Three-dimensional segmentation method of soil multi-category pores based on improved UNet-VAE network
-
摘要:
土壤不同类型孔隙结构会随生物活动和非生物作用发生形变,影响土壤孔隙整体生态功能。为研究孔隙结构与生态功能演变关系,需将不同类型的孔隙结构准确分割。目前,针对单个类别孔隙分割方法存在分割精度低、分类标准单一、鲁棒性差的问题,使得对于生物孔隙、裂隙等相交部分孔隙结构无法准确分割和判别。为此,该研究针对不同类型孔隙尺度差距大的特点,提出了一种改进UNet-VAE网络模型,首次实现土壤多类型孔隙分割。改进UNet-VAE网络引入多尺度特征融合注意力模块,以实现多尺度信息融合和冗余信息筛选。通过结合变分自动编码器生成网络(variational autoencoder,VAE),引入噪声和辅助损失函数,以增强网络的泛化能力和鲁棒性。试验结果表明:本文提出的改进UNet-VAE方法在土壤多类型孔隙(裂隙、生物孔、不规则孔隙和球状孔隙)三维分割中达到了93.83%的平均准确率,与次优VNet方法相比,平均准确率、精确率、召回率和F1值分别提升了3.32%,5.06%,8.97%和8.63%,特别是对于不规则孔隙四项指标分别提升了4.88%,15.46%,15.70%和15.50%。这证明了改进UNet-VAE法可准确分割多类型孔隙,也验证了深度学习技术在多类型孔隙判别的有效性,可为揭示土壤孔隙结构与演化研究提供有效工具。
Abstract:Soil pores plays a significant role in promoting crucial processes related to soil ecological functions. However, due to the lack of non-destructive and non-intrusive methods and systems for analyzing the spatial structure of multiple types of pores, studying the relationship between pore structure and functional evolution was extremely challenging. Among these, accurate segmentation of pore types and ranges was fundamental to the research. In this study, an improved UNet-VAE network method was proposed to achieve soil multi-category pore segmentation for the first time, providing technical support for studying the relationship between pore structure and ecological function evolution. Taking typical black soil as the research object, the Simplified Convolutional Network (SCN) method was used to segment soil pores and obtain three-dimensional data of soil pores. Based on the segmented pore dataset, a combination of automatic segmentation and manual correction was used to obtain four types of soil pore structure ground truth. Based on the 3D Unet network, a multi-scale fusion attention module was proposed to filter out redundant information generated by convolutional learning. Local attention was used to learn spatial features of small-scale pores (irregular pores and spherical pores), and global attention is used to extract feature information of large-scale pores (cracks and biological pores), to fuse multi-scale features of different categories of pores and improve the segmentation accuracy of different categories of pores. Meanwhile, commonly used segmentation networks in literature, such as 3D Unet network, Segresnet network, VNet network, and UNetR network, were used to achieve multi-category pores segmentation and compared with the proposed method. The experimental results showed that for large-scale cracks and biological pores, UNetR was difficult to learn features due to the high requirement of the Transformer for the number of datasets. Convolutional networks such as 3D UNet, Segresnet, and VNet lack the ability to learn global information and large-scale features, and classify cracks with obvious planar features as biological pores. For small-scale irregular pores, Segresnet, VNet, and UNetR all misclassified them as cracks. Except for the proposed improved network, all other networks exhibit under segmentation phenomenon. Comparing these five methods, the improved UNet-VAE method can accurately segment the pore range and determine the pore category. The improved UNet VAE achieved the best performance among all methods in four categories of pores, with average accuracy, precision, recall, and F1 values reaching 93.83%, 84.75%, 84.88%, and 84.60%, respectively. Compared with the suboptimal VNet method, the average accuracy, precision, recall, and F1 value have increased by 3.32%, 5.06%, 8.97%, and 8.63%, respectively. Especially for irregular pores, the