A red pine cone yield estimation method based on unmanned aerial vehicle image

Abstract

The application discloses a kind of red pine cone yield estimation methods based on unmanned aerial vehicle image, belongs to red pine cone yield prediction technical field.To solve the problem of improving the accuracy and timeliness of red pine cone yield estimation results.The present application includes selecting cone development period as image collection time window, collecting unmanned aerial vehicle image data of red pine forest;Lay out fixed sample plot in target stand, after cone ripening, carry out ground survey, and obtain ground sample data;Preprocessing, then divided into training set, validation set and test set;Cascade instance segmentation model based on multi-scale density attention enhancement is constructed, including successively connected multi-scale feature extraction module, multi-scale density attention enhancement module and cascade instance segmentation optimization module, the prediction result of cone is corrected, to obtain the corrected real single tree mature cone total number, the corrected real single tree small cone total number, then estimate single tree cone yield this year, predict next year's cone yield potential.

Description

Technical Field

[0001] This invention belongs to the field of Korean pine cone yield prediction technology, specifically relating to a method for predicting Korean pine cone yield based on UAV imagery. Background Technology

[0002] In the field of forestry resource surveys and economic forest yield assessment, existing technologies for estimating Korean pine cone yield mainly employ ground survey methods. This method typically involves establishing fixed standard plots, selecting representative standard trees within the stand, and, during the Korean pine seed maturity period in September each year, having professional forestry personnel harvest the cones from each target tree using climbing methods. Subsequently, the harvested cones are weighed to obtain the average cone yield per Korean pine tree. Combined with the total number of Korean pine trees in the stand, the annual Korean pine cone yield for the target area is further estimated.

[0003] However, the above-mentioned method for estimating the yield of Korean pine cones based on ground surveys still has the following shortcomings:

[0004] (1) This method mainly relies on manual climbing to pick cones, which is labor-intensive and has certain safety risks during the operation process. It is especially difficult to conduct investigations and implement the method in forest areas with tall trees and complex terrain.

[0005] (2) This method has low sampling efficiency and can usually only be carried out within a limited number of standard plots and standard trees. It is difficult to meet the needs of rapid survey and continuous monitoring of large-scale forest stands, resulting in a long data acquisition cycle and poor timeliness.

[0006] (3) This method is usually only applicable to the harvesting survey during the cone ripening period. However, the current bidding work for contracting Korean pine forests in forest areas is mostly carried out in the early spring, making it difficult to obtain information on the yield of Korean pine seeds in the current year, which is not conducive to the decision-making of contract bidding and the subsequent operation and management.

[0007] (4) This method usually uses the method of “average yield per tree × total number of trees” to estimate the regional yield. It relies on strong empirical assumptions and is difficult to accurately reflect the spatial distribution differences within the forest stand and the yield variation characteristics among individual trees. Therefore, the accuracy of the estimation results is limited.

[0008] Therefore, there is an urgent need to propose a technical method that can achieve rapid, non-contact, and automated estimation of Korean pine cone yield, so as to improve survey efficiency, reduce operational risks, and enhance the accuracy and timeliness of yield estimation results. Summary of the Invention

[0009] The problem this invention aims to solve is to improve the accuracy and timeliness of Korean pine cone yield estimation results, and proposes a method for Korean pine cone yield prediction based on UAV imagery.

[0010] To achieve the above objectives, the present invention provides the following technical solution:

[0011] A method for predicting cone yield based on UAV imagery includes the following steps:

[0012] S1. Select the cone development period as the image acquisition time window to collect UAV image data of Korean pine forest; set up fixed sample plots in the target forest stand, and conduct ground measurement surveys after the cones mature to obtain ground sample data;

[0013] S2. The UAV image data of the red pine forest obtained in step S1 is preprocessed and then divided into training set, validation set and test set;

[0014] S3. Construct a cascaded instance segmentation model based on multi-scale density attention enhancement, including a multi-scale feature extraction module, a multi-scale density attention enhancement module, and a cascaded instance segmentation optimization module connected in sequence. Use the training set obtained in step S2 to train the cascaded instance segmentation model based on multi-scale density attention enhancement, and obtain the prediction result of the sphere output by the trained cascaded instance segmentation model based on multi-scale density attention enhancement.

