Grain machinery operation area accurate extraction method based on visual remote sensing and AI

Abstract

The invention relates to the technical field of agricultural engineering, and discloses a grain machinery operation area accurate extraction method based on visual remote sensing and AI, and the method comprises the following steps: constructing a multi-source spatio-temporal data set of land parcel level alignment; reconstructing a normalized vegetation index time sequence for each plot by adopting an interpolation method and a filtering method, and generating a plot scale crop growth curve; generating a hourly improved vegetation index image through calculation, and extracting texture features in combination with a crop growth curve; constructing a change atlas based on the normalized vegetation index time sequence; inputting the remote sensing image and the crop growth curve into a semantic segmentation model; performing fusion classification on the texture features and the change atlas by using a random forest algorithm; and converting the probability graph and a binary mask result into a vector boundary, and completing precision verification by calculating an intersection-to-union ratio. Accurate data support is provided for agricultural production monitoring, agricultural machinery scheduling and the like.

Description

Technical Field

[0001] This application relates to the technical field of agricultural engineering, and in particular to a method for precise extraction of grain machinery operation areas based on visual remote sensing and AI. Background Technology

[0002] Grain machinery refers to specialized mechanical equipment used in grain production, processing, storage, transportation, and testing, covering the entire industrial chain from field harvesting to food processing. Originating from the demand for agricultural mechanization and food security, grain machinery has gradually evolved from simple threshing and rice milling equipment to intelligent and efficient complete sets of equipment with the development of industrial revolution and modern agriculture, involving key technologies such as cleaning, sorting, drying, crushing, and mixing.

[0003] The precise extraction of grain machinery operation areas stems from the modern agricultural demand for intelligent and refined operations. It aims to identify the scope of machinery operations through spatial information technology and optimize path planning and resource scheduling.

[0004] Existing technologies rely solely on the spectral features of a single image, making it difficult to effectively distinguish between mutations caused by vegetation harvesting and fluctuations caused by other disturbances. They fail to fully utilize the temporal patterns of the complete crop growth cycle to accurately locate key agricultural events such as harvesting, which can easily lead to misjudgments and omissions. Furthermore, simple classification methods are difficult to adapt to plot heterogeneity, sensor differences, and environmental noise, resulting not only in fragmented extraction results and blurred boundaries but also in the inability to conduct effective cross-validation. Summary of the Invention

[0005] To address the problem of misjudgment and omission caused by failing to fully utilize the temporal patterns of the complete crop growth cycle to accurately locate key agricultural events such as harvesting, this application provides a method for accurately extracting the operating area of ​​grain machinery based on visual remote sensing and AI.

[0006] This application provides a method for accurately extracting the operating area of ​​grain machinery based on visual remote sensing and AI. The method for accurately extracting the operating area of ​​grain machinery includes the following steps: S1. Collect remote sensing images covering the entire crop growth cycle, integrate agricultural machinery trajectory data with cultivated land plot vector boundaries, and construct a plot-level aligned multi-source spatiotemporal dataset; S2. In the multi-source spatiotemporal dataset, the remote sensing images are preprocessed, and combined with the cultivated land vector boundary, the normalized vegetation index time series is reconstructed for each plot using interpolation and filtering methods to generate plot-scale crop growth curves. S3. In remote sensing images, improve vegetation index images are generated by calculating each time phase, and texture features are extracted by combining crop growth curves. S4. Construct change maps based on normalized vegetation index time series and identify areas of abrupt changes in the land surface caused by harvesting; S5. Input the preprocessed remote sensing image and the extracted crop growth curve into the semantic segmentation model to perform end-to-end segmentation and output a binary mask of the mechanical operation area; at the same time, use the random forest algorithm to fuse and classify the texture features and change map to generate a probability map. S6. Convert the probability map and binary mask results into vector boundaries, overlay and compare them with agricultural machinery trajectory data, and complete the accuracy verification by calculating the intersection-union ratio.

[0007] Optionally, in a multi-source spatiotemporal dataset, the improved vegetation index imagery is preprocessed using remote sensing images. Combined with cultivated land vector boundaries, interpolation and filtering methods are employed to reconstruct the normalized vegetation index time series for each plot, generating plot-scale crop growth curves. This includes the following steps: S21. Perform radiometric calibration and atmospheric correction on remote sensing images in multi-source spatiotemporal datasets to obtain images of true surface reflectance. S22. Based on the digital elevation model, geometric fine correction and multi-temporal image registration are performed on the true surface reflectance image to obtain a spatially accurate multi-temporal image set. S23. Using the quality assessment band, construct cloud masks from spatially precisely aligned multi-temporal image sets, and reconstruct the normalized vegetation index time series by combining linear interpolation and smoothing filtering methods. S24. Based on the reconstructed normalized vegetation index time series, generate a complete crop growth curve for each plot.