accuracy, recall, and F1 value had increased by 4.88%, 15.46%, 15.70%, and 15.50%, respectively. In summary, the improved UNet-VAE had achieved high-precision three-dimensional segmentation of multiple categories of pores, with good feature learning ability for all four categories of pores, solving the problems of difficult classification of intersecting pores, low segmentation accuracy of single category pores, and low automation level of existing three-dimensional segmentation methods. This article will provide a data basis for the precise quantitative characterization of soil pore structure and a scientific basis for revealing the role of soil pore evolution in ecosystems.
-
Keywords:
- soils /
- pores /
- convolutional neural network /
- image segmentation /
- 3D UNet /
- attention module
-
0. 引 言
土壤孔隙是容纳水分和空气的空间,它在促进一系列土壤生态功能的重要过程中发挥着重要作用,如为微型和中型动物提供栖息地,支持水、气、营养物质的存储和运移,以及其他关键过程 [1-4]。孔隙的大小和形状对其生态功能起着决定型作用[5-6]。因此,根据形态对孔隙进行准确分类具有重要的意义。
近年来,基于计算机断层扫描(computed tomography,CT)进行孔隙分类、以及土壤孔隙结构和功能关系的研究越来越广泛[5-8]。前人基于土壤CT图像,对孔隙进行了大量的研究,如CAPOWIEZ等[9]基于体积阈值对孔隙开展了分类;张中彬等[10]进一步引入形状因子消除了大体积的非生物孔隙。然而,有研究表明,海绵状裂隙和管状生物孔隙之间存在紧密连接[11],难以通过传统图像分割方法进行分割,如ROONEY等[12]分割后仍发现与生物孔垂直相交的片状孔隙;由于外力活动生物孔隙破碎生成的小体积孔隙,具有连通性差等特征,在水气运移方面与非生物孔的功能更相似[13]。由此可见,有必要结合孔隙形态与形成原因分类,对不同的孔隙功能进行研究。
生物孔能提高透水性、溶质运移和渗透率[14-15]。耕作、冻融或干湿等非生物活动形成的非生物孔形状各异[16]:形成的呈片状的裂隙可作为优先流路径,增加农田水分和养分的流失以及地下水污染的风险[17];呈不规则形态的孔隙,有利于储水储气[18];球状孔隙由土壤干燥过程中空气截留形成[19]。基于孔隙形态将土壤孔隙分为生物孔、裂隙、不规则孔隙和球状孔隙,可以为研究孔隙结构与功能演变提供数据基础。
孔隙分割技术发展较快且取得一定成效,但也面临着若干技术层面的挑战:1)仅使用特征阈值进行多类别分割难以区分不同类别的相交孔隙,尤其是常出现相交的裂隙和生物孔[11,12,20];2)受孔隙多尺度特征影响,目前分割方法对于单个类别孔隙分割精度低;3)现有三维分割方法自动化程度低,需要手动多次调试为每个CT图像选择适当阈值以确保准确分割的重要性[21]。近年来,深度学习在土壤科学领域中,逐渐得到了广泛的应用[22-23],为解决这些挑战提供了一种有效的思路和方法。深度学习能够实现体素化分割,解决相交孔隙难以界定的问题;通过训练模型学习多尺度特征,从而提高分割精度;并且实现自动进行特征提取、和分割,无需手动调试和选择阈值。在经典卷积网络方面,韩巧玲等[24]基于简化卷积网络进行孔隙分割。傅尹开等[25]结合半监督训练模型、多尺度感受野结构提高孔隙多尺度特征信息提取能力。虽然深度学习从多个方向不断提升分割精度,但均是针对整体孔隙进行分割,无法实现多类型孔隙高精度分割,且仍存在多尺度孔隙特征难以学习及分割方法泛化性差等问题。
为解决阈值法相交孔隙难以分割、精度和自动化程度较低问题,实现土壤多类型孔隙精确分割,本文提出了一种改进UNet-VAE网络模型,以分割生物孔、裂隙、不规则孔隙和球状孔隙四类土壤孔隙。该模型在3DUNet网络基础上,提出多尺度融合注意力模块,融合多尺度信息、筛选冗余信息,并在训练中通过变分自编码器(variational autoencoders,VAE)分支引入噪声和辅助损失函数,以增强网络泛化能力、提高网络鲁棒性,以期为土壤孔隙结构精细量化表征提供数据基础,为揭示土壤孔隙演化在生态系统中的作用提供科学依据。
1. 材料与方法
1.1 图像采集与预处理
1.1.1 数据获取
土壤样品采集于中国黑龙江省克山农场,样品内径为10 cm,高度为10 cm。于层深为0~40 cm处分别修整黑土剖面。通过机械分层法,以10 cm的间距为标准使用环刀(100 cm3)逐层进行原状黑土取样。采集的黑土样本质地为黏壤土,有机质含量为65 g/kg,总大孔隙率为54.95%,平均体积含水率为31.71%。随机采取3个土样进行冻融处理,分别进行1、3、5次冰冻和融化[26]。采用黑龙江中医药大学第一附属医院CT扫描中心的螺旋CT扫描仪对样品进行扫描并获取高分辨图像(每像素点实际长度为0.236 mm)。每个土壤样本的每次处理产生220个横截面图像,共产生21组图像,详细处理方法见文献[27],并将数据集按照8:1:1划分为训练集、验证集和测试集。
1.1.2 数据预处理
土壤的原始CT图像、矫正图像和初始分割图像如图1所示。土筒和PVC玻璃管在运输过程中的接触振动,会影响土壤边界孔隙结构,因此使用内切法对土壤CT图像进行裁剪以去除干扰区域,保留图1a中的蓝框区域。通过自适应中值滤波方法去除孤立的噪声点,增强孔隙的边界特征,确保CT图像的可用性,如图1b所示。滤波后的图像集为后续土壤孔隙分割和标记任务奠定数据基础。
图 1 土壤CT孔隙分割数据集过程示意图注:蓝框代表裁剪范围;红圈代表实际扫描范围。AMF为自适应中值滤波处理。SCN为简化卷积网络。Figure 1. Diagram of soil CT pore segmentation dataset processNote: Blue frame represents the cropping range; Red circle represents the actual scanning range. AMF is adaptive median filtering. SCN is simplified convolutional network.为了获得精确的土壤孔隙结构,使用SCN方法[24]进行土壤孔隙分割并建立孔隙分割数据集。SCN方法的土壤孔隙分割平均准确率为99.61%,分割图像如图1c所示。
1.1.3 孔隙数据集标注
基于分割后的孔隙数据集,采用自动分割和手动校正相结合的方法进行4类土壤孔隙结构的标定。裂隙是弯曲的、不同宽度的片状结构[28];生物孔包括非分支、分支和网状孔结构[28];不规则孔隙呈现随机的形状和大小,外壁可能是有角度的或平滑弯曲的[29];球状孔隙尺寸较小[19]。
标定流程如图2a所示,首先,采用Python软件, 结合三维特征,体素、形状因子(孔长L与等效孔半径r的比值)进行自动分割[13]。然后,使用图2b所示三维标定软件3D Slicer,手动矫正标定相交的裂隙、生物孔及不同特征的非生物孔隙,建立了裂隙、生物孔、不规则孔隙和球状孔隙的真值数据集。其中裂隙、生物孔、不规则孔隙和球状孔隙的体素分别标记标签1、2、3和4,其余体素被视为背景标记为空。为了可视化4类孔隙空间分布及特征,对4类孔隙分配不同颜色:蓝色代表裂隙,红色代表生物孔,绿色代表不规则孔隙,黄色代表球状孔隙。
每个多类型孔隙三维标定数据都由5名具有土壤物理学知识背景人员进行重复标定,以消除主观性对标定精度的影响。
