An intelligent agricultural crop growth monitoring method based on image processing

Abstract

This invention discloses a smart agricultural crop growth monitoring method based on image processing, comprising the following steps: S1, acquiring and preprocessing target crop images at a preset frequency to construct a crop image sequence; S2, extracting crop regions to construct a crop region sequence; S3, extracting leaf morphology, color distribution, and texture structure features to generate a crop feature sequence; S4, extracting the spatial coordinates of the crop regions and performing spatial calculations to obtain a growth parameter sequence; S5, constructing a growth state sequence based on the crop feature sequence and the growth parameter sequence; S6, performing two-dimensional empirical mode decomposition and statistical analysis on the growth state sequence to construct a growth trend sequence; S7, classifying the growth trend sequence using a support vector machine to generate crop growth monitoring results. This invention comprehensively utilizes DeepLabV3+ models and other technologies, possessing advantages such as high monitoring accuracy, high automation, and strong environmental adaptability.

Description

Technical Field

[0001] This invention relates to the field of image processing technology, and in particular to a smart agricultural crop growth monitoring method based on image processing. Background Technology

[0002] With the development of smart agriculture, agricultural production is gradually shifting towards digitalization and intelligence. Crop growth status monitoring is a crucial aspect of agricultural production management. By monitoring crop morphological characteristics, growth parameters, and growth trends, it provides a basis for irrigation management, fertilization management, and pest and disease control, thereby improving agricultural production efficiency. Therefore, acquiring images of farmland crops using image acquisition equipment and analyzing crop growth status through image processing technology is gradually becoming an important research direction in the field of smart agriculture.

[0003] In existing technologies, crop growth monitoring largely relies on manual inspections, where crop growth is assessed by observing changes in crop height, leaf morphology, and leaf color. This method suffers from low monitoring efficiency, high subjectivity, and limited monitoring range, making it difficult to meet the demands of large-scale agricultural production for real-time monitoring and refined management. Therefore, some technologies are beginning to utilize image processing methods to identify and analyze farmland crops, extracting crop image features to monitor crop growth status.

[0004] However, most existing methods rely primarily on two-dimensional image features for analysis, which are easily affected by changes in light intensity, background interference, and crop shading in complex agricultural environments, leading to insufficient accuracy in crop region identification. Furthermore, existing technologies do not adequately utilize crop spatial structure information, making it difficult to accurately calculate key growth parameters such as crop height, canopy coverage, and leaf area index. In addition, regarding crop growth trend analysis, most methods employ only simple time-series statistical approaches, failing to effectively characterize the non-linear changes that occur during crop growth, thus affecting the accuracy and stability of crop growth status monitoring results.

[0005] Therefore, how to provide a smart agricultural crop growth monitoring method based on image processing is a problem that urgently needs to be solved by those skilled in the art. Summary of the Invention

[0006] One objective of this invention is to propose a smart agricultural crop growth monitoring method based on image processing. This invention comprehensively utilizes AI visual recognition algorithms, deep learning feature extraction technology, neural radiation field spatial calculation methods, and continuous-time dynamic modeling methods. It can accurately extract crop regions and calculate key growth parameters such as crop height, canopy coverage, and leaf area index. At the same time, by performing trend analysis on the growth state sequence, it improves the stability and reliability of crop growth monitoring results. It has the advantages of high monitoring accuracy, high degree of automation, and strong adaptability to complex agricultural environments.

[0007] A smart agricultural crop growth monitoring method based on image processing according to an embodiment of the present invention includes the following steps:

[0008] S1. Collect image data of the target crop at a preset frequency in continuous time steps and preprocess it to construct a crop image sequence;

[0009] S2. The YOLOv7 algorithm is used to perform multi-scale convolution and boundary prediction operations on crop image sequences to extract candidate region sets. The DeepLabV3+ model is then used to perform dilated convolution and decoding mapping operations on the candidate region sets to construct crop region sequences.

[0010] S3. Perform large convolutional kernel feature extraction and layer-by-layer feature mapping operations on the crop region sequence in the RepLKNet network, and aggregate the mapping results to generate crop feature sequences containing three types of features: leaf morphology, color distribution and texture structure.

[0011] S4. Based on the crop region sequence, extract the set of spatial coordinates of the crop region at each time step, and perform spatial calculation operations in the neural radiation field to calculate the crop height, canopy coverage and leaf area index at each time step to obtain the growth parameter sequence.

[0012] S5. Based on crop feature sequences and growth parameter sequences, perform continuous-time state evolution calculations using the Neural CDE model, and perform state propagation and hidden state update operations on the evolution results to construct a growth state sequence.

[0013] S6. Perform two-dimensional empirical mode decomposition on the growth state sequence to obtain multiple intrinsic components and residual components. Perform time difference calculation and sliding window statistical operation on the intrinsic components to extract the growth change values ​​at each time step, and construct the growth trend sequence by combining the residual components.

[0014] S7. Perform classification calculations on the growth trend sequence using a support vector machine, and generate crop growth monitoring results based on the classification results.

[0015] Optionally, the target crop refers to an individual crop or a group of crops selected as the object of growth status analysis within the farmland monitoring area; the image data refers to a set of farmland image information containing the growth morphology information of the target crop, acquired by an image acquisition device at a preset frequency; the preprocessing includes resolution unification, noise suppression, and illumination correction; the leaf morphology refers to the geometric structure and shape information of the target crop leaves; the color distribution refers to the color pixel distribution characteristics of the target crop leaves in the image; the texture structure refers to the spatial variation characteristics of the surface texture pattern of the target crop leaves in the image; the crop height refers to the spatial distance between the target crop from the ground reference position to the highest point of the plant; the canopy coverage rate refers to the spatial statistical index of the proportion of the target crop leaves in the ground surface projection area; and the leaf area index refers to the ratio index of the total area of ​​the target crop leaves per unit ground surface area.

[0016] Optionally, S2 specifically includes:

[0017] S21. Using the convolutional network structure of the YOLOv7 algorithm, multi-scale convolution calculations are performed on crop images at each time step in the crop image sequence. By using convolutional kernels of different scales, the crop images are processed layer by layer through convolutional mapping to obtain convolutional feature maps of multiple scales.

[0018] S22. Perform feature fusion operation on the convolutional feature map, and perform feature superposition and fusion calculation on the convolutional feature maps of each scale in the spatial dimension to obtain the fused feature map.

[0019] S23. Perform boundary prediction operation on the fused feature map. Use the boundary prediction structure in the YOLOv7 algorithm to perform boundary regression calculation on the fused feature map, obtain the corresponding bounding box position and construct a candidate region set.

[0020] S24. Input the candidate region set into the DeepLabV3+ model and perform dilated convolution calculation on the candidate regions. Use convolution kernels with different dilation rates to perform convolution mapping processing on the candidate regions to obtain the region feature map.

[0021] S25. Perform decoding and mapping operations on the regional feature map. Use the decoding structure in the DeepLabV3+ model to perform upsampling and pixel-level classification operations on the regional feature map. Perform Laplace operator boundary enhancement and boundary correction operations on the classification results. Extract the crop regions and arrange them in chronological order to construct a crop region sequence.

[0022] Optionally, S23 specifically includes:

[0023] S231. The fused feature map is divided into grids according to the preset size. The boundary prediction structure in the YOLOv7 algorithm is used to perform convolution calculation on each grid cell in the fused feature map. The convolution kernel is multiplied with the feature value of the corresponding region of the fused feature map, and the product result is added to obtain the feature response value of each grid cell. The center position of the bounding box of each grid cell is calculated based on the feature response value.

[0024] S232. Perform boundary regression calculation based on the feature response values ​​of each grid cell. Obtain the center coordinates of the bounding box by adding the coordinates of the corresponding grid cell in the fused feature map. At the same time, obtain the width and height of the bounding box by multiplying the feature values ​​of the corresponding grid cell with the preset bounding box size. Combine the center coordinates of the bounding box to construct the position of the bounding box.

