Aerial image-based scattered crop planting area identification method and system

Abstract

The invention relates to the technical field of information, in particular to an aerial image-based scattered crop planting area identification method and system. The method comprises the following steps: receiving a plurality of aerial images shot by an unmanned aerial vehicle, and aligning the aerial images to a regional map; according to the identification of the aerial image, partitioning the regional map and marking the purpose type of each partition to obtain the area of the partition; reading blocks corresponding to application types and planted crops, obtaining image areas of a plurality of aerial images corresponding to the blocks, and trying to identify crop types; when the crop types corresponding to the blocks are not identified, obtaining a plurality of possible crop types corresponding to the blocks; reading a pre-configured identification model corresponding to the various possible crop types, and inputting the image region into the identification model to obtain the crop types of the image region; and aggregating the identified blocks of the same crop species to obtain the planting area of each crop species.

Description

Technical Field

[0001] Several embodiments of this specification relate to the field of information technology, specifically to a method and system for identifying dispersed crop planting areas based on aerial imagery. Background Technology

[0002] Large-scale farmland monitoring primarily relies on satellite remote sensing imagery, but its limited spatial resolution and long revisit cycles make it difficult to meet the high-frequency, refined management needs of small-scale, dispersed planting areas. Drone aerial photography technology, due to its high spatiotemporal resolution, flexible deployment, and low cost, is widely used in agricultural remote sensing monitoring. However, existing crop identification technologies based on aerial images are concentrated in large-scale, contiguous planting scenarios, and still face many challenges in dealing with the complex situations of scattered farming, fragmented plots, and mixed crop types. On the one hand, aerial images need to be precisely aligned with regional maps in Geographic Information Systems (GIS) to achieve spatial positioning and area statistics; on the other hand, because different crops have similar appearances and overlapping spectral features in their early growth stages, general identification models have low confidence levels, leading to unreliable classification results. Therefore, there is an urgent need for an identification technology that can combine high-resolution aerial images, achieve precise spatial alignment, support fine-grained identification of crop types and growth stages, and is suitable for dispersed planting scenarios. Summary of the Invention

[0003] This specification describes a method and system for identifying dispersed crop planting areas based on aerial imagery through several embodiments.

[0004] Firstly, embodiments of this specification provide a method for identifying dispersed crop planting areas based on aerial imagery, including the following steps:

[0005] Receive multiple aerial images taken by a drone, and align the aerial images to a regional map based on the drone's position and the orientation parameters of the captured images;

[0006] Based on the recognition of the aerial images, the regional map is divided into blocks and the purpose type of each block is marked. The area of ​​each block is obtained based on the regional map.

[0007] Read the blocks corresponding to the usage type and the crop being planted, obtain the image regions of multiple aerial images corresponding to the blocks, and attempt to identify the crop type based on the image regions;

[0008] When there is a block whose corresponding crop type is not identified, obtain multiple possible crop types corresponding to the block;

[0009] Read the pre-configured recognition model corresponding to the various possible crop types, input the image region into the recognition model, and obtain the crop type of the image region based on the response of the recognition model;

[0010] Aggregate identified blocks of the same crop type to obtain the planting area for each crop type.

[0011] Secondly, embodiments of this specification provide a system for identifying dispersed crop planting areas based on aerial imagery, including:

[0012] The receiving module receives multiple aerial images captured by the drone and aligns the aerial images to the regional map based on the drone's position and the orientation parameters of the captured images.

[0013] The segmentation module divides the area map into blocks based on the recognition of the aerial images and marks the purpose type of each block, and obtains the area of ​​each block based on the area map.

[0014] The identification module reads the blocks corresponding to the usage type and the crop being planted, obtains the image regions of multiple aerial images corresponding to the blocks, and attempts to identify the crop type based on the image regions.

[0015] The analysis module obtains multiple possible crop types corresponding to a given block when the crop type corresponding to a block is not identified.

[0016] The secondary recognition module reads a pre-configured recognition model corresponding to the various possible crop types, inputs the image region into the recognition model, and obtains the crop type of the image region based on the response of the recognition model.

[0017] The aggregation module aggregates blocks of the same crop type to obtain the planting area for each crop type.

[0018] The beneficial effects of the technical solutions provided in some embodiments of this specification include at least the following:

[0019] In several embodiments of this specification, the provided method and system for identifying dispersed crop planting areas based on aerial images adopts a block-based approach that combines average pixel value differences with reference feature matching of surface attachments. This approach adaptively divides the area map into regional units with consistent usage types and filters out areas not used for crop planting by marking usage types, thereby improving the targeting and accuracy of crop identification.

