Unmanned aerial vehicle pesticide spraying area identification method and system combined with image recognition

Abstract

This invention provides a method and system for identifying pesticide spraying areas using drones, combining image recognition. It relates to the field of agricultural drone operation technology. First, it acquires an image set containing farmland scene images collected by the drone at different time periods, and performs regional semantic association analysis to generate a farmland region semantic association map. Next, it uses an image recognition model to identify the growth status of crop images, and constructs a spraying area adaptation model based on the semantic association map. Then, it evaluates spraying needs based on this model, divides the farmland into sub-regions, and generates a spraying area division scheme. Next, it determines the pesticide spraying parameters for each sub-region according to the spraying area division scheme, generating drone pesticide spraying instructions. Finally, it sends the instructions to the drone control system to control the drone to perform pesticide spraying operations, thereby improving the accuracy and uniformity of pesticide spraying.

Description

Technical Field

[0001] This invention relates to the field of agricultural drone operation technology, and more specifically, to a method and system for identifying pesticide spraying areas using drones that combines image recognition. Background Technology

[0002] In agricultural production, pesticide spraying is a crucial step in ensuring healthy crop growth and improving yield and quality. Traditional pesticide spraying methods rely mainly on manual operation, which is not only labor-intensive and inefficient, but also makes it difficult to guarantee the uniformity and precision of spraying, easily leading to pesticide waste and environmental pollution. With the development of drone technology, drone pesticide spraying is gradually becoming an emerging operational method, offering advantages such as high efficiency and flexibility.

[0003] However, existing drone-based pesticide spraying technologies have several shortcomings in terms of area recognition. Firstly, most current methods rely solely on simple geographic information or pre-defined fixed areas for spraying, failing to fully consider the differences in crop growth states within the farmland. Crops in different regions may exhibit varying growth states due to factors such as planting time, soil fertility, and the severity of pests and diseases, resulting in different pesticide requirements. The aforementioned one-size-fits-all spraying approach cannot meet the demands of precision agriculture. Secondly, existing methods lack comprehensive analysis of farmland images at different times, failing to capture timely changes in crop growth states over time. This makes it difficult to dynamically adjust spraying areas and parameters based on actual conditions, thus impacting the effectiveness and economic benefits of pesticide spraying. Summary of the Invention

[0004] In view of the aforementioned problems, and in conjunction with the first aspect of the present invention, embodiments of the present invention provide a method for identifying pesticide spraying areas by unmanned aerial vehicles (UAVs) using image recognition, the method comprising:

[0005] A set of farmland images collected by a drone is obtained, and regional semantic association analysis is performed on the set of farmland images to generate a farmland regional semantic association map. The set of farmland images includes farmland scene images taken at different times.

[0006] An image recognition model is invoked to identify the crop growth status of crop images in the farmland image set, and the crop growth status identification result is obtained. A spraying area adaptation model is constructed by combining the semantic association map of the farmland area.

[0007] Based on the spraying area adaptation model, the spraying demand of farmland areas is assessed, farmland sub-regions with different spraying needs are divided, and a spraying area division scheme is generated.

[0008] Based on the spraying area division scheme, determine the pesticide spraying parameters corresponding to each farmland sub-region, and integrate the farmland sub-region information with the corresponding pesticide spraying parameters to generate UAV pesticide spraying instructions;

[0009] The drone pesticide spraying command is sent to the drone control system. After receiving the drone pesticide spraying command, the drone control system controls the drone to perform the pesticide spraying operation.

[0010] In another aspect, embodiments of the present invention also provide a drone pesticide spraying area identification system that combines image recognition, including a processor and a machine-readable storage medium connected to the processor. The machine-readable storage medium is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the machine-readable storage medium to implement the above-described method.

[0011] Based on the above, this embodiment of the invention acquires a set of farmland images collected by a drone at different time periods and performs regional semantic association analysis to generate a farmland regional semantic association map. It then calls an image recognition model to identify the growth status of crop images, constructs a spraying area adaptation model based on the farmland regional semantic association map, assesses spraying needs based on this model, divides farmland into sub-regions, generates a scientifically reasonable spraying area division scheme, determines the pesticide spraying parameters corresponding to each sub-region according to the spraying area division scheme, and generates drone pesticide spraying instructions. This enables the drone to operate according to precise instructions, effectively improving the uniformity and accuracy of pesticide spraying, reducing pesticide waste and environmental pollution, and simultaneously increasing crop yield and quality. Attached Figure Description

[0012] Figure 1 This is a schematic diagram of the execution flow of the drone pesticide spraying area identification method combined with image recognition provided in an embodiment of the present invention.

[0013] Figure 2 This is a schematic diagram of exemplary hardware and software components of an image recognition-integrated drone pesticide spraying area identification system provided in an embodiment of the present invention. Detailed Implementation

[0014] The present invention will now be described in detail with reference to the accompanying drawings. Figure 1 This is a flowchart illustrating a method for identifying pesticide spraying areas using drones that combines image recognition, provided in one embodiment of the present invention. The following is a detailed description of this method for identifying pesticide spraying areas using drones that combines image recognition.

[0015] Step S110: Obtain a set of farmland images collected by the drone, perform regional semantic association analysis on the set of farmland images, and generate a farmland regional semantic association map. The set of farmland images includes farmland scene images taken at different times.

[0016] In this embodiment, the image acquisition device carried by the UAV is a multispectral camera, which can simultaneously acquire image data in the visible light and near-infrared bands. The UAV cruises and photographs contiguous farmland according to a preset flight path, which covers the entire farmland area, and there is a certain overlap between adjacent images to ensure the continuity of image stitching.

[0017] During the shooting process, the drone's flight control system records the time information of each shot in real time and obtains the geographical coordinates of the shooting location through the satellite positioning module. The above information will be associated and stored with the corresponding farmland scene images to form a set of farmland images containing image data from different time periods.

[0018] Step S111: Receive a set of farmland images transmitted by the drone through its onboard image acquisition device. Each farmland scene image in the set carries information about the shooting time and shooting location.

[0019] After image acquisition is complete, the drone transmits the farmland image set to the ground processing terminal via its wireless communication module. The receiving module of the ground processing terminal verifies the transmitted data, checking for completeness and missing information. If any data is missing or corrupted, the receiving module sends a retransmission command to the drone until a complete set of farmland images is obtained.

[0020] For each image of a farmland scene, the ground processing terminal converts the shooting time information into a standard time format and the shooting location information into a unified geographic coordinate format for subsequent location matching and correlation analysis.

[0021] Step S112: Perform region segmentation processing on each farmland scene image in the farmland image set, dividing each farmland scene image into multiple farmland image blocks of the same size, with each farmland image block corresponding to an actual small area in the farmland.

[0022] The ground processing terminal uses a sliding window method to segment each farmland scene image into regions. First, a sliding window of a fixed size is set. The size of the window is determined based on the actual area of ​​the farmland and the image resolution to ensure that the actual small area of ​​farmland corresponding to each window is consistent.

[0023] A sliding window moves across the farmland scene image at preset step sizes, and after each movement, the image region within the window is captured as a farmland image patch. During the segmentation process, for image edge regions smaller than one window size, image data is supplemented by mirroring to ensure that all segmented farmland image patches are of the same size.

[0024] Each farmland image patch corresponds to a unique identifier number, which is associated with the original farmland scene image capture information and the window's position information in the image, thereby determining the geographical location range of the corresponding actual small area of ​​farmland.

[0025] Step S113: Extract semantic features from each of the farmland image blocks. The semantic features include crop type features, soil state features, and vegetation cover features. Each semantic feature is extracted from the color information and texture information of the farmland image block.

[0026] For each segmented farmland image patch, multiple semantic features need to be extracted from its color and texture information to comprehensively describe the state of the actual small area of ​​farmland corresponding to that image patch.

[0027] Step S1131: Perform color space conversion processing on each of the farmland image blocks, converting the farmland image blocks from the RGB color space to the HSV color space, and obtain the hue parameter, saturation parameter and brightness parameter of the farmland image blocks. The hue parameter, the saturation parameter and the brightness parameter together constitute color information.

[0028] First, the farmland image patch is preprocessed to remove noise interference. A Gaussian filtering algorithm is used to smooth the image and filter out high-frequency noise signals. Then, the preprocessed RGB color space image is converted to an HSV color space image.

[0029] During the conversion process, the corresponding hue, saturation, and brightness parameters are calculated based on the pixel values ​​of the RGB three channels. For each pixel in the image, a set of HSV parameter values ​​is obtained, which in turn forms the hue parameter matrix, saturation parameter matrix, and brightness parameter matrix of the entire farmland image patch. These three matrices together constitute the color information of the farmland image patch.

[0030] Step S1132: Perform grayscale processing on the farmland image block, convert the colored farmland image block into a grayscale image block, and calculate the texture feature parameters of the grayscale image block using the grayscale co-occurrence matrix algorithm. The texture feature parameters include contrast parameters and correlation parameters, and the contrast parameters and correlation parameters together constitute texture information.

[0031] The farmland image patch in the HSV color space is converted into a grayscale image patch using a weighted average method. The weights of the three HSV parameters are assigned according to the human eye’s sensitivity to different colors, and the grayscale value of each pixel is calculated.

[0032] Texture features are extracted from grayscale image patches using the gray-level co-occurrence matrix (GLCM) algorithm. First, multiple different directions and distances are defined, and a GLCM is constructed for each direction and distance. Statistical analysis of the GLCM is then performed to calculate contrast and correlation parameters.

[0033] The contrast parameter reflects the degree of difference in grayscale values ​​among pixels in a grayscale image patch, and is obtained by calculating the product of the number of pixel pairs with different grayscale values ​​in the matrix and the square of the grayscale difference. The correlation parameter reflects the degree of linear correlation among pixel grayscale values ​​in a grayscale image patch, and is obtained by calculating the ratio of the covariance to the standard deviation of pixel grayscale values ​​in the matrix. These two parameters together constitute the texture information of the farmland image patch.