[0025] S233. Calculate the confidence level of the feature response value corresponding to each bounding box position, perform a threshold comparison operation on each bounding box position based on the confidence level, delete the bounding box positions with confidence levels lower than the preset threshold, and retain the bounding box positions with confidence levels that meet the conditions.

[0026] S234. Perform an overlap region determination operation on the retained bounding box positions, calculate the area of ​​the intersection region and the area of ​​the union region of any two bounding boxes, divide the area of ​​the intersection region by the area of ​​the union region to obtain the overlap ratio, and delete the bounding box positions whose overlap ratio exceeds the preset threshold.

[0027] S235. Based on the bounding box positions retained after the overlapping region determination, extract the corresponding image regions in the crop image sequence and construct a candidate region set.

[0028] Optionally, S25 specifically includes:

[0029] S251. Input the region feature map into the decoding structure in the DeepLabV3+ model, perform upsampling calculation on the region feature map, multiply the feature values ​​in the region feature map with the preset upsampling weights, and accumulate the product results to generate a high-resolution feature map.

[0030] S252. Perform pixel-level classification operation on the high-resolution feature map by multiplying the feature values ​​of the high-resolution feature map with the preset classification weights and accumulating the product results to obtain the classification response value at each pixel position.

[0031] S253. Generate crop segmentation results based on the classification response values, and perform Laplace operator boundary enhancement calculations on the crop segmentation results. This is achieved by performing neighborhood difference and additive accumulation operations on the classification response values ​​to obtain the boundary feature map, specifically including:

[0032] A threshold comparison operation is performed on each pixel position based on the classification response value. Pixel positions with classification response values ​​greater than the preset classification threshold are marked as crop pixels, and pixel positions with classification response values ​​less than or equal to the preset classification threshold are marked as background pixels. The marking results of all pixel positions are arranged to generate crop segmentation results.

[0033] Based on the Laplace operator, a fixed-size neighborhood window is established with each pixel position as the center in the classification response value corresponding to the crop segmentation result. The classification response value of each neighboring pixel in the neighborhood window is compared with the classification response value of the center pixel to obtain multiple neighborhood difference values.

[0034] Perform a multiplication operation between each neighborhood difference value and a preset difference weight to obtain the corresponding difference product value. Perform an addition and accumulation calculation on all difference product values ​​to obtain the boundary response value at the center pixel position.

[0035] The neighborhood difference, multiplication, and addition operations are repeatedly performed on all pixel locations to generate the corresponding set of boundary response values, which are then arranged according to the pixel spatial location to obtain the boundary feature map.

[0036] S254. Based on the boundary feature map, perform boundary correction processing on the crop segmentation results, extract the crop regions, and arrange them in chronological order to construct a crop region sequence.

[0037] Optionally, S3 specifically includes:

[0038] S31. Input the crop region sequence into the RepLKNet network, and perform layer-by-layer convolutional mapping on the crop region at each time step through a large convolutional kernel of a preset size to extract the feature information of the crop region and obtain the crop feature map.

[0039] S32. Perform layer-by-layer feature mapping operation on the crop feature map. Generate a multi-layer mapped feature map by performing continuous convolution, channel mapping and spatial response update processing on the crop feature map in the multi-layer convolutional structure of the RepLKNet network.

[0040] S33. Calculate the leaf margin outline, leaf length, leaf width and leaf extension direction in the crop region, and perform morphological response aggregation calculation on the multi-layer mapping feature map to obtain the leaf morphological feature map.

[0041] S34. Calculate the color pixel distribution in the crop region, and perform color response extraction and channel aggregation calculation on the multi-layer mapping feature map to obtain the color distribution feature map;

[0042] S35. Calculate the texture direction distribution and texture density distribution in the crop region, and perform local texture response extraction and spatial aggregation calculation on the multi-layer mapping feature map to obtain the texture structure feature map;

[0043] S36. Perform aggregation operations on the leaf morphology feature map, color distribution feature map, and texture structure feature map to generate crop feature vectors for each time step and construct crop feature sequences.

[0044] Optionally, the generation process of the leaf morphology feature map, color distribution feature map, and texture structure feature map specifically includes:

[0045] In the crop region, perform subtraction on the pixel values ​​of adjacent pixels to obtain a set of pixel differences, and perform absolute value calculation and addition accumulation on the set of pixel differences. Extract the leaf edge contour based on the accumulation result.

[0046] Extract the x and y coordinates of all pixel positions in the leaf edge contour, subtract the minimum x coordinate from the maximum x coordinate to obtain the leaf width, and subtract the minimum y coordinate from the maximum y coordinate to obtain the leaf length.

[0047] Divide the difference in the horizontal coordinates of adjacent pixels in the leaf edge contour by the difference in the vertical coordinates to obtain the leaf extension direction;

[0048] Based on the spatial position of the leaf margin contour in the multi-layer mapping feature map, the corresponding feature values ​​are extracted, and the extracted feature values ​​are multiplied with the leaf length, leaf width and leaf extension direction respectively. The multiplication results are then added together to obtain the leaf morphology feature map.

[0049] The pixel values ​​in the crop area are divided according to the color channel to obtain a red pixel set, a green pixel set and a blue pixel set. Then, the pixel values ​​in each of the three sets are added together. The sum is then divided by the number of pixels in the corresponding set to obtain the red average value, the green average value and the blue average value.

[0050] Based on the spatial location of the crop region in the multi-layer mapping feature map, the corresponding feature values ​​are extracted, and the extracted feature values ​​are multiplied by the red average value, green average value and blue average value respectively. The product results are then added together to obtain the color distribution feature map.

[0051] Establish a neighborhood window of fixed size in the crop region, and subtract the neighboring pixel values ​​from the center pixel value to obtain the pixel difference set;

[0052] The absolute value of the pixel difference set is calculated, and the texture density distribution is obtained by summing all the absolute value results.

[0053] The texture direction value is obtained by dividing the horizontal pixel difference between adjacent pixel positions by the vertical pixel difference, and then the texture direction distribution is obtained by performing an addition and accumulation calculation on all texture direction values.

[0054] Based on the spatial location of the crop region in the multi-layer mapping feature map, the corresponding feature values ​​are extracted, and the extracted feature values ​​are multiplied with the texture direction distribution and texture density distribution respectively. The multiplication results are then added together to obtain the texture structure feature map.

[0055] Optionally, S4 specifically includes:

[0056] S41. Extract the boundary contours of the crop regions at each time step in the crop region sequence, extract the horizontal and vertical coordinates of all pixel positions in the boundary contours, and arrange them according to the spatial position of the pixels to construct a set of spatial coordinates for each time step.

[0057] S42. Input the set of spatial coordinates into the neural radiation field, and perform three-dimensional spatial reconstruction calculation on the set of spatial coordinates. Generate a three-dimensional structure set of crops based on the reconstruction results.

[0058] S43. Extract the longitudinal coordinate values ​​of all spatial points in the three-dimensional structure set, and obtain the crop height at each time step by calculating the difference between the maximum and minimum values ​​of the longitudinal coordinate values.

[0059] S44. Extract the spatial distribution area of ​​crops based on the three-dimensional structure set, perform projection calculation on the horizontal plane of the spatial distribution area, count the number of crop pixels in the projection area, and perform ratio calculation between the number of crop pixels and the total number of pixels in the projection area to obtain the canopy coverage rate at each time step.

[0060] S45. Based on the three-dimensional structure set, the spatial area of ​​the crop leaf region is statistically analyzed. The total area of ​​the leaf is calculated by addition and summation of all spatial areas. The total leaf area is then compared with the crop region area to obtain the leaf area index at each time step.

[0061] S46. Arrange the crop height, canopy coverage and leaf area index calculated at each time step in chronological order to construct a growth parameter sequence.

[0062] Optionally, S5 specifically includes:

[0063] S51. Perform time alignment processing on the crop feature sequence and growth parameter sequence, and concatenate the crop feature vector and corresponding growth parameter at each time step to generate a joint feature sequence.

