Method and system for determining lodging area based on lodging monitoring spectral index image

Abstract

The invention provides a lodging area determination method and system based on a lodging monitoring spectral index image, and relates to the field of crop form prediction.The method comprises the steps that firstly, a target farmland remote sensing image and a digital surface model are obtained, radiation and geometric correction are completed in combination with GNSS positioning data and meteorological data, and a standardized orthoimage is generated; calculating a vegetation index on the image and carrying out difference to obtain change information; executing opening and closing operation by adopting a structural element adaptive to the image resolution and the crop row spacing, filtering noise according to the area of a connected domain or a pixel number threshold value, and extracting a spectral index abnormal region and a boundary thereof; and then fusing the boundary and the digital surface model under a unified coordinate reference, accumulating the area according to the pixel resolution, and outputting the area of the abnormal region. The method has parameter self-adaption and multi-source data fusion capabilities, can stably obtain a consistent abnormal region area in a multi-resolution and complex field environment, and provides reliable technical support for agricultural condition monitoring and disaster assessment.

Description

Technical Field

[0001] The present application relates to the field of crop morphology prediction, and in particular, to a method and system for determining lodging area based on lodging monitoring spectral index images. Background Art

[0002] In modern agricultural production, field crops such as rice, wheat, and corn are very likely to lodging once they encounter strong winds, heavy rains, or other extreme weather conditions, resulting in a significant decline in yield and quality. Traditional lodging detection and area statistics mainly rely on manual field inspections or conventional remote sensing image comparisons, making it difficult to obtain lodging information over a large area in real time and accurately. On the one hand, the accuracy of single spectral or RGB images in identifying lodging under a variety of complex terrain and meteorological conditions is limited; on the other hand, existing methods generally lack the comprehensive application of historical lodging data, sensor data quality, and meteorological factors, resulting in high rates of missed and misjudgment. To this end, it is necessary to use multi-source data fusion and advanced image processing algorithms to adaptively screen and dynamically fuse remote sensing images in high-noise environments to improve the accuracy and robustness of lodging monitoring. Summary of the Invention

[0003] In response to the deficiencies of the existing technology, the present application provides a method and system for determining the lodging area based on a lodging monitoring spectral index image.

[0004] In a first aspect, the present application provides a method for determining a lodging area based on a lodging monitoring spectral index image, comprising:

[0005] Acquire a remote sensing image and a digital surface model of the target farmland, and correct the remote sensing image in combination with GNSS positioning data and meteorological data to obtain a standardized orthophoto corresponding to the target farmland;

[0006] Calculating a vegetation index based on the standardized orthophoto, and determining change information of the vegetation index;

[0007] Performing morphological processing on the vegetation index change information, extracting spectral index abnormal areas and obtaining boundary information;

[0008] Wherein, the morphological processing includes:

[0009] Converting the vegetation index change result into a binary mask; using a structuring element adapted to the image resolution and crop row spacing, first performing an opening operation to remove isolated noise points, and then performing a closing operation to fill local holes; based on an area threshold or a pixel number threshold of a connected domain, deleting connected regions whose area is less than a first preset threshold, and retaining the outer boundaries of the remaining connected regions as the boundary information;

[0010] The area of ​​the abnormal region of the target farmland is determined based on the standardized orthophoto, the digital surface model and the boundary information.

[0011] As an optional implementation, it also includes:

[0012] Calculating a vegetation index based on the standardized orthophoto, and determining a first lodging candidate area based on change information of the vegetation index;

[0013] Performing morphological processing on the first lodging candidate area to determine boundary information of the first lodging candidate area;

[0014] The area of ​​the lodging region of the target farmland is determined based on the standardized orthophoto, the digital surface model and the boundary information of the first lodging candidate region.

[0015] As an optional implementation manner, the remote sensing image includes: multispectral data; and determining the area of ​​the lodging region of the target farmland includes:

[0016] For the multispectral data and the digital surface model, a preset supervised classification model is used to generate a classification result based on the boundary information;

[0017] Based on the lodging areas identified in the classification results, the area of ​​the lodging areas is counted in combination with a digital surface model to determine the area of ​​the lodging areas of the target farmland.

[0018] As an optional implementation, the construction of the supervised classification model includes:

[0019] Acquiring field survey data of the target farmland;

[0020] For the standardized orthophoto, based on the field survey data, multiple groups of training samples representing fallen crops and non-fallen crops are selected respectively;

[0021] Performing statistical analysis on the spectral characteristics, vegetation index, and texture characteristics of the training samples to construct a supervised classification model;

[0022] Also includes:

[0023] Obtain classification complexity information and accuracy requirement information for lodging monitoring;

[0024] Based on the classification complexity information and the accuracy requirement information, a classifier of the supervised classification model is determined.

[0025] As an optional implementation manner, after the classification result is generated, the method further includes:

[0026] Establishing a confusion matrix, and calculating the overall classification accuracy and / or Kappa coefficient based on the confusion matrix;

[0027] In response to the overall classification accuracy and / or the Kappa coefficient meeting a preset accuracy threshold, performing area statistics based on a digital surface model on the lodging areas identified in the classification results to determine the area of ​​the lodging areas of the target farmland;

[0028] In response to the overall classification accuracy and / or Kappa coefficient not meeting a preset accuracy threshold, at least one of the following operations is performed:

[0029] Adjusting the training samples or classifier parameters of the supervised classification model and reclassifying the multispectral data;

[0030] Re-correcting the boundary information of the first lodging candidate area in the morphological processing stage;

[0031] Output prompt information to trigger manual intervention.

[0032] As an optional implementation, it also includes:

[0033] Obtaining a knowledge graph corresponding to the target farmland;

[0034] The knowledge graph is used to store and associate the crop type, historical lodging records, historical meteorological data, soil properties, and sensor calibration parameters of the target farmland;

[0035] When identifying the first lodging candidate area and constructing the supervised classification model, retrieving reference information corresponding to the current crop growth stage and historical lodging characteristics from the knowledge graph, and adjusting the determination threshold of the lodging area;

[0036] Based on the revised judgment threshold, the revised lodging candidate area and classification result are obtained.

[0037] As an optional implementation, it also includes:

[0038] Pre-building a data quality model for each sensor in the knowledge graph, and defining an output value threshold and a cutoff value threshold respectively;

[0039] After acquiring the multispectral data and the digital surface model, calculating a quality score for each data stream in real time and comparing the score with the output value threshold and the cutoff value threshold;

[0040] Data streams with quality scores less than the cutoff threshold are excluded;

[0041] The data streams with quality scores greater than or equal to the output value threshold are fused as the main data streams to update the candidate regions of the lodging and the input of the supervised classification model.

[0042] As an optional implementation, it also includes:

[0043] When the number of data streams with qualified quality scores is less than the preset lower limit, the system enters the first mode and uses only the data stream with the highest quality score to identify the lodging area;

[0044] When the number of data streams that meet the quality score standard is greater than or equal to the preset lower limit, the second mode is entered. Based on the historical meteorological data and crop growth stage information provided by the knowledge graph, the data streams are weighted and fused according to dynamic weights to identify the lodging areas and calculate the area of ​​the lodging areas.

[0045] As an optional implementation, it also includes:

[0046] Writing the identified lodging area and the actual performance information of the data stream back to the knowledge graph;

[0047] The data quality model and threshold rules are updated, and the sensor data stream weights and the determination thresholds of the lodging areas are dynamically adjusted according to the updated threshold rules when lodging monitoring is performed next time.

