Research On Deep Learning Based Sapling Counting And Height Measurement Method

Traditional nursery sapling inspection often uses manual sampling and height measurement with rulers,which is inefficient and inaccurate,and requires a lot of human resources for nurseries that need to keep an eye on sapling growth,making it difficult to meet the fast and efficient management requirements of modern forestry.Large nurseries or forestry farms often have a large area of saplings in acres,and the environment in which they grow is susceptible to weeds,rain,wind and auxiliary growth tools,resulting in very little sapling image data available and poor quality.The large number of saplings is small and the required accuracy must be at the centimetre level,making height measurement more difficult than for larger trees;the dense planting of young saplings makes it impossible to achieve the accuracy required by normal detection.The challenge of counting and measuring the height of saplings is addressed in four main areas of this study:(1)This study creates a dataset containing a variety of sapling data based on the large data requirements for training multiple neural networks.The database was collected using a drone,binocular camera and mobile phone as data collection tools,and a total of 11 different types of sapling data were collected and collated and labelled.In order to improve the training effect and generalization ability of the network,the collected sapling data had a large difference in sparsity,and the image data and label data were expanded simultaneously using 15 expansion means such as cutting,rotating and adding noise,laying the data foundation for the neural network to automatically count and measure the height of saplings.(2)This study uses the YOLO series of neural networks YOLOv3,YOLOv4 and the latest YOLOv7 proposed in recent years to complete the detection of saplings in a large field environment,which can complete the recognition of the category and number of saplings in images and videos,and also complete the counting detection in real time,making it more suitable for nursery staff to operate.In order to meet the load-bearing capacity of embedded mobile devices and the real-time needs of low-configuration management systems,this study improves YOLOv4 so that the network can maintain efficient feature extraction while massively reducing the amount of operations to complete real-time sapling counting.(3)Considering the relatively modest economic and hardware deployment conditions of nursery managers,this study proposes to use the traditional binocular camera technique to complete the height measurement of saplings in nurseries.Through the calibration of inexpensive binocular cameras and the deployment of the system environment,the method can complete the image,video and real-time height measurement of saplings;and the integration of the method into the improved YOLOv4 makes the real-time sapling detection framework based on the improved YOLOv4 network and binocular cameras fast and efficient in providing real-time measurement results of the height and number of saplings in nurseries.(4)In response to the problem of significant degradation in the detection accuracy of highdensity saplings that emerged in the YOLO series during the study,this study proposes the use of the CCTrans neural network to output a density map of sapling images to complete the measurement of high-density saplings,and through comparison experiments with the current state-of-the-art YOLOv7 detector,the two frameworks were used for training and detection of different species and densities of saplings respectively,reasoning The detection method is suitable for saplings with different sparsity.This paper uses a variety of neural networks to detect a variety of sapling data,as well as binocular vision techniques to measure the height of saplings,filling a sapling detection problem that has only been considered in a very small amount of published work in precision forestry.In terms of the final detection results,the neural network detection method based on sapling images offers the possibility of accurate field nursery statistics and provides a successful example of the application of deep learning techniques in forestry.