[0020] A hierarchical recognition system is adopted, which combines general recognition with dedicated secondary recognition models. When the confidence level of the general crop recognition model is insufficient, a dedicated recognition model trained for pairwise combinations of candidate crop types is invoked for more refined discrimination. This is particularly suitable for distinguishing crops with similar appearances and improves the classification accuracy in complex scenarios.

[0021] By introducing crop morphological features to model growth stages, crop species identification results include a composite identifier of crop category and growth stage, providing more granular data support.

[0022] Other features and advantages of various embodiments of this specification will be further revealed in the following detailed description and accompanying drawings. Attached Figure Description

[0023] To more clearly illustrate the technical solutions in the embodiments of this specification, the accompanying drawings used in the embodiments will be briefly introduced below. Obviously, the drawings described below are only some embodiments of this specification. For those skilled in the art, other drawings can be obtained based on these drawings without creative effort.

[0024] Figure 1 This is a schematic diagram for identifying areas where crops are planted in a dispersed manner, as provided in this instruction manual.

[0025] Figure 2 This is a schematic diagram of the process for identifying dispersed crop planting areas provided in this manual.

[0026] Figure 3 The crop types provided in this instruction manual are unidentified illustrations.

[0027] Figure 4 This is a schematic diagram of secondary identification provided in this manual.

[0028] Figure 5 This is a schematic diagram of the scattered crop planting area identification system provided in this manual. Detailed Implementation

[0029] The technical solutions of the embodiments of this specification will be explained and described below with reference to the accompanying drawings. However, the following embodiments are only preferred embodiments of this specification and not all of them. Other embodiments obtained by those skilled in the art based on the embodiments in the implementation methods without creative effort are all within the protection scope of this specification.

[0030] The terms "first," "second," "third," etc., in the description, claims, and accompanying drawings are used to distinguish different objects, not to describe a specific order. Furthermore, the terms "comprising" and "having," and any variations thereof, are intended to cover non-exclusive inclusion. For example, a process, method, system, product, or apparatus that includes a series of steps or units is not limited to the listed steps or units, but may optionally include steps or units not listed, or may optionally include other steps or units inherent to such processes, methods, products, or apparatus.

[0031] In the following description, terms such as “inner,” “outer,” “upper,” “lower,” “left,” and “right” are used only to facilitate the description of the embodiments and to simplify the description, and are not intended to indicate or imply that the device or element referred to must have a specific orientation, or be constructed and operated in a specific orientation, and therefore should not be construed as a limitation of this specification.

[0032] All data involved in this application are information and data authorized by the user or fully authorized by all parties, and the collection, use and processing of the relevant data comply with the relevant laws, regulations and standards of the relevant countries and regions.

[0033] Before introducing the technical solutions described in this manual, the application scenarios and related technologies of the technical solutions will be introduced.

[0034] In the hilly and mountainous areas of the south or the small-farm clusters of the north, farmland is generally characterized by fragmentation, dispersion, and intercropping. A cultivated area of ​​several hundred acres may simultaneously contain dozens of independently managed small plots, growing various crops such as soybeans, peanuts, corn, rapeseed, and sweet potatoes, with field ridges, roads, and woodlands interspersed among them, creating blurred boundaries. Traditional manual inspection methods are time-consuming and labor-intensive, making rapid statistical analysis and accurate management difficult. Therefore, this manual provides a method and system for identifying dispersed crop planting areas based on aerial imagery. Please refer to the appendix. Figure 1 A drone is deployed to conduct low-altitude aerial photography of the target area. During its automated flight path, the drone transmits high-resolution visible light images in real time. After receiving the image stream, the ground system accurately projects and registers each frame of the image onto a pre-set electronic map of the area, based on the drone's GPS coordinates, flight altitude, attitude angles (pitch, yaw, roll), and camera intrinsic parameters, generating a geographically aligned image layer.

[0035] The imagery was then segmented into blocks (11) for analysis, identifying and filtering out non-cultivated areas such as roads, houses, and bare soil. The remaining cultivated land was divided into several geographical units with consistent uses. For areas suspected of being planted with crops, the corresponding aerial image regions were extracted, and a general crop recognition model (21) was used for preliminary judgment. The general crop recognition model (21) output a low confidence level, failing to provide a clear classification. For example, soybeans and peanuts are extremely similar in plant morphology, leaf color, and canopy density, thus making identification impossible.

[0036] Then, a refined identification process is initiated. Based on the initial judgment results, multiple candidate crops are obtained, and a pre-trained identification model 22 is automatically loaded. This identification model 22 is a dedicated binary classification model. The image region is input into the identification model 22 for secondary identification. Based on subtle leaf texture, growth posture, and spectral differences, the identification model 22 ultimately confirms that the crop planted in the plot is peanut. Further, combined with its morphological characteristics, it is determined that the peanut is currently in the flowering and pegging stage. Finally, all adjacent plots 11 identified as the same crop species are aggregated, and their total planting area is calculated.

