ArticlePDF Available

基于自筛选深度学习的滑坡易发性预测建模及其可解释性

Authors:
地球 科学 Earth Science
http://www.earth‐science.net
48 5
2 0 2 3 5
Vol. 48 No. 5
May 2 0 2 3
https://doi.org/10.3799/dqkx.2022.247
基于自筛选深度学习的滑坡易发性
预测建模及其可解释性
黄发明1 1达雄 2 1张子 1 1*
1. 南昌大学信息工程学院,江西南昌 330031
2. 昌大 学工 程建 设学 院, 西南 330031
要: -非滑坡样本可能存在误差、环境因子间非线性关系较复杂且机器学习可解
Self-
screening Bi-directional Long Short-Term Memory and Conditional Random Fields SBiLSTM-CRF.SBiLSTM-CRF 模型具
广 线
阈值区间外的错误滑坡样本 .该模型可用于解释各环境因子之间耦合关系的内部作用机制 .SBiLSTM-CRF 模型用
于陕西延长县滑坡易发性预测,并与cpLSTM-CRF .
SBiLSTM-CRF 克服了传统机器学习中存在的样本误差以及因子间复杂的非线性关系问题,具有更高的预测性
.通过该模型的可解释性能力揭示了坡度、高程和岩性等因子控制延长县的黄土滑坡发育的机制.
关键 词: 坡易发性预测深度学习双向长短时记忆网络;件随机场可解释 性;程地 .
中图分类号:P64 文章编号:1000- 2383(2023)05-1696-15 收稿日期:2022-08-18
Landslide Susceptibility Prediction Modeling and
Interpretability Based on Self-Screening Deep Learning Model
Huang Faming1 Chen Bin1 Mao Daxiong2 Liu Lekai1 Zhang Zihe1 Zhu Li1*
1. School of Information Engineering Nanchang University Nanchang 330031 China
2. School of Infrastructure Engineering Nanchang University Nanchang 330031 China
Abstract: To address the problems of landslide susceptibility prediction (LSP) modeling including possible errors in landslide and
non-landslide samples, complex non-linear relationships between environmental factors and unaddressed machine learning
interpretability, a deep learning-based Self-screening Bi-directional Long Short-Term Memory and Conditional Random Fields
(SBiLST M-CRF) model is proposed to reduce the impact of these problems on LSP and improve its confidence. The SBiLSTM -
CRF model has the advantages of deep learning network with deep layers, wide width and iterative modeling, which can predict
the non-linear relationship between environmental factors and automatically screen out the wrong landslide samples; it can select
non-landslide samples from the initial low/very low landslide susceptibility zone through iterative modeling, and finally reveal the
基金 项目 国家 自然 科学 青年基金项 No.41807285.
作者 简介 明( 1988 ), ,研 . ORCID 0000‐0002‐4428‐7133. E ‐mail
faminghuang@ncu. edu. cn
* 讯作者: 莉,E ‐mail: lizhu@ ncu. edu. cn
用格 式:黄发 明,陈彬 达雄 乐开 荷,朱莉 20 23. 基于自筛选深度学习的滑坡易发性预测建模及其可解释性 .地球科学,485
1696-1710.
CitationHuang FamingChen BinMao DaxiongLiu LekaiZhang ZiheZhu Li2023.Landslide Susceptibility Prediction M odeling and Interpret‐
ability Based on Self‐Screening Deep Learning Model.Earth Science4851696-1710.
5 发明 等: 于自 筛选 深度 学习 的滑 坡易 发性 预测 建模 及其 可解 释性
internal mechanism of the coupling of environmental factors to predict landslide susceptibility. The SBiLSTM-CRF model is used
to predict landslide susceptibility in Yanchang County of China, and compared with cpLSTM-CRF, random forest (RF), support
vector machine (SVM), stochastic gradient descent (SGD) and logistic regression (LR) models. The results show that SBiLSTM-
CRF overcomes the problems of sample error and complex nonlinear relationship between factors in traditional machine learning ,
has superior performance in modeling susceptibility than conventional machine learning, and the interpretability of the model
reveals that factors such as slope, elevation and lithology control the development of mounded landslides in Yanchang County.
Key words: landslide susceptibility prediction; deep learning; Bidirectional long short‐ term memory; conditional random field;
interpretability analysis; engineering geology.
0
广
Khan et al. 2021.滑坡易发性预测对滑坡风险
评价和潜在滑坡的准确定位具有重要作用 .
GIS 平台的方法被用
Moragues
et al. 2020数理统计模型如判别分析Eiras et
al. 2021多元线性回归Jia et al. 2021
广 2021.
能较好地揭示出易发性建模中各环境因子之间
线
.
器学习由于其不需要过多的先验知识如不需要
布)
而被广泛应用于滑坡易发性预测,如人工神经网
2020支持向量机 Support Vec‐
tor Machine SVMYao et al. 2008)、
Pradhan 2013 Random Forest
RF 2021)、 Huang et
al. 2020a Logistic Regression LR
(方 2021随机梯度下降Stochastic Gra‐
dient Descent SGDHong et al. 2020 .
而, 的机 应用 坡易 预测
还存在许多问题:1表征学习需要大量的先验知
和假 间,能充 提取 环境
线关系Huang et al. 2020b2
好地实现滑坡样本中的误差剔除,并实现有效的非
坡样 选择Yao et al. 20223模型对数据敏
失值 错误 等;4认为
非线
预测出滑坡易发性值的内部机制缺乏认识,即模型
解释 性未 到足 够关 .此研 种新 用于
滑坡易发性预测的机器学习算法非常有必要 .
深度学习在一定程度上能够克服传统机器学
的这 点,有学 力强 围广
性好 点,数据 而不 过多
知识和假设 .目前深度学习已被广泛应用于公安服
Zhou et al. 2022
Bhattacharya et al. 2021 Xu
et al. 2018 也被 广泛应于自然灾 易发
Huang et al. 2020b 2020
宝等 2022.上述深度学习预测滑坡易发性主
要还是从算法改进的角度出发,而没有从滑坡易发
性建模本身存在的问题出发来提升建模性能Zhu
et al. 2020.单纯的深度学习算法改进有时难以得
易发
非滑坡样本本身就存在误差比较大的问题黄发明
2020.这种滑坡非滑坡样本误差将会严重制约
深度学习模型的输出变量准确性,从而制约深度学
习预测滑坡易发性的精度提升 .