[0034] Step S1133: For crop type feature extraction, select the hue parameter from the color information as the main judgment basis, and combine it with the correlation parameter in the texture information to assist in the judgment, and determine the crop type features in the farmland image block to form crop type features.

[0035] Different crops exhibit significant differences in hue parameters. For example, the hue parameters of grasses typically fall within a specific range, while those of leguminous crops fall into another range. First, a database mapping crop types to hue parameter ranges is established, containing typical hue parameter intervals for common crops.

[0036] For each farmland image patch, the distribution of pixels in its tone parameter matrix is ​​statistically analyzed to determine the dominant tone parameter range for that image patch. This dominant tone parameter range is then matched with the crop type correspondence in the database to preliminarily determine the possible crop types.

[0037] Simultaneously, correlation parameters from texture information are used for auxiliary judgment. Different crops have different leaf texture structures, leading to differences in their correlation parameters. For example, crops with denser leaves typically have higher correlation parameter values, while crops with sparser leaves have lower correlation parameter values. By comparing the typical correlation parameter range corresponding to the initially determined crop type with the actual correlation parameters of the farmland image patch, the crop type is further verified and determined, forming crop type features.

[0038] Step S1134: For soil condition feature extraction, the brightness parameter is selected from the color information as the main analysis basis, and the contrast parameter in the texture information is combined to determine the roughness of soil particles, and the soil condition features are determined comprehensively.

[0039] Different levels of soil moisture result in varying intensities of reflected light, leading to differences in lightness parameters. Higher soil moisture content corresponds to lower lightness parameter values, and vice versa. First, a correlation model between soil moisture level and lightness parameter was established. This model was trained using lightness parameter measurements and moisture content detection data from a large number of soil samples.

[0040] For each farmland image patch, the average value of its brightness parameter matrix is ​​calculated. This average value is then input into a correspondence model to preliminarily determine the soil moisture level. Simultaneously, the contrast parameter from the texture information is used to determine the roughness of the soil particles. The rougher the soil particles, the higher the contrast parameter value of its surface texture; the finer the particles, the lower the contrast parameter value.

[0041] Based on the comprehensive analysis of soil moisture and particle roughness, soil condition characteristics are determined, which include information in two dimensions: soil moisture level and particle roughness level.

[0042] Step S1135: For vegetation cover feature extraction, calculate the proportion of green pixels in the farmland image block, and at the same time exclude non-crop green areas by combining the hue parameters to determine the vegetation cover features.

[0043] First, define the hue parameter range for green pixels, based on typical hue parameters of crop leaves. Then, iterate through the hue parameter matrix of the farmland image patch and count the number of pixels whose hue parameters fall within the green range.

[0044] Simultaneously, interference from non-crop green areas is eliminated, such as weeds and green plastic film in farmland. By combining the feature parameters in the texture information, the texture features of weeds are significantly different from those of crops and can be distinguished by setting preset texture feature thresholds; the brightness and saturation parameters of green plastic film are different from those of crop leaves and can be excluded by setting corresponding parameter thresholds.

[0045] The ratio of green pixels after interference removal to the total number of pixels in the farmland image patch is calculated to obtain the proportion of green pixels. This proportion reflects the density of vegetation cover in the actual small area corresponding to the farmland image patch, and the vegetation cover characteristics are determined accordingly.

[0046] Step S1136: Integrate the extracted crop type features, soil condition features, and vegetation cover features, and add corresponding extraction basis annotations to each feature to form complete semantic features.

[0047] Crop type features, soil condition features, and vegetation cover features are integrated according to a preset format. Each feature includes three parts: feature name, feature value, and feature description. Simultaneously, extraction basis annotations are added to each feature, including the specific types and analysis methods of the color information parameters (such as hue parameters and brightness parameters) and texture information parameters (such as correlation parameters and contrast parameters).

[0048] The integrated semantic features are associated with the corresponding farmland image patch identifiers and stored in the feature database for subsequent association analysis and model building.

[0049] Step S114: Based on the shooting location information of each farmland scene image, associate the farmland image blocks from different shooting time periods according to their actual geographical location correspondence, and calculate the semantic feature similarity between the farmland image blocks with corresponding geographical locations.

[0050] First, based on the shooting location information of each farmland scene image and the position of the farmland image patch in the image, the actual geographic coordinate range corresponding to each farmland image patch is determined. A geographic coordinate index table is established to associate the identifier number of each farmland image patch with its geographic coordinate range.

[0051] Based on the geographic location index table, find farmland image patches with overlapping geographic coordinate ranges from different shooting time periods. These image patches correspond to the same actual small area within the farmland. For the found corresponding farmland image patches, calculate their semantic feature similarity.

[0052] Semantic feature similarity is calculated using the cosine similarity algorithm. The semantic features of each farmland image patch are converted into feature vectors, with each dimension of the feature vector corresponding to a semantic feature parameter. The cosine angle between two corresponding feature vectors is calculated; the smaller the angle, the higher the semantic feature similarity; the larger the angle, the lower the similarity.

[0053] Step S115: Using the actual small region corresponding to each farmland image block as a node, and the semantic feature similarity as the association strength between nodes, construct an initial association graph structure.

[0054] The actual small region corresponding to each farmland image patch is abstracted into a node in the initial association graph. Each node contains the geographical coordinate range of the actual small region and the corresponding semantic feature information.

[0055] For geographically adjacent nodes, the association strength between nodes is determined based on the semantic feature similarity between their corresponding farmland image patches. If the semantic feature similarity between two adjacent nodes is high, the association strength value is large; if the similarity is low, the association strength value is small.

[0056] An initial association graph is constructed using a graph-structured data format, containing a set of nodes and a set of edges. The set of nodes consists of nodes corresponding to all actual small regions, and the set of edges consists of connections between adjacent nodes and their corresponding association strength values.

[0057] Step S116: Perform redundant node removal processing on the initial association graph structure, delete nodes that repeatedly correspond to the same actual small region, and retain the unique node corresponding to each actual small region.

[0058] Because farmland images taken at different times may overlap, the initial correlation graph may contain multiple nodes corresponding to the same small actual area, i.e., redundant nodes. Redundant nodes need to be removed from the initial correlation graph.

[0059] By comparing the geographic coordinate ranges of nodes, it is determined whether there are duplicate nodes. If the geographic coordinate ranges of two or more nodes completely overlap or the degree of overlap exceeds a preset threshold, these nodes are determined to correspond to the same actual small area and are considered redundant nodes.

[0060] For redundant nodes, the node with the most complete semantic feature information and the most recent shooting time is retained as the unique node corresponding to the actual small area, and other redundant nodes are deleted. At the same time, the association strength information between the deleted node and its adjacent nodes is merged into the retained node to ensure that the connection relationship of the association graph is not broken.

[0061] Step S117: Adjust the correlation strength values ​​between nodes according to the changing trends of the semantic features in different shooting periods, so that the correlation strength can reflect the semantic correlation changes of farmland areas in different periods.

[0062] Step S1171: Arrange the farmland image set according to the shooting time period to form a time-series farmland image sequence, with each time period corresponding to one or more farmland scene images.

[0063] The shooting time information of each image in the farmland image collection is extracted, and the images are sorted in chronological order to form a time-series farmland image sequence. For multiple farmland scene images taken in the same time period, they are sorted according to the shooting location coordinates to ensure the continuity of the time sequence.

[0064] Step S1172: For each node, extract the semantic features of the node at different shooting times for the actual small area, and form the semantic feature time sequence of the node.

[0065] Based on the geographical coordinates of the actual small area corresponding to the node, farmland image blocks corresponding to that actual small area at different shooting times are located in the time-series farmland image sequence. Semantic features are extracted from these farmland image blocks and arranged in order of shooting time to form a temporal sequence of semantic features for that node.

[0066] Each element in the semantic feature time sequence contains semantic feature parameters and corresponding shooting time information, reflecting the changes in the semantic features of the actual small area over time.

[0067] Step S1173: Calculate the change in semantic features between adjacent time periods in the semantic feature time sequence. The change is obtained by the degree of difference between the semantic features of the current time period and the semantic features of the previous time period.

[0068] For the semantic feature time series of each node, the degree of difference between the semantic features of two adjacent time periods is calculated sequentially. The difference vector is obtained by calculating the difference between the semantic feature vector of the current time period and the semantic feature vector of the previous time period. The magnitude of the difference vector is calculated, which is the amount of change in semantic features between adjacent time periods.

[0069] The greater the change, the more significant the change in the semantic features of the actual small region between two adjacent time periods; the smaller the change, the more gradual the change.

[0070] Step S1174: Calculate the semantic feature changes of all nodes, determine the average level of semantic feature changes, compare the semantic feature changes of each node with the average level, and determine whether the semantic feature changes of the node meet the set conditions.

[0071] Collect semantic feature change data for all nodes, calculate the arithmetic mean of these data, and obtain the average level of semantic feature change. Set a change threshold, which is a preset multiple of the average level.

[0072] The semantic feature change of each node is compared with a threshold. If the change is greater than or equal to the threshold, the semantic feature change of the node is determined to meet the set conditions; if the change is less than the threshold, the set conditions are determined not to be met.

[0073] Step S1175: For nodes whose semantic feature changes meet the set conditions, analyze the correlation between their semantic feature changes and those of adjacent nodes. If the semantic feature change trends of adjacent nodes are consistent, increase the correlation strength value between the nodes.

[0074] For nodes whose semantic feature changes meet the set conditions, extract the amount and direction of semantic feature changes of their neighboring nodes. If the amount of semantic feature changes of neighboring nodes is greater than or equal to the threshold, and the direction of change is the same (i.e., the semantic feature parameters all show an upward trend or all show a downward trend), then it is determined that the semantic feature change trends of these nodes are consistent.