[0008] Optionally, the improved vegetation index imagery is generated from remote sensing imagery by calculating and generating time-phase improved vegetation index images, and then combining them with crop growth curves to extract texture features, including the following steps: S31. Extract the reflectance of red light, blue light and near-infrared bands from remote sensing images, and generate an improved vegetation index image for each time phase by calculation. S32. Analyze each plot to generate a complete crop growth curve and determine the critical period for crop harvesting; S33. Based on the critical period of crop harvesting, texture features are calculated from the improved vegetation index image of the corresponding time phase using the gray-level co-occurrence matrix algorithm.

[0009] Optionally, the improved vegetation index imagery is used to construct change maps based on normalized vegetation index time series, and the identification of areas of abrupt changes in the land surface caused by harvesting includes the following steps: S41. For the normalized vegetation index time series of each plot, the change point detection algorithm is used to analyze and obtain the change point detection results; at the same time, the normalized vegetation index images of two key time phases, namely the peak crop growth season and the period after harvest, are selected, the difference between the normalized vegetation index values ​​of the two time phases is calculated, and the change map is generated. S42. Based on the change point detection results, identify the abrupt phase in which the normalized vegetation index (NVC) drops sharply in each plot sequence; S43. Map the identified mutation phases to the spatial domain, and mark all pixels that experience a sharp decline near the candidate harvest phase based on the change map to generate preliminary candidate mutation regions. S44. Combine the vector boundaries of cultivated land plots to perform spatial clustering on the preliminary candidate mutation areas, remove broken patches, and confirm them in conjunction with crop growth curves to identify surface mutation areas caused by harvesting.

[0010] Optionally, the improved vegetation index imagery inputs the preprocessed remote sensing imagery and extracted crop growth curves into the semantic segmentation model for end-to-end segmentation, outputting a binary mask of the mechanical operation area, including the following steps: Based on the preprocessed remote sensing images and the corresponding plot-scale crop growth curves, comprehensive information containing spatial spectral and temporal patterns is generated to construct a feature set for identifying the operation area. The feature set of the operation area is input into the semantic segmentation model for training. Through the encoder and decoder structure, the spatial spectral features of the image and the temporal pattern features of crop growth are learned and fused. In the output layer of the semantic segmentation model, each pixel is classified into two categories, and a binary mask with the same spatial resolution as the input image is output.

[0011] Optionally, the improved vegetation index imagery incorporates the feature set of the work area identification into a semantic segmentation model for training. Through an encoder and decoder structure, it learns and fuses the spatial spectral features of the imagery with the temporal patterns of crop growth, including the following steps: The fused feature set of the work area is input into the semantic segmentation model; In the encoder part of the semantic segmentation model, multi-level spatial spectral features are extracted and compressed from the input features through successive convolution and downsampling operations. The deepest features of the multi-level spatial spectral features are fed into the bottleneck layer and further integrated and condensed to form high-level semantic features containing global contextual information. By using skip connections, the high-level semantic features output from the bottleneck layer are fused with the corresponding shallow features in the encoder path; the decoder then upsamples the fused features to gradually restore their spatial details and output a high-resolution feature map. The high-resolution feature map is passed to the output layer, and each pixel is classified through convolution operation to generate a binary mask of the mechanical operation area with the same spatial resolution as the input image.

[0012] Optionally, the improved vegetation index image uses a random forest algorithm to fuse and classify texture features and change maps to generate a probability map, including the following steps: The extracted texture features are fused with the change map generated based on the normalized vegetation index time series to construct a unified multidimensional feature vector for each pixel in the study area and obtain a complete feature dataset. The feature dataset is combined with labeled mechanical operation area samples to train a random forest classifier. The constructed feature dataset is then input into the trained model to classify each pixel as either an operation area or a non-operation area. Based on the classification results, the probability of each pixel belonging to the mechanical operation area is output, generating a complete probability map.

[0013] Optionally, the improved vegetation index imagery transforms the probability map and binary mask results into vector boundaries, which are then overlaid and compared with agricultural machinery trajectory data. Accuracy verification is performed by calculating the intersection-union ratio (IUGR), including the following steps: S61. Perform adaptive threshold segmentation on the probability map and combine morphological operations to remove small noise or discontinuous regions generated during the segmentation process, and optimize the segmentation boundary. S62. Using a raster-to-vector algorithm, the optimized segmentation boundary is converted into boundary data in vector format to generate the vector boundary of the mechanical operation area. S63. Spatial comparison is performed between the vector boundary of the generated mechanical operation area and the agricultural machinery trajectory data. The intersection-union ratio is calculated, and the degree of overlap between the segmentation result and the actual agricultural machinery operation area is evaluated. S64. Based on the evaluation results of the intersection-union ratio, complete the accuracy verification.