[0037] This manual first provides a method for identifying scattered crop planting areas based on aerial imagery. Please refer to the appendix. Figure 2 The steps include:

[0038] Step S1) Receive multiple aerial images taken by the drone, and align the aerial images to the area map according to the drone's position and the orientation parameters of the captured images.

[0039] Methods for aligning aerial images to a regional map based on the drone's location and the orientation parameters of the captured images include:

[0040] Obtain the drone's location coordinates, flight altitude, pitch angle, yaw angle, and roll angle corresponding to the aerial images;

[0041] Read the camera parameters and build a perspective projection model to convert the image pixel coordinates into location coordinates in the geographic coordinate system;

[0042] Align the image pixels with the region map.

[0043] The drone's shooting angle during flight is not perfectly perpendicular to the ground (i.e., there is pitch, yaw, and roll). If the images are directly superimposed onto a regional map, geometric distortion and positional shift will occur. Therefore, a perspective projection model needs to be established to achieve geographic alignment.

[0044] First, obtain the UAV's GPS coordinates (latitude and longitude), flight altitude (relative to ground elevation), pitch angle, yaw angle, and roll angle for each frame of image. Read the camera's intrinsic parameters (e.g., focal length, principal point coordinates, distortion coefficients) to build a pinhole camera model. Calculate the geographic coordinates of each pixel in the image on the ground through geometric backprojection. Map all pixels to a preset regional electronic map to generate a geographically aligned image layer.

[0045] Step S2) Based on the recognition of the aerial image, the area map is divided into blocks and the usage type 12 of each block 11 is marked. The area of ​​block 11 is obtained based on the area map.

[0046] The method for dividing the area map into blocks and labeling the usage type 12 of each block 11 based on the recognition of the aerial images includes:

[0047] The aerial image is divided into squares of a preset size, the average pixel value of the squares is calculated, and the aerial image is divided into multiple blocks 11 based on the average pixel value.

[0048] Read the aerial image corresponding to each block 11 and record it as a block image, and identify the type of surface attachments in the block image;

[0049] Aggregate adjacent tiles 11 with the same type of surface attachments, or tiles 11 in the final area map;

[0050] Based on the type of surface attachments in each block 11, the use type 12 is marked.

[0051] A strategy combining grid partitioning and pixel aggregation is used to segment aerial images into semantically consistent regions.

[0052] The entire aerial image is divided into several squares of a preset size (e.g., 128×128 pixels). The average RGB value or the mean value in the HSV color space of each square is calculated. If the difference in the average pixel value of adjacent squares is less than a set threshold (e.g., ΔR<10, ΔG<10, ΔB<10), they are aggregated into a preliminary block 11.

[0053] Read the image block corresponding to block 11 and identify the types of surface attachments within the image block. First, collect labeled sample image blocks to ensure coverage of all possible types of surface attachments within the target area (such as farmland, bare soil, buildings, watercourses, cultivated land, roads, etc.). Use the sample image blocks to train the deep learning model.

[0054] For example, in a hilly farmland area of ​​approximately 500 mu (about 33 hectares), including various crop planting areas and a small amount of non-cultivated land (such as villages, roads, and rivers), a drone was used to conduct low-altitude aerial photography of the area, generating high-resolution visible light images. The aerial images were divided into 128×128 pixel grids. For each grid 11, the RGB mean, color features in HSV space, and GLCM-based texture features were calculated. The Canny operator was used to detect significant edges within the grid 11, identifying field ridges and road boundaries. For example, a continuous gray-white area was detected. By comparing it with a reference library, its average pixel value was close to the reference value for "cement road" (R≈180, G≈175, B≈170), and the texture was smooth, with a matching degree of 92%. This area was then labeled as "road," avoiding misclassification as bare soil or fallow land.

[0055] Adjacent plots of the same type (11) are aggregated to form larger geographical units. For example, multiple plots (11) labeled "Farmland - Soybeans" are merged into a single "Soybean Field," with a total area of ​​120 mu (approximately 8 hectares). Similarly, the "Road" plot (11) is aggregated into a main traffic artery running through the farmland, with a length of 3 kilometers. Through the plots (11) and aggregation strategy, the types of surface attachments on each plot can be accurately identified, and the area of ​​each type can be precisely calculated.