[0075] For adjacent nodes with consistent trends of change, the correlation strength between them is increased by a preset adjustment range. The adjustment range is determined based on the difference between the change in the semantic features of the nodes and the average level; the larger the difference, the larger the adjustment range.

[0076] Step S1176: For nodes whose semantic feature changes do not meet the set conditions, keep their association strength with neighboring nodes unchanged, or make a uniform adjustment based on the average association strength of all nodes.

[0077] For nodes whose semantic feature changes do not meet the set conditions, if their association strength with neighboring nodes is within a reasonable range (i.e., the difference between their association strength with the average association strength of all nodes is within a preset range), then the association strength value remains unchanged.

[0078] If the difference between the correlation strength value and the average correlation strength exceeds the preset range, it will be uniformly adjusted according to the average correlation strength so that the adjusted correlation strength value is close to the average level, thereby ensuring the overall stability of the correlation graph structure.

[0079] Step S1177: Based on the seasonal information corresponding to the shooting time period, the correlation strength adjustment range is corrected. The correlation strength adjustment range corresponding to the semantic feature changes during the peak growing season is greater than that during the non-peak growing season.

[0080] Obtain seasonal information corresponding to the shooting period, and determine the crop's growth stage based on the season. During the peak growing season, the crop grows rapidly, and semantic feature changes are more significant, requiring a larger adjustment range for the association strength; during the off-peak growing season, the crop grows slowly, and semantic feature changes are relatively gradual, allowing for a smaller adjustment range.

[0081] Establish a correlation between seasons and adjustment magnitude correction coefficients, with larger correction coefficients corresponding to peak growing seasons and smaller correction coefficients corresponding to off-peak growing seasons. Multiply the initially adjusted correlation strength value by the corresponding correction coefficient to obtain the final adjustment result.

[0082] Step S1178: Calculate the difference between the adjusted node association strength value and the value before adjustment, so that the difference is within a preset range, and update the adjusted association strength value to the initial association graph structure.

[0083] Calculate the difference between the adjusted correlation strength value and the original value between each node. If the difference is within the preset range, update the adjusted value directly to the initial correlation graph structure. If the difference exceeds the preset range, readjust the magnitude until the difference meets the requirements.

[0084] The updated association graph structure can more accurately reflect the changes in semantic associations of farmland areas at different times.

[0085] Step S118: Visualize the adjusted association graph structure, label the actual small area location information and semantic feature type corresponding to each node, and form a complete semantic association graph of farmland area.

[0086] Graphical visualization techniques are used to transform the adjusted relational graph structure into an intuitive image format. During the visualization process, a satellite map of the farmland is used as the background layer, and each node is labeled on the background layer according to its corresponding actual small-area geographical coordinate range.

[0087] The color of nodes is distinguished based on semantic feature type; for example, different crop types correspond to different colors. The size of nodes is determined by the density of vegetation cover features; the denser the cover, the larger the node. The thickness of the connecting lines between nodes is determined by the association strength value; the greater the strength, the thicker the line.

[0088] Simultaneously, annotation information is added to each node, including the actual geographic coordinates of the small area, crop type characteristics, soil condition characteristics, and vegetation cover characteristics. The annotation information is displayed via mouse hover interaction for easy viewing. Finally, a complete semantic association map of the farmland area is formed and stored in the image database.

[0089] Step S120: Call the image recognition model to identify the crop growth status of the crop image portion in the farmland image set, obtain the crop growth status identification result, and construct the spraying area adaptation model in combination with the farmland area semantic association map.

[0090] In this embodiment, the image recognition model is a pre-trained deep learning model specifically designed for identifying crop growth status. This model is used to process crop images within a dataset of farmland images to obtain crop growth status information. This information is then combined with a semantic association graph of the farmland area to construct a spraying area adaptation model that can accommodate different spraying needs.

[0091] Step S121: Extract the crop image portion from each farmland scene image in the farmland image set, and separate the crop area and non-crop area in the farmland scene image using an image segmentation algorithm. The crop image portion is the image content corresponding to the crop area.

[0092] A deep learning-based semantic segmentation algorithm is used to separate crop regions from non-crop regions. The input is a farmland scene image, and the output is a segmentation mask of the image. Different values ​​in the mask represent different region types (crop regions or non-crop regions).

[0093] First, the farmland scene image undergoes preprocessing, including image resizing and normalization, to ensure it meets the input requirements of the image recognition model. During resizing, the farmland scene image is scaled to a fixed pixel size while maintaining the aspect ratio. Blank areas resulting from scaling are filled with black. During normalization, the pixel value of each pixel in the image is converted to a preset numerical range to accelerate model convergence.

[0094] After preprocessing, the image is input into a semantic segmentation algorithm, which consists of an encoder and a decoder. The encoder extracts features from the image through multiple convolutional and pooling layers, gradually reducing the spatial dimension of the image and increasing the feature dimension. The decoder gradually restores the spatial dimension of the image through deconvolutional layers and upsampling operations, and fuses the features extracted by the encoder with the intermediate features of the decoder to retain more detailed information.

[0095] In the segmentation mask output by the algorithm, preset values ​​represent crop areas, while other values ​​represent non-crop areas (such as soil, field ridges, ditches, etc.). Based on the segmentation mask, the image content corresponding to the crop areas is extracted from the original farmland scene image, which is the crop image portion.

[0096] Step S122: Input the crop image portion into a pre-trained image recognition model. The image recognition model includes a feature extraction layer and a feature classification layer. The feature extraction layer extracts features from the crop image portion to obtain crop growth features, and the feature classification layer classifies the crop growth features to obtain crop growth status categories.

[0097] Step S1221: Adjust the crop image portions to a uniform size so that all crop image portions input into the image recognition model have the same pixel dimension.

[0098] Since the sizes of crop image portions extracted from different farmland scene images may vary, they need to be adjusted to a uniform size. During the adjustment process, bilinear interpolation is used to scale the crop image portions, ensuring that the scaled image does not exhibit significant distortion. The unified pixel dimensions are determined according to the design requirements of the image recognition model, ensuring that the crop image portions input to all models have the same height and width pixel values.

[0099] Step S1222: Input the resized portion of the crop image into the feature extraction layer of the image recognition model. The feature extraction layer adopts a convolutional neural network structure and performs feature extraction on the portion of the crop image through multiple convolution operations.

[0100] The feature extraction layer consists of multiple convolutional blocks, each containing a convolutional layer, a batch normalization layer, and an activation function layer. The resized crop image is first input into the first convolutional block. The convolutional layer uses a pre-sized kernel to perform a sliding convolution operation on the image, calculating the weighted sum of pixel values ​​within each convolutional window to generate a feature map. The batch normalization layer normalizes the feature map output from the convolutional layer, keeping the mean and variance within a pre-defined range and reducing internal covariate shifts. The activation function layer uses the ReLU activation function to perform a non-linear transformation on the batch-normalized feature map, enhancing the model's feature representation capability.

[0101] After each convolutional block is processed, the feature map is input into the next convolutional block. The number of convolutional kernels in subsequent convolutional blocks gradually increases to extract higher-level feature information. Through multi-layer convolutional operations, low-level features such as crop edges, textures, and shapes, as well as high-level features such as leaf arrangement and crop density, are gradually extracted from the crop image.

[0102] Step S1223: Set a pooling layer after each convolutional layer of the feature extraction layer. The pooling layer uses max pooling to downsample the feature map output by the convolutional layer.

[0103] A pooling layer is set after the activation function layer of each convolutional block, with a pooling window of preset size and stride. Max pooling is performed by selecting the maximum pixel value within each pooling window as the output, thus downsampling the feature map. This operation reduces the spatial dimensionality of the feature map, decreases the number of model parameters and computational cost, while preserving key feature information and improving the model's translation invariance.

[0104] After multiple convolution and pooling operations, the spatial dimension of the feature map output by the feature extraction layer is greatly reduced, while the feature dimension is significantly increased, containing rich feature information about the crop growth status.

[0105] Step S1224: The feature vector output by the last pooling layer of the feature extraction layer is used as the crop growth feature, which includes crop leaf morphology features, crop color distribution features and crop density features.

[0106] The feature map output from the last pooling layer of the feature extraction layer is converted into a feature vector through a flattening operation. This feature vector is the crop growth feature. Each dimension of the crop growth feature corresponds to different feature information, including crop leaf morphology features, which reflect information such as leaf size, shape, and edge smoothness; crop color distribution features, which reflect information such as the color intensity and uniformity of crop leaves; and crop density features, which reflect information such as the quantity and distribution of crops per unit area.

[0107] Crop growth characteristics comprehensively depict the growth status of crops, providing sufficient characteristic basis for subsequent classification and processing.

[0108] Step S1225: Input the crop growth characteristics into the feature classification layer of the image recognition model. The feature classification layer adopts a fully connected neural network structure.

[0109] The feature classification layer consists of multiple fully connected layers, each containing multiple neurons. Crop growth features are first input into the first fully connected layer. Each neuron in this fully connected layer is connected to all neurons in the previous layer. The output vector of this layer is obtained by calculating the weighted sum of the input feature vector and the neuron weights, plus a bias term.

[0110] The output vector of the first fully connected layer is input into the second fully connected layer. Subsequent fully connected layers use a similar calculation method to gradually integrate and abstract the feature information, mapping the crop growth features to a feature space that is more suitable for classification.

[0111] Step S1226: Set the Softmax activation function in the last layer of the feature classification layer, calculate the probability value of the crop growth feature belonging to different crop growth state categories through the Softmax activation function, select the crop growth state category with the highest probability value as the crop growth state category corresponding to the crop image part, and record the probability value as the classification confidence.

[0112] The last layer of the feature classification layer uses a Softmax activation function. This function transforms the vector output from the fully connected layer into multiple probability values, each corresponding to a crop growth state category, and the sum of all probability values ​​is 1. The probability value of a crop growth feature belonging to each crop growth state category is calculated using the Softmax activation function.