[0014] Optionally, the improved vegetation index imagery is converted from optimized segmentation boundaries to vector format boundary data using a raster-to-vector algorithm. The process for generating the vector boundary of the mechanical operation area includes the following steps: S621. The optimized binary mask is used as input grid data, where the value of each pixel represents whether it belongs to the mechanical operation area or the non-operation area. S622. Employ boundary tracing to identify connected regions in raster data and accurately extract the boundary cells of each region; S623. Using a raster-to-vector algorithm, the extracted boundary cell sequence is converted into polygon vector data composed of continuous vertices; S624. Optimize the generated polygon vector boundary, remove noisy polygons with too small area, smooth the edges to remove the grid jagged effect, and form an accurate vector boundary for the mechanical operation area.

[0015] In summary, this application includes at least one of the following beneficial technical effects: 1. This application integrates multi-source spatiotemporal data and, through steps such as remote sensing preprocessing, vegetation index temporal reconstruction, feature extraction, model classification, and accuracy verification, can accurately extract the operating area of ​​grain machinery. It effectively integrates spatial spectral and temporal features, combines semantic segmentation and random forest classification to improve recognition accuracy, and uses cross-validation to ensure the reliability of the results, providing accurate data support for agricultural production monitoring and agricultural machinery scheduling.

[0016] 2. This application ensures data accuracy by preprocessing remote sensing images such as radiometric calibration and atmospheric correction, and accurately captures crop growth patterns by reconstructing vegetation index time series and crop growth curves based on cultivated land boundaries. At the same time, by generating improved vegetation indices and extracting key period texture features, and by combining change point detection and change map to locate harvest change areas, it integrates the advantages of multi-source data and achieves performance optimization from data processing to feature extraction to target area identification.

[0017] 3. This application achieves high-precision and automated extraction of mechanical operation areas by combining multi-source data fusion and temporal feature analysis with a dual verification mechanism of deep learning and artificial intelligence, effectively improving the anti-interference capability and spatial positioning accuracy of remote sensing identification in complex agricultural environments. Attached Figure Description

[0018] Figure 1 This is a flowchart of the method in this application. Detailed Implementation

[0019] The embodiments of this application are described in detail below, and examples of the embodiments are shown in the accompanying drawings.

[0020] In the description of this specification, the references to "certain embodiments," "one embodiment," "some embodiments," "illustrative embodiment," "example," "specific example," or "some examples" refer to specific features, structures, materials, or characteristics described in connection with the described embodiment or example, which are included in at least one embodiment or example of this application. In this specification, the illustrative expressions of the above terms do not necessarily refer to the same embodiment or example. Furthermore, the specific features, structures, materials, or characteristics described may be combined in any suitable manner in one or more embodiments or examples.

[0021] This application discloses a method for accurately extracting the operating area of ​​grain machinery based on visual remote sensing and AI, referring to... Figure 1 The precise extraction method for grain machinery operation areas includes the following steps: S1. Collect remote sensing images covering the entire crop growth cycle, integrate agricultural machinery trajectory data with cultivated land plot vector boundaries, and construct a plot-level aligned multi-source spatiotemporal dataset.

[0022] It should be explained that acquiring image data of crops from sowing to harvest through remote sensing technology enables the monitoring of crop growth status and environmental changes. This image data provides crucial spatiotemporal information about the crop growth process for subsequent analysis. Agricultural machinery trajectory data records the operating paths of agricultural machinery, helping to identify the operating areas and times, and understanding the relationship between machinery operation and crop growth status. The vector boundaries of cultivated land plots are spatial data describing the extent of farmland, accurately identifying the crop status and machinery operating areas of each plot. By aligning the above data (remote sensing images, agricultural machinery trajectories, and cultivated land plot boundaries) spatially and temporally according to farmland plots, a comprehensive spatiotemporal dataset is formed, providing a reliable foundation for subsequent crop monitoring, agricultural machinery operation optimization, and other analyses.

[0023] S2. In the multi-source spatiotemporal dataset, remote sensing images are preprocessed, and combined with cultivated land vector boundaries, interpolation and filtering methods are used to reconstruct the normalized vegetation index time series for each plot, generating crop growth curves at the plot scale.

[0024] Preferably, in a multi-source spatiotemporal dataset, the remote sensing images are preprocessed, and combined with the cultivated land vector boundary, interpolation and filtering methods are used to reconstruct the normalized vegetation index time series for each plot, generating plot-scale crop growth curves, including the following steps: S21. Perform radiometric calibration and atmospheric correction on remote sensing images in multi-source spatiotemporal datasets to obtain images of true surface reflectance. S22. Based on the digital elevation model, geometric fine correction and multi-temporal image registration are performed on the true surface reflectance image to obtain a spatially accurate multi-temporal image set. S23. Using the quality assessment band, construct cloud masks from spatially precisely aligned multi-temporal image sets, and reconstruct the normalized vegetation index time series by combining linear interpolation and smoothing filtering methods. S24. Based on the reconstructed normalized vegetation index time series, generate a complete crop growth curve for each plot.

