively alleviating the problems of unclear plot edges, difficulty in integrating scattered plots, low accuracy in optimizing and identifying images of northern crop maturity periods, positional offset, and shape distortion in existing farmland plot remote sensing image optimization technologies. Specific beneficial effects include: This invention discloses a method for optimizing remote sensing images of farmland plots, comprising image preprocessing and preliminary screening, multimodal model segmentation and contour optimization, historical data fusion, precise coordinate transformation, and accuracy verification modules. This forms an end-to-end automated process: image preprocessing → exponential screening → multimodal segmentation → contour smoothing → historical data fusion → coordinate transformation → accuracy verification. Based on multimodal segmentation, targeted contour smoothing addresses the problem of unclear plot edges. Historical data fusion integrates scattered plots and effectively consolidates the attribute information of farmland plots. An improved coordinate transformation method avoids complex terrain offsets, further integrating scattered plots. Accuracy verification and exponential screening improve the optimization and recognition accuracy of images of northern crop maturity periods. The accuracy verification process addresses issues of positional offset and shape distortion. This achieves integrated crop type identification, historical plot fusion, and precise extraction of farmland plot boundaries, providing foundational layer support for "dual precision" in agricultural insurance.

[0016] The method described in this invention is applicable to scenarios such as the maturity period of northern crops, mixed cropping, and dispersed insured plots. It is suitable for the fields of precise underwriting and precise loss assessment in agricultural insurance, and meets the dual requirements of agricultural insurance for plot boundary accuracy and attributes. Attached Figure Description

[0017] The above and / or additional aspects and advantages of the present invention will become apparent and readily understood from the following description of the embodiments taken in conjunction with the accompanying drawings, wherein: Figure 1 This is a flowchart of a method for optimizing remote sensing images of farmland plots as described in Implementation Method 1; Figure 2 This is a flowchart of the historical fusion stage described in Implementation Method 1. Detailed Implementation

[0018] Various embodiments of the present invention will now be clearly and completely described with reference to the accompanying drawings. The embodiments described with reference to the drawings are exemplary and intended to explain the present invention, and should not be construed as limiting the present invention.

[0019] Implementation Method 1: This implementation method describes a method for optimizing remote sensing images of farmland plots, such as... Figure 1 As shown, it includes the following stages: Data collection phase: Collect remote sensing images of farmland plots to be optimized, remote sensing images of several farmland plots from the previous year, and corresponding attribute information from the previous year; Preliminary screening stage: Threshold screening is performed on the remote sensing images of the farmland plots to be optimized to obtain preliminary crop types; Multimodal optimization stage: The remote sensing images of the farmland plots to be optimized are used as optimization objects. The multimodal fusion segmentation model is used to perform fusion segmentation to obtain several segmented farmland plot remote sensing images and corresponding image recognition attribute information after multimodal fusion segmentation. Historical fusion stage: Spatial matching is performed on the remote sensing images of the several segmented farmland plots and the remote sensing images of several farmland plots from the previous year to obtain several related plot remote sensing images, which are then combined into a fused plot remote sensing image; The image identification attribute information and the attribute information from the previous year are fused to obtain fused attribute information; Coordinate transformation stage: The fused land parcel remote sensing images are transformed using a coordinate transformation model to obtain the corresponding corrected land parcel remote sensing images; Verification phase: If the corrected land parcel remote sensing image, the fused attribute information, and the preliminary crop type satisfy the four-layer verification system, then proceed to the optimization phase; otherwise, use the corrected land parcel remote sensing image as the optimization object for the multimodal optimization phase and proceed to the multimodal optimization phase. Optimization phase: The fused attribute information and the preliminary crop type are labeled onto the corrected remote sensing image of the plot to obtain the optimized remote sensing image of the farmland plot.

[0020] In this embodiment, the previous year's attribute information includes crop type.

[0021] In this embodiment, radiometric correction is performed on the remote sensing images of the farmland plots to be optimized during the data acquisition phase.

[0022] In this embodiment, the remote sensing images of farmland plots mentioned in the data acquisition stage are remote sensing images of farmland plots during the maturity period of northern crops, covering maturity period characteristic images of major northern crops such as corn, soybeans, wheat, and potatoes.