[0113] The crop growth status category with the highest probability value is selected as the corresponding crop growth status category for that part of the crop image, such as "healthy seedling stage," "pest and disease infection during tillering stage," and "normal maturity stage." Simultaneously, this highest probability value is recorded as the classification confidence score to assess the reliability of the classification result. If the classification confidence score is lower than a preset threshold, the classification result is marked as uncertain and requires subsequent manual review or re-identification.

[0114] Step S123: Determine crop growth status parameters according to the crop growth status category. The crop growth status parameters include crop growth stage parameters and crop health status parameters. Each crop growth status parameter corresponds one-to-one with the crop growth status category.

[0115] A correspondence table was established between crop growth status categories and crop growth status parameters. Each crop growth status category in this table corresponds to a unique crop growth stage parameter and a crop health status parameter. The crop growth stage parameters include "seedling stage," "tillering stage," "jointing stage," "heading stage," and "maturity stage," reflecting the crop's current growth and development stage. The crop health status parameters include "healthy," "mild pests and diseases," "moderate pests and diseases," "severe pests and diseases," and "poor growth," reflecting the crop's health condition.

[0116] Based on the crop growth status category output by the image recognition model, the corresponding relationship table is queried to determine the crop growth stage parameters and crop health status parameters corresponding to the crop image portion. These two parameters are then combined to form the crop growth status parameters.

[0117] For example, if the crop growth status category is "tillering stage with mild pests and diseases", then the corresponding crop growth stage parameter is "tillering stage" and the crop health status parameter is "mild pests and diseases". Together, they constitute the crop growth status parameters of the crop image portion.

[0118] Step S124: Associate the crop growth status parameters with the nodes in the semantic association graph of the farmland area to form an association graph with growth status information. The actual small area corresponding to each node is matched with the corresponding crop growth status parameters.

[0119] Obtain the identifier number of the farmland image patch corresponding to the crop image portion, and find the corresponding node in the semantic association graph of the farmland area based on the identifier number. Associate and store the determined crop growth status parameters with the node, so that the node not only contains the geographical coordinate range and semantic feature information of the actual small area, but also contains the corresponding crop growth status parameters.

[0120] The above association process is repeated across all crop image segments until each node in the semantic association graph of the farmland area matches the corresponding crop growth status parameter (if a node corresponds to a small area with no crop growth, its crop growth status parameter is marked as "no crop"). This results in an association graph with growth status information, which integrates multi-dimensional information such as semantic association, geographical location, and crop growth status of the farmland area.

[0121] Step S125: Extract the node association relationships and crop growth state parameters corresponding to the nodes from the association graph with growth state information, and use them as input features for the spraying area adaptation model.

[0122] Two core pieces of information are extracted from the association graph containing growth status information as input features. The first part is the node association relationship, including the list of neighboring nodes of each node and the corresponding association strength value. This part of the information reflects the semantic association degree and spatial adjacency relationship between different actual small regions.

[0123] The second part consists of crop growth status parameters corresponding to the nodes, including crop growth stage parameters and crop health parameters for each node. This part of the information reflects the growth status of crops in each actual small area.

[0124] These two pieces of information are encoded according to a preset feature encoding method, converting them into feature vectors that the model can recognize. Node relationships are encoded using an adjacency matrix, where rows and columns correspond to nodes, and matrix elements correspond to the numerical values ​​of the association strength between nodes. Crop growth state parameters are encoded using one-hot encoding, converting discrete parameter values ​​into binary vectors. The encoded feature vectors together constitute the input features of the spraying area adaptation model.

[0125] Step S126: Construct the core logic layer of the spraying area adaptation model. The core logic layer adopts a decision tree structure, uses the node association relationship in the input features as the branch judgment condition, and uses the crop growth state parameters as the basis for assigning branch weights.

[0126] The decision tree structure of the core logic layer is constructed in a top-down manner, with the root node serving as the initial judgment node. First, the node relationships in the input features are selected as the primary branch judgment condition. For example, whether the correlation strength between nodes is greater than a preset threshold is used as the judgment condition for the root node, dividing the input features into two child nodes.

[0127] Each child node continues to use the relationships with other nodes as judgment conditions, such as whether the number of adjacent nodes reaches a preset number, or whether the average association strength is within a preset range, to further divide the next level of child nodes. At each branch node, crop growth status parameters are used as the basis for assigning branch weights. Based on the different crop growth stage parameters and crop health parameters, corresponding weight values ​​are assigned to each branch.

[0128] For example, for a branch where the "association strength is greater than the threshold", if the corresponding crop growth status parameter is "severe pests and diseases in the mature stage", then a higher weight value is assigned to this branch; if it is "healthy seedling stage", then a lower weight value is assigned. The weight value will affect the subsequent spraying demand judgment result.

[0129] Step S127: Optimize the decision tree structure hierarchically, adjust the branch judgment order, and prioritize the node association relationship associated with the spraying demand as the judgment condition.

[0130] A large amount of historical farmland spraying data was collected, including node relationships, crop growth status parameters, and corresponding actual spraying demand results. Based on historical data, the correlation between each node relationship and the spraying demand results was calculated. The higher the correlation, the greater the impact of that node relationship on the judgment of spraying demand.

[0131] The decision tree structure is adjusted according to the degree of correlation. Nodes with a high correlation to spraying needs are prioritized as the decision conditions for upper-level branches, while those with low correlation are used as the decision conditions for lower-level branches. For example, if the "correlation strength between the node and the node in the pest and disease area" has the highest correlation with spraying needs, then this decision condition is adjusted to the decision condition of the root node or an upper-level branch near the root node.

[0132] Meanwhile, the depth of the decision tree is optimized. If the depth is too deep, the model will overfit, or the depth is too shallow, the model will underfit. The depth is adjusted by pruning or adding branches so that the decision tree structure can more accurately fit the spraying demand patterns in historical data.

[0133] Step S128: Solidify the mapping relationship between the optimized decision tree structure and the input features to form a spraying area adaptation model that can output spraying adaptation results based on the semantic association of farmland areas and crop growth status.

[0134] The optimized decision tree structure's node judgment conditions, branch weights, and hierarchical relationships are solidified and stored to determine the mapping relationship between input features and spray adaptation results. When new node associations and crop growth status parameters are input, the model will follow the decision tree's judgment process, judging layer by layer from the root node, accumulating the weight values ​​of each branch, and finally outputting the spray adaptation result.

[0135] The spraying adaptation results include information such as the necessity of spraying in the actual small area corresponding to each node and the recommended spraying method, providing a direct basis for subsequent spraying demand assessment. After the model is solidified, it still needs to be validated with new farmland data. If the accuracy of the validation results does not meet the preset standard, the decision tree structure will be re-optimized until the model performance meets the requirements.

[0136] Step S130: Based on the spraying area adaptation model, assess the spraying demand of farmland areas, divide farmland sub-areas with different spraying needs, and generate a spraying area division scheme.

[0137] Input features from the association graph with growth status information are input into the spraying area adaptation model. The spraying demand of the entire farmland area is assessed based on the spraying adaptation results output by the model. Based on the assessment results, the farmland is divided into sub-regions with different demand levels and integrated into a complete spraying area division scheme.

[0138] Step S131: Input all nodes in the semantic association graph of the farmland area and the corresponding crop growth status parameters into the spraying area adaptation model.

[0139] Extract the identifiers, node relationships (adjacency matrix form), and corresponding crop growth status parameters (unique heat encoding form) of all nodes from the semantic association graph of the farmland area. Organize them according to the input format required by the spraying area adaptation model to form batch input feature data.

[0140] Perform a consistency check on the input feature data to verify the integrity of node relationships and the absence or errors in crop growth status parameters. If any issues are found, return to the previous steps for correction; if the check passes, input the batch of feature data into the spraying area adaptation model.

[0141] Step S132: The core logic layer of the spraying area adaptation model is used to determine the spraying demand of each node, and the actual spraying demand level of each node's small area is determined based on the node association relationship and crop growth status parameters.

[0142] The core logic layer of the spraying area adaptation model performs layer-by-layer judgment on the input features of each node according to the optimized decision tree structure. First, the relationship between nodes is judged based on the judgment conditions of the root node, and then the corresponding branch node is entered. At the branch node, the judgment conditions of that node are further judged based on the weight values ​​corresponding to the crop growth status parameters, until the leaf node is reached.

[0143] The output of a leaf node represents the actual spraying demand level for the corresponding small area. Spraying demand levels are divided into five categories: "Emergency Spraying," "Priority Spraying," "Regular Spraying," "Low Demand Spraying," and "No Spraying Required." For example, if a node's association relationships show a high correlation with multiple "Emergency Spraying" level nodes, and the crop growth status parameter is "Severe Pests and Diseases at Maturity," then the node's spraying demand level is determined to be "Emergency Spraying." If the node is in a "No Crops" area, then it is determined to be "No Spraying Required."

[0144] Step S133: Count the number of nodes with the same spraying demand level and the corresponding actual small area location, merge the actual small areas that are geographically adjacent and have the same spraying demand level to form a preliminary farmland sub-region.

[0145] All nodes are categorized and statistically analyzed according to spraying demand levels, recording the number of nodes corresponding to each level and the actual geographic coordinate range of each node's small area. For nodes of the same spraying demand level, it is determined whether their corresponding small areas are geographically adjacent. If the boundaries of two small areas overlap or the distance between them is less than a preset distance, they are considered adjacent.

[0146] Adjacent small areas with the same spraying requirement level are merged to form preliminary farmland sub-regions. During the merging process, semantic feature information and crop growth status parameters of multiple nodes are integrated, and the most representative information is used as the attribute information of the preliminary farmland sub-region. For example, when merging multiple nodes with the "emergency spraying" level, the node with the worst crop health parameter is used as the health status attribute of the preliminary farmland sub-region.