提出新的深度学习算法,即基于自筛选的双向长短
时记忆网络与条件随机场Self‐screening Bidirec
tional Long Short Term Memory and Conditional
Random FieldsS BiLS TM CRF滑坡易发性预测
.该算法预期能更准确地分析出各滑坡点之间
相关 便取出 因子 更深 特征
代建模选择出准确的非滑坡样本,以便提供比原数
准确 坡空 描述Zhou et al. 2022.
SBiLSTM‐CRF 模型的优越性,本文同时
选择了 4 (即 Ran
dom ForestRF Logistic Regression
LR Stochastic Gradient Descent
SGD Support Vector Machine
SVM
网络和条件随机场Cascade‐parallel Long Short
Term Memory and Conditional Random Fields
cpLSTM‐CRF比较 Zhu et al. 2020.
1697
48
地球科学 http://www.earth‐science.net
对于滑坡易发性建模的另外一个严肃问题是
很难解释影响结果的变量或特征,即深度学习的可
Linardatos et al. 2020.
型越来越广泛地应用于关键环境下的重要预测,
机器学习的透明度也日益增高 .其危险在于创建
使
解释(成 2020
2021.其实无论是传统的机器学习还是深度学习,
都面临无法充分解释环境因子作用下滑坡易发
Gaur et al. 2020
大多数以数据驱动的深度学习目前尚处于“黑盒
.
环境因子对预测模型决策的贡献度以及与预测
结果的关联性,以获得更好的可解释性结论 .
1 易发测的 SBiLSTM‐CRF
1.1 研 思路 
滑坡易发性预测是一个复杂的非线性建模问
题,通过常规统计方法可能无法获得令人满意的结
.传统的处理机器学习和数据挖掘问题的方法都
需要很强的前提假设或者大量的先验知识,而且易
发性建模中也存在滑坡-
Hong et al. 2020.本文拟提出了 SBiLSTM
CRF 这一新的深度学习模型来尝试克服上述问题,
并首次运用于滑坡预测数据集 .算法 (图 1a
下:1滑坡影响因子的分析和筛选:基于滑坡编
(胡
2020获得各个基础环境因子的频率比 .2立空
-非滑坡的空间数据集,并划分出模
型训练集和测试集 .3自筛 模:构建
Bi‐LSTM 和全连接网络的预分类自筛选网络,
并对全区进行滑坡易发性预评估 . t
T1T0数据集中 人工标注 据采 误差
定因素导致的数据集标注结果错误数据进行筛选,
以降低易发性评价过程中的不确定性 .再从全区预
估结 果采 然断 获得 T
1T
0进行等
补充 .4特征提取建模:构建级联双向 LSTM
络对筛选后的数据集进行更深度地提取特征,获取
后的
.5序列预测建模:构建全连接二分
类网络对所提取的特征进行滑坡易发性预测,
CRF 对结果进行综合解码获得易发性预测
.6 通过假阴性、假阳性、
ROC 线衡 量模 .
1.2 滑坡易发性预测模型 
1.2.1 滑坡栅格样本自筛选 
是处理后的延长县一维地理数据,即每个样本所
12 个环境因子 .通过一个 Bi‐LSTM
.
2 060 个)
2 622 482 个)
.通过自然断点法获得全区易发性
.本研究将滑坡易发性分为
K(K= 5 )
5个等级 .对筛选阈值引入偏度以用来
度量随机变量概率分布的不对称 .
ì
í
î
ï
ïï
ï
ï
ïï
ïT1=
||
sk ×t
1 + 2
||
sk
T0=1-T1,
(1)
t为人为设置阈值.s k 为偏度, 为:
sk =1
n
i= 1
n
[( Xi-μ
σ)3] (2)
中,
μ均值
σ标准 .sk < 0 表明
为负
sk > 0 时表明整体预分类分布为正偏离,即右偏态;
sk = 0表明数据集整体预分类分布为正态分布 .
1
2将人工标签为 1而滑坡易发性概率小于
T1 0而滑坡易发性
T0 3
T
1的栅格作为新的滑坡样本,
选滑坡易发性概率小于 T
0
. T
1T
0分别是预评估中由自然
.
本研究中筛选的阈值 t设置为 0.6 T1
0.302 1筛选正样本 87 130
T
1T0'
0.659 9 0.139 1. 此筛 量补
减弱了滑坡易发性评价的不确定性,又维持了数据
集数据量一致,也保证了滑坡和非滑坡的类平衡 .
1.2.2 特征提取建模 特征提取部分采用 K(K=
5 ) Bi‐LSTM 的级联网络和全连接的前馈神经
. Bi‐LSTM 层能够更深层次地提取滑
1698
5 发明 等: 于自 筛选 深度 学习 的滑 坡易 发性 预测 建模 及其 可解 释性
.网络输入
Ri×j
{
i  |1
iIiZ
}
I
{
j  |1 jJjZ
}
J
.滑坡因子向量进入 5Bi
LSTM LSTM LSTM
单元为例,结构如图 1b 1c .正向传播
k层中第 i个栅格的第 j
LSTM
C
k,i,j= tan h
(
Wc
[
hk- 1,xij
]
+bC
)
(3)
Ck,i,j=Fk,i,j×Ck- 1 +Ik,i,j×C
k,i,j (4)
hk,i,j=Ok,i,j× tan h(Ck,i,j) (5)
hkij C
kij
先前隐藏层生成的候选状态,
Ckij是前单元状
tan h为激活函数,
Fkij为遗忘门,
Ikij为输入
Okij
ì
í
î
ï
ï
ï
ï
ï
ï
ï
ï
ï
ï
ï
ï
Fk,i,j=σ
( )
Wf
[ ]
hk- 1,xij +bf,
Ik,i,j=σ
( )
Wi
[ ]
hk- 1,xij +bi,
Ok,i,j=σ
( )
Wo
[ ]
hk- 1,xij +bo,
(6)
σ表示激活函数;
Wf
WiWo LSTM
hk- 1
xij 分别表示前一
i j
bf
bibo表示相应的偏置项;
Fkij
01之间的数字 .