[0056] On the other hand, in another embodiment, the method of dividing the aerial image into multiple blocks 11 based on the average pixel value includes:

[0057] The squares whose average pixel value difference is less than a preset threshold are grouped into a block 11;

[0058] Read multiple sample region images, each sample region image having an annotation of the extent of surface attachments;

[0059] The surface attachment image is read according to the range label, the surface attachment image is divided into squares of a preset size, and the average pixel value of the squares is calculated as the reference pixel value of the surface attachment.

[0060] The average pixel value of the squares in the aerial image is compared with the reference pixel value of the surface attachments, and the squares with a difference less than a preset threshold are aggregated into a block 11.

[0061] On the other hand, even if the average pixel values ​​differ significantly, blocks 11 that still belong to the same type also need to be aggregated. For example, blocks 11 containing predominantly white buildings and adjacent blocks 11 containing predominantly gray buildings need to be aggregated to form a larger block 11. For this purpose, it is necessary to identify surface attachments. Blocks 11 with surface attachments of the same type are then further merged.

[0062] On the other hand, in another embodiment, the method of dividing the aerial image into multiple blocks 11 based on the average pixel value includes:

[0063] The squares whose average pixel value difference is less than a preset threshold are grouped into a block 11;

[0064] Read multiple sample region images, each sample region image having an annotation of the extent of surface attachments;

[0065] The surface attachment image is read according to the range label, the surface attachment image is divided into squares of a preset size, and the average pixel value of the squares is calculated.

[0066] The average pixel value of the corresponding square in the image of surface attachments is extracted and denoted as the surface attachment pixel feature.

[0067] The average pixel value of the squares in the aerial image is compared with the pixel features of the surface attachments to obtain the matching degree. Squares with a matching degree higher than a preset threshold are aggregated into a block 11.

[0068] By introducing high-order feature extraction and matching strategies, the system not only relies on the difference in average pixel values ​​between adjacent squares for initial aggregation, but also utilizes a pre-built pixel feature library of surface attachments to achieve more accurate and semantically consistent image segmentation.

[0069] The entire aerial image is divided into squares of a preset size, and the average RGB value or the mean in the HSV color space of each square is calculated. All adjacent squares are traversed (using 4-neighborhood or 8-neighborhood). If the difference in average pixel values ​​between two adjacent squares is less than a preset threshold (e.g., ΔR<10, ΔG<10, ΔB<10), they are considered to have similar colors and are initially aggregated into the same candidate region. Multiple historical aerial images are collected as a sample dataset, each image accompanied by detailed annotations of surface attachments (e.g., "soybean field," "bare soil," "cement road," etc.). Surface attachment pixel features can employ color features, texture features, and shape features. Color features include RGB mean, standard deviation, and color moments. Texture features include contrast, entropy, and homogeneity calculated using the gray-level co-occurrence matrix. Shape features include edge density and Fourier descriptors. These features are combined into a feature vector, and vector distance is used for matching. The matching degree between each square and each type of surface is calculated. If the similarity between a grid cell and the pixel features of a surface attachment is higher than a set threshold (e.g., cosine similarity > 0.9), then the grid cell is matched with that surface attachment. Adjacent grid cells that match the same surface attachment pixel features are then aggregated.

[0070] Step S3) Read the block 11 corresponding to the crop of use type 12, obtain the image area of ​​multiple aerial images corresponding to the block 11, and attempt to identify the crop type based on the image area.

[0071] Based on the morphological characteristics of crops, their life cycle is divided into several growth stages;

[0072] The crop type includes a combination of crop category and growth stage identifier.

[0073] Crops undergo a series of observable morphological changes throughout their life cycle, which can be identified through visual features in aerial images. Plant height and canopy structure, leaf area index and color changes, flowering and fruiting characteristics, and spatial distribution patterns can be used to distinguish different growth stages of the same crop species.

[0074] For example, during the seedling stage, the plants are short, with sparse leaves and a lot of bare ground. During the branching / jointing stage, the stems elongate significantly, branches increase, and the canopy gradually closes. By the maturity stage, the canopy is completely covered, and some crops begin to wither or fall over.

[0075] The NDVI index reflects chlorophyll content. NDVI is low during the seedling stage, peaks during vigorous growth, and declines after maturity. Color changes from light green to dark green to yellowish-green, which can be used to determine the senescence stage. During flowering, distinct flower clusters are visible (e.g., yellow flowers in rapeseed, white flowers in soybeans). Pod / tuber formation stage: Indirectly inferred through changes in canopy texture (e.g., a slight bulge on the ground after peanuts "peg"). Maturity stage: Fruit discoloration, cracking, or withering of the above-ground parts.

[0076] Regarding the spatial distribution pattern, the seedlings are distributed in a dotted pattern with clear plant spacing during the emergence stage. During the canopy closure stage, the gaps between rows disappear, forming continuous green patches.