[0048] In a second aspect, the present application provides a system for determining a lodging area based on a lodging monitoring spectral index image, comprising:

[0049] An acquisition module is used to obtain remote sensing images and digital surface models of target farmland, and to correct the remote sensing images in combination with GNSS positioning data and meteorological data to obtain standardized orthophoto images corresponding to the target farmland;

[0050] an identification module, configured to calculate a vegetation index based on the standardized orthophoto, and determine change information of the vegetation index;

[0051] A processing module, configured to perform morphological processing on the vegetation index change information, extract spectral index abnormal areas and obtain boundary information;

[0052] Wherein, the morphological processing includes:

[0053] Converting the vegetation index change result into a binary mask; using a structuring element adapted to the image resolution and crop row spacing, first performing an opening operation to remove isolated noise points, and then performing a closing operation to fill local holes; based on an area threshold or a pixel number threshold of a connected domain, deleting connected regions whose area is less than a first preset threshold, and retaining the outer boundaries of the remaining connected regions as the boundary information;

[0054] A calculation module is used to determine the area of ​​the abnormal region of the target farmland based on the standardized orthophoto, the digital surface model and the boundary information.

[0055] Compared with existing technologies, this application achieves stable extraction of spectral index anomaly areas by introducing adaptive morphological opening and closing operations and connected domain screening strategies on the vegetation index difference results. This strategy dynamically adjusts the shape and size of structural elements based on image resolution and crop row spacing, and adaptively sets area or pixel thresholds based on the image's inherent noise level. This effectively suppresses isolated noise points, fills local holes, and maintains the integrity and continuity of anomaly area boundaries under varying resolutions, crop densities, and complex lighting conditions.

[0056] By directly integrating the boundaries of abnormal areas obtained through morphological processing with the digital surface model under a unified coordinate datum, the errors caused by frequent vector-to-raster conversion and resampling in traditional methods can be avoided, simplifying the area measurement process and improving the reliability of the results. The overall method framework provides adaptive adjustment mechanisms for thresholds, structural elements, and screening rules, with good scalability and portability. Without major algorithmic changes, it can flexibly adapt to the monitoring needs of different regions, different crop types, and multi-source remote sensing data, providing reliable and robust technical support for agricultural monitoring, disaster assessment, and refined agricultural management. BRIEF DESCRIPTION OF THE DRAWINGS

[0057] Figure 1 A flowchart of a method for determining a lodging area based on a lodging monitoring spectral index image provided in an embodiment of the present application;

[0058] Figure 2 A schematic diagram of rice lodging provided in an embodiment of the present application;

[0059] Figure 3 A drone image feature map of fallen rice provided in an embodiment of the present application;

[0060] Figure 4 A drone image feature map of non-lodging rice provided in an embodiment of the present application;

[0061] Figure 5 A drone image of a wheat experimental field provided in an embodiment of the present application;

[0062] Figure 6 A schematic diagram of wheat lodging provided in an embodiment of the present application;

[0063] Figure 7 A schematic diagram of a lodging area determination system based on a lodging monitoring spectral index image provided in an embodiment of the present application. DETAILED DESCRIPTION

[0064] The technical solutions in the embodiments of the present application will be clearly and completely described below in conjunction with the drawings in the embodiments of the present application. Obviously, the described embodiments are only part of the embodiments of the present application, rather than all the embodiments.

[0065] See also Figure 1 FIG. 1 is a flowchart of a method for determining a lodging area based on a lodging monitoring spectral index image according to an embodiment of the present application. The method includes steps S101 to S104, wherein:

[0066] S101: Acquire a remote sensing image and a digital surface model of a target farmland, and correct the remote sensing image in combination with GNSS positioning data and meteorological data to obtain a standardized orthophoto corresponding to the target farmland;

[0067] S102: Calculating a vegetation index based on the standardized orthophoto, and determining change information of the vegetation index;

[0068] S103: performing morphological processing on the vegetation index change information, extracting spectral index abnormal areas and obtaining boundary information;

[0069] Wherein, the morphological processing includes:

[0070] Converting the vegetation index change result into a binary mask; using a structuring element adapted to the image resolution and crop row spacing, first performing an opening operation to remove isolated noise points, and then performing a closing operation to fill local holes; based on an area threshold or a pixel number threshold of a connected domain, deleting connected regions whose area is less than a first preset threshold, and retaining the outer boundaries of the remaining connected regions as the boundary information;

[0071] S104: Determine the abnormal area of ​​the target farmland based on the standardized orthophoto, the digital surface model, and the boundary information.

[0072] In this application, an "abnormal region" refers to a collection of pixels that exhibit significantly low reflectivity or a declining trend in vegetation index differential results, contrasting with the surrounding healthy vegetation. This collection typically exhibits spectral characteristics associated with growth anomalies such as lodging, lesions, and waterlogging, but this application does not limit their specific causes. In this embodiment, lodging monitoring is the primary application scenario, so the abnormal region can be considered a suspected lodging distribution area.

[0073] As an optional method, it also includes: calculating the vegetation index based on the standardized orthophoto, and determining the first lodging candidate area based on the change information of the vegetation index; performing morphological processing on the first lodging candidate area to determine the boundary information of the first lodging candidate area; determining the lodging area of ​​the target farmland based on the standardized orthophoto, the digital surface model and the boundary information of the first lodging candidate area.

[0074] The first preset threshold can be determined based on a combination of pixel resolution, crop spacing, and data noise level, and can be obtained using empirical or statistical methods. For example, for imagery with a resolution of 0.1–0.2 m / pixel, areas with a connected domain area less than 0.5 m² (approximately 25–50 pixels) or a pixel count less than the mean minus two standard deviations can be considered noise. For other resolutions, the threshold can be scaled or adaptively set through cross-validation. The threshold can also be dynamically updated based on historical monitoring data, machine learning models, or expert experience.

[0075] The size and shape of the structuring element can be adaptively set according to the image spatial resolution and crop row spacing. For example, when the crop row spacing is about 0.3 m, a square or circular structuring element with a size not exceeding one-third of the row spacing can be preferred to balance noise removal and retention of inter-row details.

[0076] The boundaries of the abnormal areas after morphological processing and connected domain screening use the same coordinate reference as the digital surface model, so the pixel area of ​​the abnormal area can be directly accumulated at the pixel level and the area of ​​the ground feature can be converted according to the DSM grid size.

[0077] By introducing adaptive structuring elements and area threshold strategies, it is possible to maintain a balance between noise suppression and boundary accuracy under different crop densities and image resolutions, thereby improving the stability and applicability of abnormal area measurement.

[0078] In this embodiment, remote sensing images of the target farmland can be obtained using a variety of methods, such as multispectral / hyperspectral cameras, RGB visible light cameras, or satellite images carried by drones, multi-rotor or fixed-wing aircraft. Preferably, in order to balance resolution and efficiency, drones can be used to collect image data of the target farmland at a pre-set flight route and altitude, while being equipped with a laser radar (LiDAR) or using multi-view image reconstruction to generate a digital surface model (DSM). During the flight, GNSS (Global Navigation Satellite System) positioning data and meteorological data, such as wind speed, temperature, and light intensity, can also be recorded to improve data accuracy and consistency during subsequent calibration.

[0079] Regarding sensor configuration and parameter settings, the selected drone can be equipped with a multispectral camera to capture imagery in red, green, blue, and near-infrared bands. High-resolution visible light cameras with a resolution of 1 to 10 cm / pixel, or other combinations, can also be selected. Flight parameters such as altitude, speed, and overlap (lateral overlap and heading overlap) can be pre-set based on the plot area and mission accuracy requirements.

