Construction and application method of birch database based on remote sensing technology

Abstract

The application discloses a birch database construction and application method based on remote sensing technology and belongs to the technical field of birch database construction. The application solves the problem of low efficiency and poor result accuracy of traditional birch database construction caused by dependence on artificial investigation in the prior art. The application comprises the following steps: S1. Based on sub-meter satellite remote sensing data and artificial investigation data, white birch distribution extraction and overall reserve analysis and prediction are completed; S2. High-quality forest spectrum data and three-dimensional point cloud data collected by a UAV are used, supplemented by a small amount of artificial investigation data, to complete single-tree spatial distribution extraction and stand factor calculation, so that white birch species identification results and diameter at breast height estimation results are obtained; and S3. According to the white birch species identification results, the diameter at breast height estimation results and the overall reserve analysis and prediction, harvestable resource white birch juice collection planning and design are carried out. The application effectively improves the white birch identification efficiency, realizes the birch database construction and can be applied to automatic collection of birch data.

Description

Technical Field

[0001] This invention relates to methods for constructing and applying birch databases, and more particularly to methods for constructing and applying birch databases based on remote sensing technology, belonging to the field of birch database construction technology. Background Technology

[0002] Birch sap, a natural and healthy beverage, is rich in minerals, amino acids, and antioxidants. It not only possesses traditional medicinal value such as diuretic and anti-inflammatory properties but also generates economic benefits for residents in forest areas. Its development and utilization are of great significance for promoting local economic development and preserving cultural heritage. However, to achieve the sustainable use of birch sap resources, it is essential to strengthen the scientific monitoring of birch forests to accurately grasp their distribution, growth status, and health level, thereby formulating reasonable harvesting strategies and avoiding ecological damage caused by over-exploitation.

[0003] At present, there is a high demand for birch resource monitoring. Generally, the construction of a birch sap collection resource database is required. The specific requirements are as follows: (1) starting diameter class; (2) statistical requirements: birch trees in the corresponding area are counted by tree species and diameter class, and the data is filled in and reported. However, the traditional method of birch resource monitoring is still to conduct manual surveys on foot on each hill. This method has drawbacks such as low efficiency, insufficient data accuracy, high cost and difficulty in dynamic monitoring, which cannot meet the needs of modern resource management. Therefore, the current monitoring based on manual surveys has the following three main problems: (1) birch monitoring has high requirements, short time and heavy workload. (1) The degree of difficulty is large, and the investigation is more difficult and the risk factor is higher for resources that are difficult to collect; (2) The accuracy of the investigation report cannot be verified. The uncertainty of the results brought about by manual investigation and the statistical summary error that may occur in the form filling will greatly increase the uncertainty of the results of this investigation; (3) The reusability of the investigation results is low. The spatial distribution of the investigation results is not accurate enough, and it cannot provide a clear location for each birch sap collection. Moreover, the investigation is limited by technical means, and the starting diameter is a fixed value, which cannot be used for subsequent growth analysis of young birch forests. Furthermore, the investigation results obtained with a large amount of investment cannot be used for other business of resource management.

[0004] In summary, a method for constructing and applying a birch database based on remote sensing technology is needed. Summary of the Invention

[0005] A brief overview of the invention is given below to provide a basic understanding of certain aspects of it. It should be understood that this overview is not an exhaustive summary of the invention. It is not intended to identify key or essential parts of the invention, nor is it intended to limit the scope of the invention. Its purpose is merely to present certain concepts in a simplified form as a prelude to the more detailed description that follows.

[0006] In view of this, in order to solve the problems of low efficiency and poor accuracy of traditional birch database construction due to reliance on manual surveys, the present invention provides a method for constructing and applying a birch database based on remote sensing technology.

[0007] The technical solution is as follows: A method for constructing and applying a birch database based on remote sensing technology, including the following steps:

[0008] S1. Based on sub-meter level satellite remote sensing data and manual survey data, complete the extraction of birch tree distribution and overall reserve analysis and prediction;

[0009] S2. Using high-quality forest spectral data and three-dimensional point cloud data collected by UAVs, supplemented by a small amount of manual survey data, the spatial distribution of individual trees was extracted and the stand factors were calculated to obtain the birch tree species identification results and diameter at breast height estimation results.

[0010] S3. Based on the birch tree species identification results, diameter at breast height estimation results, and overall reserve analysis and prediction, conduct a planning and design for the collection of birch sap, a recoverable resource.

