Research On Forest Canopy Segmentation Based On Deep Learning

Forest canopy images can reflect various parameters of plant population growth in forest ecosystems.These parameters can be used as inputs of ecological model and global ecosystem change model,which is of great significance to the study of the health and productivity level of ecosystem and the change of global climate.Forest canopy image segmentation has high research value in the field of forest ecosystem analysis.The current forest canopy segmentation focuses on the canopy as a whole and the sky background,so it is easy to lose some detailed information inside the canopy,which will adversely affect subsequent parameter determination and 3D modeling.In recent years,deep learning has become one of the most promising techniques for image analysis.This article studies the application of deep learning in various shade factors in segmenting canopy images.The main contributions include the following:(1)A forest canopy image dataset that can be used for deep learning technology analysis is constructed.First,the raw data of forest canopy images are collected using hemisphere photography.In order to improve the operation efficiency and universality of the segmentation algorithm,the convolutional neural network CenterNet is used to remove the invalid area of the experimental sample image,and its effective imaging area is obtained as the experimental object of the segmentation.Secondly,the detected experimental objects are labeled by using a combination of the artificial and K-means clustering algorithms.Finally,the data set is augmented using data augmentation,and the data set is divided according to the 5-fold cross-validation method.(2)The segmentation algorithm based on full convolutional neural network is studied.A parity-mixed cascaded hole convolution is used to replace part of the pooling layer in the full convolutional neural network to extract features.At the same time,a multi-scale feature fusion strategy is introduced.A new fully convolutional neural network model DB-FCN is proposed.The experimental results of segmentation of forest canopy image by this model are analyzed,and the effect of batch-size on model training is discussed.The segmentation results of traditional forest canopy image segmentation algorithms(Otsu method,two-dimensional Otsu method,three-dimensional Otsu method,maximum entropy method,k-means method and fuzzy c-means method)are compared with them.(3)The optimization theory of image segmentation algorithm of full convolutional neural network is studied.Based on the model DB-FCN,a portable dual attention mechanism module is added to optimize the feature selection.At the same time,the model is optimized using second-order conditional random fields.The improved model DBC-FCN-CRFs is used to segment forest canopy images.By analyzing the experimental results,the advantages and disadvantages of the classic FCN model and the improved FCN model are compared.This paper proposes an open segmentation method of forest canopy image,which can effectively improve the impact of data collected from different channels on the segmentation model,and achieve the segmentation of branches,leaves and sky in the effective imaging area of forest canopy image,which provides necessary technical support for the research of forest ecosystem.