[0080] In addition, if a high-precision three-dimensional surface model needs to be generated, airborne LiDAR can be used or multi-view image matching can be performed in post-processing to obtain a DSM.

[0081] The drone is equipped with a high-precision GNSS module to record its position in real time during flight. Ground-based RTK or post-processed differential positioning (PPK) technology can further improve positioning accuracy. Meteorological data, such as real-time wind speed, temperature, humidity, and light intensity, can be provided simultaneously by the drone's environmental sensors or ground-based weather stations, providing a basis for subsequent radiation correction and noise analysis.

[0082] Furthermore, after the image data acquisition is completed, the original image is firstly subjected to geometric correction and radiometric correction.

[0083] For example, in terms of geometric correction, it is necessary to solve the exterior orientation elements based on the drone's route, attitude (IMU data) and GNSS positioning information, eliminate image distortion and project it into a unified coordinate system.

[0084] For example, in terms of radiation correction, the pixel brightness values ​​can be normalized or atmospherically corrected based on the collected meteorological data and the calibration parameters of the camera sensor to reduce the impact of external environmental fluctuations on the spectral reflectance.

[0085] It should be noted that if multiple images are taken, common image stitching software (such as Pix4D, Agisoft, etc.) or proprietary algorithms can be used to stitch all images into an orthophoto with a unified coordinate reference using methods such as feature point matching and Bundle Adjustment.

[0086] In this embodiment, the DSM can be obtained through LiDAR point cloud or multi-view image matching methods to record the three-dimensional information of the target farmland surface, including crop height. The local farmland area can also be cropped based on public geographic data sources or existing topographic maps to ensure that the coordinates and resolution of the DSM are consistent with those of the orthophoto image for subsequent fusion calculations.

[0087] After geometric correction, radiometric correction, and image stitching, one or more orthophotos accurately corresponding to the target farmland are obtained. Furthermore, GNSS differential data can be used to align and correct ground reference points to further improve geographic accuracy. The final orthophoto output is standardized in spatial coordinates, pixel values, and other dimensions, serving as the basis for extracting vegetation indices (such as NDVI / EVI) and identifying fallen areas.

[0088] After obtaining and generating a standardized orthophoto corresponding to the target farmland, this embodiment preferably uses the Normalized Difference Vegetation Index (NDVI) to preliminarily measure the growth status of crops. Taking multispectral images as an example, the calculation of NDVI can be expressed as:

[0089]

[0090] Among them, NIR represents the spectral reflectance of the near-infrared band, and RED represents the spectral reflectance of the visible red light band.

[0091] In addition, if the robustness to different soil backgrounds or atmospheric scattering needs to be further improved, vegetation indices such as EVI (Enhanced Vegetation Index) or MSAVI (Modified Soil Adjusted Vegetation Index) can be introduced in this embodiment.

[0092] In certain circumstances, differential analysis can be performed using multiple image periods (such as NDVI of the same plot at different time periods) to obtain "vegetation index change information" and highlight plots that have experienced sharp decline or abnormal changes. If only a single image period is available, a reference mean or threshold can be obtained within the local plot using statistical methods, and areas that are significantly below the normal vegetation index level can be identified as suspected lodging.

[0093] Specifically, a set of low vegetation index pixels can be screened out based on empirical thresholds or statistical distributions (such as the median minus two times the standard deviation), thereby preliminarily marking the "first lodging candidate area."

[0094] After mapping the calculated pixel-level vegetation index to a vector or raster area, the "suspected lodging" area can be located based on threshold screening or segmentation algorithm.

[0095] In this embodiment, if the vegetation index value of a pixel or pixel set is below a set threshold, or below a certain range below the typical value of healthy vegetation, the set is included in the "first candidate lodging area." In this case, an abnormal vegetation index alone cannot completely rule out low indices caused by diseased spots, uneven fertilization, shadows, and other factors. Therefore, subsequent steps (such as morphological processing) require further elimination of noise or pseudo-lodging areas.

[0096] The above threshold screening or differential detection results are stored as a series of connected areas or vector polygons to obtain the spatial boundary of the first lodging candidate area.

[0097] The candidate region will be further refined and confirmed in subsequent steps, such as S103.

[0098] After completing the preliminary identification of the first lodging candidate region, this embodiment further refines and removes noise on the candidate region through morphological processing to obtain accurate region boundary information.

[0099] In practice, a combination of opening and closing operations can be used. Opening, typically consisting of a sequence of erosion and dilation, can remove small noise or isolated pixels; closing, a sequence of dilation and erosion, can fill tiny holes or cracks within the target area.

[0100] If there are scattered patches or edge burrs in the candidate region, multiple dilation or erosion operations can be superimposed as needed to smooth the region contours and reduce the interference caused by pseudo-lodging areas. The effect of morphological processing is related to the size and shape of the structural element.

[0101] For example, a 3×3 or 5×5 square structuring element can be used to process images with a resolution between 0.1 and 0.5 m / pixel. If the crop row spacing is relatively large, an elliptical or circular structuring element can also be used to better match the morphology of the target area.

[0102] It should be emphasized that in actual operation, the parameters of the structural elements can be adjusted based on experiments or experience, such as adjusting the radius or shape, to strike a balance between noise removal and boundary detail preservation. This application does not limit this.

[0103] Furthermore, after completing the morphological operation, a connected component analysis can be performed: scattered areas smaller than a preset threshold (such as an area threshold or a pixel number threshold) are regarded as noise and removed; for areas with a larger area of ​​connected components, their outer boundaries (Outer Boundary) or convex hull (ConvexHull) contours are retained and marked, so that the actual lodging morphology can be more accurately segmented in the subsequent confirmation or supervision of the classification model.

[0104] By converting the binary area or raster area after morphological processing into a vector format (such as polygon or line feature), the boundary information of the "first candidate area for lodging" can be obtained; this boundary information is spatially consistent with the standardized orthophoto and digital surface model coordinate system, and can be used to extract the accurate outline of the lodging area or fuse it with three-dimensional elevation data for area calculation.

[0105] The obtained refined boundary of the “first candidate lodging area” will be passed to the next step S104 , where the final confirmation and area calculation of the lodging area will be performed after combining with the digital surface model or other data sources.

[0106] In some embodiments, if it is necessary to further correct the collapsed area using a supervised classification model or texture features, the region boundary obtained by morphological processing can be input into the subsequent algorithm as a "ROI" (Region of Interest).

[0107] In this way, through the above-mentioned morphological processing, isolated noise points can be effectively removed and boundaries can be smoothed in the suspected lodging areas that have been initially screened, thereby improving the completeness and accuracy of the recognition results and laying a more reliable foundation for the next step of area calculation or deeper classification fusion.

[0108] In the above steps, the boundary information of the "first candidate lodging area" has been represented as a vector polygon or a raster connected area, and the same coordinate reference is maintained as that of the standardized orthophoto.

[0109] At the same time, the digital surface model (DSM) is also aligned with the standardized orthophoto according to the aerial survey or ground mapping method to ensure the consistency of spatial resolution and coordinate system.

[0110] In this embodiment, vector polygons (boundaries of the lodging area) and DSM are spatially superimposed in a GIS or image processing environment, so that corresponding surface height information can be obtained for each regional pixel or vector unit.