[0011] Furthermore, in S1, the spatial distribution of birch trees at the small-scale scale across the entire region is extracted. Based on the birch sap production of different age groups over the years, the total birch sap reserves in the entire region under the current condition are assessed. The growth trend of birch sap reserves in the future is then predicted based on the stand growth model. Specifically, this includes the following steps:

[0012] S11. Multi-source features, including spectral features, texture features, and sub-block attributes, are integrated to construct a birch distribution classification model, enabling birch proportion estimation at the sub-block scale. Spectral features include bands and vegetation indices, while texture features include eight parameters of the gray-level co-occurrence matrix. Initial identification is performed using a random forest model built with remote sensing data, outputting a birch probability map. Then, the birch probability map is used as one of the input features to train U-Net for refined segmentation. A confusion matrix is ​​used to evaluate the accuracy of birch stand identification results, thus achieving birch distribution extraction.

[0013] S12. Input historical birch sap production data into the random forest model, extrapolate to birch forests in the entire region, output prediction results, generate a spatial distribution map of birch sap reserves, and realize overall reserve analysis and prediction.

[0014] Furthermore, in step S2, a birch tree identification intelligent model is established. Through the operation of "single tree segmentation - tree species classification - factor extraction", the specific location of each birch tree, as well as the data extraction of diameter at breast height (DBH) and tree height, are realized. Specifically, the following steps are included:

[0015] S21. The measured location of the birch tree was obtained by collecting data through manual assistance;

[0016] In S21, the positioning personnel use the sampling point method to locate individual birch trees and measure their diameter at breast height (DBH). First, sampling points are set up at intervals of about 100m in the birch forest. The sampling points are extracted and planned in ArcGIS software. The coordinates of the sampling points are pre-entered into the Real-Time Dynamic Differential (RTK) system. After arriving at the site, the sampling points are located one by one according to the coordinates in the RTK system. After reaching the sampling point, the birch trees within sight are located clockwise with the sampling point as the center and the tree number is marked in the RTK system.

[0017] At the same time, the personnel recorded the diameter at breast height (DBH), tree number, tree species and DBH, marked the measured individual trees to avoid duplication, and then used the real-time dynamic differential system (RTK) to stake out to the next measurement point.

[0018] S22. Data is collected using drones, and the collected drone data is preprocessed to obtain normalized point cloud data;

[0019] In S22, a multi-echo UAV lidar device is used, the UAV point cloud data density is greater than a set value, and the RGB images of the UAV in the survey area are acquired simultaneously to color the point cloud or generate orthophotos.

[0020] The raw point cloud data collected by the UAV was classified into ground points using lidar point cloud data processing and analysis software. Based on the classified ground points, all vegetation point clouds were normalized to obtain normalized point cloud data.

[0021] S23. Using normalized point cloud data combined with a point cloud-based single-tree segmentation algorithm, the crown of a single tree is segmented to obtain the point cloud data of the entire region, i.e., the single-tree segmentation result.

[0022] S24. Based on the single-tree segmentation results and the measured birch tree locations, construct a birch tree sample dataset, input it into the constructed multimodal deep learning model PointNet++, and obtain the birch tree species identification results;

[0023] In step S24, the segmentation results of individual trees are matched with the measured birch locations to extract point clouds of all birch samples, construct a birch sample dataset, and input it into the multimodal deep learning model PointNet++ to identify birch trees from the point cloud data. The specific construction process of the multimodal deep learning model PointNet++ is as follows:

[0024] The point cloud data in the input birch sample dataset includes the spatial coordinates of each point and additional RGB color information, providing location and color information for each point. The multimodal deep learning model PointNet++ extracts geometric structure and color information features by establishing relationships between points in the local neighborhood.

[0025] The multimodal deep learning model PointNet++ groups point cloud data into several local regions using farthest point sampling (FPS). Each local region is then aggregated using a multilayer perceptron (MLP) to learn local features. By progressively expanding the local regions, multi-scale local features are learned. After extracting all local features, global max pooling is used to process all local features, thus forming a global feature vector for all point cloud data. ;

[0026] global feature vector Represented as:

[0027]

[0028] in, w represents the max pooling operation, and the characteristics of each local region are: , ;

[0029] The multimodal deep learning model PointNet++ learns multi-layer patterns from local geometry to global geometry by performing group sampling within the spatial neighborhood. After completing multi-layer feature extraction, PointNet++ extracts the global feature vector of each tree point cloud. The input is fed into multiple fully connected layers in a multilayer perceptron (MLP) for classification, and the output of the classification layer is a class probability distribution. , which represents the probability that the point cloud data belongs to each category, where, Let represent the probability distribution with s categories. , This represents the total number of categories in the output layer, and the classification layer uses the Softmax function to predict the probability distribution of tree species.