[0023] In this embodiment, the previous year's attribute information mentioned in the data collection phase includes the preliminary coordinates of the land parcel boundary, crop type, insured area, insurance number, and land parcel ownership.

[0024] In this embodiment, the coordinates of the farmland remote sensing images to be optimized and the farmland remote sensing images from the previous year during the data acquisition phase are both in the WGS84 coordinate system (accuracy ≤ 0.0001°). In this embodiment, the step of standardizing the input semantic description into short noun phrases is placed before the model input in the multimodal optimization stage. This is used to simplify natural language descriptions such as "mature corn scattered insured plots" into short noun phrases such as "corn, mature period, scattered plots", which are then injected into the model semantic prompting module to improve the model's understanding and segmentation accuracy of complex scenes.

[0025] In this embodiment, during the historical fusion stage, the spatial matching involves performing an improved RPC coordinate transformation on the remote sensing images of the segmented farmland plots and the remote sensing images of the farmland plots from the previous year, and then performing spatial overlay analysis to determine plots with a matching distance of ≤0.3 meters as associated plots.

[0026] In this embodiment, during the historical fusion stage, the fusion is performed using a weighted voting method, where the weight of the image recognition attribute information is set to 0.6 and the weight of the previous year's attribute information is set to 0.4.

[0027] In this implementation, a customized weight is designed for the farmland insurance scenario. The design process is as follows: based on the characteristic that the reliability of image recognition during the maturity period of northern crops is higher than the error of manual entry of historical insurance data, the image recognition weight is set to 0.6 and the historical attribute weight is set to 0.4; the attributes such as crop type and insured area of ​​the associated plots are voted on separately, and the result with the highest vote is used as the final attribute; the weight is determined by cross-validation of 10 sets of validation sets, taking into account both the real-time performance of the images and the stability of historical data.

[0028] In this embodiment, the coordinate transformation model during the coordinate transformation stage is as follows: ,

[0029] in, For the first Column number of pixels, For the first Row number of pixels, For the first Normalized vegetation index per pixel For the first Boundary confidence of each pixel, ∈[0,100] (SAM3 output); P1, P2, P3 are longitude numerator polynomials (each containing 10 coefficients, in the form a0+a1col+a2row+a3col×row+...+a9col³row³); R1, R2, R3 are latitude numerator polynomials (each containing 10 coefficients, in the form a0+a1col+a2row+a3col×row+...+a9col³row³); Q1, Q2, Q3 are longitude denominator polynomials (in the same form as the numerator polynomial); S1, S2, S3 are latitude denominator polynomials (in the same form as the numerator polynomial).

[0030] This model is an improved RPC model, which introduces NDVI vegetation correction terms and boundary confidence correction terms. Compared with the traditional RPC model, this method solves the problems of positional offset and shape distortion caused by complex terrain such as terraced fields and slopes in the north. The geometric error is reduced from the traditional 1.5 meters to within 0.2 meters. It strengthens the pixel coordinate correction of plot boundaries to avoid boundary blurring and misalignment. It adapts to the coordinate correction of scattered small plots to prevent the loss or merging of small plots, and fully meets the geometric accuracy requirements for precise underwriting of agricultural insurance.

[0031] This embodiment describes a method for optimizing remote sensing images of farmland plots. In the historical fusion stage, such as... Figure 2 As shown, this approach integrates the graphic and attribute information of scattered land parcels from the previous year to solve the challenges of identifying and classifying these parcels, resulting in an integrated "graphic + attribute" outcome and establishing a historical land parcel fusion mechanism. Furthermore, this stage enhances the ability to integrate the segmentation model with historical insurance data, preventing the omission or merging of scattered small parcels, avoiding the loss of small parcels, and preventing the erroneous merging of adjacent parcels with different attributes. This adapts to the integrated analysis needs of scattered land parcels, especially those with scattered insurance coverage.

[0032] In the coordinate transformation stage, improving the adaptability of traditional coordinate transformation methods and avoiding positional offset and shape distortion can avoid offset in complex terrain and meet the practical requirements of agricultural insurance for the integration of scattered plots and accurate insurance coverage.

[0033] Verification phase: By combining accuracy verification with index screening, the optimization and recognition accuracy of images of crops at maturity in northern regions is improved.