[0111] Among them, if the terrain undulation is small or the focus is on plane measurement, this embodiment directly uses the area calculation result of the vector polygon in the plane coordinate system as the area of ​​the fallen area; if the surface has certain undulations or the spatial form after the vegetation is collapsed is considered, DSM data can be used to perform "surface unfolding" or triangulated network (TIN) interpolation on the polygon to obtain an area that is closer to the actual surface state.

[0112] In actual applications, the corresponding mode can be selected according to the monitoring accuracy requirements and terrain changes; if the user wants to calculate the accurate lodging area in an area with a larger slope, the three-dimensional true area algorithm is preferred.

[0113] Import the vector polygon (the boundary of the lodging area) into GIS software or a custom program and use conventional geometric formulas, such as dividing the polygon into triangles and summing the results, or use the built-in "Calculate Geometry" function in GIS to quickly obtain the planar projection area. If the lodging crop is only slightly tilted and the terrain is flat, this method of measurement often provides sufficient accuracy for field management and statistical purposes. In this embodiment, if three-dimensional accuracy is required, the DSM elevation data within the vector polygon boundary can be converted into a triangulated irregular network (TIN) or voxel structure. Using a triangulated network or surface unfolding algorithm, the area of ​​each tiny triangular facet is summed to obtain the "surface area."

[0114] For example, this process can be performed in common 3D GIS platforms, such as ArcGIS 3D Analyst, QGIS 3D plug-in, or specialized terrain analysis libraries (GDAL, CGAL). The result is closer to the area of ​​the collapsed area under the actual terrain undulations, but requires higher computational effort.

[0115] After the calculation is completed, the final result will be output as the overall area of ​​the lodging area, or the area of ​​each connected area will be counted separately and summarized; this area can be used for field management, insurance loss assessment, and can also be compared with previous data to evaluate the trend of lodging losses.

[0116] In this way, based on standardized orthophotos, digital surface models, and information on candidate lodging boundaries, this application accurately calculated the lodging area of ​​target farmland. This result can directly provide a quantitative basis for agricultural production decision-making, disaster assessment, and insurance claims.

[0117] For example, see Figure 2 , Figure 2 A schematic diagram of rice lodging provided in an embodiment of the present application; wherein the red part represents the lodged rice, and the green part represents the non-lodged rice.

[0118] For example, see Figure 5 and Figure 6 , Figure 5 A drone image of a wheat experimental field provided in an embodiment of the present application; the wheat plot surrounded by yellow lines is after edge detection. Figure 6 A schematic diagram of wheat lodging provided in an embodiment of the present application.

[0119] As an optional implementation manner, the remote sensing image includes: multispectral data; and determining the area of ​​the lodging region of the target farmland includes:

[0120] For the multispectral data and the digital surface model, a preset supervised classification model is used to generate a classification result based on the boundary information;

[0121] Based on the lodging areas identified in the classification results, the area of ​​the lodging areas is counted in combination with a digital surface model to determine the area of ​​the lodging areas of the target farmland.

[0122] In the specific implementation, after executing the above solution, the following data has been obtained:

[0123] Multispectral data: covers information of at least four bands, including red, green, blue, and near-infrared;

[0124] Digital surface model (DSM): Registered with multispectral imagery in a unified coordinate system, it can provide the distribution of farmland surface or crop height;

[0125] Boundary information of the first candidate lodging area: the outline of the suspected lodging area confirmed by morphological processing in the preceding step S103.

[0126] In addition, in the remote sensing monitoring of farmland, this embodiment pre- or externally trains a "supervised classification model", which can classify and judge the lodging status of each pixel or pixel block based on multispectral features, texture features or other prior information (such as crop type, growth stage, etc.).

[0127] To improve efficiency, this embodiment performs supervised classification only within the boundary of the "first lodging candidate area" rather than performing global classification on the entire multispectral image.

[0128] In a GIS or image processing environment, the multispectral image and the corresponding DSM are spatially cropped to a small section outside the boundary of the candidate lodging area, such as 10 to 20 pixels, to ensure that there will be no omissions or misclassifications due to edge splicing during classification.

[0129] Depending on the application scenario, the model can use a variety of spectral features (red, green, blue, near-infrared), vegetation indices (such as NDVI, EVI, MSAVI) and texture indicators (such as gray-level co-occurrence matrix, LBP, Gabor filter response), etc.; it can also combine the height information in the DSM (such as absolute elevation, crop height difference) for secondary inference to further distinguish between lodging and non-lodging areas (for example, some plots with a sudden drop in height are severe lodging areas).

[0130] In the specific implementation, the cropped multispectral data and DSM-related features are input into a preset supervised classification model (such as support vector machine SVM, random forest RF, convolutional neural network CNN, etc.), and its "lodging rice / non-lodging rice (or other crop categories)" label is predicted pixel by pixel or pixel block; the classification output result is one or more classified images (Raster) or vector polygon annotation.

[0131] In this way, we will eventually get a classification map of "fallen areas" and "non-fallen areas"; compared with the candidate fallen areas obtained from the morphological step previously, supervised classification can often more accurately eliminate interference areas such as lesions, shadows, etc. with low vegetation index but non-fallen areas, and can also identify some missed true fallen areas.

[0132] In specific implementation, the "lodging category" pixels (or vector polygons) in the classification results can be directly used as the final lodging area; or the classified image can be unioned or intersected with the previous lodging candidate area boundary, and then the final lodging range can be formed according to certain confidence rules.

[0133] Furthermore, similar to the above S104, if the terrain is flat, the calculation can be based on the projected area; if the terrain is undulating, the actual area of ​​the lodging area can be estimated on the three-dimensional surface:

[0134] For example, the geometric area of ​​the vector polygon in the plane coordinate system is used as the area of ​​the lodging area; for another example, the DSM grid within the vector polygon is constructed into a TIN, and the area of ​​the lodging area is obtained by accumulating the area of ​​the surface triangles, which is closer to the actual surface state.

[0135] In GIS or digital image processing platforms, the identified lodging area polygons are superimposed with the corresponding DSM areas to calculate the area. If there are multiple continuous or scattered lodging areas, the total area can be counted and summarized separately. In this way, the statistical results can be output for use in agricultural management, insurance loss assessment or disaster assessment.

[0136] As an optional implementation, the construction of the supervised classification model includes:

[0137] Acquiring field survey data of the target farmland;

[0138] For the standardized orthophoto, based on the field survey data, a number of training samples representing fallen crops and non-fallen crops are selected;

[0139] Statistical analysis is performed on the spectral characteristics, vegetation index and texture characteristics of the training samples to construct a supervised classification model.

[0140] See Figure 3 and Figure 4 , Figure 3 A drone image feature map of fallen rice provided in an embodiment of the present application; Figure 4 This is a drone image feature map of non-lodging rice provided in an embodiment of the present application; wherein, Figure 3 The red arrows in the figure point out the typical characteristics of lodging rice. Figure 4 The red arrows in the figure point out the typical characteristics of non-lodging rice.

[0141] In specific implementations, GPS or RTK positioning can be used to obtain the accurate geographic location and growth status records of fallen crops (such as fallen rice / wheat) and non-fallen crops, and the growth status records are obtained through ground collection.

[0142] At the same time, ground photographs, handheld spectrometers, and crop height measurements are used to confirm lodging conditions and record information such as crop variety, growth cycle, and plant density. These ground annotations and attribute information are considered field survey data and serve as "truth labels" for subsequent model training. Field survey data can be stored as vector features (such as points, lines, or surfaces) or text tables, and the locations of survey plots are matched to the coordinate system of standardized orthophotos. Each sample plot is classified and labeled ("lodged" / "not lodged") based on crop distribution and lodging degree, allowing for pixel-by-pixel comparison with remote sensing imagery.