然后用遗忘门确定丢弃的信息,与输入
LSTM
门根据当前单元状态输出和隐变量 .Bi‐LSTM
LSTM
ì
í
î
ï
ïï
ï
ï
ïï
ï
hfk,i,j=H(Whf
[ ]
hfk,i,j+xij +bhf )
hbk,i,j=H(Whb
[ ]
hbk,i,j+xij +bhb ) (7)
hfkijRi×j
hbkijRi×j 前向
ykij= [ hfkijhbkij]是这两
分的 层和 层的 被定 单个
1 算法流程及 Bi-LSTM/LSTM 结构示意
Fig.1 Algorithm flow and Bi-LSTM/LSTM structure
a. 算法流程 . FC. 全连接网络;Bi-LSTM . 双向 LSTMCRF. 条件随机场;b.Bi-LS TM 链;c.LSTM 单元结构
1699
48
地球科学 http://www.earth‐science.net
Bi‐LSTM .在完成一次前向传播过程中,从原始
据输入到 SBiLSTM ‐CRF 变化公式为
yout=
k=1
K
i=1
M
j=1
J
ok,i,j
×tan h
( )
fk,i,j
×Ck-1+ik,i,j
×C
k,i,j (8)
使 Dropout
来实现正则化效果 .
LSTM
照预设的概率随机失效 .
Dropout 设置为 0.25.
1.2.3    在通过 K Bi
LSTM
因子 间的维特征后 使用 一个 32 个神经元
组成的全连接层将所有的特征融合起来,并使用两
个神经元进行滑坡预测 .这一过程使用 Sigmoid
作为 激活 .由于滑坡数据在总体上具有一定
的空间连续性,例如,在滑坡数据的周围子区域往
.
.因此引入 CRF
.CRF 为两
一类 i个栅格点上的节点特征函
数,它只和当前栅格点的输出滑坡易发性评估结
i个栅格点 yi的节点特征函数记为:
sl
(
yi,yi
pred,i
)
,l= 1,2,...L (9)
L是定义在该栅格点上节点特征函数的总个
.第二类是定义在 yi栅格点前后的局部特征函
只和 当前 坡点 和上滑坡 有关 i
yi的局部特征函数记为:
tk
(
yi- 1,yi,yi
pred,i
)
,k= 1,2,...K (10)
K是定义在该栅格点的局部特征函数的总
.
它们的取值只能是 01即满足特征条件或者
.同时每个特征函数都被赋予一
.线
性链条件随机场的公式可表示为:
p
(
yi|yi
pred
)
=1
Z
( )
yi
pred
exp
(
i,k
λktk
(
yi- 1,yi,yi
pred,i
)
+
i,l
μlsl
( )
yi,yi
pred,i
)
(11)
tkμl分别是 λksl的权重系数,Z
(
yi
pred
)
Z
(
yi
pred
)
=
y
exp
( )
i,k
λktk
( )
yi- 1,yi,yi
pred,i+
i,l
μlsl
( )
yi,yi
pred,i
(12)
最优 的模 型, CRF用极
13 .本文使用
Viterbi Forney1973 CRF 计算过
得到 CRF 平滑处理后的最终输出序列标签:
y=
{
y1y2
}
yn.
L
(
W
)
=
i
log p
( )
yi|yi
pred . (13)
1.2.4   
.
. 使
1非滑坡概率更
0.
Lcross ‐entropy =
-1
n
i=1
n
(yi
Llog ( yi)+(1 - yi
L) log ( 1 - yi) ) , (14)
yi
yi表示预测值,
n表示滑坡样本总数 .
1.3 滑坡易发性预测结果评价 
1.3.1  计指 精度 价  本研究采用阳性预测
PPR NPR To
tal Accuracy TA
Huang et al. 2020b.PPR 通过计
.NPR
.PPR
NPR
滑坡的预测能力 .TA 被用来评估所有测试数据
集的预测准确度 .3个统计指标的计算如下:
TA = TP + TN
TP + TN + FP + FN , (15)
True Positive TP表示
坡栅 数;阴性True Negative TN表示正确
False Positive FP
表示错误分类的滑坡栅格数,即将非滑坡栅格错误
分类 滑坡格; False Negative FN
示分类错误的费滑坡栅格数,即将滑坡栅格错误的
.TA 表征模型总的预测精度,
TA 越大表示滑坡易发性预测准确度越高 .
1.3.2 ROC 曲线评价 接受者操作特征曲线Re‐
1700
5 发明 等: 于自 筛选 深度 学习 的滑 坡易 发性 预测 建模 及其 可解 释性
ceiver Operating Characteristic Curve ROC
映特异性和敏感性的相互关系,被广泛应用于评价
坡易发性模型 优劣罗路 广等,2021.特异性又
False Positive Rate FPR
阳率True Positive Rate TPR算公 为:
FPR = FP
FP + T N , (16)
TPR = TP
TP + F N , (17)
ROC 线以特 异性 纵坐 .
ROC 曲线在 y=x 上, ROC 曲线下的面
Area Under the CurveAUC一般取值范围位于
0.5 1.ROC 曲线越接近左上角, AUC
大表示该滑坡易发性预测模型的性能越好 .
1.4 深度学习预测滑坡易发性的可解释性 
近年来深度学习在研究领域取得了很多
可信度 .可解释性是指以可理解的术语向人类
Zhang et al. 2021.总体可
分为内置可解释性和事后可解释性Alvarez‐
Melis and Jaakkola 2018.
内置可解释性也称事前可解释性,是指模型
解释
用于较为简单的网络模型 . 使
Lipton
2018.事后可解释性旨在从已训练模型中提取信
息,主要运用于较为复杂的网络模型,可分为全局
可解释性和局部可解释性 .
空间
释方
从模型层面进行解释;局部可解释性是指基于输
变化
和理解模型决策,对应的方法为局部可解释方法,
通常从数据层面进行解释 .
县滑
Alvarez‐Melis and Jaakkola2018.
2 研究区及滑坡编录介绍
2.1 延 县 
延长县位于陕西省东北部,总面积为
2 368.7 k 470.6~1 383 m延长县属
温带
. 西
.