1.2 改进UNet-VAE土壤孔隙三维多类别分割模型构建
3D UNet将UNet二维卷积替换为三维卷积,实现多感受野俘获整体空间信息进行三维分割。但使用跳跃连接直接融合编码器与解码器输出的语义信息,吸取了大量冗余信息,导致其对多尺度土壤孔隙分割能力效果较差,尤其对于不规则孔隙难以辨别。为解决上述问题,本文基于3D UNet进行改进,提出改进UNet-VAE土壤多类型孔隙三维分割网络,网络结构如图3所示。编码器通过下采样卷积进行特征提取、解码器通过上采样卷积恢复图像并进行类别分割。提出了多尺度融合注意力模块来筛选由于卷积学习产生的冗余信息;通过局部注意力学习小尺度孔隙(不规则孔隙和球状孔隙)空间特征;通过全局注意力提取大尺度孔隙(裂隙和生物孔)特征信息,从而融合不同类型孔隙多尺度特征,以提高不同类型孔隙的分割精度。同时,在训练中通过VAE分支引入噪声从而避免过拟合;引入VAE辅助损失函数,通过对潜在空间中的向量进行约束,使得网络学得不同类型孔隙的特征分布,从而提高网络的泛化能力。
图 3 改进UNet-VAE网络结构注:MFA为多尺度融合注意力模块;图中数字表示空间维度数;N(μ,σ2)是均值为μ,方差为σ2的多元正态分布。Figure 3. Structure of improved UNet-VAE networkNote: MFA is the multi-scale fusion attention module; the number represents the spatial dimension; N(μ,σ2) is a multivariate normal distribution with a mean of μ and a variance of σ2.1.2.1 多尺度融合注意力模块
为融合多尺度特征信息、减少冗余信息,本研究提出多尺度融合注意力模块(multi-scale fusion attention,MFA),其结构如图4所示。该模块在跳跃连接中加入基于注意力门的多尺度特征融合注意力(attentional feature fusion,AFF)。
图 5 多尺度融合注意力模块结构图注:X是尺度为Cx×Hx×Wx×Lx的编码器侧输入特征图,G是尺度为Cg×Hg×Wg×Lg的MFA解码器侧输入特征图,X'是尺度为1×Hx×Wx×Lx的筛减冗余信息后的特征图,X"是权重注意力特征图,Z是最终输出特征图,ReLU是ReLU激活函数,Sigmoid是Sigmoid激活函数。Figure 5. Structure diagram of multi-scale fusion attention moduleNote: X is the encoder side input feature map with a scale of Cx×Hx×Wx×Lx, G is the MFA decoder side input feature map with a scale of Cg×Hg×Wg×Lg, X' is the feature map after filtering redundant information with a scale of 1×Hx×Wx×Lx, X" is the weighted attention feature map, Z is the final output feature map, ReLU is the ReLU activation function, and Sigmoid is the Sigmoid activation function.MFA解码器侧输入特征图G经上采样后尺度与编码器侧输入特征图X一致,皆为CX × HX × WX × LX(CX为输入特征图的通道数,HX为输入特征图的高度,WX为输入特征图的宽度,LX为输入特征图的深度)。通过特征相加,实现初始特征融合,并通过ReLU激活函数和逐点卷积得到筛选冗余信息后的注意力特征图XA。计算式如下:
$$ {X_A} = {\text{PWCon}}{{\text{v}}_{\text{1}}}(\delta ({\text{UPConv}}(G) + X)) $$ (1) 式中UPConv上卷积使Cg通道数减半,Hg、Wg、Lg分别扩大2倍;PWConv11×1点卷积将CX通道数缩减为1,即XA∈1×HX×WX×LX;$ \delta $表示ReLU激活函数。
注意力特征图XA与X相乘,得到筛减冗余信息后的特征图X'。
$$ X{'} = X \otimes {X_A} $$ (2) 特征图X'在通道注意力模块中分别使用局部注意力学习细节信息和全局注意力吸取上下文特征;L(X')计算式如下所示:
$$ L\left( {X{'}} \right) = B({\text{PWCon}}{{\text{v}}_3}(\delta (B({\text{PWCon}}{{\text{v}}_2}(X{'}))))) $$ (3) 式中PWConv21×1点卷积将X'通道数减少为原先的$ {1 \mathord{\left/ {\vphantom {1 \tau }} \right. } \tau } $;B表示BatchNorm层;PWConv31×1的卷积将通道数目恢复成与原输入通道数目相同;$ \tau $为通道缩放比。
相比局部特征通道注意力,全局特征通道注意力需要对X'先进行全局平均池化操作,其计算公式G(X')如下:
$$ G\left( {X{'}} \right) = B({\text{PWCon}}{{\text{v}}_3}(\delta (B({\text{PWCon}}{{\text{v}}_2}({\text{Pool}}(X{'})))))) $$ (4) 式中Pool表示全局池化。
计算之后的注意力权重分别为X和G分配权重后得到输出Z公式如下:
$$ X{'}{'} = {\text{Sigmoid}}(L(X{'}) + G(X{'})) $$ (5) $$ Z = X{'}{'} \otimes X + (1 - X{'}{'}) \otimes G $$ (6) 式中Sigmoid表示Sigmoid激活函数;X"为权重注意力特征图。
1.2.2 变分自动编码器生成网络分支
VAE是一个带有编码器和解码器的生成模型,可将输入数据x编码为潜在变量z,并将潜在变量解码为输出数据$ \bar x $。具体来说,VAE假设潜在变量来自高斯分布,VAE的编码器学习输入数据的分布参数:均值(μ)和方差(σ2)。然后通过从分布参数中采样得到潜在变量z并解码为输出数据$ \bar x $。
即VAE模型通过编码过程Q(z|x)将样本映射为潜在变量z,并假设潜在变量服从多元正态分布P(x)~N(0,I),解码器P(x|z)从隐藏变量z中抽取样本,生成指定图像$ \bar x $。通过最大化变分下界共同训练近似后验模型和生成模型。变分下界的表达式为
$$ E(q) = {E_z}[\log p(x|z)] - {D_{{\text{KL}}}}[q(z|x)||p(z)] $$ (7) VAE分支结构如图5所示,其主要作用是避免过拟合问题并提高网络的泛化能力。由图5可知,编码器端输出数据被减少到128的低维潜在空间,其中64维度用于表示平均值,64维度用于表示标准差。从具有给定均值和标准差的高斯分布中提取样本,然后按照与解码器相同的架构重建为输入图像维度。
1.2.3 损失函数
损失函数由Dice损失和VAE损失组合构成:
$$ L = {L_{{\text{Dice}}}} + 0.1({L_{{\text{Rec}}}} + {L_{{\text{KL}}}}) $$ (8) 选择0.1的超参数(正则化因子权重)以在Dice损失和VAE损失之间提供良好的平衡[30]。
1)Dice损失
设y和$ \overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{y} $分别是分割和模型预测的真值,为了避免训练数据没有标签,如$ y = \overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{y} = 0 $,将$ \varepsilon $添加到分子和分母中。Dice损失定义如下:
$$ {L_{{\text{Dice}}}}(y,\overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{y} ) = 1 - \frac{{2y\overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{y} + \varepsilon }}{{y + \overset{\lower0.5em\hbox{$\smash{\scriptscriptstyle\frown}$}}{y} + \varepsilon }} $$ (9) 通过Dice损失计算分割预测和真值之间的差距,Dice损失越小分割越准确。
2)VAE损失
VAE损失是VAE分支上的重建损失LRec和标准VAE惩罚项LKL的总损失。在本研究中,LRec是每个体素的均方误差。
$$ {L_{{\text{VAE}}}} = {L_{{Re} c}} + {L_{{\text{KL}}}} $$ (10) $$ {L_{{Re} c}} = ||{x_{{\text{rebuilding}}}} - x||_2^2 $$ (11) 式中xrebuilding和x分别表示重建图像和输入图像;LRec计算重建图像和输入图像之间的差距。LRec越小,生成图像与输入图像越相似。
由于重构是在潜在空间中采样进行生成。在网络反向传播中,生成器编码器趋势是潜在空间方差逐渐趋近于0。为了保证潜在空间的随机性,引入LKL。
LKL是估计的正态分布和先验分布之间的Kullback–Leibler散度。
$$ {L_{{\text{KL}}}} = \frac{1}{{{N_{{\text{totalvoxels}}}}}}\sum {{\mu ^2} + } {\sigma ^2} - \log {\sigma ^2} - 1 $$ (12) 式中$ {N_{{\text{totalvoxels}}}} $是图像体素的总数。LKL越小,解码器输出的和潜在空间采样分布越接近,即越接近正态分布,从而防止潜在空间方差为0。
1.2.4 评价指标
为了更好地评价网络对于多类型土壤孔隙分割性能。本研究以识别准确率(A)、精确率(P)、召回率(R)、F1分数(F1)以及四项指标的平均值进行指标评价,指标数值越大效果越好[31]。各指标计算公式如下:
$$ A = \frac{{{T_P} + {T_N}}}{{{T_P} + {F_P} + {T_N} + {F_N}}} $$ (13) $$ P = \frac{{{T_P}}}{{{T_P} + {F_P}}} $$ (14) $$ R = \frac{{{T_P}}}{{{T_P} + {F_N}}} $$ (15) $$ {F_1} = 2 \times \frac{{P \times R}}{{P + R}} $$ (16) 式中TP:实例为正确分类孔隙,分割也为正确分类孔隙;FN:实例为正确分类孔隙,分割为其他类型孔隙及背景;FP:实例为其他类型孔隙及背景,分割为正确分类孔隙;TN:实例为其他类型孔隙及背景,分割为其他类型孔隙及背景。
1.2.5 试验环境以及参数设置
本研究所有网络模型均在Featurize服务器租赁平台上linux环境下使用Pytorch进行训练。硬件配置如下,内存大小60.9 GB,CPU为30核AMD EPYC
7742 ,显卡型号为RTX A6000,显存为51.0 GB,模型训练超参数设置如下,batch size设置为1,训练本文所有方法优化器选用AdamW,学习率为0.001,权重衰减为0.001,所有网络模型训练时最大迭代次数为200。2. 结果与分析
2.1 消融试验结果对比分析
为验证各改进模块在多类型孔隙分割上的有效性,对各改进模块进行消融试验。对比原始3D UNet,增加MFA、VAE和结合2个改进点的改进UNet-VAE网络模型分割性能,分别统计不同模型识别分类的土壤孔隙(裂隙、生物孔、不规则孔隙和球状孔隙)的准确率、精确率、召回率和F1分数。结果如表1所示。
表 1 不同改进网络的多类型孔隙分割效果Table 1. Multi-category segmentation effect of differentimproved networks% 类型
Categories指标
Index模型1
Model1模型2
Model2模型3
Model3模型4
Model4裂隙
Cracks准确率 78.24±6.4 91.77±4.2 83.20±5.2 92.29±3.7 精确率 72.70±9.9 82.81±12.4 74.92±11.9 83.24±11.6 召回率 66.31±10.7 84.77±9.0 70.98±10.4 84.23±9.2 F1分数 68.95±8.7 83.13±8.4 72.39±9.4 83.40±9.1 生物孔
Biological pores准确率 75.57±6.3 90.74±4.3 81.57±5.1 91.24±3.8 精确率 76.75±9.0 89.31±7.7 79.38±10.6 88.53±8.2 召回率 77.38±10.8 87.01±10.6 83.06±9.2 89.08±8.8 F1分数 76.92±9.4 87.88±8.3 81.05±9.4 88.72±8.1 不规则孔隙
Irregular pores准确率 76.63±6.9 93.67±2.9 84.87±4.3 94.44±2.4 精确率 45.65±11.3 71.34±13.5 63.27±12.6 75.41±11.8 召回率 49.71±6.3 73.40±10.2 60.95±7.5 74.93±9.8 F1分数 47.16±8.6 72.03±11.2 61.67±9.2 74.82±9.3 球状孔隙
Spherical pores准确率 81.04±5.3 96.98±1.2 88.37±3.1 97.38±1.1 精确率 64.51±5.1 90.66±5.0 82.09±4.1 91.84±4.5 召回率 61.39±6.0 90.03±3.8 76.79±5.0 91.27±3.3 F1分数 62.73±4.3 90.20±2.7 79.26±3.7 91.45±2.5 平均值
Mean value准确率 77.87±6.2 93.29±3.1 84.50±4.4 93.83±2.7 精确率 64.90±8.8 83.83±9.6 83.16±9.8 84.75±9.0 召回率 63.70±8.4 83.97±8.4 82.78±8.0 84.87±7.8 F1分数 63.94±7.7 83.66±7.6 82.66±7.9 84.59±7.2 注:模型1:原始3D UNet网络;模型2:结合了MFA的改进UNet网络;模型3:引入了VAE的UNet-VAE网络;模型4:结合了MFA和VAE的改进UNet-VAE网络。下同。 Note: Model 1: the original 3D UNet network; Model 2: an improved UNet network that combines MFA; Model 3: a UNet-VAE network with the introduction of VAE; Model 4: an improved UNet-VAE network that combines MFA and VAE. Same as below. 从表1可以看出,在相同的训练环境下,在加入MFA后,4类孔隙分割都有了明显提升效果。4类孔隙准确率分别提升了13.53%、15.17%、17.04%和15.94%;F1分数分别提高了14.18%、10.96%、24.87%和27.47%。对体积较小、形状不规则的不规则孔隙性能提升效果最为明显,F1分数和准确率分别提升了24.87%、17.04%。小体积的球状孔隙的F1分数和准确率分别提升了27.47%、15.94%。因此,使用MFA可以提升4类孔隙特征学习能力,对于小尺寸孔隙结构具有良好的分割效果。
与原始3D UNet网络相比,增加VAE分支后的网络模型4类孔隙均提升了分割性能,4类孔隙准确率分别提升了4.96%、6.00%、8.24%和7.33%;F1分数分别提高了3.44%、4.13%、14.51%和16.53%。说明通过引入噪声和辅助损失函数,VAE有效地提高了网络模型泛化性,从而提高网络分割性能。