[0147] Step S134: Adjust the boundaries of the preliminary farmland sub-regions so that the boundaries of each preliminary farmland sub-region are consistent with the actual terrain features or crop planting boundaries in the farmland.

[0148] Step S1341: Obtain actual terrain data and crop planting planning data of farmland. The actual terrain data includes the location of field ridges and ditches in the farmland, and the crop planting planning data includes the planting area boundaries of different crops.

[0149] The actual topographic data and crop planting plan data of the contiguous farmland are obtained through the farmland management system. The actual topographic data is stored in the form of a vector map, which includes the geographical coordinates and boundary information of topographic features such as field ridges, ditches, and roads; the crop planting plan data includes the boundary coordinates of planting areas for different crops, crop types, and planting times.

[0150] The acquired data is formatted to be consistent with the coordinate system of the semantic association map of farmland areas, ensuring that terrain features and planting boundaries can be accurately matched to farmland areas.

[0151] Step S1342: Convert the actual terrain data and crop planting planning data into the same coordinate system as the semantic association map of the farmland area, so that the terrain features, planting boundaries and preliminary farmland sub-regions are displayed under the same coordinate system.

[0152] A coordinate transformation algorithm was used to convert the coordinate systems of actual terrain data and crop planting planning data into the coordinate system used by the semantic association map of farmland areas. During the transformation process, multiple common ground control points were selected, and their coordinate values ​​in both coordinate systems were obtained. The coordinate transformation parameters were then calculated through fitting.

[0153] The coordinate transformation parameters are used to transform each coordinate point in the actual terrain data and crop planting planning data to ensure that the transformed terrain features, planting boundaries and preliminary farmland sub-regions can be accurately displayed on the same coordinate plane, avoiding positional offsets or misalignments.

[0154] Step S1343: Traverse each of the preliminary farmland sub-regions and check whether its boundary intersects with the field ridges and ditch locations in the actual terrain data. If there is an intersection, mark it as a boundary segment that needs to be adjusted.

[0155] Each preliminary farmland sub-region is traversed one by one, and its boundary coordinates are compared with the boundary coordinates of terrain features such as field ridges and ditches in the actual terrain data. If a section of the boundary of a preliminary farmland sub-region overlaps with the boundary of a field ridge or ditch, or the intersection angle is less than a preset angle, then it is determined that the boundary section intersects with the terrain feature and is marked as a boundary section that needs to be adjusted.

[0156] For example, if the boundary of a preliminary farmland sub-region crosses a field ridge diagonally, then the boundary segment that crosses the field ridge is the boundary segment that needs to be adjusted.

[0157] Step S1344: For the marked boundary segments that need adjustment, move the boundary segments to the nearest field ridge or ditch location so that the boundary of the initial farmland sub-region coincides with the field ridge or ditch.

[0158] For each marked boundary segment requiring adjustment, calculate the distance between that boundary segment and the boundaries of surrounding field ridges and ditches, and select the nearest topographic feature boundary as the adjustment target. Move the boundary segment of the initial farmland sub-region to that topographic feature boundary so that the two completely overlap.

[0159] During the adjustment process, if the area of ​​the initial farmland sub-region changes by more than the preset ratio after the boundary segment is moved, the adjacent boundary segments will be adjusted appropriately to ensure that the overall area of ​​the initial farmland sub-region is basically consistent with that before the adjustment, while ensuring that the boundary coincides with the terrain features.

[0160] Step S1345: Check whether each of the preliminary farmland sub-regions crosses the crop planting area boundary in the crop planting planning data. If it does cross, the preliminary farmland sub-region is divided into multiple sub-regions according to the planting area boundary. Each sub-region contains only a single crop planting area.

[0161] The boundary coordinates of the preliminary farmland sub-region are overlaid and compared with the boundary coordinates of the crop planting area in the crop planting planning data. Each segment is checked to see if the boundary of the preliminary farmland sub-region intersects or overlaps with the boundary of the planting area of ​​different crops. If the scope of the preliminary farmland sub-region covers the planting areas of two or more crops, it is determined that the preliminary farmland sub-region crosses the boundary of the crop planting area.

[0162] For preliminary farmland sub-regions determined to cross planting area boundaries, the sub-region is divided into multiple independent sub-regions, using the crop planting area boundary as the dividing line. During the division process, it is ensured that the boundary of each sub-region is consistent with the corresponding crop planting area boundary, and that each sub-region contains only one type of crop planting area. Simultaneously, the spraying demand level, semantic features, and other attribute information of the original preliminary farmland sub-region are inherited to the divided sub-regions, ensuring the continuity of attribute information.

[0163] Step S1346: Perform boundary smoothing on the adjusted farmland sub-regions, merge isolated small regions in the adjusted farmland sub-regions into adjacent farmland sub-regions to ensure the continuity of geographical location within each farmland sub-region, and record the boundary coordinate changes of each farmland sub-region before and after adjustment to form a boundary adjustment record.

[0164] After matching and adjusting the boundaries with terrain features and planting area boundaries, all adjusted farmland sub-regions undergo boundary smoothing. A polygon smoothing algorithm is used to smooth sharp corners on the boundaries of farmland sub-regions, making the boundary lines more closely resemble the natural shape of the actual farmland terrain.

[0165] Traverse all adjusted farmland sub-regions and identify isolated small regions with an area smaller than a preset threshold. These isolated small regions are usually formed due to boundary adjustments or splits and are geographically adjacent to larger surrounding farmland sub-regions. Merge the boundaries of the isolated small regions with the boundaries of the adjacent larger farmland sub-regions to integrate the isolated small regions into the adjacent farmland sub-regions, ensuring that the geographical location within the merged farmland sub-regions is continuous without any breaks or isolated parts.

[0166] Throughout the boundary adjustment process, the original boundary coordinates before adjustment and the new boundary coordinates after adjustment for each farmland sub-region are recorded in real time. At the same time, the reasons for adjustment (such as matching field ridges, splitting planting areas, etc.), adjustment time and operator information are also recorded to form a complete boundary adjustment record, which is stored in the farmland management database for subsequent query and traceability.

[0167] Step S135: Calculate the area and center coordinates of each adjusted farmland sub-region, and record the spraying requirement level corresponding to each farmland sub-region.

[0168] Based on the adjusted boundary coordinates of the farmland sub-regions, the area of ​​each farmland sub-region is calculated using the polygon area calculation formula. During the calculation, the boundary coordinates are connected sequentially to form a closed polygon. The area enclosed by each side of the polygon and the coordinate axes is calculated using the x and y coordinate values ​​of the coordinate points. Finally, the actual area of ​​the closed polygon is obtained through algebraic summation.

[0169] Simultaneously, the center coordinates of the farmland sub-region are calculated based on its boundary coordinates. The average of the x-coordinates and y-coordinates of all boundary points is used as the center coordinates of the farmland sub-region. If the farmland sub-region has an irregular shape, a weighted average method can be used to calculate the center coordinates. This involves assigning weights based on the importance of the boundary points within the polygon, and then calculating the weighted average of the x-coordinates and y-coordinates.

[0170] The area size and center coordinates of each farmland sub-region are associated and stored with its corresponding spraying requirement level, forming a basic information table of farmland sub-regions containing location, area and requirement level.

[0171] Step S136: Assign a spraying priority to each farmland sub-region according to the spraying demand level, and assign a spraying priority that matches the level of the farmland sub-region with a higher spraying demand level.

[0172] Establish a rule for the correspondence between spraying demand levels and spraying priorities, where "emergency spraying" corresponds to the highest priority, "priority spraying" corresponds to the second highest priority, "routine spraying" corresponds to the medium priority, "low demand spraying" corresponds to the lower priority, and "no spraying required" corresponds to the lowest priority (i.e., no spraying task is assigned).

[0173] The basic information table of farmland sub-regions is traversed, and spraying priorities are assigned to each sub-region according to the spraying demand level and corresponding relationship rules. For multiple farmland sub-regions with the same spraying demand level, the priorities are further fine-tuned based on their area size and center location coordinates: farmland sub-regions with larger areas have slightly higher priorities than those with smaller areas; farmland sub-regions whose center location coordinates are closer to the farmland entrance / exit or the drone take-off / landing point can have their priorities appropriately advanced to improve spraying efficiency.

[0174] The assigned spraying priorities are added to the basic information table of farmland sub-regions to form a complete sub-region information dataset containing priority information.

[0175] Step S137: Integrate the area size, center coordinates, spraying demand level and spraying priority of each farmland sub-region to generate a spraying area division scheme with graphical annotations. The graphical annotations can intuitively display the location and attribute information of each farmland sub-region on the farmland map.

[0176] The data in the basic information table of farmland sub-regions is linked with the farmland satellite map, and the corresponding area outline is drawn on the satellite map according to the boundary coordinates of each farmland sub-region. Different colors are used to fill and mark farmland sub-regions with different spraying demand levels. For example, "emergency spraying" is filled with red, "priority spraying" with orange, "routine spraying" with yellow, "low demand spraying" with blue, and "no spraying required" with gray.

[0177] The area size, spraying priority, and center coordinates of each farmland sub-region are marked at its center. The font size of the markings is adjusted according to the area of ​​the farmland sub-region to ensure that the markings are clearly visible and do not obscure the area outline. At the same time, a legend is set at the edge of the map to explain the correspondence between colors and spraying requirement levels and priorities, so that users can quickly understand the graphical markings.

[0178] The completed graphical farmland sub-region map is integrated with the basic information table of farmland sub-regions to form a complete spraying area division scheme. This spraying area division scheme is stored in file format and supports operations such as zooming, panning, and attribute information querying on the display interface of the ground processing terminal.

[0179] Step S140: Determine the pesticide spraying parameters corresponding to each farmland sub-region according to the spraying area division scheme, and integrate the farmland sub-region information and the corresponding pesticide spraying parameters to generate a drone pesticide spraying command.