原地貌,地层依次为三叠系碎屑沉积岩、上新
.马红土分布不连
续,其上层黄土抗剪强度较差 .马兰黄土由于
.
2 82
主要类型为小型浅层覆盖滑坡 .滑坡体积以
81 98.7%),
西
2019.
2.2 滑坡环境因子分析 
本研究用到的数据源包括:
30 m 分辨率的 DEM ③1 10
例尺的延长县岩土类型分布图;30 m 分辨率
Landsat TM8 遥感影像等 .
面曲率、平面曲率、地形起伏度、地表总辐射、
湿 数(Topographic Wetness In
dex TWI)、 数(Normalized
Difference Building Index NDBI)、
Normalized Difference Vegetable Index
NDVI修正归一化差异水指数Modified
Normalized Difference Water Index MNDW I12
3.
2 延长县滑坡概况
Fig.2 Yanchang County landslide overview
1701
48
地球科学 http://www.earth‐science.net
3 滑坡易发性建模结果
3.1 滑坡相关空间数据集 
12 个环境因子作为深度学习
6 864 个栅格单
3 432 个滑坡
栅格单元以及从非滑坡区域中随机选择的
3 432 .这些滑坡非滑坡栅格
70%/30% 的比例随机划分后合并为训
(黄 2020. SBiLSTM
CRF /测试数据集预测延长县的滑
LR RF SVM SGD 等模型
.实验所需的硬件配置如表 1 .
3.2 延长县滑坡易发性结果 
3.2.1 SB iLSTM CRF 模型预测易发性 本文提
SBiLSTM‐CRF 具有对数据的自筛选功能,
具有较深的特征提取能力 .模型以 Bi‐LSTM
连接层作为预测模型对输入的数据进行筛选,在以
Bi‐LSTM 深度学习模型进行特征提取,利用
CRF 对预测结果做综合解码 .模型采用 Adam
3 延长县滑坡环境因子
Fig. 3 Environmental factors for landslides
a. 程;b . 坡度c. 坡向 d. 平面曲率;e. 剖面曲率f. 地表起伏度;g. 性;h . 地形湿度指数;i.NDVIj.NDBIk.MNDWIl. 地表总辐射
1702
5 发明 等: 于自 筛选 深度 学习 的滑 坡易 发性 预测 建模 及其 可解 释性
0.000 1 batch size 设置为 200
LSTM 单元数设置为 32 即隐藏层大小设置
32训练迭代次数为 10 000 .将训练好的
SBiLSTM CRF
Landslide
Susceptibility Mapping LSM.
Yao et al. 2008
5
12.80% 13.21 %
13.67% 19.53% 40.77%(图 4a 2.
3.2.2 CPLSTM CRF 性 
CPLST M‐C RF 模型 采用 Ad am 优化器,学习率设置
0.001batch size 设置为 200unit 设置为 32 个且训
8 000 Huang et al.2020b.
练好的 CP LST M ‐C RF
LSM .根据自然间断点法对研究
5个等级并计
8. 59% 9.33%
13.92% 22.36% 45.80% (图 4b 2.
3.2.3 RF 模型   RF
分类器,是滑坡预发性预测的常用模型 .RF
主要采用因子特征数量 m和树的数量 t等参数来
.基于袋外误差筛选法确定 RF 模型预测延长
mt的参数分别设置为 3
800 (吴 2021.同样将 RF 模型用于预测
延长县全研究区域的滑坡易发性得到 LSM .
RF
10.02% 16.86%
21.99% 24.59% 26.54%(图 4 c 2.
3.2.4 LR SVM SG D 模型预测易发性 LR
型由线性回归模型方程与 Sigmoid .
其本质是假设数据服从某一分布,然后使用极大似
然估计做参数的估计李文彬等,2021.LR 具有实
.
LR L2
C设置为 0.5
10- 4 .LR
10.90% 16.85%
21.52% 25.53% 25.19%(图 4 d 2.
SVM 模型的惩罚系数 C 1.0
用径向基;核函数系数 γ 0.3Yao et al.
2008.SVM
14 .87% 1 6.45 %
18 .76% 22.84% 27.08% 4e.SGD 模型是
广 使 SGD L2 范数作为罚
α设置为 0.000 1Hong et al.
2020.SGD
1 5.20 %
18 .46% 20 .16% 22.15% 24.03%(图 4f 2.
3.3 易发性建模结果评价 
3.3.1 统计指标精度 各滑坡预测模型的统计测
结果 3所示 . 表明 SBiLSTM ‐CRF 在阴
预测 测率 的预 上均 他传
统机器学习模型和深度学习具有更好的预测性能 .
自筛选的方式有效去除了错误样本,为模型学习提
1 实验平台软硬件环境
Table 1 Software and hardware environment of the experimental platform
实验 平台
配置
处理
(CPU)
显卡
(GPU)
RAM
参数
Intel(R) C ore(TM ) i5-7400@3.00 GHz
Nvidia GeForce GTX1080
8.00 GB DDR4
实验
平台
配置
内存
(ROM)
操作
系统
开发
环境
参数
Western Digital WDC WD10EZEX-08WN4A0
Windows10 + Ubuntu18.04
Python3.6.5 + TensorFlow1.14.0 +
Keras2.1.4 + Matlab 2018
2 滑坡易发性评估统计结果
Table 2 Statistical results of landslide susceptibility evalua‐
tion
预测 模型
SBiLSTM-CRF
cpLSTM-CRF
RF
LR
SVM
SGD
易发性等级 (%)
极高
12.80
8.59
10.02
10.9
14.87
15.2
13.21
9.33
16.86
16.85
16.45
18.46
中等
13.67
13.92
21.99
21.52
18.76
20.16
19.53
22.36
24.59
25.53
22.84
22.15
极低
40.77
45.80
26.54
25.19
27.08
24.03
1703
48
地球科学 http://www.earth‐science.net
供了高质量的数据 .级联的 L STM 可以捕获更深层次
的因子间的交互关系,使其优于其他机器学习模型 .