[0143] Furthermore, the field survey data were spatially superimposed with the standardized orthophotos to lock the pixel range around the survey point or survey area; several representative regions of interest (ROIs) were selected from the plots marked as "fallen crops" and "non-fallen crops", respectively. These regions should have spectral uniformity and typicality and avoid edge transition areas or mixed pixels as much as possible.

[0144] Furthermore, to improve the model's adaptability to diverse scenarios, it's preferable to evenly distribute sampling across plots with varying soil types and lighting conditions. To reduce model bias, this example minimizes the number of samples in the "lodged" and "non-lodged" categories, avoiding model overfitting or classification bias. Within each category, multiple, dispersed, and representative sample areas were selected, covering diverse crop growth conditions and terrain locations, to enhance the model's generalization capabilities.

[0145] Furthermore, for the selected training sample area, the average or distribution characteristics of each band (red, green, blue, near-infrared, etc.) in the multispectral data are obtained. Higher-resolution visible light bands or other sensor information can be added as needed, but this application does not limit this. NDVI, EVI, MSAVI, or other commonly used indices are calculated for the same training area. These indices are added to the feature vectors of the training samples to help the model distinguish signs of crop lodging under different growth conditions.

[0146] For texture features, gray-level co-occurrence matrix (GLCM), local binary pattern (LBP), Gabor filter or other texture operators can be introduced to quantify crop canopy texture according to ground resolution and crop row spacing; texture features have a good auxiliary role in distinguishing shadows, bare soil, lodging degree, etc., and can reduce the interference of lesions, weeds, etc. on lodging judgment.

[0147] In a specific implementation, after feature extraction is performed on each training sample area, a feature vector containing multi-dimensional indicators such as "spectral characteristics, vegetation index, and texture characteristics" is formed; combined with the annotation of "lodging / non-lodging" in the field survey data, these feature vectors are labeled with corresponding classification labels. In this embodiment, common machine learning or deep learning classification algorithms can be used, such as support vector machines (SVM), random forests (RF), maximum likelihood methods, and neural networks (CNN, MLP, etc.). By performing feature-label comparison on the training samples, supervised learning methods are used to iteratively optimize the model parameters; cross-validation or independent test sets can be used to evaluate the model's classification accuracy, recall rate, F1-score, and other indicators to determine the optimal number of training rounds or parameter settings.

[0148] After training is completed, the key parameters of the model are persisted and saved. The key parameters include the support vector of SVM, the tree structure of random forest, the weight of neural network, etc., forming a "supervised classification model" that can be directly called for subsequent lodging detection.

[0149] In actual deployment, the model can be run offline (classification is performed on a computer workstation) or applied in real time on a cloud platform or edge computing device (such as a computing module on a drone).

[0150] This approach allows for the selection of training samples based on field survey data of both lodged and intact crops on standardized orthophotos, extracting multi-source features such as spectra, vegetation indices, and textures, ultimately building a supervised classification model. This model significantly improves recognition accuracy and robustness in subsequent lodging detection, providing an efficient and scalable technical approach for large-scale farmland lodging monitoring.

[0151] As an optional implementation, the constructing of the supervised classification model further includes:

[0152] Obtain classification complexity information and accuracy requirement information for lodging monitoring;

[0153] Based on the classification complexity information and the accuracy requirement information, a classifier of the supervised classification model is determined.

[0154] In this embodiment, in order to enable the supervised classification model to obtain better recognition effect in the lodging monitoring scenario, the classification complexity information and accuracy requirement information of the lodging monitoring will be obtained first when building the model;

[0155] Among them, classification complexity information is usually determined by a combination of factors such as the crop variety diversity of the target farmland, the type and degree of lodging, the resolution of remote sensing images, and the degree of environmental background interference.

[0156] The accuracy requirement information can be set according to the user's tolerance for monitoring accuracy, missed detection rate or false detection rate, and may also depend on the strictness of business requirements such as insurance claims and disaster assessment. When the classification complexity is low and the accuracy requirements are relatively loose, this embodiment can select a classifier with a relatively simple structure, such as minimum distance, naive Bayes or linear support vector machine, to reduce the computational burden of the model training and inference process, and complete large-scale lodging identification in a shorter time. If the classification complexity is high, involving more crop types, lodging forms or environmental interference, and the monitoring party requires higher recognition accuracy, this embodiment will give priority to more advanced classification algorithms, such as random forests, multi-layer perceptrons, convolutional neural networks or other deep learning models after statistical analysis of the aforementioned features (spectral features, vegetation index, texture features, etc.), so as to obtain better recognition results even in complex scenarios.

[0157] In this way, by flexibly selecting classifiers based on classification complexity information and accuracy requirement information, this application can achieve adaptive model deployment in farmland lodging monitoring of different scales and different accuracy requirements, taking into account both monitoring efficiency and accuracy, thereby significantly improving the robustness and accuracy of lodging identification.

[0158] As an optional implementation manner, after the classification result is generated, the method further includes:

[0159] Establishing a confusion matrix, and calculating the overall classification accuracy and / or Kappa coefficient based on the confusion matrix;

[0160] In response to the overall classification accuracy and / or the Kappa coefficient meeting a preset accuracy threshold, performing area statistics based on a digital surface model on the lodging areas identified in the classification results to determine the area of ​​the lodging areas of the target farmland;

[0161] In response to the overall classification accuracy and / or Kappa coefficient not meeting a preset accuracy threshold, at least one of the following operations is performed:

[0162] Adjusting the training samples or classifier parameters of the supervised classification model and reclassifying the multispectral data;

[0163] Re-correcting the boundary information of the first lodging candidate area in the morphological processing stage;

[0164] Output prompt information to trigger manual intervention.

[0165] In the specific implementation, after the supervised classification model fuses the multispectral data with the digital surface model, it generates the corresponding classification results of the lodging area. This embodiment will first measure the performance of the model in distinguishing lodging crops from non-lodging crops by establishing a confusion matrix.

[0166] Among them, the confusion matrix is ​​a statistical form constructed based on the comparison relationship between the real label and the model classification result, which contains the sample data information of correct identification and incorrect identification. By calculating indicators such as the overall classification accuracy (OverallAccuracy) and / or the Kappa coefficient, the overall accuracy and consistency of the classification results can be intuitively evaluated. When the overall classification accuracy or the Kappa coefficient meets the pre-set accuracy threshold, this embodiment will directly perform an area statistical operation based on the digital surface model on the lodging area in the classification result to determine the lodging area of ​​the target farmland; if the result meets the requirements, the system can automatically output a lodging area report or connect the area data with insurance claims, production decision-making and other systems to achieve seamless application.

[0167] In addition, if the overall classification accuracy or Kappa coefficient does not meet expectations, this embodiment will perform at least one of the following operations:

[0168] First, the multispectral data is reclassified after adjusting the training samples or classifier parameters of the supervised classification model, thereby achieving online or semi-online model updates;

[0169] Second, the boundary information of the first lodging candidate area is re-corrected in the morphological processing stage to reduce the noise or inaccurate areas in the prior screening;

[0170] Third, the system interface can output prompt information to trigger manual intervention, and professionals can further confirm the suspected area or use manual interpretation and correction.

[0171] In this way, through this set of precision self-verification and conditional feedback mechanism, this application can continuously optimize the recognition accuracy in different environments and dynamic scenarios, ensuring that the final output of the fallen area is more reliable and practical.