[0034] In summary, this implementation method is a farmland plot boundary identification technology solution that integrates historical insurance data, adapts to crop maturity period images in northern regions, and combines contour optimization with accurate crop identification. It generates integrated data consisting of "vector boundary graphics + attributes (including crop type)," consolidating scattered insured plots into continuous / related plot units. It can directly output a basic layer adapted for agricultural insurance underwriting and loss assessment, achieving a closed loop from existing data processing to practical application, and generating a dual-precision insurance basic layer.

[0035] Implementation Method Two: This implementation method further defines the farmland remote sensing image optimization method described in Implementation Method One. In this implementation method, the threshold screening in the preliminary screening stage is as follows: The following processing is performed on each pixel in the remote sensing image of the farmland plot to be optimized to obtain several northern crop pixels as preliminary crop types: If the Normalized Difference Vegetation Index (NDVI) of the pixel is within the NDVI threshold range, the Enhanced Vegetation Index (EVI) is within the Enhanced Vegetation Index threshold range, and the Soil Adjusted Vegetation Index (SAVI) is within the Soil Adjusted Vegetation Index threshold range, then the pixel is a northern crop pixel; otherwise, it is a non-northern crop pixel.

[0036] In this embodiment, the NDVI threshold range is: The threshold range for the enhanced vegetation index is: The threshold range for the soil-adjusted vegetation index is: .

[0037] RVI (Ratio Vegetation Index) is used to assist in verifying the reliability of NDVI results. It serves as a redundant check factor in the initial screening stage. When NDVI is at a critical value, RVI>2 is used to determine effective crop pixels, thus improving screening robustness. Since the core screening has already been completed by NDVI / EVI / SAVI, it is not used as a main process step.

[0038] The aforementioned threshold range is a customized range for the maturity period of northern crops. The design process involves: collecting images of the main northern crops at maturity and statistically analyzing the distribution of NDVI / EVI / SAVI; removing index values ​​from non-crop areas such as weeds, woodlands, and bare land to determine the high-confidence range for crops; and validating the combination with 2000 sets of samples. This combination can accurately distinguish between mature northern crops and non-crop areas, solving the problem of mixing crops with similar spectra.

[0039] This implementation method further defines the preliminary screening stage and explains the threshold screening. This implementation method optimizes the image characteristics of the maturity period of the main crops in the north, combines multiple image indices to initially screen crops, improves the identification targeting, realizes the integration of maturity period adaptation of northern crops and index screening, screens northern crop types / target crop areas, and excludes non-target crops (such as weeds and woodlands).

[0040] Implementation Method 3: This implementation method further defines the farmland plot remote sensing image optimization method described in Implementation Method 1. In this implementation method, the multimodal optimization stage further includes optimizing the contours of the several segmented farmland plot remote sensing images to obtain several corresponding smooth farmland plot remote sensing images, which are then used as several segmented farmland plot remote sensing images.

[0041] In this embodiment, the contour smoothing optimization involves vectorizing the mask of the segmented farmland plot remote sensing image, correcting the contour using a Bezier curve smoothing algorithm, removing redundant vertices, and ensuring that the smoothness of the corrected contour is ≥0.9.

[0042] The smoothness threshold of the Bezier curve smoothing algorithm is [0.85, 0.95], preferably 0.92.

[0043] In this embodiment, the contour smoothing optimization preferably integrates the SAM3D model to correct the boundary offset of complex terrains such as terraces and slopes.

[0044] This embodiment further defines the multimodal optimization stage and explains the multimodal optimization stage. This embodiment addresses the problems of jagged edges and redundant vertices that easily occur in the segmented contours after multimodal fusion segmentation model by performing targeted smoothing processing.

[0045] Implementation Method Four: This implementation method further defines the farmland remote sensing image optimization method described in Implementation Method One. In this implementation method, the multimodal fusion segmentation model is obtained through the following method during the multimodal optimization stage: In the dataset construction phase, a dataset is constructed based on remote sensing images of all farmland plots from the previous year and their corresponding attribute information. During the training phase, based on the dataset, all remote sensing images of farmland plots from the previous year are used as input, and several segmented farmland plot remote sensing images and corresponding image recognition attribute information are used as output. The SAM3 model is trained using the joint loss function and the total loss function to obtain the trained SAM3 model as a multimodal fusion segmentation model. The joint loss function includes the overlap loss function, the boundary distance loss function, and the intersection-union ratio loss function; The total loss function includes a classification loss function and a smoothness constraint loss function.