[0030] Category probability distribution Represented as:

[0031]

[0032] The birch tree recognition task of the multimodal deep learning model PointNet++ is set as a binary classification problem, so that the result only represents birch trees or non-birch trees. Cross-entropy loss is used to measure the difference between the predicted class and the true class. The parameters are updated by Adam optimizer or SGD, and Dropout and Batch Normalization techniques are combined to prevent the model from overfitting.

[0033] S25. Based on the point cloud of birch samples, perform diameter at breast height (DBH) inversion for birch trees;

[0034] In S25, based on the point cloud of the birch sample, the characteristics of individual birch trees are extracted, including height variables, intensity variables, canopy closure, gap ratio, and leaf area index.

[0035] A subset of features with high importance values ​​was selected by combining correlation analysis and recursive feature elimination, and retained according to the principle of balancing the minimum number of features and the highest accuracy.

[0036] The filtered features are input into a random forest model and trained through 10-fold cross-validation to obtain a chest diameter prediction model. The prediction accuracy of the trained chest diameter prediction model is then evaluated.

[0037] All birch trees identified in the region are used to predict their diameter at breast height (DBH) using the trained DBH prediction model described above. The number of birch trees of different DBH grades in the region is calculated, and the DBH estimation results are used as a reference for birch sap production.

[0038] Finally, based on the birch species identification results and diameter at breast height (DBH) estimation results, a regional birch distribution map and a statistical chart of birch DBH analysis were drawn.

[0039] Furthermore, in S3, the work of conducting pre-collection planning, mid-collection operation guidance, and resource monitoring throughout the entire collection process for collectable resources specifically includes the following steps:

[0040] S31. Conduct preliminary data collection planning and cost estimation;

[0041] In S31, the overall situation of the available resources is clearly defined. By overlaying regional road data and terrain data, a buffer zone is set up. At the same time, historical collection data is overlaid to finally delineate the collection area for this year.

[0042] The buffer zone is set as follows: a convenient work zone is defined for road elements, low suitability is assigned to areas with terrain slope > 30°, areas collected in recent years are removed, the above information is superimposed and calculated to obtain a suitability index grid, and the collection area for this year is defined by the threshold of the suitability index grid.

[0043] For birch trees that do not meet the diameter at breast height (DBH) collection standard, the growth cycle of birch trees that do not meet the collection requirements at the current stage should be assessed based on experience and reference to known single-tree DBH growth models. The collection time for such tree species should be reasonably analyzed for medium- and long-term collection planning.

[0044] Considering the costs of consumables, transportation, and personnel handling, and taking into account the relationship between the birch tree to be collected and the nearest road, a cost model is established to evaluate the cost of collecting birch sap.

[0045] S32. Provide mid-term work guidance;

[0046] In step S32, the spatial data of birch trees that can be collected is imported into the mobile operation system using the detailed spatial distribution of each birch tree. Combined with the high-resolution remote sensing base map and road layer, a collection navigation layer is generated. Based on the positioning navigation, the front-line collection personnel can clearly locate the specific birch tree location that needs to be collected. Combined with the collection record verification, the collection status is updated to avoid missed or incorrect collection. The tree sap yield, collection time, and collection personnel information are uploaded and synchronized to the central database.

[0047] S33. Conduct full-process resource supervision;

[0048] In S33, multi-temporal remote sensing images are used to regularly monitor the collection area. The change detection adopts the vegetation index NDVI / NBR differential method, and the deep learning change detection model U-Net is used to detect surface disturbances. Based on the fusion of spectral index with meteorological and soil moisture, a health index model is established to dynamically warn of the risk of declining birch sap production.