[0180] Based on the attribute information of each farmland sub-region in the spraying area division scheme, and combined with the preset parameter calculation rules, the pesticide spraying parameters of each sub-region are determined, and the above parameters and sub-region information are integrated into spraying instructions that conform to the UAV execution standards.

[0181] Step S141: Extract the spraying demand level and area size of each farmland sub-region from the spraying area division scheme.

[0182] The system reads the identification number, spraying requirement level, and area size of each farmland sub-region from the basic information table of the spraying area division scheme through the data interface. The data is then sorted according to the identification number to form a structured sub-region parameter calculation dataset, which provides basic data support for the subsequent calculation of various spraying parameters.

[0183] During the extraction process, the data is validated to ensure that the spraying requirement level matches the preset level type and that the area size is positive and within a reasonable range. If any data anomalies are found, the process returns to the spraying area division scheme generation step for verification and correction to ensure the accuracy of the extracted data.

[0184] Step S142: Determine the pesticide concentration parameters corresponding to each farmland sub-region based on the preset correspondence between spraying requirement level and pesticide concentration.

[0185] A pre-established table mapping spraying demand levels to pesticide concentration parameters is created, based on the pest and disease control standards, growth stage characteristics, and pesticide type characteristics of different crops. For example, "emergency spraying" corresponds to the highest concentration parameter, "priority spraying" corresponds to a relatively high concentration parameter, "routine spraying" corresponds to the standard concentration parameter, and "low-demand spraying" corresponds to a relatively low concentration parameter.

[0186] For each farmland sub-region, the corresponding pesticide concentration parameters are initially determined by consulting a table based on its spraying requirement level. If a specific type of pest or disease exists in the crop within that sub-region (this information is extracted from crop type features and crop growth status parameters in semantic features), the initially determined pesticide concentration parameters are fine-tuned according to the pest or disease type. For example, for a specific disease, the concentration parameter is appropriately increased by a preset ratio based on the concentration for the corresponding requirement level to ensure control effectiveness.

[0187] The final determined pesticide concentration parameters are associated with the identification numbers of the farmland sub-regions and stored to form a pesticide concentration parameter table.

[0188] Step S143: Calculate the required amount of pesticide spraying based on the area size of each farmland sub-region, and fine-tune the amount of pesticide spraying based on the spraying demand level. The area size is positively correlated with the amount of pesticide spraying.

[0189] The basic pesticide spraying amount for each farmland sub-region is calculated by multiplying the unit area spraying amount by the area size. The unit area spraying amount is preset according to the crop type, growth stage, and pesticide type. For example, for gramineous crops in the tillering stage, the unit area spraying amount is set to a fixed standard value.

[0190] Based on the basic pesticide spraying amount, fine adjustments are made according to the spraying demand level: for farmland sub-areas at the "emergency spraying" level, the spraying amount is increased by a preset ratio on the basis of the basic spraying amount; for the "priority spraying" level, the amount is increased by a slightly lower ratio than that of "emergency spraying"; for the "routine spraying" level, the basic spraying amount remains unchanged; and for the "low demand spraying" level, the basic spraying amount is reduced by a preset ratio.

[0191] After the calculation is completed, the pesticide spraying amount of each farmland sub-region is associated with the identification number and added to the pesticide concentration parameter table to form a pesticide parameter dataset containing concentration and spraying amount.

[0192] Step S144: Determine the spraying height parameters for each farmland sub-region based on the average crop height data, and query the average crop height data at the corresponding location according to the center coordinates of the farmland sub-region.

[0193] Beforehand, a lidar device mounted on a drone is used to scan contiguous farmland to obtain crop height data corresponding to each geographical coordinate point within the farmland, thus establishing a crop height database. This crop height database stores the average crop height value at the corresponding geographical coordinates.

[0194] For each farmland sub-region, the average crop height corresponding to its center location is obtained by querying the crop height database based on its center location coordinates. If there is no direct height data at the center location coordinates, the average crop height of the sub-region is calculated by querying the crop height data of multiple surrounding coordinate points.

[0195] The spraying height parameters are determined based on the average crop height data. The spraying height parameters are set to the average crop height plus a preset safety distance. The safety distance is determined based on the spraying range of the drone nozzle and the crop growth density to ensure that the pesticide can evenly cover the top of the crop without damaging it.

[0196] Associate the spraying height parameter with the farmland sub-region identifier number and add it to the pesticide parameter dataset.

[0197] Step S145: Determine the spraying speed parameters for each farmland sub-region based on its area size and spraying priority.

[0198] A baseline spraying speed parameter is set, which is determined based on the drone's nozzle flow rate, spray width, and pesticide type. For each farmland sub-region, adjustments are made based on the baseline spraying speed parameter, combined with the area size and spraying priority.

[0199] For larger farmland sub-areas, to improve operational efficiency while ensuring spray uniformity, the spraying speed parameters should be appropriately increased. For smaller farmland sub-areas, to avoid missed or overlapping spraying, the spraying speed parameters should be appropriately decreased. Simultaneously, a secondary adjustment should be made based on spraying priority: for farmland sub-areas with higher priority, the spraying speed parameters can be appropriately decreased to ensure spraying quality; for farmland sub-areas with lower priority, the speed parameters can be appropriately increased to shorten the overall operation time.

[0200] The adjusted spraying speed parameters must be within the adjustable speed range of the drone. If the adjusted speed exceeds the range, the maximum or minimum controllable speed of the drone will be used as the final spraying speed parameter. The determined spraying speed parameters will be associated with the identification number and added to the pesticide parameter dataset.

[0201] For example, firstly, a basic system for calculating spraying speed parameters is established, which includes three core modules: a basic spraying speed benchmark value, an area adjustment coefficient library, and a priority calibration factor library. The three work together to achieve accurate determination of spraying speed.

[0202] The baseline spraying speed needs to be set by considering both the drone's hardware parameters and the pesticide's characteristics. The drone's hardware parameters include the effective spray width of the nozzle, the flow rate per nozzle, and the maximum controllable flight speed. Pesticide characteristics include the surface tension of the pesticide solution and the required atomized particle size. The baseline spraying speed is determined by calculating the ratio of the effective spray width to the flow rate per nozzle, and then combining this with the minimum flight speed threshold required for pesticide atomization. For example, if the effective spray width and the flow rate per nozzle are fixed, the baseline spraying speed, calculated using their ratio and the atomization threshold, is set to the midpoint of a fixed range.

[0203] The construction of the area adjustment coefficient library requires dividing farmland sub-regions into multiple intervals based on their area size, with each interval corresponding to a unique area adjustment coefficient. The division of area intervals is determined based on the common sub-region area distribution characteristics of contiguous farmland, such as dividing them into extremely small area intervals, small area intervals, medium area intervals, large area intervals, and extremely large area intervals. The area adjustment coefficient for each interval is obtained by fitting historical operational data, following the principle of "the larger the area, the larger the adjustment coefficient," meaning that the larger the sub-region, the closer the adjustment coefficient is to 1.2, and the smaller the sub-region, the closer the adjustment coefficient is to 0.8. For each farmland sub-region, based on its calculated area size, the corresponding interval in the area adjustment coefficient library is matched, and the area adjustment coefficient for that interval is extracted. The baseline spraying speed value is multiplied by the area adjustment coefficient to obtain the initial spraying speed value after area adaptation.

[0204] The construction of the priority calibration factor library requires five priorities corresponding to the spraying demand levels, with each priority corresponding to a unique calibration factor. Priorities decrease sequentially, with calibration factors increasing from 0.9 to 1.1. Specifically, the highest priority corresponds to a calibration factor of 0.9, the second highest to 0.95, the medium priority to 1.0, the lower priority to 1.05, and the lowest priority to 1.1. Based on the spraying priority of each farmland sub-region, the corresponding calibration factor is extracted from the priority calibration factor library. The initial spraying rate value after area adaptation is multiplied by the calibration factor to obtain the priority-calibrated candidate spraying rate value.

[0205] After completing the initial calculations, parameter boundary verification is required. The maximum and minimum controllable flight speeds of the drone are obtained as boundary thresholds. The calibrated candidate spraying speed values ​​are compared with these boundary thresholds. If the candidate value is between the maximum and minimum controllable flight speeds, it is directly determined as the final spraying speed parameter. If the candidate value is higher than the maximum controllable flight speed, the maximum controllable flight speed is used as the final spraying speed parameter, and the parameter exceeding the limit information is recorded, indicating that the spray nozzle flow rate should be adjusted or multiple sprays should be applied to ensure operational quality. If the candidate value is lower than the minimum controllable flight speed, the minimum controllable flight speed is used as the final spraying speed parameter, and the parameter exceeding the limit information is recorded, indicating that the spray width should be reduced to avoid excessive pesticide accumulation.

[0206] Finally, the determined spraying speed parameters are associated with the identification number, area size, and spraying priority of the farmland sub-regions and stored to form a detailed ledger of spraying speed parameters. At the same time, records of parameters exceeding limits are synchronized to the abnormal log of the farmland management system.

[0207] Step S146: Integrate the center coordinates, pesticide concentration parameters, pesticide spraying amount, spraying height parameters, and spraying speed parameters of each farmland sub-region into a set of pesticide spraying parameters corresponding to that farmland sub-region.

[0208] According to the identification number of the farmland sub-region, the center position coordinates of each sub-region are extracted from the basic information table of the spraying area division scheme and integrated with the corresponding pesticide concentration parameters, pesticide spraying amount, spraying height parameters and spraying speed parameters in the pesticide parameter dataset.

[0209] During integration, a structured data format is adopted. Each pesticide spraying parameter set contains six fields: identifier number, location information (center coordinates), concentration, spray volume, height, and speed, ensuring that the data type and format of each field are consistent. For example, the center position coordinates are stored as a string of "x-axis, y-axis", the concentration is stored as a mass-volume ratio, and the spray volume is stored in volume units.