[0046] In this embodiment, the joint loss function is:

[0047] in, Weights for overlap loss. Weights are assigned to the boundary distance loss. To calculate the loss weights for intersection-union comparison, This is a mask for remote sensing images of segmented farmland plots obtained from a multimodal fusion segmentation model. To accurately label the mask, For overlap loss, For boundary distance loss, To compare the losses.

[0048] Overlap Loss Weight The boundary distance loss weight is 0.4. The crossover and union ratio loss weight is 0.35. It is 0.25.

[0049] The joint loss function is used to optimize segmentation accuracy and boundary fit. The overlap loss weight, boundary distance loss weight, and intersection-union ratio loss weight are determined by cross-validation on the validation set (the optimal weight combination when the validation set loss converges).

[0050] In this embodiment, the overlap loss is

[0051] in, Let P be the sum of the element-wise products of G. Let ε be the L1 norm of P and G, and ε = 1e -6 (Avoid denominators of 0).

[0052] This overlap loss is used to optimize segmentation accuracy and boundary fit.

[0053] In this embodiment, the boundary distance loss is

[0054] in, for The number of pixels, This is the set of pixel boundaries for farmland plots from the previous year's remote sensing imagery. For Euclidean distance, for The number of pixels, This is the set of pixel boundaries for the segmented farmland plots in the remote sensing image obtained by the multimodal fusion segmentation model.

[0055] This boundary distance loss is achieved by summing the bidirectional minimum distances to ensure that the predicted boundary closely matches the true boundary, thus optimizing boundary accuracy.

[0056] In this embodiment, the crossover-union ratio loss is

[0057] The intersection-union ratio loss is used to optimize the overall regional matching degree.

[0058] In this embodiment, the total loss function is:

[0059] in, Weights for smoothness loss, For classification loss weights, For smoothness loss, For classification loss.

[0060] The smoothness loss weight The classification loss weight is 0.15. The value is 0.2, determined iteratively through loss balancing between the training and validation sets.

[0061] The total loss function is used to guide contour adjustment during model training.

[0062] In this embodiment, the smoothness loss is

[0063] in, The target smoothness threshold, For smoothness.

[0064] The target smoothness threshold The preferred value is 0.9.

[0065] The smoothness loss, when Smoothness < When the loss is positive, backpropagation drives the model to optimize the distribution of contour vertices, thereby improving smoothness.

[0066] The smoothness is

[0067] in, The number of vertices of the outline. For the first The tangent angle at each vertex.

[0068] The unit of the tangent angle is radians.

[0069] The difference in tangent angle between the first and last vertices (|θ) n - θ1|) ensures the closure of the contour; the denominator 2n×π is the sum of the maximum possible differences of the tangent angles (the maximum difference of π between each adjacent vertex, n adjacent pairs + 1 head-tail pair), Smoothness∈[0,1], the closer to 1, the smoother the contour.

[0070] In this embodiment, the classification loss is

[0071] in, The number of attribute information categories (determined based on insurance attributes, including crop type, insurance level, land ownership, etc., K=12); The tag is a one-hot encoded tag (the k-th attribute is 1, and the rest are 0); Predict the probability of the k-th class attribute for the model (by activating the output with Softmax).

[0072] The classification loss is used to guide contour adjustment during model training.

[0073] In this embodiment, the multimodal fusion segmentation model further includes a joint injection module for "image index-texture-insurance attribute", which is compressed to 1024 dimensions through 1×1 convolution and dynamically allocated weights by gating attention (crop classification scene index weight 0.5-0.6, boundary segmentation scene texture weight 0.3-0.4, and insurance attribute association scene attribute weight 0.3-0.5).

[0074] In this embodiment, the multimodal fusion segmentation model also combines "semantic prompts (such as "scattered insured plots of corn at maturity" and "insured fields of soybeans") + spatial prompts (crop areas screened based on image indices, with prompt points generated at a density of 0.02 pixels / m²)" and introduces an ambiguity module to calibrate ambiguous concepts such as "scattered insured plots" and "small insured fields".