[0025] It should be explained that remote sensing images may be affected by factors such as the sensor itself, the acquisition angle, and lighting during the acquisition process. Therefore, radiometric calibration is necessary to convert the digital values ​​of the original image into radiometric values ​​or reflectance. This can eliminate the influence of external conditions on remote sensing images and make them closer to the true radiometric characteristics of the Earth's surface. Remote sensing images are affected by scattering and absorption effects when passing through the atmosphere. The purpose of atmospheric correction is to remove atmospheric interference and make the image reflect the actual situation of ground reflection more closely. Through atmospheric correction, the true reflectance image of the Earth's surface can be obtained.

[0026] A digital elevation model (DEM) is a three-dimensional model that describes the elevation information of the Earth's surface. Through DEM, we can accurately understand the changes in the undulation of the Earth's surface. Remote sensing images acquired over time may have spatial differences due to factors such as different shooting angles and the movement of ground features. Registration is the process of spatially aligning images taken at different times to ensure that the position of the same location remains consistent in different images, thus ensuring that images from different time points can be directly compared and analyzed. Remote sensing images typically contain quality assessment bands (such as information on clouds and shadows). These bands help identify noise or cloud cover in the images, thereby filtering out valid data. Since clouds can significantly affect the analysis results in remote sensing images, it is necessary to construct cloud masks using quality assessment bands to remove interference from cloud areas. Linear interpolation and smoothing filtering methods are commonly used to handle missing values ​​in Normalized Difference Vegetation Index (NDVI) time series. Linear interpolation predicts missing values ​​using known data from previous and subsequent time points, while smoothing filtering is used to smooth noise in the time series and remove irregular fluctuations.

[0027] Normalized Difference Vegetation Index (NDVI) is a vegetation index calculated from remote sensing imagery to reflect vegetation cover and health status. By applying cloud masking, interpolation, and filtering to time-series images, a continuous and smooth NDVI time series can be reconstructed for subsequent crop growth analysis. Crop growth curves reflect the growth status of crops at different time points during the growth process. By using the NDVI time series, crop growth changes can be tracked, thereby generating crop growth curves for each plot.

[0028] S3. In remote sensing images, improve vegetation index images are generated by calculating each time phase, and texture features are extracted by combining crop growth curves.

[0029] Preferably, in remote sensing imagery, generating a time-phase improved vegetation index image by calculation and extracting texture features by combining it with crop growth curves includes the following steps: S31. Extract the reflectance of red light, blue light and near-infrared bands from remote sensing images, and generate an improved vegetation index image for each time phase by calculation. S32. Analyze each plot to generate a complete crop growth curve and determine the critical period for crop harvesting; S33. Based on the critical period of crop harvesting, texture features are calculated from the improved vegetation index image of the corresponding time phase using the gray-level co-occurrence matrix algorithm.

[0030] It should be explained that remote sensing images typically contain multiple bands, among which red, blue, and near-infrared reflectance (NIR) are key bands used for vegetation index calculation. By extracting the reflectance of these bands, an improved vegetation index can be calculated. The expression for the improved vegetation index is: In the formula, Indicates the improved vegetation index; This represents the gain factor, which is usually taken as 2.5; This represents the adjustment factor, which is usually set to 1; and This represents the correction factor, which is usually taken as 6 or 7.5.

[0031] Based on improved vegetation index image data of each time phase, vegetation changes of each plot can be tracked throughout the entire growth cycle; The improved vegetation index value of each plot changes over time to form a crop growth curve, which usually shows a trend of "growth - vigorous growth - decline". Gray-level co-occurrence matrix (GLCM) is a commonly used method for calculating texture features. It extracts structural information by analyzing the spatial combination frequency of gray values ​​in an image.

[0032] S4. Construct change maps based on the normalized vegetation index time series and identify areas of abrupt changes in the land surface caused by harvesting.

[0033] Preferably, constructing a change map based on the normalized vegetation index time series and identifying areas of abrupt changes in the land surface caused by harvesting includes the following steps: S41. For the normalized vegetation index time series of each plot, the change point detection algorithm is used to analyze and obtain the change point detection results; at the same time, the normalized vegetation index images of two key time phases, namely the peak crop growth season and the period after harvest, are selected, the difference between the normalized vegetation index values ​​of the two time phases is calculated, and the change map is generated. S42. Based on the change point detection results, identify the abrupt phase in which the normalized vegetation index (NVC) drops sharply in each plot sequence; S43. Map the identified mutation phases to the spatial domain, and mark all pixels that experience a sharp decline near the candidate harvest phase based on the change map to generate preliminary candidate mutation regions. S44. Combine the vector boundaries of cultivated land plots to perform spatial clustering on the preliminary candidate mutation areas, remove broken patches, and confirm them in conjunction with crop growth curves to identify surface mutation areas caused by harvesting.