[0172] As an optional implementation, the present application also includes:

[0173] Obtaining a knowledge graph corresponding to the target farmland;

[0174] The knowledge graph is used to store and associate the crop type, historical lodging records, historical meteorological data, soil properties, and sensor calibration parameters of the target farmland;

[0175] When identifying the first lodging candidate area and constructing the supervised classification model, retrieving reference information corresponding to the current crop growth stage and historical lodging characteristics from the knowledge graph, and adjusting the determination threshold of the lodging area;

[0176] Based on the revised judgment threshold, the revised lodging candidate area and classification result are obtained.

[0177] In this embodiment, to further improve the accuracy and adaptability of lodging area identification, a knowledge graph corresponding to the target farmland is preferentially acquired before executing the above classification and determination process. This knowledge graph is used to store and associate information such as the target farmland's crop type, historical lodging records, meteorological data, soil properties, and sensor calibration parameters.

[0178] In actual deployment, the knowledge graph can be a data management framework based on a relational database or triples, which contains a unified semantic description of farmland plots, crop varieties, performance characteristics of each sensor, and lodging conditions over the years.

[0179] Specifically, when identifying the first candidate lodging area and constructing a supervised classification model, this embodiment will retrieve reference information corresponding to the current crop growth stage and historical lodging characteristics from the knowledge graph, and dynamically adjust the original values ​​in the lodging area determination and classification threshold setting links.

[0180] For example, if the knowledge graph records that "the local plot has experienced partial lodging many times in the late growth period of previous years, and the soil has poor water retention", the system will use a more relaxed threshold when calculating the vegetation index or performing morphological processing to avoid missing mature or semi-lodged crops as healthy growth; if the sensor calibration parameters or soil properties show that the plot is extremely susceptible to strong winds, the possibility of wind-induced lodging can be given a higher weight in the supervised classification model to improve the recognition success rate in strong wind-induced lodging scenarios.

[0181] In this embodiment, the "knowledge graph" is not limited to the management of static information, but also includes the indexing and association of time series data, such as meteorological evolution, crop growth period progress, and noise levels of previous monitoring sensors. By retrieving the meteorological data and historical lodging records of the current farmland location, it is possible to automatically identify periods of high rain or strong winds, thereby appropriately increasing the threshold for lodging judgment or the weight of lodging features in the supervised classification model. Accordingly, if the knowledge graph prompts that "this stage is usually the straw hardening period, and the probability of lodging is low", the system will relatively increase the lodging judgment threshold to reduce false alarms.

[0182] In this way, the lodging candidate areas or the results output by the classification model are corrected based on the information of the knowledge graph, forming lodging identification results that are more in line with the actual agricultural conditions.

[0183] After graph retrieval and threshold correction, this embodiment will regenerate or update the boundary information of the first lodging candidate area and the classification results of the supervised classification model. If the historical lodging situation indicated by the knowledge graph is consistent with the current spectral and texture features, the lodging judgment can be further consolidated; if a conflict is found, the system may make a secondary fine-tuning of the threshold or prompt manual intervention for verification. In certain extreme cases (such as rare pests and diseases or meteorological conditions that have never been seen in history), the knowledge graph can assist the system in identifying anomalies and outputting reasonable threshold recommendations. Through the "human-machine collaboration" approach, the present invention still has a high degree of robustness in diverse farmland environments.

[0184] In this way, by introducing the retrieval and reference of the knowledge graph in the process of identifying the first candidate lodging area and building the supervised classification model, this application can dynamically adjust the lodging judgment threshold based on factors such as crop growth stage, historical lodging data, and sensor calibration parameters, thereby obtaining the revised candidate area boundary and classification results. This approach not only effectively improves the adaptability of lodging detection to different farmland environments, sensor characteristics, and historical priors, but also enables the system to output more accurate and reliable lodging identification and area calculation results even when faced with complex meteorological or highly heterogeneous soil conditions.

[0185] As an optional implementation, it also includes:

[0186] Pre-building a data quality model for each sensor in the knowledge graph, and defining an output value threshold and a cutoff value threshold respectively;

[0187] After acquiring the multispectral data and the digital surface model, calculating a quality score for each data stream in real time and comparing the score with the output value threshold and the cutoff value threshold;

[0188] Data streams with quality scores less than the cutoff threshold are excluded;

[0189] The data streams with quality scores greater than or equal to the output value threshold are fused as the main data streams to update the candidate regions of the lodging and the input of the supervised classification model.

[0190] In this embodiment, to effectively exclude low-quality or noisy data streams during multi-source sensor data fusion, a data quality model for each sensor is pre-built within the knowledge graph before acquiring multispectral data and the digital surface model. Output and cutoff thresholds are defined for each sensor. This data quality model can be developed based on historical monitoring records, sensor calibration parameters, and field validation results, and is used to quantitatively characterize the performance of different sensors under various environmental conditions.

[0191] Specifically, the knowledge graph not only stores the basic characteristics of the sensor, such as resolution, spectral range, and signal-to-noise ratio, but also includes the actual measured accuracy of the sensor under different weather, time periods, or terrain conditions, as well as the cumulative accuracy and failure rate in previous lodging identification tasks.

[0192] Through unified semantic management of these heterogeneous data, this embodiment can define a set of objective "output value thresholds" and "cutoff value thresholds" for each data stream. That is, when the real-time quality score exceeds the output value threshold, the data stream can be regarded as the main data stream and activated for fusion; when the quality score is lower than the cutoff value threshold, the data stream is considered to have too much noise or obvious failure, and should be excluded to avoid affecting the subsequent identification of candidate areas of collapse or the accuracy of the supervised classification model.

[0193] After acquiring multispectral data and the digital surface model, the system calculates key metrics for each data stream (such as spectral consistency, signal-to-noise ratio, and resolution matching) in real time based on the sensor's data quality model. This quality score is then compared against predefined output and cutoff thresholds in the knowledge graph. If the quality score is below the cutoff threshold, the data stream is no longer valid or has excessive noise under the current environmental conditions and should be excluded. If the quality score is greater than or equal to the output threshold, the data stream is of high quality and can be included as the primary data stream in the fusion process for lodging identification.

[0194] At the same time, this embodiment can also perform secondary weight fusion or temporary shelving processing on data streams whose quality scores are between the threshold intervals. The specific strategy can be determined by the inference rules or expert experience in the knowledge graph.

[0195] Through this adaptive screening mechanism, this application can retain only high-quality sensor information when fusing multi-source data, thereby improving the overall accuracy of the candidate area for lodging and the supervised classification model, and dynamically updating the lodging area or model input corresponding to the main data stream. When a data stream is considered the main data stream and enters the subsequent identification step, the quality score can be further fed back into the knowledge graph based on its historical performance and the current measured results, forming a cumulative understanding of sensor performance and making more accurate threshold settings for the next monitoring task.

[0196] In this way, when faced with sensor failures, drastic environmental changes, or crop growth conditions in different seasons, the system can utilize the semantic management of the knowledge graph and the adaptive capabilities of the data quality model to automatically exclude low-quality data streams or noise streams and prioritize the integration of high-quality data streams, continuously updating the inputs of candidate lodging areas and supervised classification models, and enhancing the robustness and accuracy of the entire process.