[0049] The beneficial effects of this invention are as follows: This invention adopts an integrated air-ground approach to construct a birch resource monitoring and database system, which is of great significance for addressing local forest resource monitoring needs and promoting the development of the birch sap industry. This invention combines advanced technologies such as remote sensing, big data, and artificial intelligence to construct an efficient and accurate birch resource monitoring system, ensuring that while protecting the ecological environment, birch sap resources are scientifically and rationally developed and utilized, achieving coordinated development of ecological and economic benefits. This invention adopts a technical approach that primarily relies on remote sensing supplemented by manual surveys, fully utilizing remote sensing and drone technologies, supplemented by a small number of personnel conducting on-site verification, to establish and optimize intelligent models. This allows for the inversion of birch tree distribution and diameter class, providing a birch database with detailed spatial distribution and more survey indicators. This invention is used for birch sap collection planning, cost accounting, and collection business management. It improves the accuracy and efficiency of baseline data surveys, liberates manpower, and enhances the scientific nature of birch sap collection management, providing a reference for the informatization of other under-forest economies. When using the solution of this invention to conduct birch tree identification and diameter at breast height (DBH) estimation verification in a forest farm, it uses drones to collect high-precision orthophotos and lidar point cloud data covering the entire forest farm. Combined with a small amount of ground measurement, it was able to successfully identify a total of 10,397 trees within a 33-hectare area of ​​the forest farm, including 308 birch trees, with an overall accuracy of 95.37%. It accurately identified the DBH of all birch trees, with a verification accuracy of 90.94%. In the overall experiment, the drone aerial photography took 1 day, and the data analysis took 1.5 days, which greatly improves efficiency compared to traditional manual surveys. Attached Figure Description

[0050] The accompanying drawings, which are included to provide a further understanding of the invention and form part of this invention, illustrate exemplary embodiments of the invention and are used to explain the invention, but do not constitute an undue limitation of the invention. In the drawings:

[0051] Figure 1 This is a flowchart illustrating the construction and application methods of a birch database based on remote sensing technology.

[0052] Figure 2 This is a schematic diagram of single-tree segmentation and birch identification based on lidar point clouds;

[0053] Figure 3 This is a schematic diagram of the architecture of the multimodal deep learning model PointNet++.

[0054] Figure 4 This is a schematic diagram showing the distribution of birch trees;

[0055] Figure 5 This is a bar chart showing the estimated chest diameter. Detailed Implementation

[0056] To make the technical solutions and advantages of the embodiments of the present invention clearer, the exemplary embodiments of the present invention will be further described in detail below with reference to the accompanying drawings. Obviously, the described embodiments are only a part of the embodiments of the present invention, and not an exhaustive list of all embodiments. It should be noted that, unless otherwise specified, the embodiments and features in the embodiments of the present invention can be combined with each other.

[0057] refer to Figures 1-5 This embodiment details the method for constructing and applying a birch database based on remote sensing technology, specifically including the following steps:

[0058] S1. Based on sub-meter level satellite remote sensing data (Gaofen-2) and manual survey data, complete the extraction of birch tree distribution and overall reserve analysis and prediction;

[0059] S2. Using high-quality forest spectral data and three-dimensional point cloud data collected by UAVs, supplemented by a small amount of manual survey data, the spatial distribution of individual trees was extracted and the stand factors were calculated to obtain the birch tree species identification results and diameter at breast height estimation results.

[0060] S3. Based on the birch tree species identification results, diameter at breast height estimation results, and overall reserve analysis and prediction, conduct a planning and design for the collection of birch sap, a recoverable resource.

[0061] Furthermore, in S1, the spatial distribution of birch trees at the small-scale scale across the entire region is extracted. Based on the birch sap production of different age groups over the years, the total birch sap reserves in the entire region under the current condition are assessed. The growth trend of birch sap reserves in the future is then predicted based on the stand growth model. Specifically, this includes the following steps:

[0062] S11. Multi-source features, including spectral features, texture features, and sub-block attributes, are integrated to construct a birch distribution classification model, enabling birch proportion estimation at the sub-block scale. Spectral features include bands and vegetation indices (NIR, NDVI, RVI), and texture features include eight parameters of the gray-level co-occurrence matrix. Initial identification is performed using a random forest model built with remote sensing data, outputting a birch probability map. Then, the birch probability map is used as one of the input features to train U-Net for refined segmentation. A confusion matrix is ​​used to evaluate the accuracy of birch stand identification results, thus achieving birch distribution extraction.

[0063] S12. Input historical birch sap production data into the random forest model, extrapolate to birch forests in the entire region, output prediction results, generate a spatial distribution map of birch sap reserves, and realize overall reserve analysis and prediction.