[0064] S52. In the Neural CDE model, continuous-time driven computation is performed on the joint feature sequence. A continuous-time signal is constructed based on the change trajectory of the joint feature sequence in the time dimension. Then, a state evolution operation is performed on the continuous-time signal in the Neural CDE model to generate an initial state set, specifically including:

[0065] Perform a difference operation on the joint feature vectors of adjacent time steps in the joint feature sequence, and perform a division operation on the difference result and the corresponding time interval to generate a continuous time-driven sequence as a continuous time signal;

[0066] The feature mapping calculation of the continuous-time signal is performed by the control function structure in the Neural CDE model. The intermediate vector is generated by multiplying the feature values ​​in the continuous-time signal with the preset control weights and accumulating the multiplication results.

[0067] In the Neural CDE model, state evolution calculation is performed by multiplying the feature values ​​of each time step in the continuous time signal with the corresponding intermediate vector, and then performing addition and accumulation calculation on the product results to generate the state change value of the corresponding time step.

[0068] Continuous-time update calculations are performed on the intermediate vector based on the state change values. The initial state set is obtained by adding the state change values ​​to the intermediate vector.

[0069] S53. Perform recursive update calculation on the initial state of each time step in the initial state set. Generate the propagation state set by performing multiplication operation on the initial state and the corresponding joint feature vector, and then performing addition and accumulation calculation on the product result.

[0070] S54. Perform hidden state update operation on the propagation state set. Perform channel mapping calculation and spatial response update calculation on the propagation state vector at each time step, and arrange the update results in time order to generate a hidden state sequence.

[0071] S55. Perform Z-Score normalization on the hidden state sequence to obtain the growth state sequence arranged in chronological order.

[0072] Optionally, S6 specifically includes:

[0073] S61. Using the two-dimensional empirical mode decomposition algorithm, local extremum detection is performed on each dimension of the growth state vector in the growth state sequence to extract the local maxima and local minima of each dimension of the state value in the time dimension, and construct the set of maxima points and the set of minima points respectively.

[0074] S62. Based on the set of maximum and minimum points, perform interpolation connection operation on the state values ​​of the same dimension in the time dimension to construct the upper envelope sequence and the lower envelope sequence, and perform mean calculation on the upper envelope and lower envelope at the same time step to obtain the local mean sequence of each time step.

[0075] S63. Perform subtraction operation on the growth state vector of each time step and the corresponding local mean sequence to obtain the screening sequence. Repeat the local extremum detection, interpolation connection, local mean calculation and subtraction operation on the screening sequence until the screening sequence meets the preset screening conditions in the time dimension. Extract each screening sequence as an eigencomponent.

[0076] S64. Subtract each intrinsic component from the growth state sequence in turn to obtain the remaining component. Repeat the local extremum detection, envelope construction, local mean calculation and sieving operations on the remaining component to extract multiple intrinsic components one by one. When the number of local maxima and local minima in the remaining component is less than the preset number, the current remaining component is taken as the residual component.

[0077] S65. Perform time difference calculation on each intrinsic component. By performing subtraction operation on the values ​​of the same intrinsic component at adjacent time steps, obtain the difference results of the corresponding intrinsic component at each time step, and arrange all difference results in time order to construct a difference sequence.

[0078] S66. Perform sliding window statistical operation on the difference sequence, establish a fixed-length sliding window in the time dimension, extract the difference results in each sliding window in turn, perform average calculation and maximum value extraction operations on all difference results in the same sliding window, and construct the growth change value of the corresponding time step based on the average and maximum values.

[0079] S67. Perform addition operations on the growth change values ​​and corresponding residual components at each time step, and arrange them in chronological order to construct a growth trend sequence.

[0080] The beneficial effects of this invention are:

[0081] First, this invention performs target detection and semantic segmentation calculations on crop image sequences, utilizes multi-scale convolution and boundary prediction operations to achieve accurate extraction of crop regions, and combines a large convolution kernel feature extraction structure to extract features of leaf morphology, color distribution and texture structure in crop regions. This effectively improves the accuracy of crop region recognition in complex agricultural environments, reduces the impact of light changes, background interference and crop occlusion on image recognition results, and makes the acquisition of crop growth features more stable and reliable.

[0082] Secondly, this invention achieves the acquisition of crop spatial structure information by performing three-dimensional spatial reconstruction calculations on the set of spatial coordinates of the crop area in a neural radiation field, and further calculates key growth parameters such as crop height, canopy coverage, and leaf area index. Compared with monitoring methods that only analyze two-dimensional image features, this invention can more comprehensively reflect the true growth status of crops, thereby improving the accuracy of crop growth parameter calculations and the reliability of monitoring results.

[0083] Finally, this invention performs continuous-time state evolution calculations on crop feature sequences and growth parameter sequences using a Neural CDE model to construct a growth state sequence. It then combines two-dimensional empirical mode decomposition to perform trend decomposition and change analysis on the growth state sequence, extracts growth change values, and constructs a growth trend sequence. Finally, it performs classification calculations on the growth trend sequence using a support vector machine, thereby achieving automatic determination of crop growth status, effectively improving the crop growth trend analysis capability, and making the crop growth monitoring results more stable and accurate. Attached Figure Description

[0084] The accompanying drawings are provided to further illustrate the invention and form part of the specification. They are used in conjunction with embodiments of the invention to explain the invention and do not constitute a limitation thereof. In the drawings:

[0085] Figure 1 This is a flowchart of an image processing-based smart agriculture crop growth monitoring method proposed in this invention;

[0086] Figure 2 This is a flowchart of crop region identification and feature extraction in an image processing-based smart agriculture crop growth monitoring method proposed in this invention;

[0087] Figure 3 This is a flowchart illustrating the crop growth trend analysis and status determination of a smart agricultural crop growth monitoring method based on image processing proposed in this invention. Detailed Implementation

[0088] The present invention will now be described in further detail with reference to the accompanying drawings. These drawings are simplified schematic diagrams, illustrating only the basic structure of the invention, and therefore only show the components relevant to the invention.

[0089] refer to Figures 1-3 A smart agricultural crop growth monitoring method based on image processing includes the following steps:

[0090] S1. Collect image data of the target crop at a preset frequency in continuous time steps and preprocess it to construct a crop image sequence;

[0091] S2. The YOLOv7 algorithm is used to perform multi-scale convolution and boundary prediction operations on crop image sequences to extract candidate region sets. The DeepLabV3+ model is then used to perform dilated convolution and decoding mapping operations on the candidate region sets to construct crop region sequences.

[0092] S3. Perform large convolutional kernel feature extraction and layer-by-layer feature mapping operations on the crop region sequence in the RepLKNet network, and aggregate the mapping results to generate crop feature sequences containing three types of features: leaf morphology, color distribution and texture structure.

[0093] S4. Based on the crop region sequence, extract the set of spatial coordinates of the crop region at each time step, and perform spatial calculation operations in the neural radiation field to calculate the crop height, canopy coverage and leaf area index at each time step to obtain the growth parameter sequence.

[0094] S5. Based on crop feature sequences and growth parameter sequences, perform continuous-time state evolution calculations using the Neural CDE model, and perform state propagation and hidden state update operations on the evolution results to construct a growth state sequence.

[0095] S6. Perform two-dimensional empirical mode decomposition on the growth state sequence to obtain multiple intrinsic components and residual components. Perform time difference calculation and sliding window statistical operation on the intrinsic components to extract the growth change values ​​at each time step, and construct the growth trend sequence by combining the residual components.

[0096] S7. Perform classification calculations on the growth trend sequence using a support vector machine, and generate crop growth monitoring results based on the classification results.

[0097] In this embodiment, the target crop refers to an individual crop or group of crops selected as the object of growth status analysis within the farmland monitoring area. The image data refers to a set of farmland image information containing the growth morphology information of the target crop, acquired by an image acquisition device at a preset frequency. The preprocessing includes resolution unification, noise suppression, and illumination correction. The leaf morphology refers to the geometric structure and shape information of the target crop leaves. The color distribution refers to the color pixel distribution characteristics of the target crop leaves in the image. The texture structure refers to the spatial variation characteristics of the surface texture pattern of the target crop leaves in the image. The crop height refers to the spatial distance between the target crop from the ground reference position to the highest point of the plant. The canopy coverage rate refers to the spatial statistical index of the proportion of the target crop leaves in the ground projection area. The leaf area index refers to the ratio index of the total area of ​​the target crop leaves per unit ground surface area.