与次优指标表现的改进UNet网络相比,改进UNet-VAE网络对于4类孔隙准确率分别提升了0.52%、0.50%、0.77%和0.40%;F1分数分别提高了0.27%、0.84%、2.79%和1.25%,具有最佳的分割效果。
为了直观体现不同模块对于改进网络的影响,使用4类孔隙平均分割指标,在表1平均值中可以看出MFA和VAE分支都对4类孔隙分割效果有明显提升。
结合MFA的改进UNet网络分别在平均准确率、精确率、召回率和F1分数上提升了15.42%、18.93%、20.27%和19.72%。引入VAE的UNet-VAE网络分别在平均准确率、精确率、召回率和F1分数上提升了6.63%、18.26%、19.08%和18.72%。本文提出的同时结合MFA和VAE分支的方法对于4类孔隙分割具有最优性能,平均准确率、精确率、召回率和F1分数分别达到93.83%、84.75%、84.87%和84.50%,说明了2个改进点的有效性。
2.2 不同三维分割模型对比分析
3D UNet、Segresnet、VNet和UNetR作为三维分割中的经典模型,近些年来受到广泛认可和使用。因此,在孔隙多类型分割效果试验中,选取以上网络与本研究提出的改进UNet-VAE网络模型进行分割效果对比。随机选取土壤图像样本进行结果展示,5种网络对于多类别孔隙的分割结果如图6所示和表2所示。
图 8 不同网络分割效果对比注:样本一圈中是代表性裂隙和生物孔,样本二圈中是代表性裂隙和不规则孔隙。红色为生物孔,蓝色为裂隙,绿色为不规则孔隙,黄色为球状孔隙。Figure 8. Comparison of different network segmentation effectsNote: In sample 1, there are representative fractures and biogenic pores, while in sample 2, there are representative fractures and irregular pores. The red represents biological pores, the blue represents cracks, the green represents irregular pores, and the yellow represents spherical pores.表 2 不同网络的多类型分割效果Table 2. Multi-category segmentation effects of different networks% 网络
Networks类型
Categories准确率
Accuracy精确率
Precision召回率
RecallF1分数
F1-score3D UNet 裂隙 78.24±6.4 72.70±9.9 66.31±10.7 68.95±8.7 生物孔 75.57±6.3 76.75±9.0 77.38±10.8 76.92±9.4 不规则孔隙 76.63±6.9 45.65±11.3 49.71±6.3 47.16±8.6 球状孔隙 81.04±5.3 64.51±5.1 61.39±6.0 62.73±4.3 平均指标 77.87±6.2 64.90±8.8 63.70±8.4 63.94±7.7 Segresnet 裂隙 87.96±3.4 71.39±17.1 77.88±5.3 73.60±20.7 生物孔 82.16±4.0 78.10±10.9 80.82±7.0 78.96±7.6 不规则孔隙 83.49±3.1 38.23±9.0 42.25±5.5 39.67±6.9 球状孔隙 91.59±1.6 82.23±4.5 65.60±3.9 72.81±2.2 平均指标 86.30±3.0 67.49±10.3 66.64±5.4 66.26±9.3 VNet 裂隙 90.06±3.6 81.53±11.1 70.45±16.8 74.29±12.9 生物孔 87.44±3.3 83.60±7.2 85.31±9.3 84.32±7.7 不规则孔隙 89.56±2.8 59.95±8.5 59.23±4.4 59.32±5.6 球状孔隙 94.99±1.5 83.68±5.3 88.66±1.9 85.96±2.0 平均指标 90.51±2.8 79.69±8.0 75.91±8.1 75.97±7.0 UNetR 裂隙 84.56±4.3 69.23±13.6 71.80±23.0 64.11±7.3 生物孔 79.06±2.7 74.02±12.9 81.11±16.2 81.73±11.5 不规则孔隙 82.90±3.2 37.31±9.6 59.17±6.6 60.02±6.6 球状孔隙 93.76±1.2 82.63±4.8 87.44±2.5 87.77±1.9 平均指标 85.07±2.8 65.80±10.2 74.88±12.1 65.35±6.8 改进UNet-VAE 裂隙 92.29±3.7 83.24±11.6 84.23±9.2 83.40±9.1 生物孔 91.24±3.8 88.53±8.2 89.08±8.8 88.72±8.1 不规则孔隙 94.44±2.4 75.41±11.8 74.93±9.8 74.82±9.3 球状孔隙 97.38±1.1 91.84±4.5 91.27±3.3 91.45±2.5 平均指标 93.83±2.7 84.75±9.0 84.88±7.8 84.60±7.2 对于大体积的裂隙和生物孔,如样本一(图6a~图6f),改进UNet-VAE网络可以准确分割类型及范围。3D UNet、Segresnet、VNet和UNetR对于裂隙(蓝色)和生物孔(红色)难以分辨,其中UNetR由于Transformer对数据集数量有着较高要求,难以学习特征,各类孔隙欠分割现象明显。3D UNet、Segresnet和VNet将平面特征明显的裂隙分类为生物孔,说明了卷积网络对于全局信息和大尺寸特征学习的欠缺。
对于小体积的不规则孔隙,如样本二(图6 g~图6l),Segresnet、VNet和UNetR均将其误分为裂隙。且其余网络均出现产生欠分割现象,导致连通的裂隙被误分为不规则孔隙及球状孔隙,而改进UNet-VAE网络可以准确将断裂形成的不规则孔隙进行分类。说明了改进对于小尺寸孔隙特征学习及网络泛化能力的有效性。
由表2可知,3D UNet的分割精度最低,Segresnet比起3D UNet有一定提升,平均准确率、精确率、召回率和F1分数分别提升了8.43%、2.59%、2.94%和2.32%。VNet对于4类孔隙特征指标均有较大提升,平均准确率、精确率、召回率和F1分数分别提升了12.64%、14.79%、12.21%和12.03%。UNetR作为典型Transformer网络,与VNet相比平均准确率、精确率、召回率和F1分数分别降低了5.44%、13.89%、1.03%和10.62%。改进UNet-VAE在4类孔隙中均达到了所有方法中的最佳指标,平均准确率、精确率、召回率和F1值分别达到了93.83%,84.75%,84.88%和84.60%。与次优VNet方法相比,平均准确率、精确率、召回率和F1值分别提升了3.32%,5.06%,8.97%和8.63%,特别是对于不规则孔隙准确率、精确率、召回率和F1值分别提升了4.88%,15.46%,15.70%和15.50%。综上,改进UNet-VAE实现了高精度多类型孔隙三维分割,对4类孔隙均有良好的特征学习能力,与标定图最为接近,方法泛化能力强。
3. 结 论
为进一步了解孔隙形态、形成原因和孔隙功能的联系,本研究依据孔隙形态分类定义、结合多阈值分割及人工矫正标注建立4类孔隙(裂隙、生物孔、不规则孔隙和球状孔隙)真值数据集。针对不同类型孔隙演变中产生相交的现象,引入深度学习进行土壤多类型孔隙三维分割。针对不同类型孔隙分割精度低、分割方法自动化程度低、鲁棒性差等问题,提出了一种改进UNet-VAE土壤多类型孔隙三维分割方法,通过融合多尺度特征并结合生成网络提高模型泛化性。通过消融试验和定性定量比较分析证明:
1)基于提出的MFA的局部和全局注意力机制,改进UNet-VAE网络提高了多尺度特征信息的融合能力,并筛减冗余信息。通过VAE引入噪声信息和辅助损失函数,改进UNet-VAE网络通过学习不同类型孔隙特征,自动分割多类型孔隙,有效提高了分割方法的鲁棒性,增强了网络的泛化能力。