3.3.2 ROC 线   ROC 线
AUC ROC 曲线下的
.ROC 线
5 . SBiLSTM‐CRF 在精确率
和召回率都明显优于其他模型 . SBiL
STM‐CRF
LSTM 明显增强了模型的非线性表达能力 .
4
4.1 模型迭代分析 
SB iLST M CRF 训练损失值和测试准确率随
迭代次数增加的变化曲线如图 6 .可以看出迭
500 次内 率停 0.5
意味着模型没有更新 .相反,模型在停滞过程中 loss
.模型迭代次数在 2 000 次内 4
速下 降到 0.5 后缓降低 趋于 .时,
4 滑坡易发性图
Fig. 4 Landslide susceptibility maps
a.SBiLSTM-CRFb.cpLSTM-CRFc.RFd. LR e. SVM f.SGD
1704
5 发明 等: 于自 筛选 深度 学习 的滑 坡易 发性 预测 建模 及其 可解 释性
0.5 0. 8
.这表明模型具有较快的收敛
.
4.2 SBiLST M⁃CRF 模型分析 
滑坡易发性建模结果表明,SBiLSTM CRF
模型性能优于以往的 CP LST M ‐C RF 和传统的
RF LRSVMSGD 等机器学习模型 .这是因为
SBiLSTM‐CRF 模型在易发性建模过程中表现出
.比如该模型使用自筛选模块以解决
使 Bi‐LSTM 以增加网络宽
度,并以级联的方式增加网络的深度;CRF
一步分析栅格点之间的双能量关系使得滑坡预
.SBiLSTM‐CRF 能更充分地拟合
数据以便学习环境因子间的非线性关系,具有自
.SBiL
STM‐CRF
.然而在网络拥有良好的特征提取能
SBiLSTM CRF
其他的深度学习和机器学习网络相比更为复杂 .
因此,是否能在尽量不牺牲精度的前提下对模型
进行剪枝和优化,将成为下一步改进的方向 .
4.3 SBiLST M⁃CRF 的易发性建模可解释性 
文分 因子 滑坡 性、子交
度等
测滑坡易发性进行解释Linardatos et al. 2020.
4.3.1   
得滑坡易发性概率,统计环境单因子的预测分布图
解释 自然 层面 揭示 同单
2021.对测试集中 2 060 个样本进行统计可以得出
个环境因 在各区间的分 如图 7 所示 .以坡度
例,坡度 布,每个 进行
坡度在各个区间的预测分布如图所示 .上半部分的
箱型图表示各区间滑坡易发性概率的分布情况,
Q2 部分 柱状 表示 区间 样本的数
统计情况状图中的 数字表示该区 间样 本个 .
7a以看出,度在 0°~8.8°8.8°~12.
间内 的样 量较 少,别为 140 221
6.8% 10.7%21.9°~4区间内的样
量最大, 582例,约占总量的 2 8.3% . 坡度 0°~
12.1°呈现出非常低的滑坡易发性概率,均小于 0.2
两个区间内的 Q2 0.085 19.1°~21.9°
41.0°~50.3°区间 内的 概率 集中 0.5~
0.9Q2 0.671 21.9° ~41°
6 L oss acc uracy 随迭代次数曲线
Fig.6 Loss and accuracy curves with number of iterations
3 不同模型对延长县滑坡预测的性能对比
Table 3 Comparison of landslide prediction performance of
different models in Yanchang County
模型
TP
TN
FP
FN
PPR (%)
NPR(%)
TA (%)
SBiLSTM-
CRF
849
857
173
181
83.07
82.56
82.82
CPLSTM -
CRF
809
701
329
221
71.09
76.03
73.30
RF
813
771
259
217
75.84
78.04
76.89
SVM
767
729
301
263
71.82
73.49
72.62
LR
729
724
306
301
70.43
70.63
70.53
SGD
813
616
414
217
66.26
73.95
69.37
5 各滑坡易发性预测模型的 R OC 曲线
Fig.5 ROC curves for each landslide susceptibility prediction
model
1705
48
地球科学 http://www.earth‐science.net
0.6~0.9Q2 0.845表现出了具有
(郭 2013.坡度作
为易发性评价中重要的环境因子,其大小直接影
~41.0°
41.0°~50.区间内滑坡易发性概率较21.9°~41.0°
.这是因为坡度较小通常坡体比较稳
定,而坡度大通常滑坡风化层不易累计、人类活动
良好 发生 滑坡 .
7b 可看 NDVI 0~0.0990.449~1
的数较少本数 3例, 0.15%NDVI
0.113~0.1220.237~0.2430.259~0.270
0.347~ 0.398 间的 据量 样较 少, 本数 118
比约 5.7% . 在随 NDVI 的增大,并没有呈现
NDVI 0.113~0.1220.237~
0.2430.259~0.2700.347~0.398 之间时,超过一半
0~0.5 Q2
0.397 0.157~0.197
易发性概率都落在了 0.5~0.9Q2 0.599.
Moragues et
al. 2020
.NDVI 较低时,通常为水系、
市和裸地,通常没有滑坡发育的条件;NDVI
人类观测,因此有记录的滑坡频次较少 .
4.3.2 双因子交互作用 
一步统计双因子交互作用,以下选取相关性较高
6个结果做分析 .DEM Slope
DEM Slope 分别的结果进行统计,得到 DEM
Slope 双因子交互如图 8a 所示,图中展示每个
各因子每个区间的分布情况,圆中的数字分别表
示当前区间下的样本量和平均滑坡易发性概率,
圆的大小表征样本量的多少,圆的颜色深浅表征
滑坡易发性概率的大小成科扬等,2020.
8a 坡度 ~12.1° 积物
性能大于下滑性能,形成不了滑坡 .
情况下,滑坡易发性概率随着坡度的增大先增后
21°~41.0°了最 下随
大先 增后 减, 866.6~979.4 m
. 866.6~979.4 mSlope
21°~41.0°区间内的滑坡易发性概率0.855
.说明高程在 866.6~979.4 m 区间内利于滑坡堆
成物 育,接受 坡积 洪积
.岩石经过风化以后,容易从高程为 979.4~
1 369.8 m 的区域通过坡积过程搬运到高程为
866.6~979.4 m 的区域,此时再遇到 19.1°~50.3°
坡度即沉积下来的堆积物就容易发育成滑坡 .