[0075] In this embodiment, the remote sensing images of farmland plots cover 12 major northern crops such as corn, soybeans, and wheat, including strip fields, scattered small plots, and mixed crop scenes, and contain 5,000 low-resolution images.

[0076] In this embodiment, the remote sensing image of the farmland plot and the attribute information, such as the boundary coordinates, crop type, and insured area of ​​the insured plot, are accurately associated with the image sample to achieve a three-in-one labeling of "image-plot-attribute".

[0077] In this embodiment, the remote sensing images of the farmland plots and the attribute information are optimized and labeled in a multi-dimensional manner, including "crop type + maturity period + plot dispersion + insurance attribute", and exclusive labels such as "dispersed insured plots" and "small insured plots" are added to adapt to the SAM3 concept prompting mechanism.

[0078] In this implementation, the dataset construction stage and retrograde data enhancement strategies include rotation (0-360°), scaling (0.8-1.2 times), texture perturbation, semantic consistency enhancement (maintaining the matching of crop type and image index value), and insurance attribute correlation enhancement (ensuring that the plot boundary and attribute label correspond one-to-one), to improve the model's adaptability and generalization ability to existing data.

[0079] In this embodiment, the training and segmentation process of the multimodal fusion segmentation model is as follows: Initialize the SAM3 model weights (using pre-trained weights for fine-tuning); Training parameters: Batch size = 8, learning rate = 1e -5 (AdamW optimizer), iterations = 100 rounds, early stopping strategy (training stops if the validation set loss does not decrease for 5 consecutive rounds); Substitute the joint training formula and update the model parameters through backpropagation to simultaneously optimize segmentation accuracy, contour smoothness and crop classification accuracy. Input the preprocessed data to generate preliminary boundary masks as remote sensing images of several segmented farmland plots.

[0080] This implementation further defines the multimodal optimization stage and describes the multimodal fusion segmentation model. In this implementation, SAM3 (SegmentAnythingModel3) is selected as the core segmentation model, and the following methods are used to optimize and adapt to the scene requirements.

[0081] In the dataset construction phase, a dedicated dataset is built based on the previous year's mature-period images and historical plot data to achieve integrated "image-plot-attribute" annotation, improving the model's scene adaptability and achieving deep adaptation to existing data. In the training phase, segmentation, boundary, smoothing, and classification losses are integrated to form targeted training objectives, taking into account boundary accuracy, contour quality, and crop classification accuracy, and realizing a multi-dimensional joint training formula system.

[0082] During the training phase, a joint loss function and a total loss function are introduced. The joint loss function simultaneously optimizes overlap, boundary distance, and intersection-union ratio to solve the problems of blurred plot edges and internal voids. The smoothness loss function forces the contour smoothness to be ≥0.9 to eliminate jaggedness. The classification loss function covers all attribute outputs to solve the problem of spectral mixing of crops in the northern mature period. The total loss function achieves four-dimensional optimization of "precise segmentation + clear boundaries + smooth contours + accurate classification" to adapt to the "discontinuous, multi-segmented, and scattered" characteristics of farmland plots.

[0083] Implementation Method 5: This implementation method further defines the farmland plot remote sensing image optimization method described in Implementation Method 1. In this implementation method, during the verification stage, the four-layer verification system includes a boundary verification layer, a semantic verification layer, a fusion verification layer, and a coordinate verification layer. The boundary verification layer is used to verify the farmland boundaries in the corrected remote sensing image of the farmland plot. The verification conditions are: the overlap rate between the farmland boundary and the texture feature index abrupt change point of the remote sensing image of the farmland plot to be optimized is ≥95%, and the overlap rate between the farmland boundary and the farmland boundary of the remote sensing image of the farmland plot in the previous year is ≥92%. The semantic verification layer is used to verify the matching degree between the crop type in the fused attribute information and the preliminary crop type. If the matching degree is ≥85%, the verification passes. The fusion verification layer is used to verify the matching degree between the fused attribute information and the previous year's attribute information. If the matching degree is ≥90%, the verification passes. The coordinate verification layer is used to verify the error between the corrected land parcel remote sensing image and the fused land parcel remote sensing image. If the error is ≤0.2 meters, the verification is passed.

[0084] In this embodiment, in the boundary verification layer, the abrupt change in the texture feature index is more than 50%.