2)改进UNet-VAE网络模型可以精确分割不同类型孔隙,针对4类孔隙分割准确率分别达到92.29%、91.24%、94.44%和97.38,平均准确率达到93.83%。实现了高精度土壤4类孔隙分割,可为研究土壤物理结构提供先进的技术手段。
本文方法实现了对黑土4类孔隙的精准分割,而对于其他土壤不同功能孔隙的分类与分割和不同孔隙的结构演变研究,将是进一步研究的重点。
-
图 1 土壤CT孔隙分割数据集过程示意图
注:蓝框代表裁剪范围;红圈代表实际扫描范围。AMF为自适应中值滤波处理。SCN为简化卷积网络。
Figure 1. Diagram of soil CT pore segmentation dataset process
Note: Blue frame represents the cropping range; Red circle represents the actual scanning range. AMF is adaptive median filtering. SCN is simplified convolutional network.
图 3 改进UNet-VAE网络结构
注:MFA为多尺度融合注意力模块;图中数字表示空间维度数;N(μ,σ2)是均值为μ,方差为σ2的多元正态分布。
Figure 3. Structure of improved UNet-VAE network
Note: MFA is the multi-scale fusion attention module; the number represents the spatial dimension; N(μ,σ2) is a multivariate normal distribution with a mean of μ and a variance of σ2.
图 5 多尺度融合注意力模块结构图
注:X是尺度为Cx×Hx×Wx×Lx的编码器侧输入特征图,G是尺度为Cg×Hg×Wg×Lg的MFA解码器侧输入特征图,X'是尺度为1×Hx×Wx×Lx的筛减冗余信息后的特征图,X"是权重注意力特征图,Z是最终输出特征图,ReLU是ReLU激活函数,Sigmoid是Sigmoid激活函数。
Figure 5. Structure diagram of multi-scale fusion attention module
Note: X is the encoder side input feature map with a scale of Cx×Hx×Wx×Lx, G is the MFA decoder side input feature map with a scale of Cg×Hg×Wg×Lg, X' is the feature map after filtering redundant information with a scale of 1×Hx×Wx×Lx, X" is the weighted attention feature map, Z is the final output feature map, ReLU is the ReLU activation function, and Sigmoid is the Sigmoid activation function.
图 8 不同网络分割效果对比
注:样本一圈中是代表性裂隙和生物孔,样本二圈中是代表性裂隙和不规则孔隙。红色为生物孔,蓝色为裂隙,绿色为不规则孔隙,黄色为球状孔隙。
Figure 8. Comparison of different network segmentation effects
Note: In sample 1, there are representative fractures and biogenic pores, while in sample 2, there are representative fractures and irregular pores. The red represents biological pores, the blue represents cracks, the green represents irregular pores, and the yellow represents spherical pores.
表 1 不同改进网络的多类型孔隙分割效果
Table 1 Multi-category segmentation effect of differentimproved networks
% 类型
Categories指标
Index模型1
Model1模型2
Model2模型3
Model3模型4
Model4裂隙
Cracks准确率 78.24±6.4 91.77±4.2 83.20±5.2 92.29±3.7 精确率 72.70±9.9 82.81±12.4 74.92±11.9 83.24±11.6 召回率 66.31±10.7 84.77±9.0 70.98±10.4 84.23±9.2 F1分数 68.95±8.7 83.13±8.4 72.39±9.4 83.40±9.1 生物孔
Biological pores准确率 75.57±6.3 90.74±4.3 81.57±5.1 91.24±3.8 精确率 76.75±9.0 89.31±7.7 79.38±10.6 88.53±8.2 召回率 77.38±10.8 87.01±10.6 83.06±9.2 89.08±8.8 F1分数 76.92±9.4 87.88±8.3 81.05±9.4 88.72±8.1 不规则孔隙
Irregular pores准确率 76.63±6.9 93.67±2.9 84.87±4.3 94.44±2.4 精确率 45.65±11.3 71.34±13.5 63.27±12.6 75.41±11.8 召回率 49.71±6.3 73.40±10.2 60.95±7.5 74.93±9.8 F1分数 47.16±8.6 72.03±11.2 61.67±9.2 74.82±9.3 球状孔隙
Spherical pores准确率 81.04±5.3 96.98±1.2 88.37±3.1 97.38±1.1 精确率 64.51±5.1 90.66±5.0 82.09±4.1 91.84±4.5 召回率 61.39±6.0 90.03±3.8 76.79±5.0 91.27±3.3 F1分数 62.73±4.3 90.20±2.7 79.26±3.7 91.45±2.5 平均值
Mean value准确率 77.87±6.2 93.29±3.1 84.50±4.4 93.83±2.7 精确率 64.90±8.8 83.83±9.6 83.16±9.8 84.75±9.0 召回率 63.70±8.4 83.97±8.4 82.78±8.0 84.87±7.8 F1分数 63.94±7.7 83.66±7.6 82.66±7.9 84.59±7.2 注:模型1:原始3D UNet网络;模型2:结合了MFA的改进UNet网络;模型3:引入了VAE的UNet-VAE网络;模型4:结合了MFA和VAE的改进UNet-VAE网络。下同。 Note: Model 1: the original 3D UNet network; Model 2: an improved UNet network that combines MFA; Model 3: a UNet-VAE network with the introduction of VAE; Model 4: an improved UNet-VAE network that combines MFA and VAE. Same as below. 表 2 不同网络的多类型分割效果
Table 2 Multi-category segmentation effects of different networks
% 网络
Networks类型
Categories准确率
Accuracy精确率
Precision召回率
RecallF1分数