[0077] Methods for attempting to identify crop types based on the image region include:

[0078] The image region is input into a pre-connected general crop recognition model 21, and the top N crop types with the highest confidence are output as candidate types.

[0079] If the highest confidence level is lower than a preset threshold, the crop type of block 11 is determined to be unidentified, and the candidate type is associated with block 11.

[0080] If the highest confidence level is not lower than a preset threshold, then the crop type corresponding to the highest confidence level is marked as block 11.

[0081] For each use type 12 and the corresponding crop block 11, the image area is cropped out by back-mapping to the coordinate system of the original aerial image based on its geographical boundary in the regional map.

[0082] Image regions undergo standardization preprocessing, including size normalization (e.g., scaling to 224×224 pixels to fit the input of deep learning models), illumination normalization (using histogram equalization or white balance correction to reduce the effects of shadows or backlighting), and noise reduction (e.g., Gaussian filtering) to improve image quality.

[0083] A pre-trained general crop recognition model 21 is connected, which can be deployed in the cloud or on edge devices. The general crop recognition model 21 is a multi-class deep neural network, trained on a large-scale aerial image dataset containing various common crops (such as soybeans, peanuts, corn, rapeseed, wheat, rice, sweet potatoes, etc.). The pre-processed image region is input into the model, and the predicted probability distribution of each crop category is output. The top N crops with the highest probabilities are selected as the candidate category list (e.g., N=3, output: "soybeans (0.52), peanuts (0.28), mung beans (0.15)"). If the highest confidence score is not lower than a preset threshold (e.g., when the preset threshold is 0.5), the crop category corresponding to the highest confidence score is labeled as block 11. For example, if the classification is soybean, then block 11 is labeled as soybean.

[0084] The top N crops with the highest probabilities are selected as candidate crop types (e.g., N=3, output: "soybean (0.35), peanut (0.31), broad bean (0.15)"). When the highest confidence level is lower than a preset threshold (e.g., the preset threshold is 0.65), the crop type in block 11 is determined to be unidentified, and the candidate crop types are associated with block 11, as shown in the appendix. Figure 3 As shown.

[0085] Step S4) When there is a crop type corresponding to block 11 that is not identified, obtain multiple possible crop types corresponding to block 11.

[0086] As a recommended approach, candidate species are directly used as the various possible crop species corresponding to block 11. For example, [soybean (0.35), peanut (0.31), broad bean (0.15)].

[0087] Step S5) Read the pre-configured recognition model 22 corresponding to the various possible crop types, input the image region into the recognition model 22, and obtain the crop type of the image region according to the response of the recognition model 22.

[0088] The methods for pre-configuring the recognition model 22 include:

[0089] Obtain the set of values ​​for crop types, obtain pairwise combinations of crop types in the set, and establish a recognition model for each combination 22;

[0090] Read historical aerial images and obtain patch images labeled with crop types as sample images;

[0091] Based on the crop types included in each combination, obtain the corresponding sample images;

[0092] The recognition model 22 is trained using the sample images, and the configuration of the recognition model 22 is completed based on the trained recognition model 22.

[0093] Obtain a list of possible crop types corresponding to block 11, such as: [soybean (0.35), peanut (0.31), broad bean (0.15)]. Sort the combinations in descending order of confidence output by the general crop identification model 21, and process the most likely combinations first. If the candidate set contains more than two crop types, select pairwise combinations. For each combination, find and load the pre-trained dedicated binary classification model, i.e., identification model 22, such as the "soybean-peanut classification model".

[0094] Based on the planting structure of the target area, the main local crop sets are determined, and then pairwise combinations are generated. Historical aerial images from multiple years and seasons are collected, and the crop types in the cultivated land patches 11 are labeled. For each combination (e.g., soybeans and peanuts), patch images containing only these two crops are selected from the labeled images as sample images. The sample images are ensured to cover different lighting conditions, growth stages, and soil backgrounds, and can be used for data augmentation. The recognition model 22 is trained using these sample images.

[0095] On one hand, in one embodiment, the method of reading a pre-configured recognition model 22 corresponding to the various possible crop types, inputting the image region into the recognition model 22, and obtaining the crop type of the image region based on the response of the recognition model 22 includes:

[0096] Obtain a set of multiple possible crop types, and exhaustively obtain pairwise combinations of crop types;

[0097] Read the recognition model 22 corresponding to the crop types in the pairwise combinations respectively, and input the image region into the recognition model 22;

[0098] Obtain the classification results and confidence scores of all 22 pairs of recognition models, and obtain the classification result corresponding to the highest confidence score as the crop type of the image region.