[0064] Furthermore, in step S2, a birch tree identification intelligent model is established. Through the operation of "single tree segmentation - tree species classification - factor extraction", the specific location of each birch tree, as well as the data extraction of diameter at breast height (DBH) and tree height, are realized. For different regions, the accuracy of tree species classification and factor extraction can be improved by adjusting the measured data parameters and optimizing the model. Specifically, the following steps are included:

[0065] S21. The measured location of the birch tree was obtained by collecting data through manual assistance;

[0066] In step S21, the positioning personnel use the sampling point method to locate individual birch trees and measure their diameter at breast height (DBH). First, sampling points are set up at intervals of about 100m in the birch forest. The sampling points are extracted and planned in ArcGIS software. The coordinates of the sampling points (WGS84_UTM projection coordinate system) are pre-entered into the Real-Time Dynamic Differential (RTK) system. After arriving at the site, the sampling points are searched one by one according to the coordinates in the RTK system. After reaching the sampling point, the birch trees within sight are located clockwise with a radius of about 10m as the center. The tree number is marked in the RTK system. If there are not many birch / black birch trees within a 10m radius, the search range can be expanded.

[0067] Meanwhile, the recorder measures the diameter at breast height (DBH), records the tree number, tree species and DBH, and uses spray paint to mark the measured individual trees to avoid duplication. After completion, the real-time dynamic differential system (RTK) is used to lay out the sample to the next sample point. The recorder can stand at the center of the sample point to guide the positioning personnel. In this embodiment, the final target number of birch samples is more than 300 trees.

[0068] S22. Data is collected using drones, and the collected drone data is preprocessed to obtain normalized point cloud data;

[0069] In S22, a multi-echo drone lidar device, such as DJI Zenmuse L2, is used. The drone point cloud data density is greater than a set value, i.e., 200pt / m2 is suitable. Simultaneously, RGB images of the drone in the survey area are acquired to color the point cloud or generate orthophotos.

[0070] The raw point cloud data collected by the UAV is classified into ground points using lidar point cloud data processing and analysis software (such as LiDAR360), and all vegetation point clouds are normalized based on the classified ground points to obtain normalized point cloud data.

[0071] S23. Using normalized point cloud data combined with a point cloud-based single-tree segmentation algorithm, the crown of a single tree is segmented to obtain the point cloud data of the entire region, i.e., the single-tree segmentation result.

[0072] S24. Based on the single-tree segmentation results and the measured birch tree locations, construct a birch tree sample dataset, input it into the constructed multimodal deep learning model PointNet++, and obtain the birch tree species identification results;

[0073] In step S24, the segmentation results of individual trees are matched with the measured birch locations to extract point clouds of all birch samples, construct a birch sample dataset, and input it into the multimodal deep learning model PointNet++ to identify birch trees from the point cloud data. The specific construction process of the multimodal deep learning model PointNet++ is as follows:

[0074] The point cloud data in the input birch sample dataset includes the spatial coordinates (horizontal coordinate X, vertical coordinate Y, height coordinate Z) of each point and additional RGB color information (red, green, blue), providing location and color information for each point. The multimodal deep learning model PointNet++ extracts geometric structure and color information features by establishing relationships between points in local neighborhoods (such as distance, direction, and angle), such as the straightness of the trunk, the shape of the crown (conical, umbrella, etc.), the branching hierarchy, and the color differences between the trunk and crown. These geometric features reflect the differences in morphology and color among tree species.

[0075] The multimodal deep learning model PointNet++ groups point cloud data into several local regions using farthest point sampling (FPS). Each local region is then aggregated using a multilayer perceptron (MLP) to learn local features. By progressively expanding the local regions, multi-scale local features are learned. After extracting all local features, global max pooling is used to process all local features, thus forming a global feature vector for all point cloud data. ;

[0076] global feature vector Represented as:

[0077]

[0078] in, w represents the max pooling operation, and the characteristics of each local region are: , ;

[0079] To better capture local structural features, the multimodal deep learning model PointNet++ performs grouped sampling within the spatial neighborhood to learn multi-layer patterns from local geometry to global geometry. After completing multi-layer feature extraction, the multimodal deep learning model PointNet++ extracts the global feature vector of each tree point cloud. The input is fed into multiple fully connected layers in a multilayer perceptron (MLP) for classification, and the output of the classification layer is a class probability distribution. , which represents the probability that the point cloud data belongs to each category, where, Let represent the probability distribution with s categories. , This represents the total number of categories in the output layer, and the classification layer uses the Softmax function to predict the probability distribution of tree species.

[0080] Category probability distribution Represented as:

[0081]

[0082] The birch tree recognition task of the multimodal deep learning model PointNet++ is set as a binary classification problem, so that the result can only represent birch trees or non-birch trees. Cross-entropy loss is used to measure the difference between the predicted class and the true class. The parameters are updated by Adam optimizer or SGD, and techniques such as Dropout and Batch Normalization are combined to prevent the model from overfitting.