[0210] The pesticide spraying parameters of all farmland sub-regions are collected and summarized to form a master table of pesticide spraying parameters, which is used to generate subsequent spraying instructions.

[0211] Step S147: Sort the pesticide spraying parameter sets of all the farmland sub-regions according to the spraying priority order to form an ordered spraying parameter sequence.

[0212] The spraying priority of each farmland sub-region is extracted from the basic information table of the spraying area division scheme and associated with the parameter set in the pesticide spraying parameter master table through an identification number. The pesticide spraying parameter set is sorted in descending order of spraying priority, and parameter sets with the same priority are further sorted in ascending order of distance from the center coordinates to the UAV take-off and landing point.

[0213] After sorting, an ordered sequence of spraying parameters is formed. Each element in the sequence represents a set of pesticide spraying parameters for a specific farmland sub-region. The order of the sequence determines the order in which the drone performs the spraying operation. This sorting ensures that higher-priority farmland sub-regions receive spraying treatment first, while also optimizing flight paths and reducing the drone's unnecessary flight distance.

[0214] Step S148: Combine the ordered spraying parameter sequence with farmland map coordinate information to generate a drone pesticide spraying command containing time sequence, location coordinates and parameter information.

[0215] The ordered sequence of spraying parameters is calibrated against the coordinate system of the farmland satellite map to ensure that the center coordinates of each parameter set in the sequence accurately correspond to the map coordinates. Based on the drone's flight speed, spraying speed, and the size of each farmland sub-region, the estimated operation time of the drone in each farmland sub-region, as well as the flight time from the current sub-region to the next sub-region, are calculated.

[0216] Following the sequence of spraying parameters, the estimated start time, center coordinates, pesticide concentration parameters, pesticide spraying volume, spraying height parameters, spraying speed parameters, and estimated operation duration for each farmland sub-region are encoded according to a command format recognizable by the UAV control system. The command format uses a preset binary or text format, including a command header, command body, and checksum. The command header identifies the command type, the command body contains specific parameter information, and the checksum ensures data integrity during command transmission.

[0217] The encoded instructions are combined in sequence to form a complete drone pesticide spraying instruction. This drone pesticide spraying instruction can clearly tell the drone when, where, and with what parameters to perform pesticide spraying operations.

[0218] Step S150: The drone pesticide spraying command is sent to the drone control system. After receiving the drone pesticide spraying command, the drone control system controls the drone to perform the pesticide spraying operation.

[0219] The generated pesticide spraying commands from the drone are sent to the drone's control system via a wireless communication module. During transmission, a data fragmentation method is used, where the large command file is split into multiple data fragments and transmitted sequentially. After each fragment is transmitted, the drone control system returns a confirmation message to the ground processing terminal. The ground processing terminal then transmits the next fragment after receiving the confirmation message, until all command data transmission is complete.

[0220] After receiving all command data, the UAV control system verifies and decodes the commands, checking their integrity and validity. If the verification passes, the control system parses the time sequence, position coordinates, and parameter information from the commands. Combining this with its own satellite positioning module, attitude sensor, and nozzle control module, the system plans a specific flight path according to the command requirements.

[0221] During flight, the drone adjusts its altitude according to the spraying height parameters in the instructions. After reaching the center of the target farmland sub-area, it controls the nozzles to open according to the spraying speed and pesticide concentration parameters, and controls the spraying duration according to the pesticide spraying volume. After completing the spraying operation in that sub-area, it automatically flies to the next sub-area to continue the operation. During the operation, the drone control system transmits information such as flight status and spraying parameter execution status back to the ground processing terminal in real time, so that operators can monitor the operation progress in real time.

[0222] Figure 2 The illustration shows exemplary hardware and software components of an image recognition-integrated drone pesticide spraying area identification system 100, which can implement the ideas of this application, according to some embodiments of this application. For example, a processor 120 can be used in the image recognition-integrated drone pesticide spraying area identification system 100 and to perform the functions in this application.

[0223] The drone pesticide spraying area identification system 100, which combines image recognition, can be a general-purpose server or a special-purpose server. Both can be used to implement the drone pesticide spraying area identification method combining image recognition of this application. Although only one server is shown in this application, for convenience, the functions described in this application can be implemented in a distributed manner on multiple similar platforms to balance the processing load.

[0224] For example, the drone pesticide spraying area identification system 100 incorporating image recognition may include a network port 110 connected to a network, one or more processors 120 for executing program instructions, a communication bus 130, and various forms of storage media 140, such as a disk, ROM, or RAM, or any combination thereof. Exemplarily, the drone pesticide spraying area identification system 100 incorporating image recognition may also include program instructions stored in ROM, RAM, or other types of non-transitory storage media, or any combination thereof. The methods of this application can be implemented according to these program instructions. The drone pesticide spraying area identification system 100 incorporating image recognition also includes an I / O interface 150 between the computer and other input / output devices.

[0225] For ease of explanation, only one processor is described in the image recognition-integrated drone pesticide spraying area recognition system 100. However, it should be noted that the image recognition-integrated drone pesticide spraying area recognition system 100 of this application may also include multiple processors. Therefore, the steps performed by one processor described in this application may also be performed jointly or individually by multiple processors. For example, if the processor of the image recognition-integrated drone pesticide spraying area recognition system 100 performs steps A and B, it should be understood that steps A and B may also be performed jointly by two different processors or individually by one processor. For example, the first processor performs step A, the second processor performs step B, or the first processor and the second processor jointly perform steps A and B.

[0226] Furthermore, this embodiment of the invention also provides a readable storage medium, which has computer-executable instructions pre-set in it. When the processor executes the computer-executable instructions, the above-mentioned method for identifying pesticide spraying areas by drones combined with image recognition is realized.

[0227] It should be noted that, in order to simplify the description of the present invention and thus help to understand one or more embodiments of the invention, multiple features may sometimes be grouped into one embodiment, drawing or description thereof in the foregoing description of the embodiments of the present invention.

Claims

1. A method for identifying pesticide spraying areas using unmanned aerial vehicles (UAVs) by combining image recognition, characterized in that, The method includes: A set of farmland images collected by a drone is obtained, and regional semantic association analysis is performed on the set of farmland images to generate a farmland regional semantic association map. The set of farmland images includes farmland scene images taken at different times. An image recognition model is invoked to identify the crop growth status of crop images in the farmland image set, and the crop growth status identification result is obtained. A spraying area adaptation model is constructed by combining the semantic association map of the farmland area. Based on the spraying area adaptation model, the spraying demand of farmland areas is assessed, farmland sub-regions with different spraying needs are divided, and a spraying area division scheme is generated. Based on the spraying area division scheme, determine the pesticide spraying parameters corresponding to each farmland sub-region, and integrate the farmland sub-region information with the corresponding pesticide spraying parameters to generate UAV pesticide spraying instructions; The drone pesticide spraying command is sent to the drone control system. After receiving the drone pesticide spraying command, the drone control system controls the drone to perform the pesticide spraying operation. The step involves calling an image recognition model to identify the crop growth status of crop images in the farmland image set, obtaining crop growth status identification results, and constructing a spraying area adaptation model based on the farmland area semantic association map, including: The crop image portion is extracted from each farmland scene image in the farmland image set. The crop area and non-crop area in the farmland scene image are separated by an image segmentation algorithm. The crop image portion is the image content corresponding to the crop area. The crop image portion is input into a pre-trained image recognition model, which includes a feature extraction layer and a feature classification layer. The feature extraction layer extracts features from the crop image portion to obtain crop growth features, and the feature classification layer classifies the crop growth features to obtain crop growth status categories. Crop growth status parameters are determined based on the crop growth status category. These parameters include crop growth stage parameters and crop health status parameters, with each parameter corresponding to a specific crop growth status category. The crop growth status parameters are associated with the nodes in the semantic association graph of the farmland area to form an association graph with growth status information. The actual small area corresponding to each node is matched with the corresponding crop growth status parameters. Extract the node relationships and crop growth state parameters corresponding to the nodes from the association graph with growth state information, and use them as input features for the spraying area adaptation model. The core logic layer for constructing the spraying area adaptation model adopts a decision tree structure, using the node association relationship in the input features as the branch judgment condition and the crop growth state parameter as the basis for assigning branch weights. The decision tree structure is hierarchically optimized by adjusting the order of branch judgments, so that the association relationship of nodes related to spraying needs is given priority as the judgment condition. The mapping relationship between the optimized decision tree structure and the input features is solidified to form a spraying area adaptation model that can output spraying adaptation results based on the semantic association of farmland areas and crop growth status.

2. The method for identifying pesticide spraying areas using unmanned aerial vehicles (UAVs) combined with image recognition according to claim 1, characterized in that, The process of acquiring a set of farmland images collected by a drone, performing regional semantic association analysis on the farmland image set, and generating a farmland regional semantic association map includes: The system receives a set of farmland images transmitted by a drone through its onboard image acquisition equipment. Each farmland scene image in the set contains information about the shooting time and location. Each farmland scene image in the farmland image set is segmented into multiple farmland image blocks of the same size, and each farmland image block corresponds to a small actual area in the farmland. Semantic features are extracted from each of the farmland image blocks. The semantic features include crop type features, soil condition features, and vegetation cover features. Each semantic feature is extracted from the color information and texture information of the farmland image block. Based on the shooting location information of each farmland scene image, the farmland image blocks from different shooting time periods are associated according to the actual geographical location correspondence, and the semantic feature similarity between the farmland image blocks with corresponding geographical locations is calculated; Using the actual small region corresponding to each farmland image block as a node, and the semantic feature similarity as the association strength between nodes, an initial association graph structure is constructed. Redundant nodes are removed from the initial association graph structure, deleting nodes that correspond to the same actual small region repeatedly, and retaining the unique node corresponding to each actual small region. Based on the changing trends of the semantic features at different shooting times, the correlation strength values ​​between nodes are adjusted so that the correlation strength can reflect the semantic correlation changes of farmland areas at different times. The adjusted association graph structure is visualized, and the actual small area location information and semantic feature type corresponding to each node are labeled to form a complete semantic association graph of farmland area.