[0015] S4. Correct the predicted cones obtained in step S3 to obtain the corrected total number of real mature cones per tree and the corrected total number of real small cones per tree;

[0016] S5. Estimate the annual yield of individual tree cones based on the corrected total number of mature cones per tree obtained in step S4.

[0017] S6. Based on the corrected total number of real single-tree red pine cones obtained in step S4, predict the cone yield potential for the following year.

[0018] Furthermore, the specific implementation method of step S1 includes the following steps:

[0019] S1.1. During the cone development period, use a DJI drone platform equipped with a Zenmuse P1 camera to conduct aerial photography of the target Korean pine forest, acquire high-resolution RGB image data, and take pictures at noon on a clear and windless day to obtain drone image data of the Korean pine forest;

[0020] S1.2. Fixed sample plots were set up within the target forest stand. After the cones matured, the cone yield of the sample trees in the sample plots was measured. The two-year-old mature cones on the sample trees were harvested and counted, and the one-year-old small cones were visually counted. The harvested mature cones were weighed to obtain the yield of cones per tree. At the same time, RTK equipment was used to locate all the measured sample trees and obtain their spatial coordinates.

[0021] Furthermore, in step S2, the acquired UAV image data of the Korean pine forest is cropped and divided into image blocks of fixed size; the one-year-old small cones and two-year-old mature cones in the image blocks are manually labeled using a labeling tool, and the labeled sample data is divided into training set, validation set and test set.

[0022] Furthermore, the specific implementation method of step S3 includes the following steps:

[0023] S3.1. Establish a multi-scale feature extraction module, which extracts features from the input image through a deep convolutional network and a feature pyramid (FPN) structure to obtain a set of multi-scale feature maps;

[0024] S3.2. Establish a multi-scale density attention enhancement module, and extract features from the multi-scale feature maps obtained in step S3.1 to obtain spatial attention weights and channel attention weights;

[0025] The multi-scale feature maps obtained in step S3.1 are aggregated across channels to obtain a two-dimensional response map. Then, local window statistics are performed on the two-dimensional response map to obtain a density response map. A 3×3 convolution operation is then performed on the density response map to fuse local receptive field information, and a boundary padding strategy is applied to maintain spatial dimension consistency. Finally, a spatial attention weight map is obtained by passing the sigmoid activation function.

[0026] The multi-scale feature map obtained in step S3.1 is subjected to global average pooling (GAP) to obtain the channel description vector; then the channel description vector is processed by a two-layer fully connected network and activation function to extract the channel attention weights.

[0027] Then, the spatial attention weights and channel attention weights are applied to the multi-scale feature maps obtained in step S3.1 to obtain the enhanced multi-scale feature map set;

[0028] S3.3. Establish a cascaded instance segmentation optimization module. Input the enhanced multi-scale feature map set obtained in step S3.2 into the cascaded instance segmentation network. First, the Region Generation Network (RPN) generates candidate boxes (RoIs) on the enhanced multi-scale feature maps; let the... The candidate boxes input in each cascade stage are Extract the region features corresponding to the candidate boxes and calculate the offset through a regression branch. This yields the updated candidate boxes:

[0029]

[0030] in, For the first Candidate boxes for each cascading stage:

[0031] By setting progressively increasing Intersection over Union (IoU) thresholds, the target candidate boxes are refined in multiple stages to obtain region features based on the optimal candidate boxes.

[0032] The network's segmentation head, Mask Head, predicts pixel-level masks of the target based on the regional features of the optimal candidate boxes. These masks characterize the spatial distribution of the target within the candidate region. Each pixel-level sphere boundary mask corresponds to a sphere instance, which describes the contour range of the sphere.