[0083] S25. Based on the point cloud of birch samples, perform diameter at breast height (DBH) inversion for birch trees;

[0084] In S25, based on the point cloud of the birch sample, the characteristics of individual birch trees are extracted, including height variables, intensity variables, canopy closure, gap ratio, leaf area index, etc.

[0085] The subset of features with the highest importance values ​​was selected by combining correlation analysis (r>0.9) and recursive feature elimination, and retained according to the principle of balancing the minimum number of features and the highest accuracy.

[0086] The filtered features are input into a random forest model and trained through 10-fold cross-validation to obtain a chest diameter prediction model. The prediction accuracy of the trained chest diameter prediction model is then evaluated.

[0087] All birch trees identified in the region are used to predict their diameter at breast height (DBH) using the trained DBH prediction model described above. The number of birch trees of different DBH grades in the region is calculated, and the DBH estimation results are used as a reference for birch sap production.

[0088] Finally, based on the birch species identification results and diameter at breast height (DBH) estimation results, a regional birch distribution map and a statistical chart of birch DBH analysis were drawn.

[0089] Specifically, during the training phase of the multimodal deep learning model PointNet++, random enhancements such as rotation, translation, and color perturbation are applied to the point cloud, improving the model's generalization ability under different observation conditions. The multimodal deep learning model PointNet++ is a variant based on PointNet++, specifically designed to process point cloud data containing additional RGB information. Compared to the traditional PointNet++, the design of this invention's framework mainly integrates the use of RGB information, further enhancing the expressive power of point cloud data, especially for the recognition and classification of objects in complex environments.

[0090] Furthermore, in S3, the work of conducting pre-collection planning, mid-collection operation guidance, and resource monitoring throughout the entire collection process for collectable resources specifically includes the following steps:

[0091] S31. Conduct preliminary data collection planning and cost estimation;

[0092] In S31, the overall situation of the mineable resources (diameter at breast height ≥ 16cm) is clearly defined. By overlaying regional road data and terrain data, a buffer zone is set up. At the same time, historical collection data is overlaid to finally delineate the collection area for this year.

[0093] The buffer zone is set as follows: a buffer zone (50–200 m) is set for road elements to define a convenient operation zone. Areas with a terrain slope > 30° are assigned low suitability. Areas that have been collected in recent years are removed (historical collection buffer zone 30 m). The above information is superimposed and calculated to obtain a suitability index grid. The collection area for this year is defined by the threshold of the suitability index grid (e.g., suitability ≥ 0.6).

[0094] For birch trees that do not meet the diameter at breast height (DBH) collection standard, the growth cycle of birch trees that do not meet the collection requirements at the current stage should be assessed based on experience and reference to known single-tree DBH growth models. The collection time for such tree species should be reasonably analyzed for medium- and long-term collection planning.

[0095] Considering the costs of consumables, transportation, and personnel handling, and combining the relationship between the birch trees to be collected and the nearest road, a cost model is established to evaluate the cost of birch sap collection, providing data support for collection planning and contract pricing.

[0096] S32. Provide mid-term work guidance;

[0097] In step S32, the spatial data of birch trees that can be collected is imported into the mobile operation system using the detailed spatial distribution of each birch tree. Combined with the high-resolution remote sensing base map and road layer, a collection navigation layer is generated. Based on the positioning navigation, the front-line collection personnel can clearly locate the specific birch tree location that needs to be collected. Combined with the collection record verification, the collection status (not collected / collected / unsuitable for collection) is updated to avoid missed or incorrect collection. The tree sap yield, collection time, and collection personnel information are uploaded and synchronized to the central database.

[0098] S33. Conduct full-process resource supervision;

[0099] In S33, multi-temporal remote sensing images (Sentinel-2 / GF-2) are used to regularly monitor the collection area. Change detection adopts the vegetation index NDVI / NBR differential method. The deep learning change detection model U-Net is used to detect surface disturbances. Based on the fusion of spectral indices (NDVI, PSRI, NIRv) with meteorological and soil moisture, a health index model is established to dynamically warn of the risk of birch sap yield decline.