[0197] As an optional implementation, it also includes:

[0198] When the number of data streams with qualified quality scores is less than the preset lower limit, the system enters the first mode and uses only the data stream with the highest quality score to identify the lodging area;

[0199] When the number of data streams that meet the quality score standard is greater than or equal to the preset lower limit, the second mode is entered. Based on the historical meteorological data and crop growth stage information provided by the knowledge graph, the data streams are weighted and fused according to dynamic weights to identify the lodging areas and calculate the area of ​​the lodging areas.

[0200] In this embodiment, after calculating the quality score of each data stream in real time and comparing the output value threshold with the cutoff value threshold, the system not only excludes data streams with lower quality scores and retains data streams with higher quality scores, but also determines the different operating modes to enter based on the number of data streams with qualified quality scores (i.e., quality score ≥ output value threshold). If the number of data streams with qualified quality scores is less than a preset lower threshold, this embodiment enters the first mode, called "exclusive mode," in which only the data stream with the highest quality score is used for downed area identification and processing. This ensures the minimum accuracy requirement of the identification process in relatively extreme situations (such as poor quality of most sensors or failure of some sensors), avoiding distortion of the results caused by the fusion of large amounts of low-quality information.

[0201] In this mode, the system will select the data stream with the highest quality score as the main data source, and use the previous lodging candidate area determination rules and supervised classification model to perform further area identification, area statistics and other steps.

[0202] On the contrary, when the number of data streams with qualified quality scores is greater than or equal to the preset lower limit, the system enters the second mode, also known as the "sharing mode" or "multi-source fusion mode".

[0203] In this mode, the system will assign dynamic weights to data streams that meet the quality score standards based on the historical meteorological data and crop growth stage information stored in the knowledge graph, and then perform weighted fusion to identify lodging areas and calculate the area of ​​lodging areas.

[0204] For example, if the knowledge graph records show that strong winds frequently occur during a certain period of time and the hyperspectral sensor performs well in this environment, the system can give a higher weight to the hyperspectral data stream; if another sensor can still provide stable infrared band information under cloudy or low-light conditions, its proportion weight can be increased in rainy environments.

[0205] In this way, the fused multi-source data can more comprehensively reflect the actual situation of the crop and the characteristics of lodging, significantly improving the accuracy and robustness of recognition. In addition, this embodiment can also combine the aforementioned morphological processing, supervised classification model and DSM elevation information to perform detailed segmentation of the lodging area in a shared mode and calculate the lodging area in three dimensions. If the crop is in a sensitive growth stage or has already experienced local lodging, dynamic weighted fusion can particularly highlight dominant factors, such as near-infrared channels or crop height differences, providing strong support for lodging judgment and accurate area calculation.

[0206] Through this mode switching mechanism, the system can fully utilize multi-source information when the sensor quality is good or the data resources are sufficient; when most sensors perform poorly, it avoids excessive noise and directly uses the optimal data stream to complete recognition.

[0207] Through this approach, even under adverse environmental conditions, the basic stability of lodging identification can be maintained; when conditions permit, high-quality data integration can be maximized to improve accuracy and completeness. Combining prior information from the knowledge graph with feedback on sensor performance, this application can adaptively and seamlessly switch between exclusive and shared modes, significantly improving the reliability of lodging area determination and adaptability to complex farmland scenarios.

[0208] As an optional implementation, it also includes:

[0209] Writing the identified lodging area and the actual performance information of the data stream back to the knowledge graph;

[0210] The data quality model and threshold rules are updated, and the sensor data stream weights and the determination thresholds of the lodging areas are dynamically adjusted according to the updated threshold rules when lodging monitoring is performed next time.

[0211] In this embodiment, after the data stream fusion with the quality score meeting the standard is completed and the final lodging area is identified in the first mode or the second mode, the actual performance information of the monitoring process will be further written back to the knowledge graph to continuously improve the understanding and management of each sensor, fusion rules and threshold settings.

[0212] Specifically, the following information can be recorded and uploaded to the knowledge graph:

[0213] First, the spatial range and area of ​​the identified lodging area;

[0214] Second, during this fusion process, the real-time quality score of each data stream and the weight assigned to it in the fusion model;

[0215] Third, if ground acceptance is carried out later or manual labeling verification results are obtained, the comparison situation (such as recognition accuracy, missed detection, and false detection statistics) will also be written into the knowledge graph to form a historical performance record.

[0216] Since the knowledge graph described in this application can store and associate multi-source information such as crop types, meteorological data, sensor calibration parameters, etc., these "actual performance information" will, together with existing historical lodging records and ground survey data, provide an adaptive decision-making basis for future lodging detection under the same or similar conditions.

[0217] After successfully writing back the above information, this embodiment will update the data quality model and threshold rules in a timely manner based on the performance data accumulated in the knowledge graph.

[0218] For example, if a sensor is found to have a high quality score in low-light environments but poor actual recognition performance during multiple monitoring sessions, the system can automatically lower its score weight in low-light periods or increase the cutoff threshold. Conversely, if another sensor repeatedly demonstrates good classification performance under strong wind conditions, its output value threshold can be automatically adjusted under similar conditions for rapid activation.

[0219] At the same time, the judgment threshold for lodging areas will also be fine-tuned based on the false alarm / missed alarm analysis in recent monitoring sessions: if lodging is often misjudged in a certain plot of land during the seedling or mature stage, it means that the spectral or texture characteristics of the plot are special, and the system can lower the lodging judgment threshold or combine additional judgment rules; if a large amount of low signal-to-noise ratio data leads to overall recognition deviation, the knowledge graph can also set a stricter sensor quality score threshold in the next task to reduce noise flow.

[0220] In this way, through the above update mechanism, the present application can dynamically adjust the weight distribution of each sensor data stream based on the updated rules when performing lodging monitoring next time, and adaptively optimize the lodging judgment threshold. Such continuous iteration forms a self-learning and self-evolving farmland remote sensing monitoring process. Whether it is repeated monitoring under similar meteorological conditions or new lodging identification in completely different terrain or crop stages, the system can refer to the historical performance in the knowledge graph and quickly set more reasonable fusion strategies and threshold parameters to provide higher accuracy and robustness for the final lodging area judgment and area estimation. This adaptive method of accumulating real-time quality scoring results and ground inspection feedback information into the knowledge graph and synchronously updating the data quality model and threshold rules not only allows the system to continue to improve in various agricultural scenarios, but also can cope with unforeseen factors such as sensor aging and climate anomalies, ultimately significantly improving the scalability and long-term effectiveness of lodging monitoring.

[0221] Based on the same inventive concept, the embodiments of the present application also provide a system for determining the lodging area based on the lodging monitoring spectral index image, which corresponds to the method for determining the lodging area based on the lodging monitoring spectral index image. Since the principle of solving the problem by the system in the embodiments of the present application is similar to the above-mentioned method for determining the lodging area based on the lodging monitoring spectral index image in the embodiments of the present application, the implementation of the system can refer to the implementation of the method, and the repeated parts will not be repeated.

[0222] Reference Figure 7 FIG. 1 is a schematic diagram of a system for determining a lodging area based on a lodging monitoring spectral index image according to an embodiment of the present application, wherein the system includes:

[0223] The acquisition module 10 is used to obtain remote sensing images and digital surface models of the target farmland, and to correct the remote sensing images in combination with GNSS positioning data and meteorological data to obtain standardized orthophoto images corresponding to the target farmland;

[0224] an identification module 20 for calculating a vegetation index based on the standardized orthophoto and determining change information of the vegetation index;

[0225] The processing module 30 is used to perform morphological processing on the vegetation index change information, extract the spectral index abnormal area and obtain boundary information;

[0226] Wherein, the morphological processing includes:

[0227] Converting the vegetation index change result into a binary mask; using a structuring element adapted to the image resolution and crop row spacing, first performing an opening operation to remove isolated noise points, and then performing a closing operation to fill local holes; based on an area threshold or a pixel number threshold of a connected domain, deleting connected regions whose area is less than a first preset threshold, and retaining the outer boundaries of the remaining connected regions as the boundary information;

[0228] The calculation module 40 is configured to determine the abnormal area of ​​the target farmland based on the standardized orthophoto, the digital surface model, and the boundary information.