[0034] It should be explained that change point detection is used to identify points of trend change in a time series. It is often used to analyze whether there are abrupt changes in growth during the crop growth cycle. Change points usually mean environmental changes, the implementation of farmland management measures, or significant changes in crop growth. By selecting normalized vegetation index images of two key time phases, namely the peak crop growth season (usually the period when vegetation is most vigorous) and the period after harvest, the difference in normalized vegetation index values ​​between the two time phases is calculated to generate a change map. This helps to identify areas of change in vegetation growth status. Change point detection algorithms help identify sharp decline points in time series. These decline points usually represent crop harvesting, disasters, or other sudden events. Through change point detection, the phases in which the NDVI value drops sharply can be automatically identified, namely abrupt phases. Abrupt phases reflect changes caused by crop harvesting or other human operations (such as tillage, weeding, etc.). By using the change map (generated from the NDVI difference between two key time phases), pixels (spatial pixels) that change drastically near the abrupt change phase can be marked. All pixel regions that decrease drastically near the abrupt change phase will be aggregated to form preliminary candidate abrupt change regions. Spatial clustering (e.g., using clustering algorithms such as DBSCAN) is performed on the preliminary candidate mutation regions to merge adjacent or spatially close mutation regions into a larger overall region, reducing noise and isolated small regions, and removing regions that are too small in space or do not conform to the actual plot shape. These regions may be due to errors caused by noise, errors, or irregular objects. Combined with the previously generated crop growth curves, it is confirmed whether the mutation regions conform to the natural harvest period or sudden event pattern in the crop growth process. For example, if the growth curve of a region drops sharply at a certain time and coincides with the crop harvest period, the region can be confirmed as a surface mutation region caused by harvesting.

[0035] S5. Input the preprocessed remote sensing image and the extracted crop growth curve into the semantic segmentation model to perform end-to-end segmentation and output a binary mask of the mechanical operation area; at the same time, use the random forest algorithm to fuse and classify the texture features and change map to generate a probability map.

[0036] Preferably, the process of inputting the preprocessed remote sensing image and the extracted crop growth curve into the semantic segmentation model for end-to-end segmentation and outputting a binary mask of the mechanical operation area includes the following steps: Based on the preprocessed remote sensing images and the corresponding plot-scale crop growth curves, comprehensive information containing spatial spectral and temporal patterns is generated to construct a feature set for identifying the operation area. The feature set of the operation area is input into the semantic segmentation model for training. Through the encoder and decoder structure, the spatial spectral features of the image and the temporal pattern features of crop growth are learned and fused. In the output layer of the semantic segmentation model, each pixel is classified into two categories, and a binary mask with the same spatial resolution as the input image is output.

[0037] Preferably, the feature set of the work area is input into the semantic segmentation model for training. Through the encoder and decoder structure, the spatial spectral features of the image and the temporal pattern features of crop growth are learned and fused, including the following steps: The fused feature set of the work area is input into the semantic segmentation model; In the encoder part of the semantic segmentation model, multi-level spatial spectral features are extracted and compressed from the input features through successive convolution and downsampling operations. The deepest features of the multi-level spatial spectral features are fed into the bottleneck layer and further integrated and condensed to form high-level semantic features containing global contextual information. By using skip connections, the high-level semantic features output from the bottleneck layer are fused with the corresponding shallow features in the encoder path; the decoder then upsamples the fused features to gradually restore their spatial details and output a high-resolution feature map. The high-resolution feature map is passed to the output layer, and each pixel is classified through convolution operation to generate a binary mask of the mechanical operation area with the same spatial resolution as the input image.

[0038] It should be explained that the spatial spectral features of remote sensing images are fused with the temporal variation features of crop growth curves to construct a unified high-dimensional input feature. Feature fusion methods include: stitching, weighted averaging, feature attention mechanisms, etc.

[0039] Encoder section: Extracts deep spatial semantic features.

[0040] The specific steps are as follows: 1. Continuous convolution: Extracts local features of an image; 2. Downsampling: Reduce spatial resolution, expand the receptive field, and preserve important features; For a single convolution operation: In the formula, Indicates the first l Feature map of the layer; Indicates the first l Layer convolution kernel; This represents the feature map of the previous layer; Indicates the first l Layer bias; Indicates the activation function; The bottleneck layer is located between the encoder and decoder and is used to further compress and integrate features to extract high-level semantic representations with global context. Decoder section: Gradually restores the spatial resolution of the image and outputs fine boundary information.

[0041] The specific steps are as follows: 1. Upsampling: Restore the spatial dimensions of the feature map (this can be achieved using methods such as deconvolution or bilinear interpolation); 2. Skip connection: Fuse shallow detail features of the encoder with current layer features of the decoder to enhance spatial boundary recognition capability; Output: One category per pixel (job area = 1, non-job area = 0).