[0033] S3.4. Thresholding and spatial mapping are performed on the pixel-level cone boundary mask to obtain the instance segmentation results of the cones in the original image; the final cone instance set is obtained by filtering and deduplicating the instance segmentation results; the cone instance set is spatially matched with the canopy area of ​​a single tree, and the cone identification and quantity statistics at the single tree scale are realized according to the spatial position relationship between the cone instances and the canopy area. The predicted results of the cones output above include the detection box position of the cones, the pixel-level cone boundary mask, and the category attribute of the cones. The category attribute of the cones is either one-year-old small cones or two-year-old mature cones.

[0034] Furthermore, in step S4, a quantitative relationship is established between the number of mature cones and small cones visible to the drone and the actual total number, respectively. The mathematical expression is:

[0035]

[0036]

[0037] in, This represents the corrected total number of mature cones per tree. This represents the corrected total number of true individual cones per tree. The number of mature cones detected by the drone. The number of small cones detected by the drone. This refers to the projected area of ​​a single tree crown extracted from drone imagery. The mean value of canopy texture features calculated based on the gray-level co-occurrence matrix; and The coefficients are the partial regression coefficients of the multiple linear regression for mature cones and small cones, respectively, obtained by fitting the data from ground sample trees using the least squares method. .

[0038] Furthermore, in step S5, based on the measured number of mature cones and the corresponding total weight of cones from ground sample trees, a univariate linear quantitative relationship model between the number of mature cones per tree and the weight of the cones is established.

[0039] Then, the corrected total number of mature cones per tree obtained in step S4 is substituted into the univariate linear quantitative relationship model between the number of mature cones per tree and the cone weight, to achieve a non-destructive estimation of the predicted cone weight per tree, resulting in:

[0040]

[0041] in, To predict the weight of mature cones from a single tree, and These are the intercept and slope model parameters obtained by fitting measured data from sample trees.

[0042] Furthermore, in step S6, based on the corrected total number of actual cones per tree obtained in step S4, and using the average weight of mature cones collected that year as a reference benchmark, the potential cone yield per tree for the following year is predicted and evaluated, resulting in:

[0043]

[0044] in, To predict the potential yield of single-tree cones for the next year, This represents the average weight of a single mature cone in that year.

[0045] The beneficial effects of this invention are:

[0046] This invention discloses a method for estimating cone yield based on UAV imagery, enabling early acquisition of cone quantity and improving the timeliness of yield estimation. This invention utilizes UAVs to acquire forest stand imagery data during the cone development period, achieving early identification of Korean pine cone quantity. Compared to the traditional method of obtaining yield data only through ground harvesting during the September maturity period, this method allows for earlier acquisition of cone quantity information, facilitating earlier implementation of Korean pine forest management and contracting decisions.

[0047] This invention discloses a method for estimating cone yield based on UAV imagery, which significantly reduces labor costs and operational risks. Traditional methods require manual climbing to harvest and weigh cones one by one, which is labor-intensive and poses safety hazards. This invention uses UAV remote sensing data and combines it with automated recognition algorithms to extract the number of cones and estimate yield, significantly reducing manual intervention and lowering operational risks.

[0048] This invention presents a method for estimating cone yield based on UAV imagery, enabling rapid estimation over a large area and improving survey efficiency. Traditional ground surveys are limited by the number of sample plots, making it difficult to cover large forest stands. This invention uses UAV imagery to acquire regional-scale data and achieves stand-scale yield estimation through individual tree-level modeling, enabling rapid estimation of Korean pine cone yield over a large area and significantly improving survey efficiency.

[0049] This invention presents a method for predicting cone yield based on UAV imagery, which improves the accuracy of cone identification under complex canopy conditions. Addressing the issues of small-scale, densely distributed, and severely occluded Korean pine cones in UAV imagery, this invention constructs an identification model based on a cascaded instance segmentation framework and introduces a multi-scale density attention mechanism. This effectively enhances the feature representation ability of dense small targets, suppresses interference from branches, leaves, and background, and improves the accuracy of cone identification and segmentation.