[0229] Those skilled in the art will appreciate that the units and algorithm steps of each example described in conjunction with the embodiments disclosed in this application can be implemented in electronic hardware, or a combination of computer software and electronic hardware. Whether these functions are performed in hardware or software depends on the specific application and design constraints of the technical solution. Professional and technical personnel can use different methods to implement the described functions for each specific application, but such implementation should not be considered beyond the scope of this application.

Claims

1. A method for determining lodging area based on a lodging monitoring spectral index image, characterized in that: include: Acquire a remote sensing image and a digital surface model of the target farmland, and correct the remote sensing image in combination with GNSS positioning data and meteorological data to obtain a standardized orthophoto corresponding to the target farmland; Calculating a vegetation index based on the standardized orthophoto, and determining change information of the vegetation index; Performing morphological processing on the vegetation index change information, extracting spectral index abnormal areas and obtaining boundary information; Wherein, the morphological processing includes: Converting the vegetation index change result into a binary mask; using a structuring element adapted to the image resolution and crop row spacing, first performing an opening operation to remove isolated noise points, and then performing a closing operation to fill local holes; based on an area threshold or a pixel number threshold of a connected domain, deleting connected regions whose area is less than a first preset threshold, and retaining the outer boundaries of the remaining connected regions as the boundary information; The area of ​​the abnormal region of the target farmland is determined based on the standardized orthophoto, the digital surface model and the boundary information.

2. The method for determining the lodging area based on the lodging monitoring spectral index image according to claim 1, characterized in that: Also includes: Calculating a vegetation index based on the standardized orthophoto, and determining a first lodging candidate area based on change information of the vegetation index; Performing morphological processing on the first lodging candidate area to determine boundary information of the first lodging candidate area; The area of ​​the lodging region of the target farmland is determined based on the standardized orthophoto, the digital surface model and the boundary information of the first lodging candidate region.

3. The method for determining the lodging area based on the lodging monitoring spectral index image according to claim 2, characterized in that: The remote sensing image includes: multispectral data; the determination of the lodging area of ​​the target farmland includes: For the multispectral data and the digital surface model, a preset supervised classification model is used to generate a classification result based on the boundary information; Based on the lodging areas identified in the classification results, the area of ​​the lodging areas is counted in combination with a digital surface model to determine the area of ​​the lodging areas of the target farmland.

4. The method for determining the lodging area based on the lodging monitoring spectral index image according to claim 3 is characterized in that: The construction of the supervised classification model includes: Acquiring field survey data of the target farmland; For the standardized orthophoto, based on the field survey data, multiple groups of training samples representing fallen crops and non-fallen crops are selected respectively; Performing statistical analysis on the spectral characteristics, vegetation index, and texture characteristics of the training samples to construct a supervised classification model; Also includes: Obtain classification complexity information and accuracy requirement information for lodging monitoring; Based on the classification complexity information and the accuracy requirement information, a classifier of the supervised classification model is determined.

5. The method for determining the lodging area based on the lodging monitoring spectral index image according to claim 4 is characterized in that: After the classification result is generated, the method further includes: Establishing a confusion matrix, and calculating the overall classification accuracy and / or Kappa coefficient based on the confusion matrix; In response to the overall classification accuracy and / or the Kappa coefficient meeting a preset accuracy threshold, performing area statistics based on a digital surface model on the lodging areas identified in the classification results to determine the area of ​​the lodging areas of the target farmland; In response to the overall classification accuracy and / or Kappa coefficient not meeting a preset accuracy threshold, at least one of the following operations is performed: Adjusting the training samples or classifier parameters of the supervised classification model and reclassifying the multispectral data; Re-correcting the boundary information of the first lodging candidate area in the morphological processing stage; Output prompt information to trigger manual intervention.

6. The method for determining the lodging area based on the lodging monitoring spectral index image according to claim 5, characterized in that: Also includes: Obtaining a knowledge graph corresponding to the target farmland; The knowledge graph is used to store and associate the crop type, historical lodging records, historical meteorological data, soil properties, and sensor calibration parameters of the target farmland; When identifying the first lodging candidate area and constructing the supervised classification model, retrieving reference information corresponding to the current crop growth stage and historical lodging characteristics from the knowledge graph, and adjusting the determination threshold of the lodging area; Based on the revised judgment threshold, the revised lodging candidate area and classification result are obtained.

7. The method for determining the lodging area based on the lodging monitoring spectral index image according to claim 6, characterized in that: Also includes: Pre-building a data quality model for each sensor in the knowledge graph, and defining an output value threshold and a cutoff value threshold respectively; After acquiring the multispectral data and the digital surface model, calculating a quality score for each data stream in real time and comparing the score with the output value threshold and the cutoff value threshold; Data streams with quality scores less than the cutoff threshold are excluded; The data streams with quality scores greater than or equal to the output value threshold are fused as the main data streams to update the candidate regions of the lodging and the input of the supervised classification model.

8. The method for determining the lodging area based on the lodging monitoring spectral index image according to claim 7, characterized in that: Also includes: When the number of data streams with qualified quality scores is less than the preset lower limit, the system enters the first mode and uses only the data stream with the highest quality score to identify the lodging area; When the number of data streams that meet the quality score standard is greater than or equal to the preset lower limit, the second mode is entered. Based on the historical meteorological data and crop growth stage information provided by the knowledge graph, the data streams are weighted and fused according to dynamic weights to identify the lodging areas and calculate the area of ​​the lodging areas.

9. The method for determining the lodging area based on the lodging monitoring spectral index image according to claim 8, characterized in that: Also includes: Writing the identified lodging area and the actual performance information of the data stream back to the knowledge graph; The data quality model and threshold rules are updated, and the sensor data stream weights and the determination thresholds of the lodging areas are dynamically adjusted according to the updated threshold rules when lodging monitoring is performed next time.

10. A system for determining a lodging area based on a lodging monitoring spectral index image, for implementing the method for determining a lodging area based on a lodging monitoring spectral index image according to any one of claims 1 to 9, characterized in that: include: An acquisition module is used to obtain remote sensing images and digital surface models of target farmland, and to correct the remote sensing images in combination with GNSS positioning data and meteorological data to obtain standardized orthophoto images corresponding to the target farmland; an identification module, configured to calculate a vegetation index based on the standardized orthophoto, and determine change information of the vegetation index; A processing module, configured to perform morphological processing on the vegetation index change information, extract spectral index abnormal areas and obtain boundary information; Wherein, the morphological processing includes: Converting the vegetation index change result into a binary mask; using a structuring element adapted to the image resolution and crop row spacing, first performing an opening operation to remove isolated noise points, and then performing a closing operation to fill local holes; based on an area threshold or a pixel number threshold of a connected domain, deleting connected regions whose area is less than a first preset threshold, and retaining the outer boundaries of the remaining connected regions as the boundary information; A calculation module is used to determine the area of ​​the abnormal region of the target farmland based on the standardized orthophoto, the digital surface model and the boundary information.

Images