8b 地形 18.0~84.6 m 且不变的情
下,易发 率随 形湿 的增 增大
7 滑坡易发性预测单因子可解释性结果
Fig.7 One-way interpretable results for landslide susceptibility prediction
a. 度;b.NDVI
1706
5 发明 等: 于自 筛选 深度 学习 的滑 坡易 发性 预测 建模 及其 可解 释性
.在地形起伏度为 33.1~84.6 m 和地形湿
0~0.123 0.757
.在地形湿度较大的区域边坡堆积层
坡(冯 2022.
湿
使
.
4.3.3 滑坡环境因子贡献度解释 本文对逐条样
本根据各滑坡环境因子做积分梯度并取期望值,
.为简化计算基线设置为 0
似方法使用的步数设置为 50.
SBiLSTM‐CRF 做决策的
1 地形地貌和基础地质
0.31 0.18.2 地表覆
-0.03
0.02.3
0.14 0.12 0.08这也与前面的
.4 NDVI 为负贡献度- 0.04
MNDWI 0NDBI 的贡献度不足
0.01. 上述结果表明 NDVIMNDWI NDBI
.
.
5
本文提出了一种基于 Bi‐LSTM 筛选
联双 LSTM‐CRF 网络模型开展滑坡易发性预
.模型与 CPLSTM‐CRF 和传统的机器学习
8 滑坡易发性预测的双因子交互可解释性结果
Fig.8 Two-factor interactive interpretability results
a. 高程与坡度;b. 地形起伏度与 TWI
1707
48
地球科学 http://www.earth‐science.net
RF SVM SGD LR
有显著的优越性,建模结果均优于其他模型 .
使
使 S BiL STM CR F 从以往的黑盒
.结果表明就滑坡易发性数据集
而言,坡度、高程、岩性、地表起伏度和坡向等
.
866.6~979.4 m 21.~41°
33.1~84.6 m T3y地形湿度为
0~0.123 . SBiLSTM
CRF 模型因为其非线性表征能力和可解释性
而拥有显著的滑坡易发性预测建模实用性 .
References
Alvarez ‐Melis, D., Jaakkola, T. S., 2018. Towards Robust
Interpretability with Self Explaining Neural Networks.
Advances in Neural Information Processing Systems 31
(NeurIPS 2018). https://arxiv.org/abs/1806.07538
Bhattacharya, P., Tanwar, S., Bodke, U., et al., 2021.
BinDaaS: Blockchain‐Based Deep ‐Learning as‐a‐Service
in Healthcare 4.0 Applications. IEEE Transactions on
Network Science and Engineering, 8: 1242-1255.
Cheng, K. Y., Wang, N., Shi, W. X., et al., 2020. Re‐
search Advances in the Interpretability of Deep Learn‐
ing. Journal of Computer Research and Development, 57
(6): 1208-1217 (in Chinese with English abstract).
Eiras, C., Souza, J., Freitas, R., et al., 2021. Discriminant
Analysis as an Efficient Method for Landslide Suscepti‐
bility Assessment in Cities with the Scarcity of Predispo
sition Data. N atural Hazards, 107: 1427- 1442.
Fang, R. K., Liu, Y. H ., Su, Y. C., et al., 2021. A Early
Warning Model of Regional Landslide in Qingchuan
County, Sichuan Province Based on Logistic Regres
sion. Hydrogeology & Engineering Geology, 48(1):
181-187 (in Chinese with English abstract).
Feng, X., Wang, Y., Liu, Y., et al., 2022. Susceptibility
Assessment of a Translational Rockslide Considering
the Control M echanism and Spatial U ncertainty of a
Weak Interlayer: Application Study in Tiefeng Town
ship, Wanzhou District. Bulletin of Geological Science
and Technology, 41(2): 254- 266 (in Chinese with
English abstract).
Forney, G.D., 1973. The Viterbi Algorithm. Proceedings of
the IEEE, 61: 268-278.
Gaur, M., Faldu, K., Sheth, A., 2020. Semantics of the
Black Box: Can Knowledge Graphs Help Make Deep
Learning Systems More Interpretable and Explainable?.
IEEE Internet Computing, 25: 51-59.
Guo, G., Chen, Y., Li, M.H., et al., 2013. Statistic Rela
tionship between Slope Gradient and Landslide Proba‐
bility in Soil Slopes around Reservoir. Journal of Engi‐
neering Geology, 21(4): 607-612 (in Chinese with
English abstract).
Guo, T. S., Zhang, J. Q., Han, Y., et al., 2019. Evaluation
of Landslide Susceptibility in Yanchang County Based
on Particle Swarm Optim ization Based Support Vector
Machine. Geological Science and Technology Informa
tion, 38(3): 236-243 (in Chinese with English abstract).
Hong, H. Y., Tsangaratos, P., Ilia, I., et al., 2020. In
troducing a Novel Multi Layer Perceptron Network
Based on Stochastic Gradient Descent Optimized by a
MetaHeuristic Algorithm for Landslide Susceptibility
Mapping. The Science of the Total Environment,
742: 140549. https://doi. org/10.1016/j. scito‐
tenv.2020.140549
Hu, T., Fan, X., Wang, S., et al., 2020. Landslide Sus
ceptibility Evaluation of Sinan County Using Logistics
Regression Model and 3S Technology. Bulletin of Geo‐
logical Science and T echnology, 39(2): 113-121 (in
Chinese with English abstract).
Huang, F., Cao, Z., Jiang, S. H., et al., 2020a. Landslide
Susceptibility Prediction Based on a Semi Supervised
Multiple Layer Perceptron Model. L andslides, 17:
2919-2930.
Huang, F., Zhang, J., Zhou, C., et al., 2020b. A Deep
Learning Algorithm Using a Fully Connected Sparse Au‐
toencoder Neural Network for Landslide Susceptibility
Prediction. Landslides, 17: 217- 229.
Huang, F. M., Ye, Z., Yao, C., et al., 2020. Uncertainties
of Landslide Susceptibility Prediction: Different Attri‐
bute Interval Divisions of Environmental Factors and
Different Data Based Models. Earth Science, 45(12):
4535-4549 (in Chinese with English abstract).