[0050] This invention discloses a method for estimating cone yield based on UAV imagery, which establishes a quantity correction mechanism to improve the accuracy of cone quantity estimation. By introducing variables such as canopy projection area and texture features, this invention establishes a cone quantity correction model under UAV observation conditions, effectively compensating for missed detection errors caused by occlusion and observation angle, and improving the reliability of cone quantity estimation per tree.

[0051] The present invention provides a method for estimating cone yield based on UAV imagery, which decouples quantity and yield, thereby improving the applicability of the model. The present invention processes "cone quantity acquisition" and "yield estimation" in stages, and establishes a quantitative relationship model between quantity and yield through a small number of sample trees, thereby enabling yield estimation using the number of cones acquired by UAVs. This avoids dependence on large-scale harvesting data and improves the applicability and scalability of the method.

[0052] The present invention provides a method for estimating cone yield based on UAV imagery, which has the ability to estimate the yield of the current year and predict the yield of the following year. Based on the estimation of the yield of mature cones in the current year, the present invention further utilizes the number of small cones identified by UAVs to predict the yield potential of cones in the following year, thereby realizing the cross-year assessment of red pine cone yield and expanding the application scope of the method. Attached Figure Description

[0053] Figure 1 This is a flowchart of a method for estimating cone yield based on UAV imagery, as described in this invention.

[0054] Figure 2 This is a structural block diagram of a cone yield prediction method based on UAV imagery according to the present invention. Detailed Implementation

[0055] To make the objectives, technical solutions, and advantages of this invention clearer, the invention will be further described in detail below with reference to the accompanying drawings and specific embodiments. It should be understood that the specific embodiments described herein are only for explaining the invention and are not intended to limit the invention; that is, the described specific embodiments are merely a part of the embodiments of the invention, and not all of them. The components of the specific embodiments of the invention described and shown in the accompanying drawings can generally be arranged and designed in various different configurations, and the invention may also have other embodiments.

[0056] Therefore, the following detailed description of specific embodiments of the invention provided in the accompanying drawings is not intended to limit the scope of the claimed invention, but merely to illustrate selected specific embodiments of the invention. All other specific embodiments obtained by those skilled in the art based on these specific embodiments without inventive effort are within the scope of protection of this invention.

[0057] To further understand the invention's content, features, and effects, the following specific embodiments are provided, along with accompanying drawings. Figure 1 -Appendix Figure 2 Detailed explanation is as follows:

[0058] Example 1:

[0059] A method for predicting cone yield based on UAV imagery includes the following steps:

[0060] S1. Select the cone development period as the image acquisition time window to collect UAV image data of Korean pine forest; set up fixed sample plots in the target forest stand, and conduct ground measurement surveys after the cones mature in September to obtain ground sample data;

[0061] Furthermore, the specific implementation method of step S1 includes the following steps:

[0062] S1.1. In late June, during the cone development period, a DJI drone platform equipped with a Zenmuse P1 camera was used to conduct aerial photography of the target Korean pine forest to obtain high-resolution RGB image data. The images were taken at noon on a clear, windless day to obtain drone image data of the Korean pine forest.

[0063] Furthermore, the flight altitude was set at 30 meters above the tree canopy;

[0064] S1.2. Fixed sample plots were set up within the target forest stand. After the cones matured in September each year, the cone yield of the sample trees in the sample plots was measured. The two-year-old mature cones on the sample trees were harvested and counted, and the one-year-old small cones were visually counted. The harvested mature cones were weighed to obtain the cone yield per tree. At the same time, RTK equipment was used to locate all the measured sample trees and obtain their spatial coordinates.

[0065] Furthermore, 20 fixed sample plots of 50 m × 50 m were evenly distributed in the forest stand to ensure that there were no less than 30 Korean pine trees in each sample plot; and 5 standard trees were selected from each sample plot as actual test sample trees using the equal sectional area method.

[0066] S2. The UAV image data of the red pine forest obtained in step S1 is preprocessed and then divided into training set, validation set and test set;

[0067] Furthermore, in step S2, the acquired UAV image data of the Korean pine forest is cropped and divided into image blocks of fixed size; the one-year-old small cones and two-year-old mature cones in the image blocks are manually labeled using a labeling tool, and the labeled sample data is divided into training set, validation set and test set.

