Tree Ring Segmentation And Measurement Method Based On Random Forest And Convolutional Neural Network

Dendrochronological research requires tree-age and tree-ring width to estimate the infonnation of environmental changes and tree growth,so it is important to accurately extract the characteristics information such as the early wood,late wood,and bark parts in the tree-ring images for further analysis.Traditional tree ring research methods mainly recognize tree rings by eyes,and use tools to collect the information of tree rings.Multi-person cooperative work is required to reduce measurement errors leads to inefficiency.Recently,computer technology has been widely used in forestry research,which has better replaced some manual measurement work,especially the progress of computer vision and image processing technology,making it possible to develop a system for automatically extracting tree ring information.At present,the commercial tree ring analysis software is WinDENDRO of Canadian REGENT company and LINTAB of Frank Rinn company of Germany.Although these two software can realize the measurement of tree ring parameters,their automation is low,and they require a lot of manual interaction and are expensive.Precision and high quality image segmentation results of early wood,late wood and bark of tree ring is a necessary prerequisite for annual ring number counting and the ring width measurement.Unfortunately,it is difficult to obtain the desired effect using traditional image segmentation algorithm such as threshold segmentation algorithm,region growing segmentation algorithm and other segmentation algorithm because there are defects such as fuzzy interface between the early and late woods,knots and pseudo-annual rings during growth and burrs and noise spots on the image of the tree ring disc during the cutting and collecting process.Combining the multi-color space and texture features of the tree ring image,this paper uses random forest algorithm to construct a pixel classifier to segment the tree ring image.Considering the superior performance of convolutional neural network in image feature learning and classification and its powerful model generalization ability,this paper constructs a tree ring image segmentation model based on convolutional neural network,which realizes the precise segmentation of tree ring early-wood,late-wood and bark.Then the pith of the segmented tree ring image is automatically located,and the ring parameters such as tree age,ring width and late wood rate are measured by annular scanning method.The main research contents are as follows:(1)Image preprocessing,feature extraction of tree ring images and tree ring image segmentation based on random forest algorithm.Based on the analysis of the characteristics of the tree ring image,an approach to segment the tree ring image is proposed by using random forest algorithm,which combines the multi-color space features and texture features of the image.Firstly,9 color component of sample images data in RGB,HSV and L*a*b model and 8 texture features,namely mean values and standard deviations of contrast,correlation,energy and entropy from sample tree ring images based on Gray Level Co-occurrence Matrix,are extracted.And then an image pixel classifier based on random forest algorithm is constructed for tree ring segmentation using the training features which combined color and texture features randomly.A method named morphological was used to eliminate burrs and miscellaneous points to obtain a more accurate segmentation image.(2)Data augmentation and tree ring image segmentation based on convolution neural network.Depth neural network requires a large number of training samples for training and learning,but very few tree ring samples are available and labeling work is heavy,it is necessary to expand the number of training samples through data augmentation.In addition to traditional image rotation,perspective and color transformation,this paper also implements image deformation algorithm based on moving least squares,which meets the need of training depth neural network with small samples.The typical image segmentation model U-Net and DeepLab-v3-plus are implemented,and an improved U-shaped convolution neural network named I-UNet is proposed to solve the identifying difficult problems of rings near the pith and narrow rings.I-UNet deepen the depth of U-Net,adopt residual connection to avoid the gradient disappearance and use multiple losses calculation to retain the image bottom information.Similarly,batch normalization and dropout are also added to accelerate network training,enhance the network robustness and avoid the over-fitting phenomenon.Experiments show that the I-UNet model improves the tree ring segmentation effect and better solves the identifying difficult problems of rings near the pith and narrow rings.(3)Tree ring parameters automatic measurement based on circular scanning algorithm.Scanning lines are drawn from the image center in horizontal and vertical directions,and the tree pith coordinates are located by the intersection of the scanning lines which passing through the most number of annual rings.Radiused lines are drawn every 15°,the coordinates of every tree ring lines are recorded in all directions,tree age is the mode of tree ring lines in all radiused lines.Selecting the right radiused lines of tree age,calculating the tree ring width and late wood rate and other parameters.Experiments show that the ring scanning algorithm not only has high accuracy but also can retain the relevant data in tree ring parameter measurement.Based on the above research,this paper develops a system for tree ring parameters automatic measurement by using Python language and pyQT Library,which provides convenience for researchers to analyze and extract parameters from tree ring images.