F1-score3D UNet 裂隙 78.24±6.4 72.70±9.9 66.31±10.7 68.95±8.7 生物孔 75.57±6.3 76.75±9.0 77.38±10.8 76.92±9.4 不规则孔隙 76.63±6.9 45.65±11.3 49.71±6.3 47.16±8.6 球状孔隙 81.04±5.3 64.51±5.1 61.39±6.0 62.73±4.3 平均指标 77.87±6.2 64.90±8.8 63.70±8.4 63.94±7.7 Segresnet 裂隙 87.96±3.4 71.39±17.1 77.88±5.3 73.60±20.7 生物孔 82.16±4.0 78.10±10.9 80.82±7.0 78.96±7.6 不规则孔隙 83.49±3.1 38.23±9.0 42.25±5.5 39.67±6.9 球状孔隙 91.59±1.6 82.23±4.5 65.60±3.9 72.81±2.2 平均指标 86.30±3.0 67.49±10.3 66.64±5.4 66.26±9.3 VNet 裂隙 90.06±3.6 81.53±11.1 70.45±16.8 74.29±12.9 生物孔 87.44±3.3 83.60±7.2 85.31±9.3 84.32±7.7 不规则孔隙 89.56±2.8 59.95±8.5 59.23±4.4 59.32±5.6 球状孔隙 94.99±1.5 83.68±5.3 88.66±1.9 85.96±2.0 平均指标 90.51±2.8 79.69±8.0 75.91±8.1 75.97±7.0 UNetR 裂隙 84.56±4.3 69.23±13.6 71.80±23.0 64.11±7.3 生物孔 79.06±2.7 74.02±12.9 81.11±16.2 81.73±11.5 不规则孔隙 82.90±3.2 37.31±9.6 59.17±6.6 60.02±6.6 球状孔隙 93.76±1.2 82.63±4.8 87.44±2.5 87.77±1.9 平均指标 85.07±2.8 65.80±10.2 74.88±12.1 65.35±6.8 改进UNet-VAE 裂隙 92.29±3.7 83.24±11.6 84.23±9.2 83.40±9.1 生物孔 91.24±3.8 88.53±8.2 89.08±8.8 88.72±8.1 不规则孔隙 94.44±2.4 75.41±11.8 74.93±9.8 74.82±9.3 球状孔隙 97.38±1.1 91.84±4.5 91.27±3.3 91.45±2.5 平均指标 93.83±2.7 84.75±9.0 84.88±7.8 84.60±7.2 -
[1] KRAVCHENKO A N, GUBER A K. Soil pores and their contributions to soil carbon processes[J]. Geoderma, 2017, 287: 31-39. DOI: 10.1016/j.geoderma.2016.06.027
[2] RABOT E, WIESMEIER M, SCHLÜTER S, et al. Soil structure as an indicator of soil functions: A review[J]. Geoderma, 2018, 314: 122-137. DOI: 10.1016/j.geoderma.2017.11.009
[3] VOGEL H J, EBERHARDT E, FRANKO U, et al. Quantitative evaluation of soil functions: Potential and state[J]. Frontiers in Environmental Science, 2019, 7: 164. DOI: 10.3389/fenvs.2019.00164
[4] 丁天宇,郭自春,钱泳其,等. 秸秆还田方式对砂姜黑土有机碳组分和孔隙结构的影响[J]. 农业工程学报,2023,39(16):71-78. DOI: 10.11975/j.issn.1002-6819.202305110 DING Tianyu, GUO Zichun, QIAN Yongqi, et al. Effects of straw return methods on the soil organic carbon fractions and pore structure characteristics of Shajiang black soil (Vertisol)[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2023, 39(16): 71-78. (in Chinese with English abstract) DOI: 10.11975/j.issn.1002-6819.202305110
[5] PFEIFER J, KIRCHGESSNER N, COLOMBI T, et al. Rapid phenotyping of crop root systems in undisturbed field soils using X-ray computed tomography[J]. Plant Methods, 2015, 11: 1-8. DOI: 10.1186/s13007-015-0043-0
[6] 彭珏,陈家赢,王军光,等. 中国典型地带性土壤团聚体稳定性与孔隙特征的定量关系[J]. 农业工程学报,2022,38(18):113-121. PENG Jue, CHEN Jiaying, WANG Junguang, et al. Linking aggregate stability to the characteristics of pore structure in different soil types along a climatic gradient in China[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2022, 38(18): 113-121. (in Chinese with English abstract)
[7] 邱琛,韩晓增,陈旭,等. CT扫描技术研究有机物料还田深度对黑土孔隙结构影响[J]. 农业工程学报,2021,37(14):98-107. QIU Chen, HAN Xiaozeng, CHEN Xu, et al. Effects of organic amendment depths on black soil pore structure using CT scanning technology[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2021, 37(14): 98-107. (in Chinese with English abstract)
[8] 徐洋洋,张兴,左西宇,等. 再生水灌溉对土壤表层大孔隙的影响[J]. 农业工程学报,2023,39(23):113-122. XU Yangyang, ZHANG Xing, ZUO Xiyu, et al. Effects of reclaimed water irrigation with different water quality on surface soil macro-pores[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2023, 39(23): 113-122. (in Chinese with English abstract)
[9] CAPOWIEZ Y, SAMMARTINO S, MICHEL E. Using X-ray tomography to quantify earthworm bioturbation non-destructively in repacked soil cores[J]. Geoderma, 2011, 162(1/2): 124-131.
[10] ZHANG Z B, ZHOU H, ZHAO Q G, et al. Characteristics of cracks in two paddy soils and their impacts on preferential flow[J]. Geoderma, 2014, 228: 114-121.