[0099] If the set of multiple possible crop types is: {soybean (0.35), peanut (0.31), broad bean (0.15)}, then three combinations are generated: (soybean, peanut), (soybean, broad bean), and (peanut, broad bean). These three combinations are then input into the corresponding dedicated binary classification model, i.e., recognition model 22. The obtained output results are soybean (0.88), soybean (0.95), and peanut (0.91), respectively. Finally, soybean (0.95) is selected as the crop type for the image region.

[0100] On the other hand, in another embodiment, the method of reading a pre-configured recognition model 22 corresponding to the various possible crop types, inputting the image region into the recognition model 22, and obtaining the crop type of the image region based on the response of the recognition model 22 includes:

[0101] Multiple possible crop types are sorted in descending order of confidence, and a sliding window that can accommodate two crop types is constructed. The initial position of the sliding window is located at the first two crop types in the sort.

[0102] Read the recognition model 22 corresponding to the crop type in the sliding window, and input the image region into the recognition model 22;

[0103] When the highest confidence level output by the recognition model 22 exceeds a preset reference threshold, the crop type corresponding to the highest confidence level is taken as the crop type of the image region. When the highest confidence level output by the recognition model 22 does not exceed the preset reference threshold, the crop type with lower confidence level is deleted and the next crop type is added to the sliding window. The previous step is executed again until the crop type of the image region is obtained. When the sliding window slides to the end of the arrangement and the crop type of the image region is still not obtained, the image region is marked as unrecognized.

[0104] For example, consider a set of possible crop types: {soybeans (0.35), peanuts (0.31), broad beans (0.15)}. A sliding window is initially used, covering the first two crops in the sorted list, i.e., {soybeans (0.35), peanuts (0.31)}. A dedicated binary classification model, i.e., recognition model 22, is found that corresponds to the combination of the two crops in the current window. The aerial image region corresponding to block 11 is input into the dedicated binary classification model. The dedicated binary classification model outputs the predicted probabilities of the two categories, and the highest value is taken as the output confidence score for this classification. It is then determined whether this exceeds a preset reference threshold (e.g., 0.90). If it does, the classification result is considered reliable, and the crop corresponding to the highest confidence score is taken as the final recognition result. Example: If the model outputs "peanuts: 0.93", which is greater than the reference threshold of 0.9, then peanuts are confirmed as the crop type of the image region.

[0105] If the score does not exceed the reference threshold of 0.9, it is considered that the current model cannot clearly distinguish between the two crops, and the candidate range needs to be further narrowed. Low-priority crop types in the sliding window are eliminated, and the next candidate crop type is introduced. This process is repeated until a result is obtained or the number of candidate crop types is exhausted. Successful crop identification occurs when the confidence score of a specific binary classification model exceeds the reference threshold. If the sliding window has reached the end of the sorting process and no more candidate crops can be introduced, then block 11 is marked as "unidentified," which can trigger manual review or subsequent re-shooting.

[0106] On the other hand, in another implementation, the method of pre-configuring the recognition model includes:

[0107] Obtain the set of values ​​for crop types, obtain pairwise combinations of crop types in the set, and build a recognition model for each combination;

[0108] Read historical aerial images, crop image areas, and label the crop type and the preset angle range where the shooting angle is located, as sample images;

[0109] Sample images of the same crop species are grouped according to angle intervals. The recognition model is trained using the sample images of each group. The recognition model with the highest accuracy is used as the final recognition model, and the corresponding angle interval is marked.

[0110] When obtaining the image regions of multiple aerial images corresponding to the block, the shooting angles of the multiple image regions corresponding to each block are marked.

[0111] When the accuracy of identifying two crops is low when images are taken from a near-vertical downward angle, images taken from other angles, primarily non-vertical downward angles, can more clearly show plant height and morphology from a non-downward perspective, resulting in higher differentiation accuracy. The drone aerial photography described in this manual uses low-altitude aerial photography. This manual focuses on a specific area, such as the crop planting area statistics within a township, thus the total area is relatively small, allowing for low-altitude photography. While a large angle between the shooting angle and the vertical direction can cause image blurring, the low-altitude shooting method ensures sufficient image clarity within a certain range for the recognition model. Recommended flight altitudes are between 30 and 60 meters, suitable for crop species identification and growth status assessment (e.g., wheat, corn, rice), with a pixel scale of 3-5 cm / pixel. Larger-scale aerial photography typically uses flight altitudes between 60 and 100 meters, with a pixel scale of 5-10 cm / pixel, for regional crop distribution mapping. As a recommendation, the aerial images in this manual can be taken at an altitude of 30-40 meters.

