A deep learning-based forest tree species and number recognition method and system

Abstract

The application discloses a kind of forest species and quantity identification method and system based on deep learning, comprising the following steps: 1) data collection is carried out to the forest sample site to be investigated, obtains the forest image covering the whole sample site scene and makes training set;2) construct CBC-YOLOV5 target detection model, and the model constructed is trained using training set;3) using the CBC-YOLOV5 target detection model trained processes forest image collected in real time by camera, and the result of identification is further processed, obtains the species and quantity of sample site forest.The application solves the problems that the current forest resource investigation process is highly dependent on manual, high in cost, low in investigation efficiency, long in cycle, and the accuracy of the investigation results cannot be verified, thereby improving the efficiency and accuracy of forest resource investigation, and reducing the investigation cost of forest resources.

Description

TECHNICAL FIELD

[0001] The present application belongs to the technical field of forestry investigation, and particularly relates to a forest tree species and quantity identification method and system based on deep learning. BACKGROUND

[0002] Forests account for 75% of the total primary productivity of the earth's biosphere and 80% of the plant biomass, and also have many ecological functions such as water conservation, soil conservation, wind prevention and sand fixation. Therefore, forests are not only a valuable asset of mankind, but also play an irreplaceable role in protecting the living environment of mankind. How to make good use of forest resources while protecting the living environment of mankind is a common problem faced by all mankind. Forestry resource investigation mainly takes forest management units or county-level administrative regions as the investigation object, and checks the types, distribution, quantity and quality of forest, tree and forest land resources to objectively reflect the forest management situation in the investigation region. The results of forest resource investigation reflect the total amount and ecological condition of forests, and provide an important basis for rational management of forest resources and protection of the sustainable development of forests.

[0003] In order to better improve the growth quality of seedlings, facilitate the transplanting and warehouse management of seedlings, identify dead trees for compensation, and maximize the benefits of the nursery, it is of great significance to count the number and growth of each tree species in each plot of the nursery. At present, the nursery organizes the counting of the number of seedlings of each tree species only once a year, and the counting method is manual counting, taking the average of the number counted by workers, supervisors and owners respectively. Owners also occasionally conduct manual spot checks. This method has low accuracy, consumes time and effort, and is difficult to meet the basic requirements of information management of the nursery. Therefore, reducing the work content and difficulty of manual work in the process of forest resource investigation and improving the speed and accuracy of forest resource investigation are urgent problems in the process of forest resource investigation.

[0004] The current forest tree identification method and problems are as follows:

[0005] (1) Video shooting, the shooting process is the same as the manual counting process, which needs to enter the forest land on foot, the complex terrain of the forest land causes inconvenience to the shooting personnel and equipment, and the workload is huge. From the video effect, the visible range of regular forest land is wide, and the farthest clear tree can be seen to nearly ten rows; the effect of irregular forest land is very poor, multiple tree species are mixed together, which is difficult to distinguish on the video, and the arrangement is simultaneous, the size is different, and even there are trees with branches starting from the bottom, which are difficult to distinguish on the video.

[0006] (2) RFID technology, each tree is labeled, and the system counts the labels for convenient management. However, a large number of equipment is needed, the procurement cost is high, and once the subsequent inventory is involved, there will be asset loss. The owner mainly solves the counting problem of the number and types of trees, and the maintenance mainly focuses on personnel management, so the electronic tag method has little significance.

[0007] (3) The backpack mobile vision device is generally installed in a backpack, and needs to be carried by a person for shooting, needs to be walked into a forest land, the forest land has complex terrain, causes inconvenience to the shooting person and the device, the process is relatively laborious, and the labor intensity of the operator is increased. SUMMARY

[0008] To solve the above problems, the application discloses a kind of forest tree species and quantity identification method and system based on deep learning, solve the current forest resource investigation process in which artificial dependence is big, cost is high, investigation efficiency is low, period is long, the accuracy of investigation result cannot be verified and other problems, to improve the efficiency and accuracy of forest tree resource investigation, reduce the investigation cost of forest resources.

[0009] The specific scheme is as follows:

[0010] A kind of forest tree species and quantity identification method based on deep learning, it is characterized in that, it include the following steps:

[0011] 1) data collection is carried out to the forest sample plot to be investigated, and forest tree image covering the whole sample plot scene is acquired to make training set;

[0012] 2) CBC-YOLOV5 target detection model is constructed, and the model constructed is trained using training set;

[0013] 3) the CBC-YOLOV5 target detection model trained is used to process forest tree image collected in real time by camera, and the species and quantity of sample plot forest tree are obtained by further processing the result of identification.