[11] GARBOUT A, MUNKHOLM L J, HANSEN S B. Tillage effects on topsoil structural quality assessed using X-ray CT, soil cores and visual soil evaluation[J]. Soil and Tillage Research, 2013, 128: 104-109. DOI: 10.1016/j.still.2012.11.003
[12] ROONEY E C, BAILEY V L, PATEL K F, et al. Soil pore network response to freeze-thaw cycles in permafrost aggregates[J]. Geoderma, 2022, 411: 115674. DOI: 10.1016/j.geoderma.2021.115674
[13] ZHANG Z, LIU K, ZHOU H, et al. Three dimensional characteristics of biopores and non-biopores in the subsoil respond differently to land use and fertilization[J]. Plant and Soil, 2018, 428: 453-467. DOI: 10.1007/s11104-018-3689-3
[14] BOTTINELLI N, ZHOU H, CAPOWIEZ Y, et al. Earthworm burrowing activity of two non-Lumbricidae earthworm species incubated in soils with contrasting organic carbon content (Vertisol vs. Ultisol)[J]. Biology and fertility of soils, 2017, 53: 951-955. DOI: 10.1007/s00374-017-1235-8
[15] KOESTEL J, LARSBO M. Imaging and quantification of preferential solute transport in soil macropores[J]. Water Resources Research, 2014, 50(5): 4357-4378. DOI: 10.1002/2014WR015351
[16] RINGROSE-VOASE A J. Measurement of soil macropore geometry by image analysis of sections through impregnated soil[J]. Plant and Soil, 1996, 183: 27-47. DOI: 10.1007/BF02185563
[17] 张中彬,彭新华. 土壤裂隙及其优先流研究进展[J]. 土壤学报,2015,52(3):477-488. ZHANG Zhongbin, PENG Xinhua. A review of researches on soil cracks and their impacts on preferential flow[J]. Journal of Soil Science, 2015, 52(3): 477-488. (in Chinese with English abstract)
[18] CHEN M, LI Y, JIANG X, et al. Study on soil physical structure after the bioremediation of Pb pollution using Microbial-induced carbonate precipitation methodology[J]. Journal of Hazardous Materials, 2021, 411: 125103. DOI: 10.1016/j.jhazmat.2021.125103
[19] PAGLIAI M. Pore Morphology And Soil Function. In: GLINSKI J, HORABIK J. , LIPIEC J. (eds) Encyclopedia of Agrophysics. Encyclopedia of Earth Sciences Series. Dordrecht: Springer, 2011: 640-645.
[20] 周云艳,徐琨,陈建平,等. 基于CT扫描与细观力学的植物侧根固土机理分析[J]. 农业工程学报,2014,30(1):1-9. ZHOU Yunyan, XU Kun, CHEN Jianping, et al. Mechanism of plant lateral root reinforcing soil based on CT scan and mesomechanics analysis[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2014, 30(1): 1-9. (in Chinese with English abstract)
[21] WANG H, QIAN H, GAO Y. Characterization of macropore structure of remolded loess and analysis of hydraulic conductivity anisotropy using X-ray computed tomography technology[J]. Environmental Earth Sciences, 2021, 80: 1-15. DOI: 10.1007/s12665-020-09327-2
[22] EBRAHIMI M K V, LEE H, WON J, et al. Estimation of soil texture by fusion of near-infrared spectroscopy and image data based on convolutional neural network[J]. Computers and Electronics in Agriculture, 2023, 212: 108117. DOI: 10.1016/j.compag.2023.108117
[23] MENG C, YANG W, BAI Y, et al. Research of soil surface image occlusion removal and inpainting based on GAN used for estimation of farmland soil moisture content[J]. Computers and Electronics in Agriculture, 2023, 212: 108155. DOI: 10.1016/j.compag.2023.108155
[24] 韩巧玲,赵玥,赵燕东,等. 基于全卷积网络的土壤断层扫描图像中孔隙分割[J]. 农业工程学报,2019,35(2):128-133. HAN Qiao, ZHAO Yue, ZHAO Yandong, et al. Pore segmentation in soil tomography images based on fully convolutional network[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2019, 35(2): 128-133. (in Chinese with English abstract)
[25] FU Y, ZHAO Y, ZHAO Y, et al. Semi-supervised segmentation of multi-scale soil pores based on a novel receptive field structure[J]. Computers and Electronics in Agriculture, 2023, 212: 108071. DOI: 10.1016/j.compag.2023.108071
[26] 韩巧玲. 基于CT图像的黑土大孔隙精细分割与重构方法研究[D]. 北京:北京林业大学,2020. HAN Qiaoling. Fine Segmentation and Reconstruction of Black Soil Macropore Based on CT Image[D]. Beijing: Beijing Forestry University, 2020. (in Chinese with English abstract)
[27] 韩巧玲,周希博,宋润泽,等. 基于序列信息的土壤CT图像超分辨率重建[J]. 农业工程学报,2021,37(17):90-96. HAN Qiaoling, ZHOU Xibo, SONG Runze, et al. Super-resolution reconstruction of soil CT images using sequence information[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2021, 37(17): 90-96. (in Chinese with English abstract)
[28] RINGROSE-VOASE A J. Measurement of soil macropore geometry by image analysis of sections through impregnated soil[J]. Plant and Soil, 1996, 183: 27-47. DOI: 10.1007/BF02185563
[29] 高朝侠,徐学选,赵娇娜,等. 土壤大孔隙流研究现状与发展趋势[J]. 生态学报,2014,34(11):2801-2811. GAO Chaoxia, XU Xuexuan, ZHAO Jiaona, et al. Review on macropore flow in soil[J]. Acta Ecologica Sinica, 2014, 34(11): 2801-2811. (in Chinese with English abstract)
[30] MYRONENKO A. 3D MRI brain tumor segmentation using autoencoder regularization[C]//Brainlesion: Glioma, Multiple Sclerosis, Stroke and Traumatic Brain Injuries: 4th International Workshop, BrainLes 2018, Held in Conjunction with MICCAI 2018, Granada, Spain, September 16, 2018, Revised Selected Papers, Part II 4. Springer International Publishing, 2019: 311-320.
[31] 韩巧玲,柏浩,赵玥,等. 采用染色示踪技术的土壤优先流自动分割与量化系统[J]. 农业工程学报,2021,37(6):127-134. HAN Qiaoling, BAI Hao, ZHAO Yue, et al. A soil priority flow automatic segmentation and quantification system using dye tracing technology[J]. Transactions of the Chinese Society of Agricultural Engineering(Transactions of the CSAE), 2021, 37(6): 127-134. (in Chinese with English abstract)
-
期刊类型引用(2)
1. 程傲,任子由,张承明,李峰,吴门新,李红英,段金馈,刘一笑. 先验知识融合语义特征的冬小麦田块精细提取方法. 农业工程学报. 2025(04): 164-174 . 本站查看
2. 赵璐,王佳妮,殷婕,袁京,李国学,周海宾,马若男. 不同源生物炭对多元物料协同堆肥腐熟度和腐殖化的影响. 农业工程学报. 2025(04): 249-259 . 本站查看
其他类型引用(0)