[0112] The method for obtaining the crop type of the image region by reading a pre-configured recognition model corresponding to the various possible crop types, inputting the image region into the recognition model, and obtaining the crop type of the image region based on the response of the recognition model includes:

[0113] Obtain a set of multiple possible crop types, and exhaustively obtain pairwise combinations of crop types;

[0114] Read the recognition model corresponding to the crop type in each pairwise combination, and input the image region whose shooting angle matches the angle range marked by the recognition model into the recognition model;

[0115] Obtain the classification results and confidence scores of all pairwise combinations of the recognition models, and obtain the classification result corresponding to the highest confidence score as the crop type of the image region.

[0116] Since the recognition model is trained separately on sample images from different angle ranges, and the model with the highest accuracy is selected as the final recognition model, inputting images of regions with matching shooting angles into the recognition model during recognition can yield more accurate classification results.

[0117] Step S6) Aggregate the identified blocks 11 of the same crop type to obtain the planting area of ​​each crop type.

[0118] After completing the identification of crop types in the segments, including direct identification using a general model and refined identification using a specialized model, the process moves to the statistical aggregation stage. Geographically dispersed segments (11) with the same crop types are spatially aggregated, and the total planting area for each crop is calculated.

[0119] For example, a drone aerial survey covered 5 townships and identified a total of 237 farmland plots.

[0120] The technology described in this embodiment identifies 12 plots under the "Peanut_Seedling Stage" category, totaling 1248 mu (approximately 82.5 hectares). One plot originally labeled "Rapeseed" was found to actually be labeled "Peanut," prompting an alert indicating a correction is needed. The total planting area for "Soybeans" is 963 mu (approximately 64.5 hectares), representing an 18% increase compared to the same period last year, providing a basis for relevant institutions to predict annual soybean yields.

[0121] The embodiments described above are merely preferred embodiments of this specification and are not intended to limit the scope of this specification. Any modifications and improvements made by those skilled in the art to the technical solutions of this specification without departing from the spirit of this specification should fall within the protection scope defined by the claims of this specification.

Claims

1. A method for identifying dispersed crop planting areas based on aerial imagery, characterized in that, Including the following steps: Receive multiple aerial images taken by a drone, and align the aerial images to a regional map based on the drone's position and the orientation parameters of the captured images; Based on the recognition of the aerial images, the regional map is divided into blocks and the purpose type of each block is marked. The area of ​​each block is obtained based on the regional map. Read the blocks corresponding to the usage type and the crop being planted, obtain the image regions of multiple aerial images corresponding to the blocks, and attempt to identify the crop type based on the image regions; When there is a block whose corresponding crop type is not identified, obtain multiple possible crop types corresponding to the block; Read the pre-configured recognition model corresponding to the various possible crop types, input the image region into the recognition model, and obtain the crop type of the image region based on the response of the recognition model; Aggregate identified blocks of the same crop type to obtain the planting area for each crop type.

2. The method for identifying dispersed crop planting areas based on aerial images according to claim 1, characterized in that, Methods for aligning aerial images to a regional map based on the drone's location and the orientation parameters of the captured images include: Obtain the drone's location coordinates, flight altitude, pitch angle, yaw angle, and roll angle corresponding to the aerial images; Read the camera parameters and build a perspective projection model to convert the image pixel coordinates into location coordinates in the geographic coordinate system; Align the image pixels with the region map.

3. The method for identifying dispersed crop planting areas based on aerial images according to claim 2, characterized in that, A method for dividing a region map into blocks and labeling the purpose type of each block based on the recognition of the aerial images includes: The aerial image is divided into squares of a preset size, the average pixel value of the squares is calculated, and the aerial image is divided into multiple blocks based on the average pixel value. Read the aerial image corresponding to each block and record it as a block image, and identify the type of surface attachments in the block image; Aggregate adjacent blocks of the same type of surface attachment, or divide the final area map into blocks; The use type is marked according to the type of surface attachments in each block.

4. The method for identifying dispersed crop planting areas based on aerial images according to claim 3, characterized in that, The method for dividing an aerial image into multiple blocks based on the average pixel value includes: Group adjacent squares whose average pixel value difference is less than a preset threshold into a block; Read multiple sample region images, each sample region image having an annotation of the extent of surface attachments; The surface attachment image is read according to the range label, the surface attachment image is divided into squares of a preset size, and the average pixel value of the squares is calculated as the reference pixel value of the surface attachment. The average pixel value of the squares in the aerial image is compared with the reference pixel value of the surface attachments, and the squares with a difference less than a preset threshold are aggregated into a block.