Jia, G., Gariano, S.L., Tang, Q., 2021. Advancing Rainfall‐
Induced Landslide Detection Using Homogeneous Slope
Units and Distributed Rainfall Thresholds. EGU Gener‐
al Assembly Conference Abstracts, EGU21-6145.
Khan, R., Yousaf, S., Haseeb, A., et al., 2021. Exploring
a Design of Landslide Monitoring System. Complexity,
2021: 5552417. https://doi.org/10.1155/2021/5552417
Li, W. B., Fan, X. M ., Huang, F. M., et al., 2021. Uncer‐
tainties of Landslide Susceptibility Modeling under Dif‐
ferent Environm ental Factor Connections and Prediction
Models. Earth Science, 46(10): 3777-3795 (in Chinese
1708
5 发明 等: 于自 筛选 深度 学习 的滑 坡易 发性 预测 建模 及其 可解 释性
with English abstract).
Luo, L. G., Pei, X. J., Huang, R. Q., et al. 2021. Evalua‐
tion of Landslide Susceptibility in Jiuzhaigou Scenic Spot
by CF and Logistic Regression Coupled with GIS. Jour‐
nal of Engineering Geology, 29: 526-535 (in Chinese
with English abstract).
Linardatos, P., Papastefanopoulos, V., Kotsiantis, S.,
2020. Explainable AI: A Review of Machine Learning
Interpretability M ethods. Entropy (Basel, Switzerland),
23(1): 18. https://doi.org/10.3390/e23010018
Lipton, Z.C., 2018. In Machine Learning, the Concept of In‐
terpretability is Both Important and Slippery. Queue,
16: 28.
Moragues, S., Lenzano, M. G., Lanfri, M., et al., 2020.
Analytic Hierarchy Process Applied to Landslide Suscep
tibility Mapping of the North Branch of Argentino Lake,
Argentina. Natural Hazards, 105: 915-941.
Pradhan, B., 2013. A Comparative Study on the Predictive
Ability of the Decision Tree, Support Vector Machine
and Neuro Fuzzy Models in Landslide Susceptibility
Mapping Using GIS ‐Science Direct. Computers & Geo
sciences, 51: 350- 365.
Tian, N. M., Lan, H. X., Wu, Y. M., et al., 2020. Perfor‐
mance Comparison of BP Artificial Neural Network and
CART Decision Tree M odel in Landslide Susceptibility
Prediction. Journal of Geo ‐Information Science, 22(12):
2304-2316 (in Chinese with English abstract).
Wei, Z.J., 2020. Multi‐Stage Series Deep Convolutional Neu
ral Network Medical Image Registration Method (Disser‐
tation). Hunan University, Changsha (in Chinese with
English abstract).
Wu, R. Z., Hu, X. D., Mei, H. B., et al., 2021. Spatial
Susceptibility Assessment of Landslides Based on Ran
dom Forest: A Case Study from Hubei Section in the
Three Gorges Reservoir Area. Earth Science, (1): 321-
330 (in Chinese with English abstract).
Xu, M ., Feng, Q., Mei, Q., et al., 2018. Deep Type: On
Device Deep Learning for Input Personalization Service
with Minimal Privacy Concern. Proceedings of the ACM
on Interactive Mobile Wearable and Ubiquitous Tech
nologies, 2: 1-26.
Yao, J., Qin, S., Qiao, S., et al., 2022. Application of a
Two Step Sampling Strategy Based on Deep Neural
Network for Landslide Susceptibility Mapping. Bulletin
of Engineering Geology and the Environment, 81:
1-20.
Yao, X ., Tham, L., Dai, F., 2008. Landslide Susceptibility
Mapping Based on Support Vector Machine: A Case
Study on Natural Slopes of Hong Kong, China. Geomor‐
phology, 101: 572-582.
Zeng, C.Y., Yan, K., Wang, Z.F., et al., 2021. Survey of
Interpretability Research on Deep Learning Models.
Computer Engineering and Applications, 57(8): 1- 9
(in Chinese with English abstract).
Zhang, Y., Tino, P., Leonardis, A., et al., 2021. A Survey
on Neural Network Interpretability. IEEE Transactions
on Emerging Topics in Computational Intelligence, 5:
1-17.
Zhao, H. B., Liu, R., Liu, Y. H., et al., 2022. Research on
Classification and Identification of M ine Microseismic
Signals Based on Deep Learning Method. Journal of
Mining Science and Technology, 7(2): 166-174 (in Chi‐
nese with English abstract).
Zhou, T., Wu, W., Peng, L., et al., 2022. Evaluation of
Urban Bus Service Reliability on Variable Time Hori‐
zons Using a Hybrid Deep Learning Method. Reliability
Engineering and System Safety, 217(3):108090.
Zhu, L., Huang, L. H., Fan, L. Y., et al., 2020. Landslide
Susceptibility Prediction Modeling Based on Remote
Sensing and a Novel Deep Learning Algorithm of a
Cascade Parallel Recurrent Neural Network. Sensors
(Basel, Sw itzerland), 20(6): 1576. https://doi. org/
10.3390/s20061576
中文参考 文献
成科扬, , 师文喜, , 2020. 深度学习可解释性研究进
. 算机研究与发, 57(6): 1208-1217.
方然可, 艳辉 , 永超 , , 2021. 于逻辑回归的四川
川县区域滑坡灾害预警模型 . 水文地质工程地质, 48
(1): 181-187.
冯霄 , , 刘洋 , , 2022. 考虑软弱夹层控滑机制及其空
不确 顺层 滑坡 性评 : 铁峰
乡应用研究 . 质科技通, 41(2): 254-266.
郭果 , , 李明惠, , 2013. 质滑坡发育概率与坡度
关系研究 . 程地质学, 21(4): 607-612.
郭天颂, 张菊清, 韩煜 , , 2019. 基于 子群优化支持向量
. 地质科技情报, 38(3):
236-243.
, , , , 2020. 基于逻辑回归模型和 3S
. 地质科技通报, 39(2):
113-121.
黄发明, 叶舟 , 姚池 , , 2020. 滑坡 性预 确定 :
环境因子不同属性区间划分和不同数据驱动模型的影
. 地球科学, 45(12): 4535-4549.