[0085] In this embodiment, the texture feature index is obtained by extracting texture features from the remote sensing image of the farmland plot to be optimized using GLCM. The texture feature index includes contrast, entropy, and energy.

[0086] Implementation Method Six: The farmland remote sensing image optimization method used in this implementation method is based on the farmland remote sensing image optimization method described in Implementation Method One, combined with the farmland remote sensing image optimization methods optimized in Implementation Methods Two to Five.

[0087] In this embodiment, the following dataset was used to test this embodiment. (1) Existing images of the previous year's northern crop maturity period: Selected remote sensing images of maize (late September), soybean (early October), and wheat (mid-July) maturity periods in Heilongjiang, Jilin, and Inner Mongolia. Figure 2 10,000 images, 0.5 meters resolution; (2) Existing data on insured land parcels from the previous year: There are 12,000 scattered insured land parcels in the above-mentioned areas, including crop type, insured area, ownership information and preliminary boundary coordinates (WGS84 coordinate system). (3) Data labeling: Based on existing images and insured land data, crop type, maturity period, land dispersion and insurance attribute labels are labeled to form a training set (16,000 images), a validation set (2,000 images) and a test set (2,000 images).

[0088] This implementation method was tested using the hardware environment of the following implementation environment. The CPU is an Intel Core i9-13900K, the GPU is an NVIDIA RTX 4090, and the memory is 64GB. The software environment is as follows: The operating system is Ubuntu 22.04, the deep learning framework is PyTorch 2.0, the programming language is Python 3.9, and the GIS tool is ArcGIS Pro 3.1.

[0089] The experimental results of this implementation method are as follows: The accuracy rate of crop identification in northern regions was 93.5%, and the overlap rate between the boundary demarcation and the existing insured land plots from the previous year was ≥96%. The outline smoothness is 0.93, the fusion matching rate of scattered plots is 96%, and the area calculation error is ≤0.8%. The model training converged and stabilized, and the validation lumped loss tended to plateau after 30 rounds (final loss value ≤ 0.12). The processing time for a single image is ≤3 seconds, and the generated base layer fully meets the practical needs of accurate underwriting and loss assessment in agricultural insurance.

Claims

1. A method for optimizing remote sensing images of farmland plots, characterized in that, Includes the following stages: Data collection phase: Collect remote sensing images of farmland plots to be optimized, remote sensing images of several farmland plots from the previous year, and corresponding attribute information from the previous year; Preliminary screening stage: Threshold screening is performed on the remote sensing images of the farmland plots to be optimized to obtain preliminary crop types; Multimodal optimization stage: The remote sensing images of the farmland plots to be optimized are used as optimization objects. The multimodal fusion segmentation model is used to perform fusion segmentation to obtain several segmented farmland plot remote sensing images and corresponding image recognition attribute information after multimodal fusion segmentation. Historical fusion stage: Spatial matching is performed on the remote sensing images of the several segmented farmland plots and the remote sensing images of several farmland plots from the previous year to obtain several related plot remote sensing images, which are then combined into a fused plot remote sensing image; The image identification attribute information and the attribute information from the previous year are fused to obtain fused attribute information; Coordinate transformation stage: The fused land parcel remote sensing images are transformed using a coordinate transformation model to obtain the corresponding corrected land parcel remote sensing images; Verification phase: If the corrected remote sensing image of the plot, the fused attribute information, and the preliminary crop type meet the four-layer verification system, then proceed to the optimization phase; Otherwise, the corrected land parcel remote sensing image will be used as the optimization object in the multimodal optimization stage, and the multimodal optimization stage will begin. Optimization phase: The fused attribute information and the preliminary crop type are labeled onto the corrected plot remote sensing image to obtain the optimized farmland plot remote sensing image; During the verification phase, the four-layer verification system includes a boundary verification layer, a semantic verification layer, a fusion verification layer, and a coordinate verification layer; The boundary verification layer is used to verify the farmland boundaries in the corrected land parcel remote sensing image. The verification conditions are: the overlap rate between the farmland boundary and the texture feature index abrupt change point of the farmland parcel remote sensing image to be optimized is ≥95%, and the overlap rate between the farmland boundary and the farmland boundary of the previous year's farmland parcel remote sensing image is ≥92%. The semantic verification layer is used to verify the matching degree between the crop type in the fused attribute information and the preliminary crop type. If the matching degree is ≥85%, the verification passes. The fusion verification layer is used to verify the matching degree between the fused attribute information and the previous year's attribute information. If the matching degree is ≥90%, the verification passes. The coordinate verification layer is used to verify the error between the corrected land parcel remote sensing image and the fused land parcel remote sensing image. If the error is ≤0.2 meters, the verification is passed.