[0098] In this embodiment, S2 specifically includes:

[0099] S21. Using the convolutional network structure of the YOLOv7 algorithm, multi-scale convolution calculations are performed on crop images at each time step in the crop image sequence. By using convolutional kernels of different scales, the crop images are processed layer by layer through convolutional mapping to obtain convolutional feature maps of multiple scales.

[0100] S22. Perform feature fusion operation on the convolutional feature map, and perform feature superposition and fusion calculation on the convolutional feature maps of each scale in the spatial dimension to obtain the fused feature map.

[0101] S23. Perform boundary prediction operation on the fused feature map. Use the boundary prediction structure in the YOLOv7 algorithm to perform boundary regression calculation on the fused feature map, obtain the corresponding bounding box position and construct a candidate region set.

[0102] S24. Input the candidate region set into the DeepLabV3+ model and perform dilated convolution calculation on the candidate regions. Use convolution kernels with different dilation rates to perform convolution mapping processing on the candidate regions to obtain the region feature map.

[0103] S25. Perform decoding and mapping operations on the regional feature map. Use the decoding structure in the DeepLabV3+ model to perform upsampling and pixel-level classification operations on the regional feature map. Perform Laplace operator boundary enhancement and boundary correction operations on the classification results. Extract the crop regions and arrange them in chronological order to construct a crop region sequence.

[0104] In this embodiment, S23 specifically includes:

[0105] S231. The fused feature map is divided into grids according to the preset size. The boundary prediction structure in the YOLOv7 algorithm is used to perform convolution calculation on each grid cell in the fused feature map. The convolution kernel is multiplied with the feature value of the corresponding region of the fused feature map, and the product result is added to obtain the feature response value of each grid cell. The center position of the bounding box of each grid cell is calculated based on the feature response value.

[0106] S232. Perform boundary regression calculation based on the feature response values ​​of each grid cell. Obtain the center coordinates of the bounding box by adding the coordinates of the corresponding grid cell in the fused feature map. At the same time, obtain the width and height of the bounding box by multiplying the feature values ​​of the corresponding grid cell with the preset bounding box size. Combine the center coordinates of the bounding box to construct the position of the bounding box.

[0107] S233. Calculate the confidence level of the feature response value corresponding to each bounding box position, perform a threshold comparison operation on each bounding box position based on the confidence level, delete the bounding box positions with confidence levels lower than the preset threshold, and retain the bounding box positions with confidence levels that meet the conditions.

[0108] S234. Perform an overlap region determination operation on the retained bounding box positions, calculate the area of ​​the intersection region and the area of ​​the union region of any two bounding boxes, divide the area of ​​the intersection region by the area of ​​the union region to obtain the overlap ratio, and delete the bounding box positions whose overlap ratio exceeds the preset threshold.

[0109] S235. Based on the bounding box positions retained after the overlapping region determination, extract the corresponding image regions in the crop image sequence and construct a candidate region set.

[0110] In this embodiment, S25 specifically includes:

[0111] S251. Input the region feature map into the decoding structure in the DeepLabV3+ model, perform upsampling calculation on the region feature map, multiply the feature values ​​in the region feature map with the preset upsampling weights, and accumulate the product results to generate a high-resolution feature map.

[0112] S252. Perform pixel-level classification operation on the high-resolution feature map by multiplying the feature values ​​of the high-resolution feature map with the preset classification weights and accumulating the product results to obtain the classification response value at each pixel position.

[0113] S253. Generate crop segmentation results based on the classification response values, and perform Laplace operator boundary enhancement calculations on the crop segmentation results. This is achieved by performing neighborhood difference and additive accumulation operations on the classification response values ​​to obtain the boundary feature map, specifically including:

[0114] A threshold comparison operation is performed on each pixel position based on the classification response value. Pixel positions with classification response values ​​greater than the preset classification threshold are marked as crop pixels, and pixel positions with classification response values ​​less than or equal to the preset classification threshold are marked as background pixels. The marking results of all pixel positions are arranged to generate crop segmentation results.

[0115] Based on the Laplace operator, a fixed-size neighborhood window is established with each pixel position as the center in the classification response value corresponding to the crop segmentation result. The classification response value of each neighboring pixel in the neighborhood window is compared with the classification response value of the center pixel to obtain multiple neighborhood difference values.

[0116] Perform a multiplication operation between each neighborhood difference value and a preset difference weight to obtain the corresponding difference product value. Perform an addition and accumulation calculation on all difference product values ​​to obtain the boundary response value at the center pixel position.

[0117] The neighborhood difference, multiplication, and addition operations are repeatedly performed on all pixel locations to generate the corresponding set of boundary response values, which are then arranged according to the pixel spatial location to obtain the boundary feature map.

[0118] S254. Based on the boundary feature map, perform boundary correction processing on the crop segmentation results, extract the crop regions, and arrange them in chronological order to construct a crop region sequence.

[0119] In this embodiment, S3 specifically includes:

[0120] S31. Input the crop region sequence into the RepLKNet network, and perform layer-by-layer convolutional mapping on the crop region at each time step through a large convolutional kernel of a preset size to extract the feature information of the crop region and obtain the crop feature map.

[0121] S32. Perform layer-by-layer feature mapping operation on the crop feature map. Generate a multi-layer mapped feature map by performing continuous convolution, channel mapping and spatial response update processing on the crop feature map in the multi-layer convolutional structure of the RepLKNet network.

[0122] S33. Calculate the leaf margin outline, leaf length, leaf width and leaf extension direction in the crop region, and perform morphological response aggregation calculation on the multi-layer mapping feature map to obtain the leaf morphological feature map.

[0123] S34. Calculate the color pixel distribution in the crop region, and perform color response extraction and channel aggregation calculation on the multi-layer mapping feature map to obtain the color distribution feature map;

[0124] S35. Calculate the texture direction distribution and texture density distribution in the crop region, and perform local texture response extraction and spatial aggregation calculation on the multi-layer mapping feature map to obtain the texture structure feature map;

[0125] S36. Perform aggregation operations on the leaf morphology feature map, color distribution feature map, and texture structure feature map to generate crop feature vectors for each time step and construct crop feature sequences.

[0126] In this embodiment, S32 specifically includes:

[0127] S321. In the convolutional structure of the RepLKNet network, convolution calculation is performed on the crop feature map. By multiplying the feature values ​​at each position in the crop feature map with the weight values ​​at the corresponding positions in the convolution kernel, and then adding all the multiplication results, the convolution response value is obtained. The convolution response values ​​are arranged according to their spatial positions to form a convolution feature map.

[0128] S322. Perform channel mapping calculation on the convolutional feature map. Multiply the feature values ​​of each channel in the convolutional feature map with the preset channel mapping weights respectively. Add the multiplication results to obtain the channel mapping results and arrange them in the channel order to form a mapped feature map.

[0129] S323. In the convolutional structure of the RepLKNet network, perform convolution calculation again on the mapped feature map, multiply the feature values ​​at each position in the mapped feature map with the weight values ​​of the convolution kernel, and add the multiplication results to obtain a new convolutional feature map.

[0130] S324. Perform spatial response update calculation on the new convolutional feature map, and perform addition operation on the feature values ​​of each position in the new convolutional feature map and the feature values ​​of the corresponding positions in the crop feature map to obtain the updated feature map.

[0131] S325. Iteratively perform convolution calculation, channel mapping calculation and spatial response update calculation on the updated feature map until the preset iteration round is reached to obtain multiple levels of updated feature maps. Arrange all updated feature maps in hierarchical order to generate multi-layer mapping feature maps.