[0042] Use 1×1 convolution for final classification: This represents the probability that pixel x belongs to the machine operation area; This represents the feature values ​​that were not activated after the output layer convolution; Represents input features; It represents the base of the natural logarithm.

[0043] The probability that pixel x belongs to the mechanical work area The binary mask is obtained by comparing it with a threshold (e.g., 0.5). The final output is a mask image with the same spatial resolution as the input remote sensing image; In the mask image, each pixel has a value of 1 indicating a mechanical operation area and a value of 0 indicating a non-operation area.

[0044] Preferably, the process of using a random forest algorithm to fuse and classify texture features and variation maps to generate a probability map includes the following steps: The extracted texture features are fused with the change map generated based on the normalized vegetation index time series to construct a unified multidimensional feature vector for each pixel in the study area and obtain a complete feature dataset. The feature dataset is combined with labeled mechanical operation area samples to train a random forest classifier. The constructed feature dataset is then input into the trained model to classify each pixel as either an operation area or a non-operation area. Based on the classification results, the probability of each pixel belonging to the mechanical operation area is output, generating a complete probability map.

[0045] It needs to be explained that the process involves extracting texture features (such as contrast and energy) and change maps (vegetation changes calculated based on NDVI time series) from remote sensing imagery. These two types of information are then fused to form a multi-dimensional feature vector for each pixel, which serves as the input data for the model. Using labeled samples (e.g., knowing which areas are work areas and which are not), these feature vectors and their corresponding labels (work area or non-work area) are fed into a random forest classifier for training. Through training, the model learns the relationship between texture features, change maps, and work areas. The trained random forest model then accepts new feature datasets, classifies each pixel, and outputs the probability value of each pixel belonging to a work area. This probability value represents the confidence level that each pixel is a work area. Finally, a probability map is generated based on the classification results, where the value of each pixel represents its probability of belonging to a work area.

[0046] S6. Convert the probability map and binary mask results into vector boundaries, overlay and compare them with agricultural machinery trajectory data, and complete the accuracy verification by calculating the intersection-union ratio.

[0047] Preferably, the probability map and binary mask results are converted into vector boundaries, overlaid and compared with agricultural machinery trajectory data, and the accuracy is verified by calculating the intersection-union ratio (IUU). This includes the following steps: S61. Perform adaptive threshold segmentation on the probability map, and combine morphological operations to remove small noise or discontinuous regions generated during the segmentation process, and optimize the segmentation boundary.

[0048] S62. Using a raster-to-vector algorithm, the optimized segmentation boundary is converted into boundary data in vector format to generate the vector boundary of the mechanical operation area.

[0049] Preferably, the optimized segmentation boundary is converted into vector format boundary data using a raster-to-vector algorithm to generate the vector boundary of the mechanical operation area, comprising the following steps: S621. The optimized binary mask is used as input grid data, where the value of each pixel represents whether it belongs to the mechanical operation area or the non-operation area. S622. Employ boundary tracing to identify connected regions in raster data and accurately extract the boundary cells of each region; S623. Using a raster-to-vector algorithm, the extracted boundary cell sequence is converted into polygon vector data composed of continuous vertices; S624. Optimize the generated polygon vector boundary, remove noisy polygons with too small area, smooth the edges to remove the grid jagged effect, and form an accurate vector boundary for the mechanical operation area.

[0050] It should be explained that extracting boundary cells from raster data typically involves identifying the boundary between the target area and the background using edge detection algorithms (such as Sobel or Canny edge detection). Connected component analysis is then used to group these boundary cells, identifying continuous boundary regions. A contour tracking algorithm is then employed to convert these continuous boundary cells into polygons, forming vector data. For each polygon, vertices are generated by scanning the extracted boundary cells. Finally, a simplification algorithm (such as the Douglas-Peucker algorithm) is used to simplify and optimize the polygon vertices, removing noise and improving efficiency. The processed vector data can be further analyzed and visualized using GIS tools and exported in a standard format.

[0051] S63. Spatial comparison is performed between the vector boundary of the generated mechanical operation area and the agricultural machinery trajectory data. The intersection-union ratio is calculated, and the degree of overlap between the segmentation result and the actual agricultural machinery operation area is evaluated.

[0052] S64. Based on the evaluation results of the intersection-union ratio, complete the accuracy verification.

[0053] It should be explained that the Intersection over Union (IoU) = the area of ​​the intersection between the predicted region and the actual region ÷ the area of ​​the union between the predicted region and the actual region.

