| China is the country with the most bamboo resources in the world,and moso bamboo has great economic,ecological,edible and medicinal values.The growth and development of bamboo shoots has a unique high growth phenomenon or high growth stage,stem elongation is very rapid,DBH growth rate and shoot rate are also very high.The high growth stage is very important for bamboo cultivation,management and growth and development research of bamboo.Due to the uncertain timing and location of moso bamboo shoot emergence and height growth rate,plus the limitations of bamboo forest environment,there is no real-time,efficient and accurate monitoring method.This study proposes a method to obtain the high growth information of moso bamboo shoots based on deep learning.Firstly,a web camera is built to capture and preprocess bamboo images,then the static detection of moso bamboo shoots is studied from the perspective of target detection and instance segmentation,and a dynamic tracking method is proposed based on the static detection.It can provide decision support for yield estimation,irrigation and fertilizer management,shoot withdrawal treatment and other similar problems in the field of agriculture and forestry.The main work and results are as follows:1)The acquisition and pre-processing of moso bamboo shoot dataset.Obtaining high-definition images is a prerequisite for growth studies.Firstly,a forest camera is built to capture images in real time and store them in the cloud server,the camera is powered by solar panels and batteries,and the local machine is downloaded from the server for data augmentation to enhance diversity,and then data annotation is performed to prepare for subsequent experiments.2)Static Detection of high growth of bamboo shoots with deep learning.Taking the typical models YOLOv4 and Mask R-CNN as examples,the models are optimized and trained from the perspective of prior box,prediction box and detection box to obtain the final models Bamboo-YOLOv4 and Bamboo-Mask,which improve the detection accuracy and reduce the missed detection rate.Based on the detection box generated by the object detection algorithm,the number of moso bamboo shoots and the corresponding pixel height in the image can be obtained.The instance-based segmentation algorithm can also combine the obtained semantic information with Graham algorithm to calculate the pixel ground diameter of each moso bamboo shoot.According to the results,it is clear that the corresponding evaluation indexes m AP,F,and Io U of Bamboo-YOLOv4 are higher than the corresponding values of Bamboo-Mask.3)Dynamic tracking of moso bamboo shoots.A dynamic tracking method based on sorting and screening the coordinates of detection boxes is proposed,and Bamboo-YOLOv4 and Bamboo-Mask are deployed to the server for experimental verification.The results show that the former is more suitable for monitoring the growth of moso bamboo shoots throughout the growth stage.4)Development of online monitoring software for high growth of moso bamboo shoots.Based on Qt library and integrating Bamboo-YOLOv4 static detection and dynamic tracking algorithm,the human-computer interaction interface for online monitoring the growth of moso bamboo shoots can facilitate bamboo forest managers to monitor the growth of moso bamboo shoots online,and provide functions such as querying historical data and drawing. |