[0068] Furthermore, the ratio of training set, validation set, and test set data is 8:1:1.

[0069] S3. Construct a cascaded instance segmentation model based on multi-scale density attention enhancement, including a multi-scale feature extraction module, a multi-scale density attention enhancement module, and a cascaded instance segmentation optimization module connected in sequence. Use the training set obtained in step S2 to train the cascaded instance segmentation model based on multi-scale density attention enhancement, and obtain the prediction result of the sphere output by the trained cascaded instance segmentation model based on multi-scale density attention enhancement.

[0070] Furthermore, the specific implementation method of step S3 includes the following steps:

[0071] S3.1. Establish a multi-scale feature extraction module, which extracts features from the input image through a deep convolutional network and a feature pyramid (FPN) structure to obtain a set of multi-scale feature maps;

[0072] Furthermore, let the input drone image be:

[0073]

[0074] in, For input images; and These are the height and width of the image, respectively;

[0075] The resulting set of multi-scale feature maps is as follows:

[0076]

[0077] in, Indicates the first Layer feature map; , and These represent the spatial height, width, and number of channels of the feature map of this layer, respectively.

[0078] S3.2. Establish a multi-scale density attention enhancement module, and extract features from the multi-scale feature maps obtained in step S3.1 to obtain spatial attention weights and channel attention weights;

[0079] The multi-scale feature maps obtained in step S3.1 are aggregated across channels to obtain a two-dimensional response map. Then, local window statistics are performed on the two-dimensional response map to obtain a density response map. A 3×3 convolution operation is then performed on the density response map to fuse local receptive field information, and a boundary padding strategy is applied to maintain spatial dimension consistency. Finally, a spatial attention weight map is obtained by passing the sigmoid activation function.

[0080] Furthermore, firstly, regarding the feature map Cross-channel aggregation yields the following two-dimensional response graph:

[0081]

[0082] in, For the first Layer feature map at location The channel aggregation response value at that location. Then, local window statistics are performed on the two-dimensional response map to obtain the density response map as follows:

[0083]

[0084] in, For position Density response value at; window side length With window radius satisfy Since red pine cones are often densely clustered in UAV imagery, the local high-response regions of the feature map can reflect the degree of cone aggregation. By performing local window statistics as described above, this spatial aggregation feature can be explicitly captured, thereby guiding the model to focus on densely occluded areas.

[0085] To generate spatial attention weights, a composite mapping is performed on the density response map. Specifically, a 3×3 convolution operation is performed on the density response map to fuse local receptive field information, and a boundary padding strategy is applied to maintain spatial dimensionality consistency. Finally, the spatial attention weight map is obtained by passing it through a sigmoid activation function:

[0086]

[0087] in, Indicates position Spatial attention weights at the location This is the Sigmoid activation function.

[0088] The multi-scale feature map obtained in step S3.1 is subjected to global average pooling (GAP) to obtain the channel description vector; then the channel description vector is processed by a two-layer fully connected network and activation function to extract the channel attention weights.

[0089] Furthermore, global average pooling (GAP) is performed on the feature map to obtain the statistical values ​​for each channel:

[0090] ;

[0091] The Constructing a channel description vector This is used to characterize the global response features of the feature map across each channel dimension. Further, the channel attention is obtained through a two-layer fully connected network and activation function:

[0092]

[0093] in, For the first Layer The weight of each channel; and For learnable dimensionality reduction and dimensionality increase parameter matrices; It is a non-linear activation function (such as ReLU);

[0094] Then, the spatial attention weights and channel attention weights are applied to the multi-scale feature maps obtained in step S3.1 to obtain the enhanced multi-scale feature map set;

[0095] Furthermore, the enhanced features are:

[0096] ;

[0097] The density attention enhancement (MDA) operation is performed on the feature maps at each scale output by the feature pyramid to obtain an enhanced multi-scale feature map set, which is then input into the subsequent cascaded instance segmentation module. The density attention enhancement mechanism (MDA) shares or independently sets parameters among feature maps at different scales, thereby adaptively enhancing the features of spherical targets in a multi-scale space and improving the detection capability of small targets and densely occluded targets at different resolutions.