3. The method for identifying pesticide spraying areas using unmanned aerial vehicles (UAVs) combined with image recognition according to claim 2, characterized in that, The extraction of semantic features from each of the farmland image blocks includes: Each of the farmland image blocks is subjected to color space conversion processing, converting the farmland image block from RGB color space to HSV color space, and obtaining the hue parameter, saturation parameter and brightness parameter of the farmland image block. The hue parameter, the saturation parameter and the brightness parameter together constitute color information. The farmland image patch is converted to grayscale by performing grayscale processing. The texture feature parameters of the grayscale image patch are calculated by the grayscale co-occurrence matrix algorithm. The texture feature parameters include contrast parameters and correlation parameters. The contrast parameters and correlation parameters together constitute the texture information. For crop type feature extraction, hue parameters are selected from color information as the main judgment criterion, and correlation parameters in texture information are combined to assist in the judgment, thereby determining the crop type features in the farmland image block and forming crop type features. Different crop types correspond to specified hue parameter ranges. For soil condition feature extraction, the brightness parameter is selected from color information as the main analysis basis, and the contrast parameter in texture information is combined to determine the roughness of soil particles. The soil condition features are determined comprehensively, and different brightness parameters correspond to different soil moisture levels. For vegetation cover feature extraction, the proportion of green pixels in the farmland image block is calculated, and non-crop green areas are excluded by combining hue parameters to determine vegetation cover features. The proportion of green pixels reflects the density of vegetation cover. The extracted crop type features, soil condition features, and vegetation cover features are integrated, and corresponding extraction basis annotations are added to each feature. The annotation content includes the color information parameters and texture information parameters used, forming a complete semantic feature.

4. The method for identifying pesticide spraying areas using unmanned aerial vehicles (UAVs) combined with image recognition according to claim 2, characterized in that, The step of adjusting the correlation strength values ​​between nodes based on the changing trends of the semantic features during different shooting periods includes: The farmland image set is arranged in the order of shooting time to form a time-series farmland image sequence, with each time period corresponding to one or more farmland scene images; For each node, the semantic features of the node are extracted at different shooting times to form a temporal sequence of semantic features of the node. Calculate the change in semantic features between adjacent time periods in the semantic feature time sequence. The change is obtained by the degree of difference between the semantic features of the current time period and the semantic features of the previous time period. The greater the degree of difference, the greater the change. Statistically analyze the semantic feature changes of all nodes, determine the average level of semantic feature changes, compare the semantic feature changes of each node with the average level, and determine whether the semantic feature changes of that node meet the set conditions. For nodes whose semantic feature changes meet the set conditions, analyze the correlation between their semantic feature changes and those of adjacent nodes. If the semantic feature change trends of adjacent nodes are consistent, increase the correlation strength value between the nodes. For nodes whose semantic feature changes do not meet the set conditions, keep their association strength with neighboring nodes unchanged, or make a uniform adjustment based on the average association strength of all nodes. Based on the seasonal information corresponding to the shooting time, the adjustment range of the association strength is corrected. The adjustment range of the association strength corresponding to the semantic feature changes in the peak growing season is greater than that in the non-peak growing season. Calculate the difference between the adjusted node association strength values ​​and the original values, ensuring the difference is within a preset range. Then, update the adjusted association strength values ​​to the initial association graph structure, replacing the original association strength values, thus forming an association graph structure that reflects the temporal semantic association changes.

5. The method for identifying pesticide spraying areas by drones combined with image recognition according to claim 1, characterized in that, The crop image portion is input into a pre-trained image recognition model. The image recognition model includes a feature extraction layer and a feature classification layer. The feature extraction layer extracts features from the crop image portion to obtain crop growth features, and the feature classification layer classifies the crop growth features to obtain crop growth state categories, including: The crop image portions are adjusted to a uniform size so that all crop image portions input into the image recognition model have the same pixel dimension; The resized portion of the crop image is input into the feature extraction layer of the image recognition model. The feature extraction layer adopts a convolutional neural network structure and extracts features from the crop image portion through multiple convolution operations. A pooling layer is set after each convolutional layer of the feature extraction layer. The pooling layer uses max pooling to downsample the feature map output by the convolutional layer. The feature vector output by the last pooling layer of the feature extraction layer is used as the crop growth feature, which includes crop leaf morphology features, crop color distribution features, and crop density features. The crop growth characteristics are input into the feature classification layer of the image recognition model, and the feature classification layer adopts a fully connected neural network structure. A Softmax activation function is set in the last layer of the feature classification layer. The probability value of the crop growth feature belonging to different crop growth state categories is calculated by the Softmax activation function. The crop growth state category with the highest probability value is selected as the crop growth state category corresponding to the crop image part. At the same time, the probability value is recorded as the classification confidence.

6. The method for identifying pesticide spraying areas by drones combined with image recognition according to claim 1, characterized in that, The process of assessing the spraying demand of farmland areas based on the spraying area adaptation model, dividing farmland sub-regions with different spraying needs, and generating a spraying area division scheme includes: Input all nodes in the semantic association graph of the farmland area and the corresponding crop growth status parameters into the spraying area adaptation model; The core logic layer of the spraying area adaptation model determines the spraying demand of each node and determines the actual spraying demand level of each node's small area based on the node's association relationship and crop growth status parameters. The number of nodes with the same spraying demand level and their corresponding actual small area locations are counted. The actual small areas that are geographically adjacent and have the same spraying demand level are merged to form preliminary farmland sub-regions. The boundaries of the preliminary farmland sub-regions are adjusted so that the boundaries of each preliminary farmland sub-region are consistent with the actual terrain features or crop planting boundaries in the farmland. Calculate the area and center coordinates of each adjusted farmland sub-region, and record the spraying requirement level corresponding to each farmland sub-region; Spraying priorities are assigned to each of the farmland sub-regions based on the spraying demand level, and farmland sub-regions with higher spraying demand levels are assigned spraying priorities that match the level. The area size, center coordinates, spraying demand level, and spraying priority of each farmland sub-region are integrated to generate a spraying area division scheme with graphical annotations. The graphical annotations can intuitively display the location and attribute information of each farmland sub-region on the farmland map.

7. The method for identifying pesticide spraying areas by drones combined with image recognition according to claim 6, characterized in that, The step of adjusting the boundaries of the preliminary farmland sub-regions to ensure that the boundaries of each preliminary farmland sub-region are consistent with the actual terrain features or crop planting boundaries in the farmland includes: Acquire actual terrain data and crop planting planning data for farmland. The actual terrain data includes the location of field ridges and ditches in the farmland, and the crop planting planning data includes the boundaries of planting areas for different crops. The actual terrain data and crop planting planning data are converted into the same coordinate system as the semantic association map of the farmland area, so that the terrain features, planting boundaries and preliminary farmland sub-regions are displayed under the same coordinate system; Traverse each of the preliminary farmland sub-regions and check whether its boundaries intersect with the positions of field ridges and ditches in the actual terrain data. If there is an intersection, mark it as a boundary segment that needs to be adjusted. For the marked boundary segments that need adjustment, move the boundary segments to the nearest field ridge or ditch location so that the boundary of the initial farmland sub-region coincides with the field ridge or ditch. Check whether each of the preliminary farmland sub-regions crosses the crop planting area boundary in the crop planting planning data. If it does cross, the preliminary farmland sub-region is divided into multiple sub-regions according to the planting area boundary. Each sub-region contains only a single crop planting area. The adjusted farmland sub-regions are subjected to boundary smoothing processing. Isolated small areas in the adjusted farmland sub-regions are merged into adjacent farmland sub-regions to ensure the continuity of geographical location within each farmland sub-region. The boundary coordinate changes of each farmland sub-region before and after adjustment are recorded to form a boundary adjustment record.

8. The method for identifying pesticide spraying areas by drones combined with image recognition according to claim 1, characterized in that, The step of determining the pesticide spraying parameters corresponding to each farmland sub-region according to the spraying area division scheme, and integrating the farmland sub-region information with the corresponding pesticide spraying parameters to generate UAV pesticide spraying instructions includes: Extract the spraying requirement level and area size of each farmland sub-region from the spraying area division scheme; Based on the preset correspondence between spraying demand levels and pesticide concentrations, determine the pesticide concentration parameters corresponding to each of the farmland sub-regions; The required amount of pesticide spraying is calculated based on the area of ​​each farmland sub-region, and the amount of pesticide spraying is fine-tuned in combination with the level of spraying demand. The area size is positively correlated with the amount of pesticide spraying. The spraying height parameters for each farmland sub-region are determined based on the average crop height data, and the average crop height data at the corresponding location is queried based on the center coordinates of the farmland sub-region. The spraying speed parameters for each farmland sub-region are determined based on the area size and spraying priority of the farmland sub-region. The center coordinates, pesticide concentration parameters, pesticide spraying amount, spraying height parameters, and spraying speed parameters of each farmland sub-region are integrated into a set of pesticide spraying parameters corresponding to that farmland sub-region. The pesticide spraying parameter sets of all the farmland sub-regions are sorted according to the spraying priority to form an ordered spraying parameter sequence; By combining the ordered sequence of spraying parameters with farmland map coordinates, a drone pesticide spraying command containing time sequence, location coordinates, and parameter information is generated.

9. A drone pesticide spraying area identification system combining image recognition, characterized in that, The image recognition-integrated drone pesticide spraying area identification system includes a processor and a memory, the memory and the processor being connected. The memory is used to store programs, instructions or code, and the processor is used to execute the programs, instructions or code in the memory to implement the image recognition-integrated drone pesticide spraying area identification method according to any one of claims 1-8.

Images