[0132] In this embodiment, the generation process of the leaf morphology feature map, color distribution feature map, and texture structure feature map specifically includes:

[0133] In the crop region, perform subtraction on the pixel values ​​of adjacent pixels to obtain a set of pixel differences, and perform absolute value calculation and addition accumulation on the set of pixel differences. Extract the leaf edge contour based on the accumulation result.

[0134] Extract the x and y coordinates of all pixel positions in the leaf edge contour, subtract the minimum x coordinate from the maximum x coordinate to obtain the leaf width, and subtract the minimum y coordinate from the maximum y coordinate to obtain the leaf length.

[0135] Divide the difference in the horizontal coordinates of adjacent pixels in the leaf edge contour by the difference in the vertical coordinates to obtain the leaf extension direction;

[0136] Based on the spatial position of the leaf margin contour in the multi-layer mapping feature map, the corresponding feature values ​​are extracted, and the extracted feature values ​​are multiplied with the leaf length, leaf width and leaf extension direction respectively. The multiplication results are then added together to obtain the leaf morphology feature map.

[0137] The pixel values ​​in the crop area are divided according to the color channel to obtain a red pixel set, a green pixel set and a blue pixel set. Then, the pixel values ​​in each of the three sets are added together. The sum is then divided by the number of pixels in the corresponding set to obtain the red average value, the green average value and the blue average value.

[0138] Based on the spatial location of the crop region in the multi-layer mapping feature map, the corresponding feature values ​​are extracted, and the extracted feature values ​​are multiplied by the red average value, green average value and blue average value respectively. The product results are then added together to obtain the color distribution feature map.

[0139] Establish a neighborhood window of fixed size in the crop region, and subtract the neighboring pixel values ​​from the center pixel value to obtain the pixel difference set;

[0140] The absolute value of the pixel difference set is calculated, and the texture density distribution is obtained by summing all the absolute value results.

[0141] The texture direction value is obtained by dividing the horizontal pixel difference between adjacent pixel positions by the vertical pixel difference, and then the texture direction distribution is obtained by performing an addition and accumulation calculation on all texture direction values.

[0142] Based on the spatial location of the crop region in the multi-layer mapping feature map, the corresponding feature values ​​are extracted, and the extracted feature values ​​are multiplied with the texture direction distribution and texture density distribution respectively. The multiplication results are then added together to obtain the texture structure feature map.

[0143] In this embodiment, S4 specifically includes:

[0144] S41. Extract the boundary contours of the crop regions at each time step in the crop region sequence, extract the horizontal and vertical coordinates of all pixel positions in the boundary contours, and arrange them according to the spatial position of the pixels to construct a set of spatial coordinates for each time step.

[0145] S42. Input the set of spatial coordinates into the neural radiation field, and perform three-dimensional spatial reconstruction calculation on the set of spatial coordinates. Generate a three-dimensional structure set of crops based on the reconstruction results.

[0146] S43. Extract the longitudinal coordinate values ​​of all spatial points in the three-dimensional structure set, and obtain the crop height at each time step by calculating the difference between the maximum and minimum values ​​of the longitudinal coordinate values.

[0147] S44. Extract the spatial distribution area of ​​crops based on the three-dimensional structure set, perform projection calculation on the horizontal plane of the spatial distribution area, count the number of crop pixels in the projection area, and perform ratio calculation between the number of crop pixels and the total number of pixels in the projection area to obtain the canopy coverage rate at each time step.

[0148] S45. Based on the three-dimensional structure set, the spatial area of ​​the crop leaf region is statistically analyzed. The total area of ​​the leaf is calculated by addition and summation of all spatial areas. The total leaf area is then compared with the crop region area to obtain the leaf area index at each time step.

[0149] S46. Arrange the crop height, canopy coverage and leaf area index calculated at each time step in chronological order to construct a growth parameter sequence.

[0150] In this embodiment, S42 specifically includes:

[0151] S421. Input the set of spatial coordinates into the neural radiation field. Perform numerical normalization on the horizontal and vertical coordinates in the set of spatial coordinates. Subtract the minimum value of the horizontal coordinate from the horizontal coordinate value and divide by the difference between the maximum and minimum values ​​of the horizontal coordinate to obtain the horizontal normalized value. At the same time, subtract the minimum value of the vertical coordinate from the vertical coordinate value and divide by the difference between the maximum and minimum values ​​of the vertical coordinate to obtain the vertical normalized value. Arrange the horizontal and vertical normalized values ​​according to their spatial positions to construct a set of normalized coordinates.

[0152] S422. Perform spatial feature calculation on the normalized coordinate set in the neural radiation field. Multiply the horizontal normalized value and the vertical normalized value in the normalized coordinate set with the preset spatial weights respectively, and perform addition and accumulation calculation on the product results to obtain the spatial feature values ​​of the corresponding spatial positions. Arrange the spatial feature values ​​according to the spatial positions to generate a spatial feature set.

[0153] S423. Perform a three-dimensional spatial position calculation operation on the normalized coordinate set based on the spatial feature set. Perform multiplication operations on each spatial feature value with the corresponding horizontal normalized value and vertical normalized value, and perform addition accumulation calculation on the multiplication results to obtain the three-dimensional height coordinate value of each spatial position.

[0154] S424. Based on the three-dimensional height coordinates, horizontal normalized values ​​and vertical normalized values ​​corresponding to each spatial location, arrange and combine them according to the spatial location to construct a three-dimensional structure set for each time step.

[0155] In this embodiment, S5 specifically includes:

[0156] S51. Perform time alignment processing on the crop feature sequence and growth parameter sequence, and concatenate the crop feature vector and corresponding growth parameter at each time step to generate a joint feature sequence.

[0157] S52. In the Neural CDE model, continuous-time driven computation is performed on the joint feature sequence. A continuous-time signal is constructed based on the change trajectory of the joint feature sequence in the time dimension. Then, a state evolution operation is performed on the continuous-time signal in the Neural CDE model to generate an initial state set, specifically including:

[0158] Perform a difference operation on the joint feature vectors of adjacent time steps in the joint feature sequence, and perform a division operation on the difference result and the corresponding time interval to generate a continuous time-driven sequence as a continuous time signal;

[0159] The feature mapping calculation of the continuous-time signal is performed by the control function structure in the Neural CDE model. The intermediate vector is generated by multiplying the feature values ​​in the continuous-time signal with the preset control weights and accumulating the multiplication results.

[0160] In the Neural CDE model, state evolution calculation is performed by multiplying the feature values ​​of each time step in the continuous time signal with the corresponding intermediate vector, and then performing addition and accumulation calculation on the product results to generate the state change value of the corresponding time step.

[0161] Continuous-time update calculations are performed on the intermediate vector based on the state change values. The initial state set is obtained by adding the state change values ​​to the intermediate vector.

[0162] S53. Perform recursive update calculation on the initial state of each time step in the initial state set. Generate the propagation state set by performing multiplication operation on the initial state and the corresponding joint feature vector, and then performing addition and accumulation calculation on the product result.

[0163] S54. Perform hidden state update operation on the propagation state set. Perform channel mapping calculation and spatial response update calculation on the propagation state vector at each time step, and arrange the update results in time order to generate a hidden state sequence.

[0164] S55. Perform Z-Score normalization on the hidden state sequence to obtain the growth state sequence arranged in chronological order.

[0165] In this embodiment, S6 specifically includes:

[0166] S61. Using the two-dimensional empirical mode decomposition algorithm, local extremum detection is performed on each dimension of the growth state vector in the growth state sequence to extract the local maxima and local minima of each dimension of the state value in the time dimension, and construct the set of maxima points and the set of minima points respectively.

[0167] S62. Based on the set of maximum and minimum points, perform interpolation connection operation on the state values ​​of the same dimension in the time dimension to construct the upper envelope sequence and the lower envelope sequence, and perform mean calculation on the upper envelope and lower envelope at the same time step to obtain the local mean sequence of each time step.