[0098] S3.3. Establish a cascaded instance segmentation optimization module. Input the enhanced multi-scale feature map set obtained in step S3.2 into the cascaded instance segmentation network. First, the Region Generation Network (RPN) generates candidate boxes (RoIs) on the enhanced multi-scale feature maps; let the... The candidate boxes input in each cascade stage are Extract the region features corresponding to the candidate boxes and calculate the offset through a regression branch. This yields the updated candidate boxes:

[0099]

[0100] in, For the first Candidate boxes for each cascading stage:

[0101] By setting progressively increasing Intersection over Union (IoU) thresholds, the target candidate boxes are refined in multiple stages to obtain region features based on the optimal candidate boxes.

[0102] The network's segmentation head, Mask Head, predicts pixel-level masks of the target based on the regional features of the optimal candidate boxes. These masks characterize the spatial distribution of the target within the candidate region. Each pixel-level sphere boundary mask corresponds to a sphere instance, which describes the contour range of the sphere.

[0103] S3.4. Thresholding and spatial mapping are performed on the pixel-level cone boundary mask to obtain the instance segmentation results of the cones in the original image; the final cone instance set is obtained by filtering and deduplicating the instance segmentation results; the cone instance set is spatially matched with the canopy area of ​​a single tree, and the cone identification and quantity statistics at the single tree scale are realized according to the spatial position relationship between the cone instances and the canopy area. The predicted results of the cones output above include the detection box position of the cones, the pixel-level cone boundary mask, and the category attribute of the cones. The category attribute of the cones is either one-year-old small cones or two-year-old mature cones.

[0104] S4. Correct the predicted cones obtained in step S3 to obtain the corrected total number of real mature cones per tree and the corrected total number of real small cones per tree;

[0105] Furthermore, based on UAV imagery, existing region merging algorithms are used for individual tree canopy segmentation to extract the independent canopy boundaries and spatial distribution range of each target tree. A sliding window slicing inference strategy combined with the constructed instance segmentation model is employed to detect cones within the individual tree canopy regions, improving the detection completeness of small target cones and thus obtaining predicted values ​​for the number of mature and small cones per tree. The cone instance segmentation results are then spatially matched with the individual tree canopy regions to count the number of cones at the individual tree scale.

[0106] Furthermore, in step S4, a quantitative relationship is established between the number of mature cones and small cones visible to the drone and the actual total number, respectively. The mathematical expression is:

[0107]

[0108]

[0109] in, This represents the corrected total number of mature cones per tree. This represents the corrected total number of true individual cones per tree. The number of mature cones detected by the drone. The number of small cones detected by the drone. This refers to the projected area of ​​a single tree crown extracted from drone imagery. The mean value of canopy texture features calculated based on the gray-level co-occurrence matrix; and The coefficients are the partial regression coefficients of the multiple linear regression for mature cones and small cones, respectively, obtained by fitting the data from ground sample trees using the least squares method. .

[0110] S5. Estimate the annual yield of individual tree cones based on the corrected total number of mature cones per tree obtained in step S4.

[0111] Furthermore, in step S5, based on the measured number of mature cones and the corresponding total weight of cones from ground sample trees, a univariate linear quantitative relationship model between the number of mature cones per tree and the weight of the cones is established.

[0112] Then, the corrected total number of mature cones per tree obtained in step S4 is substituted into the univariate linear quantitative relationship model between the number of mature cones per tree and the cone weight, to achieve a non-destructive estimation of the predicted cone weight per tree, resulting in:

[0113]

[0114] in, To predict the weight of mature cones from a single tree, and These are the intercept and slope model parameters obtained by fitting measured data from sample trees.

[0115] S6. Based on the corrected total number of real single-tree red pine cones obtained in step S4, predict the cone yield potential for the following year.