[0100] Although the invention has been described with reference to a limited number of embodiments, those skilled in the art will understand from the foregoing description that other embodiments are conceivable within the scope of the invention described herein. Furthermore, it should be noted that the language used in this specification has been chosen primarily for readability and instructional purposes, and not for the purpose of interpreting or limiting the subject matter of the invention. Therefore, many modifications and variations will be apparent to those skilled in the art without departing from the scope and spirit of the appended claims. The disclosure of the invention is illustrative and not restrictive, and the scope of the invention is defined by the appended claims.

Claims

1. A method for constructing and applying a birch database based on remote sensing technology, characterized in that, Includes the following steps: S1. Based on sub-meter level satellite remote sensing data and manual survey data, complete the extraction of birch tree distribution and overall reserve analysis and prediction; S2. Using high-quality forest spectral data and three-dimensional point cloud data collected by UAVs, supplemented by a small amount of manual survey data, the spatial distribution of individual trees was extracted and the stand factors were calculated to obtain the birch tree species identification results and diameter at breast height estimation results. S3. Based on the birch tree species identification results, diameter at breast height estimation results, and overall reserve analysis and prediction, conduct a planning and design for the collection of birch sap, a recoverable resource. In S1, the spatial distribution of birch trees at the small-scale scale across the entire region is extracted. Based on the birch sap production of different age groups over the years, the total birch sap reserves in the entire region under the current condition are assessed. Furthermore, the growth trend of birch sap reserves in the future is predicted based on the stand growth model. Specifically, this includes the following steps: S11. Multi-source features, including spectral features, texture features, and sub-block attributes, are integrated to construct a birch distribution classification model, enabling birch proportion estimation at the sub-block scale. Spectral features include bands and vegetation indices, while texture features include eight parameters of the gray-level co-occurrence matrix. Initial identification is performed using a random forest model built with remote sensing data, outputting a birch probability map. Then, the birch probability map is used as one of the input features to train U-Net for refined segmentation. A confusion matrix is ​​used to evaluate the accuracy of birch stand identification results, thus achieving birch distribution extraction. S12. Input historical birch sap yield data into the random forest model, extrapolate it to birch forests in the entire region, output the prediction results, generate a spatial distribution map of birch sap reserves, and realize overall reserve analysis and prediction. In step S2, a birch tree identification intelligent model is established. Through the operation of "single tree segmentation - tree species classification - factor extraction", the specific location, diameter at breast height (DBH), and tree height of each birch tree are extracted. Specifically, the following steps are included: S21. The measured location of the birch tree was obtained by collecting data through manual assistance; In S21, the positioning personnel use the sampling point method to locate individual birch trees and measure their diameter at breast height (DBH). First, sampling points are set up at intervals of about 100m in the birch forest. The sampling points are extracted and planned in ArcGIS software. The coordinates of the sampling points are pre-entered into the Real-Time Dynamic Differential (RTK) system. After arriving at the site, the sampling points are located one by one according to the coordinates in the RTK system. After reaching the sampling point, the birch trees within sight are located clockwise with the sampling point as the center and the tree number is marked in the RTK system. At the same time, the personnel recorded the diameter at breast height (DBH), tree number, tree species and DBH, marked the measured individual trees to avoid duplication, and then used the real-time dynamic differential system (RTK) to stake out to the next measurement point. S22. Data is collected using drones, and the collected drone data is preprocessed to obtain normalized point cloud data; In S22, a multi-echo UAV lidar device is used, the UAV point cloud data density is greater than a set value, and the RGB images of the UAV in the survey area are acquired simultaneously to color the point cloud or generate orthophotos. The raw point cloud data collected by the UAV was classified into ground points using lidar point cloud data processing and analysis software. Based on the classified ground points, all vegetation point clouds were normalized to obtain normalized point cloud data. S23. Using normalized point cloud data combined with a point cloud-based single-tree segmentation algorithm, the crown of a single tree is segmented to obtain the point cloud data of the entire region, i.e., the single-tree segmentation result. S24. Based on the single-tree segmentation results and the measured birch tree locations, construct a birch tree sample dataset, input it into the constructed multimodal deep learning model PointNet++, and obtain the birch tree species identification results; In step S24, the segmentation results of individual trees are matched with the measured birch locations to extract point clouds of all birch samples, construct a birch sample dataset, and input it into the multimodal deep learning model PointNet++ to identify birch trees from the point cloud data. The specific construction process of the multimodal deep learning model PointNet++ is as follows: The point cloud data in the input birch sample dataset includes the spatial coordinates of each point and additional RGB color information, providing location and color information for each point. The multimodal deep learning model PointNet++ extracts geometric structure and color information features by establishing relationships between points in the local neighborhood. The multimodal deep learning model PointNet++ groups point cloud data into several local regions using farthest point sampling (FPS). Each local region is then aggregated using a multilayer perceptron (MLP) to learn local features. By progressively expanding the local regions, multi-scale local features are learned. After extracting all local features, global max pooling is used to process all local features, thus forming a global feature vector for all point cloud data. ; global feature vector Represented as: in, w represents the max pooling operation, and the characteristics of each local region are: , ; The multimodal deep learning model PointNet++ learns multi-layer patterns from local geometry to global geometry by performing group sampling within the spatial neighborhood. After completing multi-layer feature extraction, PointNet++ extracts the global feature vector of each tree point cloud. The input is fed into multiple fully connected layers in a multilayer perceptron (MLP) for classification, and the output of the classification layer is a class probability distribution. , which represents the probability that the point cloud data belongs to each category, where, Let represent the probability distribution with s categories. , This represents the total number of categories in the output layer, and the classification layer uses the Softmax function to predict the probability distribution of tree species. Category probability distribution Represented as: The birch tree recognition task of the multimodal deep learning model PointNet++ is set as a binary classification problem, so that the result only represents birch trees or non-birch trees. Cross-entropy loss is used to measure the difference between the predicted class and the true class. The parameters are updated by Adam optimizer or SGD, and Dropout and Batch Normalization techniques are combined to prevent the model from overfitting. S25. Based on the point cloud of birch samples, perform diameter at breast height (DBH) inversion for birch trees; In S25, based on the point cloud of the birch sample, the characteristics of individual birch trees are extracted, including height variables, intensity variables, canopy closure, gap ratio, and leaf area index. A subset of features with high importance values ​​was selected by combining correlation analysis and recursive feature elimination, and retained according to the principle of balancing the minimum number of features and the highest accuracy. The filtered features are input into a random forest model and trained through 10-fold cross-validation to obtain a chest diameter prediction model. The prediction accuracy of the trained chest diameter prediction model is then evaluated. All birch trees identified in the region are used to predict their diameter at breast height (DBH) using the trained DBH prediction model described above. The number of birch trees of different DBH grades in the region is calculated, and the DBH estimation results are used as a reference for birch sap production. Finally, based on the birch species identification results and diameter at breast height (DBH) estimation results, a regional birch distribution map and a statistical chart of birch DBH analysis were drawn.