[0168] S63. Perform subtraction operation on the growth state vector of each time step and the corresponding local mean sequence to obtain the screening sequence. Repeat the local extremum detection, interpolation connection, local mean calculation and subtraction operation on the screening sequence until the screening sequence meets the preset screening conditions in the time dimension. Extract each screening sequence as an eigencomponent.

[0169] S64. Subtract each intrinsic component from the growth state sequence in turn to obtain the remaining component. Repeat the local extremum detection, envelope construction, local mean calculation and sieving operations on the remaining component to extract multiple intrinsic components one by one. When the number of local maxima and local minima in the remaining component is less than the preset number, the current remaining component is taken as the residual component.

[0170] S65. Perform time difference calculation on each intrinsic component. By performing subtraction operation on the values ​​of the same intrinsic component at adjacent time steps, obtain the difference results of the corresponding intrinsic component at each time step, and arrange all difference results in time order to construct a difference sequence.

[0171] S66. Perform sliding window statistical operation on the difference sequence, establish a fixed-length sliding window in the time dimension, extract the difference results in each sliding window in turn, perform average calculation and maximum value extraction operations on all difference results in the same sliding window, and construct the growth change value of the corresponding time step based on the average and maximum values.

[0172] S67. Perform addition operations on the growth change values ​​and corresponding residual components at each time step, and arrange them in chronological order to construct a growth trend sequence.

[0173] In this embodiment, S7 specifically includes:

[0174] S71. Input the growth trend sequence into the support vector machine in chronological order, extract the growth change value and residual component at each time step, and perform numerical splicing operation to generate trend value.

[0175] S72. In the support vector machine, perform weighted calculation on each trend value. By performing multiplication operation on the trend value of each time step with the preset weight, and performing addition and accumulation calculation on the multiplication result, the judgment value of the corresponding time step is obtained.

[0176] S73. Perform a threshold comparison operation on the judgment value, and obtain the judgment difference by performing a difference operation between the judgment value at each time step and the preset judgment threshold.

[0177] S74. The growth trend sequence is marked with a state based on the judgment difference at each time step. When the judgment difference is greater than zero, it is marked as a normal growth state. When the judgment difference is less than or equal to zero, it is marked as an abnormal growth state. All marking results are arranged in chronological order to generate crop growth monitoring results.

[0178] Example 1: To verify the feasibility of this invention in practice, it was applied to a smart agriculture crop growth monitoring scenario. In this scenario, agricultural managers need to continuously monitor the changes in crop growth status throughout the entire growth cycle in order to promptly detect growth anomalies and take corresponding management measures. However, in traditional agricultural management, crop growth status mainly relies on manual inspection and recording. This method is not only labor-intensive but also difficult to continuously quantify and statistically analyze crop height, canopy coverage, and leaf area index. Furthermore, agricultural environments commonly present problems such as varying light levels, leaf shading, and complex backgrounds, making it easy for ordinary image recognition methods to encounter recognition errors in practical applications, thus affecting the accuracy of crop growth status assessment. Therefore, a crop growth monitoring method that combines image processing, spatial computing, and deep learning analysis capabilities is needed to achieve continuous monitoring and accurate analysis of the crop growth process.

[0179] In this embodiment, image acquisition devices are deployed to continuously acquire image data of the target crop at consecutive time steps, and a crop image sequence is constructed according to a fixed acquisition frequency. The acquired images first undergo resolution unification, noise suppression, and color normalization to ensure stable image quality. After obtaining the crop image sequence, the YOLOv7 algorithm is used to perform multi-scale convolution and boundary prediction calculations on the target crop in the images to extract a candidate region set. Subsequently, the DeepLabV3+ model is used to perform semantic segmentation calculations on the candidate region set, and a crop region sequence is generated through dilated convolution and decoding mapping operations to achieve precise localization of the crop regions.

[0180] After obtaining the crop region sequence, a RepLKNet network is used to perform large convolutional kernel feature extraction calculations on the images within the crop region. This extracts three key features—leaf morphology, color distribution, and texture structure—from the images, forming a crop feature sequence. Subsequently, based on the spatial coordinates of each time step in the crop region sequence, three-dimensional spatial calculations are performed in the neural radiation field to continuously calculate crop height, canopy coverage, and leaf area index, yielding a growth parameter sequence. After obtaining the crop feature sequence and growth parameter sequence, a Neural CDE model is used to perform continuous-time state evolution calculations on the crop growth process, resulting in a growth state sequence.

[0181] Based on this, a two-dimensional empirical mode decomposition (EMD) calculation is performed on the growth state sequence. Multiple intrinsic components are extracted through local extremum detection, envelope construction, and sieving operations, and growth change values ​​are obtained through time difference calculation. Subsequently, a sliding window statistical calculation is used to obtain the crop's change trend at different time stages, and a growth trend sequence is constructed by combining the residual components. Finally, a support vector machine (SVM) is used to perform classification calculations on the growth trend sequence, marking the crop growth status at each time step, thereby generating crop growth monitoring results. This enables agricultural managers to promptly grasp the crop growth status and manage abnormal areas.

[0182] To further verify the practical effect of the method of the present invention, crops in the monitoring area were continuously monitored, and the method of the present invention was compared with manual inspection and ordinary image recognition methods. Statistical analysis was performed on several key indicators, and the results are shown in the table below:

[0183] Table 1. Comparison of Comprehensive Performance of Crop Growth Monitoring

[0184] index Manual inspection method Common image recognition methods Method of the present invention Crop region identification accuracy 78.4% 86.9% 96.3% Crop region positioning error 0.42m 0.27m 0.08m Crop height calculation error 0.38m 0.25m 0.07m Canopy coverage error 19.6% 13.2% 4.5% Leaf area index error 17.3% 11.4% 3.8% Growth status identification accuracy 72.6% 83.7% 94.7% Abnormal growth recognition rate 68.3% 80.4% 92.6% Accuracy of anomaly area location 64.7% 78.6% 91.8% Average processing time (per batch) 95min 42min 18min Daily monitoring capacity 4 batches 9 batches 22 batches Continuous monitoring stability Low middle high Human involvement high middle Low

[0185] As can be seen from the data in Table 1, in the manual inspection method, since it mainly relies on manual observation and judgment, the accuracy rate of crop area identification is only 78.4%. At the same time, the error in crop height calculation reaches 0.38 meters, the error in canopy coverage reaches 19.6%, and the error in leaf area index reaches 17.3%, making it difficult to achieve accurate monitoring of growth status.

[0186] While conventional image recognition methods can improve recognition efficiency to some extent, they are still easily affected by changes in light and background interference in complex agricultural environments, resulting in a crop area positioning error of 0.27 meters and a leaf area index error of 11.4%.

[0187] In contrast, the method of this invention achieves accurate crop region identification through target detection, semantic segmentation, and large convolutional kernel feature extraction structures, improving the crop region identification accuracy to 96.3% and reducing the crop region localization error to 0.08 meters. Simultaneously, it uses neural radiation fields to perform spatial calculations to obtain crop height, canopy coverage, and leaf area index, reducing the height calculation error to 0.07 meters, the canopy coverage error to 4.5%, and the leaf area index error to 3.8%. Furthermore, by using a Neural CDE model and two-dimensional empirical mode decomposition analysis to analyze crop growth trends, the accuracy of growth status identification reaches 94.7%, and the abnormal growth identification rate reaches 92.6%.

[0188] The above data demonstrates that the method of this invention is significantly superior to existing methods in terms of crop region identification accuracy, spatial growth parameter calculation accuracy, and growth status identification capability. It also greatly improves the efficiency of crop growth monitoring, enabling agricultural managers to grasp crop growth conditions more timely and accurately, thereby improving the level of agricultural production management.

[0189] The above description is only a preferred embodiment of the present invention, but the scope of protection of the present invention is not limited thereto. Any equivalent substitutions or modifications made by those skilled in the art within the scope of the technology disclosed in the present invention, based on the technical solution and inventive concept of the present invention, should be covered within the scope of protection of the present invention.