[0014] As a further improvement of the application, the step 1) data collection is carried out to the forest sample plot to be investigated, and forest tree image covering the whole sample plot scene is acquired to make training set, the specific process is as follows:

[0015] 2.1) forest tree image covering the whole sample plot scene is acquired by using camera fixed on unmanned aerial vehicle to collect the forest sample plot to be investigated, constitutes sample data, and the sample data is collated, cleaned and labeled;

[0016] 2.2) the sample data is classified, the image is data enhanced, more training samples are generated, and training set is formed;

[0017] 2.3) the training set is labeled using labelme software.

[0018] As a further improvement of the application, the CBC-YOLOV5 target detection model constructed in step 2) is an improved model of the YOLOV5 model, the original Neck module of the YOLOV5 model is replaced with a CSP-PAN module, the PAN structure is used to obtain multi-level feature maps, the CSP network is used for feature connection and fusion between adjacent feature maps, and the CSP-PAN is further down-sampled once to add a smaller feature scale to improve the detection effect of large objects; on the basis of the new Neck module CSP-PAN module, a coordinate attention mechanism module CBAM is added to enhance the useful feature information weight while suppressing invalid feature information; an ASFF module is added between the Neck module and the Head module to correctly fuse feature information at different levels.

[0019] A deep learning-based forest species and quantity identification system, characterized by comprising the following modules:

[0020] Image input module: used for acquiring regional image data and sending the image data to the forest feature extraction calculation module;

[0021] Forest feature extraction calculation module: using the CBC-YOLOV5 target detection model to analyze the image data and extract the forest features in the image data to obtain the identification result of the forest;

[0022] Field investigation module: using a UAV mobile terminal to conduct field investigation and sampling of the forest in the region to be investigated;

[0023] Data management module: managing the forest region data obtained by field investigation and sampling, and sending the management result to the statistical analysis module

[0024] Statistical analysis module: statistically analyzing the forest species and quantity and related data information to form a technical analysis report of the investigation result.

[0025] The beneficial effects of the application are:

[0026] (1) Based on the YOLOV5 target detection model, the CBC-YOLOV5 target detection model is proposed to improve the detection accuracy of multi-scale objects.

[0027] (2) A method and system for identifying forest species and quantity using a UAV carrying a monocular camera are proposed.

[0028] The use of UAV technology can dynamically monitor the forest situation and effectively investigate the distribution of forest tree species, the quantity and quality of trees, and the growth situation. The application of UAV technology in forestry investigation can collect more objective and authentic data than traditional manual investigation, which can effectively improve the accuracy and efficiency of forestry investigation. BRIEF DESCRIPTION OF DRAWINGS

[0029] Figure 1 YOLOV5 model schematic diagram.

[0030] Figure 2 CBC-YOLOV5 model schematic diagram.

[0031] Figure 3 CBAM module schematic diagram. DETAILED DESCRIPTION

[0032] The present application will be further illustrated below in conjunction with the drawings and specific embodiments, and it should be understood that the following specific embodiments are only used to illustrate the present application and not to limit the scope of the present application.

[0033] In forestry investigation, due to the difficulty of surveying personnel to reach the complex environment on site, the traditional manual surveying is not comprehensive, the information is incomplete, the data results are single, the unmanned aerial vehicle is not limited by the terrain and the shelter, the evidence collection is comprehensive, the operation difficulty is reduced, the data collection efficiency is more than 10 times of manual operation, the photo and evidence map spot can be automatically associated, and the labor cost is reduced, therefore, compared with the traditional forestry investigation, the scientific and reasonable application of unmanned aerial vehicle technology is more time-saving, labor-saving and scientific and accurate, the timeliness and reliability of the forestry investigation are effectively improved, a scientific and accurate data basis is laid for the forestry construction and development, and an irreplaceable role is played in improving the forestry investigation efficiency and promoting the rapid and stable development of the forestry in China.

[0034] The traditional measurement method has low manual measurement efficiency, poor data objectivity, difficult standard control unification, low measurement equipment precision and low automation degree. Under complex background conditions, deep learning algorithm has self-learning ability and powerful feature extraction ability, and has been widely concerned in various industries, and has achieved certain results in remote sensing image target recognition and extraction tasks.

[0035] As shown in the figure, the present application provides a kind of forest species and quantity identification method and system based on deep learning, comprising the following steps:

[0036] 1) data collection is carried out on the forest sample site to be investigated, forest images covering the whole sample site scene are obtained, and a training set is made;

[0037] 2) a CBC-YOLOV5 target detection model is constructed, and the model constructed is trained by using the training set;

[0038] 3) the CBC-YOLOV5 target detection model trained is used to process the forest images collected by camera in real time, the results of identification are further processed, and the species and quantity of sample site forest are obtained.

[0039] In the embodiment, the step 1) is to collect data of the forest sample site to be investigated, obtain forest images covering the whole sample site scene and make a training set, and the specific process is as follows:

[0040] 2.1) Obtain forest images covering the whole sample site scene by collecting the forest sample site to be investigated by using a camera fixed on a UAV, and constitute sample data, and the sample data is arranged, cleaned and labeled;

[0041] 2.2) Classify the sample data, perform data enhancement on the images, generate more training samples, and form a training set;

[0042] 2.3) Label the training set by using the labelme software.