[0116] Furthermore, in step S6, based on the corrected total number of actual cones per tree obtained in step S4, and using the average weight of mature cones collected that year as a reference benchmark, the potential cone yield per tree for the following year is predicted and evaluated, resulting in:

[0117]

[0118] in, To predict the potential yield of single-tree cones for the next year, This represents the average weight of a single mature cone in that year.

[0119] It should be noted that relational terms such as "first" and "second" are used merely to distinguish one entity or operation from another, and do not necessarily require or imply any such actual relationship or order between these entities or operations. Furthermore, the terms "comprising," "including," or any other variations thereof are intended to cover non-exclusive inclusion, such that a process, method, article, or apparatus that comprises a list of elements includes not only those elements but also other elements not expressly listed, or elements inherent to such a process, method, article, or apparatus. Without further limitations, an element defined by the phrase "comprising one..." does not exclude the presence of other identical elements in the process, method, article, or apparatus that includes said element.

[0120] Although this application has been described above with reference to specific embodiments, various modifications can be made and components can be replaced with equivalents without departing from the scope of this application. In particular, as long as there is no structural conflict, the features in the specific embodiments disclosed in this application can be combined with each other in any way. The lack of an exhaustive description of these combinations in this specification is merely for the sake of brevity and resource conservation. Therefore, this application is not limited to the specific embodiments disclosed herein, but includes all technical solutions falling within the scope of the claims.

Claims

1. A method for predicting cone yield based on UAV imagery, characterized in that, Includes the following steps: S1. Select the cone development period as the image acquisition time window to collect UAV image data of Korean pine forest; set up fixed sample plots in the target forest stand, and conduct ground measurement surveys after the cones mature to obtain ground sample data; S2. The UAV image data of the red pine forest obtained in step S1 is preprocessed and then divided into training set, validation set and test set; S3. Construct a cascaded instance segmentation model based on multi-scale density attention enhancement, including a multi-scale feature extraction module, a multi-scale density attention enhancement module, and a cascaded instance segmentation optimization module connected in sequence. Use the training set obtained in step S2 to train the cascaded instance segmentation model based on multi-scale density attention enhancement, and obtain the prediction result of the sphere output by the trained cascaded instance segmentation model based on multi-scale density attention enhancement. S4. Correct the predicted cones obtained in step S3 to obtain the corrected total number of real mature cones per tree and the corrected total number of real small cones per tree; S5. Estimate the annual yield of individual tree cones based on the corrected total number of mature cones per tree obtained in step S4. S6. Based on the corrected total number of real single-tree red pine cones obtained in step S4, predict the cone yield potential for the following year.

2. The method for predicting cone yield based on UAV imagery according to claim 1, characterized in that, The specific implementation method of step S1 includes the following steps: S1.

1. During the cone development period, use a DJI drone platform equipped with a Zenmuse P1 camera to conduct aerial photography of the target Korean pine forest, acquire high-resolution RGB image data, and take pictures at noon on a clear and windless day to obtain drone image data of the Korean pine forest; S1.

2. Fixed sample plots were set up within the target forest stand. After the cones matured, the cone yield of the sample trees in the sample plots was measured. The two-year-old mature cones on the sample trees were harvested and counted, and the one-year-old small cones were visually counted. The harvested mature cones were weighed to obtain the yield of cones per tree. At the same time, RTK equipment was used to locate all the measured sample trees and obtain their spatial coordinates.

3. The method for predicting cone yield based on UAV imagery according to claim 2, characterized in that, In step S2, the acquired UAV image data of the Korean pine forest is cropped and divided into image blocks of fixed size. The one-year-old small cones and two-year-old mature cones in the image blocks are manually labeled using a labeling tool. The labeled sample data is then divided into training set, validation set and test set.

4. The method for predicting cone yield based on UAV imagery according to claim 3, characterized in that, The specific implementation method of step S3 includes the following steps: S3.

1. Establish a multi-scale feature extraction module, which extracts features from the input image through a deep convolutional network and a feature pyramid (FPN) structure to obtain a set of multi-scale feature maps; S3.