Claims

1. A smart agricultural crop growth monitoring method based on image processing, characterized in that, Includes the following steps: S1. Collect image data of the target crop at a preset frequency in continuous time steps and preprocess it to construct a crop image sequence; S2. The YOLOv7 algorithm is used to perform multi-scale convolution and boundary prediction operations on crop image sequences to extract candidate region sets. The DeepLabV3+ model is then used to perform dilated convolution and decoding mapping operations on the candidate region sets to construct crop region sequences. S3. Perform large convolutional kernel feature extraction and layer-by-layer feature mapping operations on the crop region sequence in the RepLKNet network, and aggregate the mapping results to generate crop feature sequences containing three types of features: leaf morphology, color distribution, and texture structure. S4. Based on the crop region sequence, extract the set of spatial coordinates of the crop region at each time step, and perform spatial calculation operations in the neural radiation field to calculate the crop height, canopy coverage and leaf area index at each time step to obtain the growth parameter sequence. S5. Based on crop feature sequences and growth parameter sequences, perform continuous-time state evolution calculations using the Neural CDE model, and perform state propagation and hidden state update operations on the evolution results to construct a growth state sequence. S6. Perform two-dimensional empirical mode decomposition on the growth state sequence to obtain multiple intrinsic components and residual components. Perform time difference calculation and sliding window statistical operation on the intrinsic components to extract the growth change value at each time step, and construct the growth trend sequence by combining the residual components. S7. Perform classification calculations on the growth trend sequence using a support vector machine, and generate crop growth monitoring results based on the classification results.

2. The intelligent agricultural crop growth monitoring method based on image processing according to claim 1, characterized in that, The target crop refers to an individual crop or group of crops selected as the object of growth status analysis within the farmland monitoring area. The image data refers to a set of farmland image information containing the growth morphology information of the target crop, acquired by an image acquisition device at a preset frequency. The preprocessing includes resolution unification, noise suppression, and illumination correction. The leaf morphology refers to the geometric structure and shape information of the target crop leaves. The color distribution refers to the color pixel distribution characteristics of the target crop leaves in the image. The texture structure refers to the spatial variation characteristics of the surface texture pattern of the target crop leaves in the image. The crop height refers to the spatial distance between the target crop from the ground reference position to the highest point of the plant. The canopy coverage rate refers to the spatial statistical index of the proportion of the target crop leaves in the ground projection area. The leaf area index refers to the ratio of the total area of ​​the target crop leaves per unit ground surface area.

3. The intelligent agricultural crop growth monitoring method based on image processing according to claim 1, characterized in that, S2 specifically includes: S21. Using the convolutional network structure of the YOLOv7 algorithm, multi-scale convolution calculations are performed on crop images at each time step in the crop image sequence. By using convolutional kernels of different scales, the crop images are processed layer by layer through convolutional mapping to obtain convolutional feature maps of multiple scales. S22. Perform feature fusion operation on the convolutional feature map, and perform feature superposition and fusion calculation on the convolutional feature maps of each scale in the spatial dimension to obtain the fused feature map. S23. Perform boundary prediction operation on the fused feature map. Use the boundary prediction structure in the YOLOv7 algorithm to perform boundary regression calculation on the fused feature map, obtain the corresponding bounding box position and construct a candidate region set. S24. Input the candidate region set into the DeepLabV3+ model and perform dilated convolution calculation on the candidate regions. Use convolution kernels with different dilation rates to perform convolution mapping processing on the candidate regions to obtain the region feature map. S25. Perform decoding and mapping operations on the regional feature map. Use the decoding structure in the DeepLabV3+ model to perform upsampling and pixel-level classification operations on the regional feature map. Perform Laplace operator boundary enhancement and boundary correction operations on the classification results. Extract the crop regions and arrange them in chronological order to construct a crop region sequence.

4. The intelligent agricultural crop growth monitoring method based on image processing according to claim 3, characterized in that, S23 specifically includes: S231. The fused feature map is divided into grids according to the preset size. The boundary prediction structure in the YOLOv7 algorithm is used to perform convolution calculation on each grid cell in the fused feature map. The convolution kernel is multiplied with the feature value of the corresponding region of the fused feature map, and the product result is added to obtain the feature response value of each grid cell. The center position of the bounding box of each grid cell is calculated based on the feature response value. S232. Perform boundary regression calculation based on the feature response values ​​of each grid cell. Obtain the center coordinates of the bounding box by adding the coordinates of the corresponding grid cell in the fused feature map. At the same time, obtain the width and height of the bounding box by multiplying the feature values ​​of the corresponding grid cell with the preset bounding box size. Combine the center coordinates of the bounding box to construct the position of the bounding box. S233. Calculate the confidence level of the feature response value corresponding to each bounding box position, perform a threshold comparison operation on each bounding box position based on the confidence level, delete the bounding box positions with confidence levels lower than the preset threshold, and retain the bounding box positions with confidence levels that meet the conditions. S234. Perform an overlap region determination operation on the retained bounding box positions, calculate the area of ​​the intersection region and the area of ​​the union region of any two bounding boxes, divide the area of ​​the intersection region by the area of ​​the union region to obtain the overlap ratio, and delete the bounding box positions whose overlap ratio exceeds the preset threshold. S235. Based on the bounding box positions retained after the overlapping region determination, extract the corresponding image regions in the crop image sequence and construct a candidate region set.

5. The intelligent agricultural crop growth monitoring method based on image processing according to claim 3, characterized in that, Specifically, S25 includes: S251. Input the region feature map into the decoding structure in the DeepLabV3+ model, perform upsampling calculation on the region feature map, multiply the feature values ​​in the region feature map with the preset upsampling weights, and accumulate the product results to generate a high-resolution feature map. S252. Perform pixel-level classification operation on the high-resolution feature map by multiplying the feature values ​​of the high-resolution feature map with the preset classification weights and accumulating the product results to obtain the classification response value at each pixel position. S253. Generate crop segmentation results based on the classification response values, and perform Laplace operator boundary enhancement calculations on the crop segmentation results. This is achieved by performing neighborhood difference and additive accumulation operations on the classification response values ​​to obtain the boundary feature map, specifically including: A threshold comparison operation is performed on each pixel position based on the classification response value. Pixel positions with classification response values ​​greater than the preset classification threshold are marked as crop pixels, and pixel positions with classification response values ​​less than or equal to the preset classification threshold are marked as background pixels. The marking results of all pixel positions are arranged to generate crop segmentation results. Based on the Laplace operator, a fixed-size neighborhood window is established with each pixel position as the center in the classification response value corresponding to the crop segmentation result. The classification response value of each neighboring pixel in the neighborhood window is compared with the classification response value of the center pixel to obtain multiple neighborhood difference values. Perform a multiplication operation between each neighborhood difference value and a preset difference weight to obtain the corresponding difference product value. Perform an addition and accumulation calculation on all difference product values ​​to obtain the boundary response value at the center pixel position. The neighborhood difference, multiplication, and addition operations are repeatedly performed on all pixel locations to generate the corresponding set of boundary response values, which are then arranged according to the pixel spatial location to obtain the boundary feature map. S254. Based on the boundary feature map, perform boundary correction processing on the crop segmentation results, extract the crop regions, and arrange them in chronological order to construct a crop region sequence.

6. The intelligent agricultural crop growth monitoring method based on image processing according to claim 1, characterized in that, S3 specifically includes: S31. Input the crop region sequence into the RepLKNet network, and perform layer-by-layer convolutional mapping on the crop region at each time step through a large convolutional kernel of a preset size to extract the feature information of the crop region and construct a crop feature map. S32. Perform layer-by-layer feature mapping operation on the crop feature map. Generate a multi-layer mapped feature map by performing continuous convolution, channel mapping and spatial response update processing on the crop feature map in the multi-layer convolutional structure of the RepLKNet network. S33. Calculate the leaf margin outline, leaf length, leaf width and leaf extension direction in the crop region, and perform morphological response aggregation calculation on the multi-layer mapping feature map to obtain the leaf morphological feature map. S34. Calculate the color pixel distribution in the crop region, and perform color response extraction and channel aggregation calculation on the multi-layer mapping feature map to obtain the color distribution feature map; S35. Calculate the texture direction distribution and texture density distribution in the crop region, and perform local texture response extraction and spatial aggregation calculation on the multi-layer mapping feature map to obtain the texture structure feature map; S36. Perform aggregation operations on the leaf morphology feature map, color distribution feature map, and texture structure feature map to generate crop feature vectors for each time step and construct crop feature sequences.