2. The method for optimizing remote sensing images of farmland plots according to claim 1, characterized in that, The threshold screening mentioned in the initial screening stage is as follows: The following processing is performed on each pixel in the remote sensing image of the farmland plot to be optimized to obtain several northern crop pixels as preliminary crop types: If the NDVI of the pixel is within the NDVI threshold range, the Enhanced Vegetation Index (EVI) is within the Enhanced Vegetation Index threshold range, and the Soil Adjusted Vegetation Index (SAVI) is within the Soil Adjusted Vegetation Index threshold range, then the pixel is a northern crop pixel; otherwise, it is a non-northern crop pixel.

3. The method for optimizing remote sensing images of farmland plots according to claim 1, characterized in that, The multimodal optimization stage also includes optimizing the contours of the several segmented farmland remote sensing images to obtain several smooth farmland remote sensing images, which are then used as several segmented farmland remote sensing images.

4. A remote sensing image optimization system for farmland plots, characterized in that, Includes the following modules: Module 1 is used to collect remote sensing images of farmland plots to be optimized, remote sensing images of several farmland plots from the previous year, and corresponding attribute information from the previous year. Module 2 is used to perform threshold screening on the remote sensing images of the farmland plots to be optimized to obtain preliminary crop types; Module 3 is used to take the remote sensing image of the farmland plot to be optimized as the optimization object, and perform fusion segmentation through a multimodal fusion segmentation model to obtain several segmented farmland plot remote sensing images and corresponding image recognition attribute information after multimodal fusion segmentation. Module 4 is used to spatially match the remote sensing images of the several segmented farmland plots with the remote sensing images of several farmland plots from the previous year to obtain several related plot remote sensing images, and combine them into a fused plot remote sensing image; and to fuse the image recognition attribute information and the attribute information from the previous year to obtain fused attribute information. Module 5 is used to obtain the corresponding corrected remote sensing image of the land parcel through a coordinate transformation model; Module 6 is used to proceed to Module 7 if the corrected land parcel remote sensing image, the fused attribute information, and the preliminary crop type satisfy the four-layer verification system; otherwise, the corrected land parcel remote sensing image is used as the optimization object of Module 3 and proceeds to Module 3. Module 7 is used to label the fused attribute information and the preliminary crop type onto the corrected plot remote sensing image to obtain an optimized farmland plot remote sensing image; In module six, the four-layer verification system includes a boundary verification layer, a semantic verification layer, a fusion verification layer, and a coordinate verification layer; The boundary verification layer is used to verify the farmland boundaries in the corrected land parcel remote sensing image. The verification conditions are: the overlap rate between the farmland boundary and the texture feature index abrupt change point of the farmland parcel remote sensing image to be optimized is ≥95%, and the overlap rate between the farmland boundary and the farmland boundary of the previous year's farmland parcel remote sensing image is ≥92%. The semantic verification layer is used to verify the matching degree between the crop type in the fused attribute information and the preliminary crop type. If the matching degree is ≥85%, the verification passes. The fusion verification layer is used to verify the matching degree between the fused attribute information and the previous year's attribute information. If the matching degree is ≥90%, the verification passes. The coordinate verification layer is used to verify the error between the corrected land parcel remote sensing image and the fused land parcel remote sensing image. If the error is ≤0.2 meters, the verification is passed.

5. An electronic device, characterized in that, It includes a processor, a communication interface, a memory, and a communication bus, wherein the processor, the communication interface, and the memory communicate with each other through the communication bus; Memory, used to store computer programs; A processor, when executing a computer program stored in memory, implements the method of any one of claims 1-3.

6. A computer-readable storage medium, characterized in that, The computer-readable storage medium stores a computer program that, when executed by a processor, implements the method of any one of claims 1-3.

Images