2. Establish a multi-scale density attention enhancement module, and extract features from the multi-scale feature maps obtained in step S3.1 to obtain spatial attention weights and channel attention weights; The multi-scale feature maps obtained in step S3.1 are aggregated across channels to obtain a two-dimensional response map. Then, local window statistics are performed on the two-dimensional response map to obtain a density response map. A 3×3 convolution operation is then performed on the density response map to fuse local receptive field information, and a boundary padding strategy is applied to maintain spatial dimension consistency. Finally, a spatial attention weight map is obtained by passing the sigmoid activation function. The multi-scale feature map obtained in step S3.1 is subjected to global average pooling (GAP) to obtain the channel description vector; then the channel description vector is processed by a two-layer fully connected network and activation function to extract the channel attention weights. Then, the spatial attention weights and channel attention weights are applied to the multi-scale feature maps obtained in step S3.1 to obtain the enhanced multi-scale feature map set; S3.

3. Establish a cascaded instance segmentation optimization module. Input the enhanced multi-scale feature map set obtained in step S3.2 into the cascaded instance segmentation network. First, the Region Generation Network (RPN) generates candidate boxes (RoIs) on the enhanced multi-scale feature maps; let the... The candidate boxes input in each cascade stage are Extract the region features corresponding to the candidate boxes and calculate the offset through a regression branch. This yields the updated candidate boxes: in, For the first Candidate boxes for each cascading stage: By setting progressively increasing Intersection over Union (IoU) thresholds, the target candidate boxes are refined in multiple stages to obtain region features based on the optimal candidate boxes. The network's segmentation head, Mask Head, predicts pixel-level masks of the target based on the regional features of the optimal candidate boxes. These masks characterize the spatial distribution of the target within the candidate region. Each pixel-level sphere boundary mask corresponds to a sphere instance, which describes the contour range of the sphere. S3.

4. Thresholding and spatial mapping are performed on the pixel-level cone boundary mask to obtain the instance segmentation results of the cones in the original image; the final cone instance set is obtained by filtering and deduplicating the instance segmentation results; the cone instance set is spatially matched with the canopy area of ​​a single tree, and the cone identification and quantity statistics at the single tree scale are realized according to the spatial position relationship between the cone instances and the canopy area. The predicted results of the cones output above include the detection box position of the cones, the pixel-level cone boundary mask, and the category attribute of the cones. The category attribute of the cones is either one-year-old small cones or two-year-old mature cones.

5. The method for predicting cone yield based on UAV imagery according to claim 4, characterized in that, In step S4, a quantitative relationship is established between the number of mature cones and small cones per tree visible to the drone and the actual total number, respectively. The mathematical expression is: in, This represents the corrected total number of mature cones per tree. This represents the corrected total number of true individual cones per tree. The number of mature cones detected by the drone. The number of small cones detected by the drone. This refers to the projected area of ​​a single tree crown extracted from drone imagery. The mean value of canopy texture features calculated based on the gray-level co-occurrence matrix; and The coefficients are the partial regression coefficients of the multiple linear regression for mature cones and small cones, respectively, obtained by fitting the data from ground sample trees using the least squares method. .

6. The method for predicting cone yield based on UAV imagery according to claim 5, characterized in that, In step S5, based on the measured number of mature cones and the corresponding total weight of cones from ground sample trees, a univariate linear quantitative relationship model between the number of mature cones per tree and the weight of the cones is established. Then, the corrected total number of mature cones per tree obtained in step S4 is substituted into the univariate linear quantitative relationship model between the number of mature cones per tree and the cone weight, to achieve a non-destructive estimation of the predicted cone weight per tree, resulting in: in, To predict the weight of mature cones from a single tree, and These are the intercept and slope model parameters obtained by fitting measured data from sample trees.

7. The method for predicting cone yield based on UAV imagery according to claim 6, characterized in that, In step S6, based on the corrected total number of actual cones per tree obtained in step S4, and using the average weight of mature cones collected that year as a reference benchmark, the potential cone yield per tree for the following year is predicted and evaluated, resulting in: in, To predict the potential yield of single-tree cones for the next year, This represents the average weight of a single mature cone in that year.

Images