7. The intelligent agricultural crop growth monitoring method based on image processing according to claim 6, characterized in that, The generation process of the leaf morphology feature map, color distribution feature map, and texture structure feature map specifically includes: In the crop region, perform subtraction on the pixel values ​​of adjacent pixels to obtain a set of pixel differences, and perform absolute value calculation and addition accumulation on the set of pixel differences. Extract the leaf edge contour based on the accumulation result. Extract the x-coordinate and y-coordinate of all pixel positions in the leaf edge contour, subtract the minimum x-coordinate from the maximum x-coordinate to obtain the leaf width, and subtract the minimum y-coordinate from the maximum y-coordinate to obtain the leaf length. Divide the difference in the horizontal coordinates of adjacent pixels in the leaf edge contour by the difference in the vertical coordinates to obtain the leaf extension direction; Based on the spatial position of the leaf margin contour in the multi-layer mapping feature map, the corresponding feature values ​​are extracted, and the extracted feature values ​​are multiplied with the leaf length, leaf width and leaf extension direction respectively. The multiplication results are then added together to obtain the leaf morphology feature map. The pixel values ​​in the crop area are divided according to the color channel to obtain a red pixel set, a green pixel set and a blue pixel set. Then, the pixel values ​​in each of the three sets are added together. The sum is then divided by the number of pixels in the corresponding set to obtain the red average value, the green average value and the blue average value. Based on the spatial location of the crop region in the multi-layer mapping feature map, the corresponding feature values ​​are extracted, and the extracted feature values ​​are multiplied by the red average value, green average value and blue average value respectively. The product results are then added together to obtain the color distribution feature map. Establish a fixed-size neighborhood window in the crop region, and subtract the neighborhood pixel values ​​from the center pixel value to obtain the pixel difference set; The absolute value of the pixel difference set is calculated, and the texture density distribution is obtained by summing all the absolute value results. The texture direction value is obtained by dividing the horizontal pixel difference between adjacent pixel positions by the vertical pixel difference, and then the texture direction distribution is obtained by performing an addition and accumulation calculation on all texture direction values. Based on the spatial location of the crop region in the multi-layer mapping feature map, the corresponding feature values ​​are extracted, and the extracted feature values ​​are multiplied with the texture direction distribution and texture density distribution respectively. The multiplication results are then added together to obtain the texture structure feature map.

8. The intelligent agricultural crop growth monitoring method based on image processing according to claim 1, characterized in that, S4 specifically includes: S41. Extract the boundary contours of the crop regions at each time step in the crop region sequence, extract the horizontal and vertical coordinates of all pixel positions in the boundary contours, and arrange them according to the spatial position of the pixels to construct a set of spatial coordinates for each time step. S42. Input the set of spatial coordinates into the neural radiation field, and perform three-dimensional spatial reconstruction calculation on the set of spatial coordinates. Generate a three-dimensional structure set of crops based on the reconstruction results. S43. Extract the longitudinal coordinate values ​​of all spatial points in the three-dimensional structure set, and obtain the crop height at each time step by calculating the difference between the maximum and minimum values ​​of the longitudinal coordinate values. S44. Extract the spatial distribution area of ​​crops based on the three-dimensional structure set, perform projection calculation on the horizontal plane of the spatial distribution area, count the number of crop pixels in the projection area, and perform ratio calculation between the number of crop pixels and the total number of pixels in the projection area to obtain the canopy coverage rate at each time step. S45. Based on the three-dimensional structure set, the spatial area of ​​the crop leaf region is statistically analyzed. The total area of ​​the leaf is calculated by addition and summation of all spatial areas. The total leaf area is then compared with the crop region area to obtain the leaf area index at each time step. S46. Arrange the crop height, canopy coverage and leaf area index calculated at each time step in chronological order to construct a growth parameter sequence.

9. The intelligent agricultural crop growth monitoring method based on image processing according to claim 1, characterized in that, S5 specifically includes: S51. Perform time alignment processing on the crop feature sequence and growth parameter sequence, and concatenate the crop feature vector and corresponding growth parameter at each time step to generate a joint feature sequence. S52. In the Neural CDE model, continuous-time driven computation is performed on the joint feature sequence. A continuous-time signal is constructed based on the change trajectory of the joint feature sequence in the time dimension. Then, a state evolution operation is performed on the continuous-time signal in the Neural CDE model to generate an initial state set, specifically including: Perform a difference operation on the joint feature vectors of adjacent time steps in the joint feature sequence, and perform a division operation on the difference result and the corresponding time interval to generate a continuous time-driven sequence as a continuous time signal; The feature mapping calculation of the continuous-time signal is performed by the control function structure in the Neural CDE model. The intermediate vector is generated by multiplying the feature values ​​in the continuous-time signal with the preset control weights and accumulating the multiplication results. In the Neural CDE model, state evolution calculation is performed by multiplying the feature values ​​of each time step in the continuous time signal with the corresponding intermediate vector, and then performing addition and accumulation calculation on the product results to generate the state change value of the corresponding time step. Continuous-time update calculations are performed on the intermediate vector based on the state change values. The initial state set is obtained by adding the state change values ​​to the intermediate vector. S53. Perform recursive update calculation on the initial state of each time step in the initial state set. Generate the propagation state set by performing multiplication operation on the initial state and the corresponding joint feature vector, and then performing addition and accumulation calculation on the product result. S54. Perform hidden state update operation on the propagation state set. Perform channel mapping calculation and spatial response update calculation on the propagation state vector at each time step, and arrange the update results in time order to generate a hidden state sequence. S55. Perform Z-Score normalization on the hidden state sequence to obtain the growth state sequence arranged in chronological order.

10. The intelligent agricultural crop growth monitoring method based on image processing according to claim 1, characterized in that, S6 specifically includes: S61. Using a two-dimensional empirical mode decomposition algorithm, local extremum detection is performed on each dimension of the growth state vector in the growth state sequence to extract the local maxima and local minima of each dimension of the state value in the time dimension, and construct the set of maxima points and the set of minima points respectively. S62. Based on the set of maximum and minimum points, perform interpolation connection operation on the state values ​​of the same dimension in the time dimension to construct the upper envelope sequence and the lower envelope sequence, and perform mean calculation on the upper envelope and lower envelope at the same time step to obtain the local mean sequence of each time step. S63. Perform subtraction operation on the growth state vector of each time step and the corresponding local mean sequence to obtain the screening sequence. Repeat the local extremum detection, interpolation connection, local mean calculation and subtraction operation on the screening sequence until the screening sequence meets the preset screening conditions in the time dimension. Extract each screening sequence as an eigencomponent. S64. Subtract each intrinsic component from the growth state sequence in turn to obtain the remaining component. Repeat the local extremum detection, envelope construction, local mean calculation and sieving operations on the remaining component to extract multiple intrinsic components one by one. When the number of local maxima and local minima in the remaining component is less than the preset number, the current remaining component is taken as the residual component. S65. Perform time difference calculation on each intrinsic component. By performing subtraction operation on the values ​​of the same intrinsic component at adjacent time steps, obtain the difference results of the corresponding intrinsic component at each time step, and arrange all difference results in time order to construct a difference sequence. S66. Perform sliding window statistical operation on the difference sequence, establish a fixed-length sliding window in the time dimension, extract the difference results in each sliding window in turn, perform average calculation and maximum value extraction operations on all difference results in the same sliding window, and construct the growth change value of the corresponding time step based on the average and maximum values. S67. Perform addition operations on the growth change values ​​and corresponding residual components at each time step, and arrange them in chronological order to construct a growth trend sequence.

Images