[0043] In the embodiment, the CBC-YOLOV5 target detection model constructed in the step 2) is an improved model of the YOLOV5 model, the original Neck module of the YOLOV5 model is changed into a CSP-PAN module, the PAN structure is used to obtain multi-level feature maps, the CSP network is used for feature connection and fusion between adjacent feature maps, and the CSP-PAN is down-sampled once again, a smaller feature scale is added to improve the detection effect of large objects; on the basis of the new Neck module CSP-PAN module, a coordinate attention mechanism module CBAM is added to enhance the useful feature information weight and suppress invalid feature information; an ASFF module is added between the Neck module and the Head module to correctly fuse feature information of different levels.

[0044] The application also provides a forest species and quantity recognition system based on deep learning, comprising the following modules:

[0045] An image input module is used to obtain regional image data and send the image data to a forest feature extraction calculation module;

[0046] The forest feature extraction calculation module is used to analyze the image data by using the CBC-YOLOV5 target detection model, extract forest features in the image data, and obtain a forest recognition result;

[0047] An outdoor investigation module is used to perform forest field investigation and sampling in the region to be investigated by using a UAV mobile terminal;

[0048] A data management module is used to manage forest region data obtained by field investigation and sampling, and send the management result to a statistical analysis module;

[0049] The statistical analysis module is used to count forest species and quantity, analyze related data information, and form a technical analysis report of the investigation result.

[0050] The technical means disclosed in the present application are not limited to the technical means disclosed in the above embodiments, and include technical solutions composed of any combination of the above technical features. It should be noted that, for those skilled in the art, without departing from the principles of the present application, a number of improvements and refinements can be made, and these improvements and refinements are also considered within the scope of protection of the present application.

Claims

1. A method for identifying forest species and quantities based on deep learning, characterized in that, Includes the following steps: 1) Collect data from the forest plots to be investigated, obtain images of trees covering the entire plot scene, and create a training set; 2) Construct a CBC-YOLOV5 object detection model and train it using the training set. The CBC-YOLOV5 object detection model is an improved version of the YOLOV5 model. The original YOLOV5 model's Neck module is replaced with a CSP-PAN module. The PAN structure is used to acquire multi-level feature maps, and the CSP network performs feature connection and fusion between adjacent feature maps. Based on the CSP-PAN, it is downsampled again, adding a smaller feature scale to improve the detection effect of large objects. Based on the new Neck module CSP-PAN, a coordinate attention mechanism module CBAM is added to enhance the weight of useful feature information while suppressing invalid feature information. An ASFF module is added between the Neck module and the Head module to correctly fuse feature information at different levels. 3) The trained CBC-YOLOV5 target detection model is used to process the real-time forest images captured by the camera. The recognition results are further processed to obtain the species and quantity of trees in the sample plot.

2. The method for identifying forest species and quantities based on deep learning according to claim 1, characterized in that, Step 1) involves collecting data from the forest plots to be surveyed, acquiring images of trees covering the entire plot area, and creating a training set. The specific process is as follows: 2.1) Use cameras fixed on drones to collect images of trees covering the entire forest plot to be investigated, forming sample data. The sample data is then organized, cleaned, and labeled. 2.2) Classify the sample data, perform data augmentation on the images, generate more training samples, and form a training set; 2.3) Use labelme software to label the training set.

3. A deep learning-based system for identifying forest species and quantities, characterized in that, Includes the following modules: Image input module: used to acquire regional image data and send the image data to the forest feature extraction and calculation module; The forest feature extraction and calculation module analyzes image data using the CBC-YOLOV5 object detection model to extract forest features and obtain forest identification results. The CBC-YOLOV5 object detection model is an improved version of the YOLOV5 model. The original YOLOV5 Neck module is replaced with a CSP-PAN module. The PAN structure is used to acquire multi-level feature maps, and the CSP network performs feature connections and fusion between adjacent feature maps. Based on the CSP-PAN module, a further downsampling is performed, adding a smaller feature scale to improve the detection effect of large objects. A coordinate attention mechanism (CBAM) module is added to the new Neck module (CSP-PAN) to enhance the weight of useful feature information while suppressing invalid feature information. An ASFF module is added between the Neck and Head modules to correctly fuse feature information at different levels. Field survey module: Using a drone mobile terminal, conduct on-site surveys and sampling of trees in the area to be surveyed; Data Management Module: Manages the forest area data collected from field surveys and sends the management results to the Statistics Module; Statistical Analysis Module: This module collects statistics on the types and quantities of trees, analyzes relevant data, and generates a technical analysis report of the survey results.

Images