5. The method for identifying dispersed crop planting areas based on aerial images according to claim 3, characterized in that, The method for dividing an aerial image into multiple blocks based on the average pixel value includes: Group adjacent squares whose average pixel value difference is less than a preset threshold into a block; Read multiple sample region images, each sample region image having an annotation of the extent of surface attachments; The surface attachment image is read according to the range label, the surface attachment image is divided into squares of a preset size, and the average pixel value of the squares is calculated. The average pixel value of the corresponding square in the image of surface attachments is extracted and denoted as the surface attachment pixel feature. The average pixel value of the squares in the aerial image is compared with the pixel features of the surface attachments to obtain the matching degree. Squares with a matching degree higher than a preset threshold are aggregated into a block.

6. The method for identifying dispersed crop planting areas based on aerial images according to any one of claims 1 to 5, characterized in that, Methods for attempting to identify crop types based on the image region include: The image region is input into a pre-connected general crop recognition model, and the top N crop types with the highest confidence are output as candidate types. If the highest confidence level is lower than a preset threshold, the crop type of the block is determined to be unidentified, and the candidate type is associated with the block. If the highest confidence level is not lower than a preset threshold, then the crop type corresponding to the highest confidence level is marked as the block.

7. The method for identifying dispersed crop planting areas based on aerial images according to claim 6, characterized in that, Methods for pre-configuring recognition models include: Obtain the set of values ​​for crop types, obtain pairwise combinations of crop types in the set, and build a recognition model for each combination; Read historical aerial images and obtain patch images labeled with crop types as sample images; Based on the crop types included in each combination, obtain the corresponding sample images; The recognition model is trained using the sample images, and the configuration of the recognition model is completed based on the trained recognition model.

8. The method for identifying dispersed crop planting areas based on aerial images according to claim 7, characterized in that, The method for obtaining the crop type of the image region by reading a pre-configured recognition model corresponding to the various possible crop types, inputting the image region into the recognition model, and obtaining the crop type of the image region based on the response of the recognition model includes: Multiple possible crop types are sorted in descending order of confidence, and a sliding window that can accommodate two crop types is constructed. The initial position of the sliding window is located at the first two crop types in the sort. Read the recognition model corresponding to the crop type in the sliding window, and input the image region into the recognition model; When the highest confidence level output by the recognition model exceeds a preset reference threshold, the crop type corresponding to the highest confidence level is taken as the crop type of the image region. When the highest confidence level output by the recognition model does not exceed the preset reference threshold, the crop type with lower confidence level is deleted and the next crop type is added to the sliding window. The previous step is repeated until the crop type of the image region is obtained. When the sliding window slides to the end of the arrangement and the crop type of the image region is still not obtained, the image region is marked as unrecognized.

9. The method for identifying dispersed crop planting areas based on aerial images according to claim 6, characterized in that, Methods for pre-configuring recognition models include: Obtain the set of values ​​for crop types, obtain pairwise combinations of crop types in the set, and build a recognition model for each combination; Read historical aerial images, crop image areas, and label the crop type and the preset angle range where the shooting angle is located, as sample images; Sample images of the same crop type are grouped according to angle intervals. The recognition model is trained using the sample images of each group. The recognition model with the highest accuracy is used as the final recognition model, and the corresponding angle interval is labeled. When obtaining the image regions of multiple aerial images corresponding to the block, mark the shooting angles of the multiple image regions corresponding to each block; The method for obtaining the crop type of the image region by reading a pre-configured recognition model corresponding to the various possible crop types, inputting the image region into the recognition model, and obtaining the crop type of the image region based on the response of the recognition model includes: Obtain a set of multiple possible crop types, and exhaustively obtain pairwise combinations of crop types; Read the recognition model corresponding to the crop type in each pairwise combination, and input the image region whose shooting angle matches the angle range marked by the recognition model into the recognition model; Obtain the classification results and confidence scores of all pairwise combinations of the recognition models, and obtain the classification result corresponding to the highest confidence score as the crop type of the image region.

10. A system for identifying dispersed crop planting areas based on aerial imagery, characterized in that, include: The receiving module receives multiple aerial images captured by the drone and aligns the aerial images to the regional map based on the drone's position and the orientation parameters of the captured images. The segmentation module divides the area map into blocks based on the recognition of the aerial images and marks the purpose type of each block, and obtains the area of ​​each block based on the area map. The identification module reads the blocks corresponding to the usage type and the crop being planted, obtains the image regions of multiple aerial images corresponding to the blocks, and attempts to identify the crop type based on the image regions. The analysis module obtains multiple possible crop types corresponding to a given block when the crop type corresponding to a block is not identified. The secondary recognition module reads a pre-configured recognition model corresponding to the various possible crop types, inputs the image region into the recognition model, and obtains the crop type of the image region based on the response of the recognition model. The aggregation module aggregates blocks of the same crop type to obtain the planting area for each crop type.

Images