2. The method for constructing and applying a birch database based on remote sensing technology according to claim 1, characterized in that, In S3, the work of planning for the early stage of collection, providing operational guidance during the collection process, and supervising the resource throughout the entire collection process for recoverable resources specifically includes the following steps: S31. Conduct preliminary data collection planning and cost estimation; In S31, the overall situation of the available resources is clearly defined. By overlaying regional road data and terrain data, a buffer zone is set up. At the same time, historical collection data is overlaid to finally delineate the collection area for this year. The buffer zone is set as follows: a convenient work zone is defined for road elements, low suitability is assigned to areas with terrain slope > 30°, areas collected in recent years are removed, the above information is superimposed and calculated to obtain a suitability index grid, and the collection area for this year is defined by the threshold of the suitability index grid. For birch trees that do not meet the diameter at breast height (DBH) collection standard, the growth cycle of birch trees that do not meet the collection requirements at the current stage should be assessed based on experience and reference to known single-tree DBH growth models. The collection time for such tree species should be reasonably analyzed for medium- and long-term collection planning. Considering the costs of consumables, transportation, and personnel handling, and taking into account the relationship between the birch tree to be collected and the nearest road, a cost model is established to evaluate the cost of collecting birch sap. S32. Provide mid-term work guidance; In step S32, the spatial data of birch trees that can be collected is imported into the mobile operation system using the detailed spatial distribution of each birch tree. Combined with the high-resolution remote sensing base map and road layer, a collection navigation layer is generated. Based on the positioning navigation, the front-line collection personnel can clearly locate the specific birch tree location that needs to be collected. Combined with the collection record verification, the collection status is updated to avoid missed or incorrect collection. The tree sap yield, collection time, and collection personnel information are uploaded and synchronized to the central database. S33. Conduct full-process resource supervision; In S33, multi-temporal remote sensing images are used to regularly monitor the collection area. The change detection adopts the vegetation index NDVI / NBR differential method, and the deep learning change detection model UNet is used to detect surface disturbances. Based on the fusion of spectral index with meteorological and soil moisture, a health index model is established to dynamically warn of the risk of declining birch sap production.

Images