[0054] Specific examples are as follows: A county in the North China Plain is a major winter wheat producing area. In May and June of 20XX, the wheat was ready for harvest, and the local mechanized harvesting rate exceeded 95%. This study used Sentinel-2 remote sensing imagery (5-day revisit cycle, 12 scenes in total), county-level cultivated land vector boundaries, digital elevation models, and agricultural machinery Beidou trajectory data to accurately identify mechanized harvesting areas and verify the accuracy of the data. First, radiometric calibration and atmospheric correction were performed on the original Sentinel-2 images to remove haze interference. Then, geometric fine correction was completed in conjunction with the DEM to ensure accurate alignment of multi-temporal images. The cloud layer was masked using image quality bands, missing values ​​were filled with linear interpolation, and noise was removed by smoothing filtering. The NDVI time series of each cultivated land plot was reconstructed, and finally, a winter wheat growth curve was generated, which peaked in mid-May (peak growing season), dropped sharply at the end of May (harvest started), and dropped to the trough in early June (harvest completed). Calculate time-phase EVI images (avoiding NDVI saturation), and extract texture features using gray-level co-occurrence matrix during the critical harvest period (May 25 - June 5)—the texture of the plots after mechanical harvesting is more uniform and the contrast is significantly reduced; at the same time, construct an NDVI change map (difference between the peak growing season and the post-harvest image), and combine it with a change point detection algorithm to accurately locate the harvesting abrupt change areas where NDVI drops sharply. After removing small-area noise patches, the preliminary harvesting area range is obtained; The preprocessed remote sensing images and land parcel growth curves are input into the U-Net semantic segmentation model, which outputs a binary mask of the mechanical operation area (1 for operation area and 0 for non-operation area). The texture features and NDVI variation map are then fused, and a classifier is trained using the random forest algorithm to generate a probability map of the operation area (value 0-1, higher values ​​indicate stronger confidence). The probability map and binary mask were converted into vector boundaries and overlaid with the Beidou trajectory data of the local agricultural machinery cooperative. The intersection-union ratio (IoU) was calculated to be 0.86, indicating that the identification results are highly consistent with the actual mechanical harvesting area and the accuracy meets the needs of agricultural production monitoring.

[0055] It should be noted that the calculation formulas and all parameters involved in the calculations in this application have been dimensionless beforehand. The process of dimensionless processing is well known in the industry and will not be described here.

[0056] Although embodiments of this application have been shown and described above, it is understood that the above embodiments are exemplary and should not be construed as limiting this application. Those skilled in the art can make changes, modifications, substitutions and variations to the above embodiments within the scope of this application.

Claims

1. A method for precise extraction of a grain machinery operation area based on visual remote sensing and AI, characterized in that, The grain machine operation area accurate extraction method comprises the following steps: S1, collecting remote sensing images covering the complete growth period of crops, integrating agricultural machinery track data and cultivated land plot vector boundaries, and constructing a multi-source spatio-temporal data set aligned at the plot level; S2, in the multi-source spatio-temporal data set, pre-processing the remote sensing images, combining the cultivated land vector boundaries, and reconstructing the normalized vegetation index time series for each plot using interpolation and filtering methods to generate crop growth curves at the plot scale; S3, in the remote sensing images, generating an improved vegetation index image for each time phase by calculation, and extracting texture features in combination with the crop growth curve; S4, constructing a change map based on the normalized vegetation index time series, and identifying surface mutation areas caused by harvesting; S5, inputting the pre-processed remote sensing images and the extracted crop growth curve into a semantic segmentation model for end-to-end segmentation, and outputting a binary mask of the mechanical operation area; at the same time, using a random forest algorithm to fuse and classify the texture features and the change map to generate a probability map; S6, converting the probability map and the binary mask result into a vector boundary, superimposing and comparing with the agricultural machinery track data, and completing accuracy verification by calculating the intersection-over-union ratio.

2. The method according to claim 1, wherein, The pre-processing of the remote sensing images in the multi-source spatio-temporal data set, combining the cultivated land vector boundaries, and reconstructing the normalized vegetation index time series for each plot using interpolation and filtering methods to generate crop growth curves at the plot scale comprises the following steps: S21, performing radiometric calibration and atmospheric correction on the remote sensing images in the multi-source spatio-temporal data set to obtain ground true reflectance images; S22, based on the digital elevation model, performing geometric precise correction and multi-temporal image registration on the ground true reflectance images to obtain a spatially accurately aligned multi-temporal image set; S23, using a quality evaluation band, constructing a cloud mask for the spatially accurately aligned multi-temporal image set, and combining linear interpolation and smoothing filtering to reconstruct the normalized vegetation index time series; S24, based on the reconstructed normalized vegetation index time series, generating a complete crop growth curve for each plot. 3.The method of claim 1, wherein, The generation of an improved vegetation index image for each time phase in the remote sensing images, and the extraction of texture features in combination with the crop growth curve comprises the following steps: S31, extracting red, blue and near-infrared band reflectance from the remote sensing images, and generating an improved vegetation index image for each time phase by calculation; S32, analyzing the complete crop growth curve for each plot to determine the key period of crop harvesting; S33, according to the key period of crop harvesting, calculating the texture features from the improved vegetation index image of the corresponding time phase using a gray level co-occurrence matrix algorithm.

4. The method according to claim 3, wherein, The construction of a change map based on the normalized vegetation index time series, and the identification of surface mutation areas caused by harvesting comprises the following steps: S41, analyzing the normalized vegetation index time series of each plot using a change point detection algorithm to obtain change point detection results; at the same time, selecting the normalized vegetation index images of the two key time phases of the crop growth season and the post-harvest period, calculating the difference between the normalized vegetation index values of the two time phases, and generating a change map. S42, based on the change point detection result, identifying a mutation phase in which the normalized vegetation index in each land block sequence sharply decreases; S43, mapping the identified mutation phase to the spatial domain, and based on the change map, marking all pixels that sharply decrease near the candidate harvesting phase to generate a preliminary candidate mutation region; S44, combining the cultivated land block vector boundary, performing spatial clustering on the preliminary candidate mutation region, removing broken patches, and confirming with the crop growth curve to identify the ground mutation region caused by harvesting.

5. The method according to claim 1, wherein, The step of inputting the preprocessed remote sensing image and the extracted crop growth curve into the semantic segmentation model for end-to-end segmentation and outputting a binary mask of the mechanical operation region includes the following steps: Based on the preprocessed remote sensing image and the corresponding land block scale crop growth curve, the comprehensive information containing spatial spectrum and time sequence rules is fused to construct the operation region identification feature set; The operation region identification feature set is input into the semantic segmentation model for training, and the spatial spectrum features of the image and the time sequence rule features of crop growth are learned and fused through the encoder and decoder structure; In the output layer of the semantic segmentation model, each pixel is classified into two categories, and a binary mask with the same spatial resolution as the input image is output.

6. The method according to claim 5, wherein, The step of inputting the operation region identification feature set into the semantic segmentation model for training, and learning and fusing the spatial spectrum features of the image and the time sequence rule features of crop growth through the encoder and decoder structure includes the following steps: The fused operation region identification feature set is input into the semantic segmentation model; In the encoder part of the semantic segmentation model, the multi-level spatial spectrum features are extracted and compressed from the input features through successive convolution and downsampling operations; The deepest features in the multi-level spatial spectrum features are sent to the bottleneck layer, and further integrated and condensed to form high-level semantic features containing global context information; Through the jump connection, the high-level semantic features output by the bottleneck layer are fused with the corresponding shallow features in the encoder path; the decoder then upsamples the fused features to gradually restore their spatial details and output high-resolution feature maps; The high-resolution feature maps are passed to the output layer, where each pixel is classified through convolution operation to generate a binary mask of the mechanical operation region with the same spatial resolution as the input image.

7. The method according to claim 6, wherein, The step of using the random forest algorithm to fuse and classify the texture features and change maps to generate a probability map includes the following steps: The extracted texture features and the change maps generated based on the normalized vegetation index time series are fused to construct a unified multi-dimensional feature vector for each pixel in the study area, and a complete feature dataset is obtained; The feature dataset is combined with the labeled mechanical operation region samples to train the random forest classifier, and the constructed feature dataset is input into the trained model to classify each pixel as an operation region or a non-operation region; Based on the classification results, the probability of each pixel belonging to the mechanical operation region is output to generate a complete probability map.

8. The method according to claim 7, wherein, The step of converting the probability map and the binary mask result into a vector boundary, superimposing and comparing with the agricultural machinery track data, and verifying the accuracy by calculating the intersection over union ratio includes the following steps: S61, adaptive threshold segmentation is performed on the probability map, and morphological operations are combined to remove small noise points or discontinuous regions generated in the segmentation process, and the segmentation boundary is optimized; S62, the optimized segmentation boundary is converted into vector format boundary data by a grid-to-vector algorithm, and a vector boundary of the mechanical operation area is generated; S63, the generated vector boundary of the mechanical operation area is compared with the agricultural machinery track data in space, the intersection-over-union ratio is calculated, and the overlap degree of the segmentation result and the actual agricultural machinery operation area is evaluated; S64, according to the evaluation result of the intersection-over-union ratio, the precision verification is completed. 9.The method of claim 8, wherein, The grid-to-vector algorithm is used to convert the optimized segmentation boundary into vector format boundary data, and the vector boundary of the mechanical operation area is generated, including the following steps: S621, the optimized binary mask is used as input grid data, and the value of each pixel represents whether it belongs to the mechanical operation area or the non-operation area; S622, boundary tracking is used to identify the connected regions in the grid data, and the boundary pixels of each region are accurately extracted; S623, the extracted boundary pixel sequence is converted into polygon vector data composed of continuous vertices by a grid-to-vector algorithm; S624, the generated polygon vector boundary is optimized, small noise polygons are removed, the edges are smoothed to remove the grid sawtooth effect, and an accurate mechanical operation area vector boundary is formed.

Images