李文彬, 宣梅 , 发明 , , 2021. 同环境因子联接和
测模型的滑坡易发性建模不确定性 . 地球科学, 46
1709
48
地球科学 http://www.earth‐science.net
(10): 3777-3795.
路广 , 向军 , 黄润秋, 2021. GIS 支持下 CF Lo‐
gistic 回归模型耦合的九寨沟景区滑坡易发性评价 .
程地质学报, 29526-535.
田乃满, 恒星 , 宇明 , , 2020. 工神经网络和决策
型在 滑坡 易发 分析 中的 性能 . 球信
学报 , 22(12): 2304-2316.
志军 2020. 多级串联深度卷积神经网络医学图像配准方
(硕士 学位 论文 ). 沙:南大 .
吴润泽, 旭东 , 红波 , , 2021. 于随机森林的滑坡
间易发性评价: 以三峡库区湖北段为例 . 地球科学, 46
(1): 321-330.
曾春艳, , 王志锋, , 2021. 深度学习模型可解释性研
究综述 . 算机工程与应, 57(8): 1-9.
赵洪宝, , 刘一洪, , 2022. 基于深度学习方法的矿山
. 矿业科学学报, 7(2):
166-174.
1710
Article
Full-text available
Landslide susceptibility mapping (LSM) is the foundation and critical part of landslide risk assessment. The bibliometric analysis of LSM literature can be applied to quantitatively analyze the research progress and development trend. The result will provide references for geological hazard risk assessment in China. In this study, based on the Web of Science and CNKI databases, the CiteSpace visual knowledge graph analysis tool has been used to carry out bibliometric analysis of LSM literature from 1985 to 2022. Moreover, the LDA analysis has been conducted on the abstract to subdivide the research in this field. The results showed that: (1) LSM is still a research hotspot at present. In China, there are a large number of studies and international cooperation about LSM. (2) Four of the top 10 authors in the number of published papers on LSM are from China. The institution that published the most papers on LSM is the Chinese Academy. The Chinese Journal of Geological Hazard and Control is the most popular Chinese journal and the Natural Hazards is the most popular English journals to publish LSM papers. The research on the subject of LSM has been greatly funded by the National Natural Science Foundation of China and the National Land and Resources Survey Project. (3) In the past five years, machine learning models (including deep learning, etc.) have been widely used as the most popular LSM models. (4) In order to achieve the simplification and intelligence of landslide susceptibility modeling and to improve the accuracy and practicability of the LSM results, the following parts of LSM, including the landslide inventory, conditioning factors, assessment unit, assessment model, connection methods and accuracy verification, need to be deeply explored in further studies.
Article
Full-text available
In view of the lack of research about landslide intensity prediction in the current quantitative vulnerability analysis of buildings, this paper innovatively proposes a quantitative analysis method of based on the combination of intensity empirical curve based on InSAR technology and spatial displacement prediction of secondary development of ABAQUS.Taking the Shimongmen landslide in the Three Gorges Reservoir Area as an example, the PS-InSAR method was adopted to calculate the cumulative displacement of the landslide in 2017-2020 and obtained the empirical curve of landslide intensity. The ABAQUS software was used to compile the subroutine of load and pore water pressure to calculate the cumulative displacement under extreme scenarios (reservoir water level drop with heavy rainfall) and predicted the vulnerability of buildings. The evaluation system of resistance was constructed by weighting eight indicators of PSO-Fuzzy AHP model, which can be combined with the landslide intensity to quantitatively evaluate the vulnerability of buildings. The results indicate that: (1) The evaluation system of resistance proposed in this paper can well present the structural characteristics of rural buildings in the Three Gorges Reservoir area, and has high evaluation accuracy. (2) The retrieved upper-intensity curve obtained by PS-InSAR is Ipu = 0.065 * Dtot 0.236which has higher prediction accuracy and effectively reduces false-negative errors. (3) The landslide intensity of extreme conditions simulated by ABAQUS increases with the increase of rainfall, the predicted vulnerability level of buildings increases, and the buildings with obvious deformation in the previous investigation are successfully warned. It is concluded that the landslide intensity prediction method and vulnerability analysis method proposed in this paper has high spatial identification and early warning accuracy, and realtime vulnerability mapping of buildings can be obtained through landslide intensity information.
Article
Full-text available
Translational rockslide is one of the important disasters that endanger the safety of mountain towns. The dip slopes with weak interlayers are the areas prone to a translational rockslide. The regional translational rockslide susceptibility assessment should consider the sliding mechanism and spatial distribution uncertainty analysis of weak interlayers. Taking Tiefeng Town of Wanzhou District as the study area, based on the detailed investigation of the material, structure, and spatial distribution of weak interlayer, this paper analyzed the evolution mechanism of shale and mudstone developing into the sliding surface under the action of primary deposition, tectonic deformation and supergene transformation, and summarized the deformation and failure mechanism of the translational rockslide. Considering the spatial distribution uncertainty of weak interlayer, a calculation model of the vertical distribution of weak interlayer and the contribution of the weak interlayer to slip control in the effective control depth are proposed. The key factors that characterize the sliding structure, including the type of weak interlayer and the contribution degree of the weak interlayer, are selected as susceptibility assessment factors. Besides, four elements of topography, slope structure, hydrogeology, and human activities are considered. The susceptibility of a translational rockslide in the study area was assessed with slope units adopting the analytic hierarchy process. The investigation and assessment results show that the mud interlayers of the Jurassic Zhenzhuchong Formation and the shale layer of the Ziliujing Formation are the main potential slip surface of the translational rockslide in the study area. The extreme high susceptibility area and high susceptibility area account for 9.7 % and 25.8 % respectively. The distribution of the underlying weak interlayer and the excavation of the slope units are the main factors affecting the susceptibility to landslide disasters. Human activities such as house building and road excavation are the main triggering factors of the translational rockslide. Compared with the susceptibility assessment results without considering the factors of the weak layer, the results of the method proposed in this paper are more in accord with the fact.
Article
Full-text available
Landslide is a critical natural geological hazard that causes severe damage to property, infrastructure, and humans. In general, some location-specific factors trigger a landslide. Wireless sensor network (WSN) is an enabling technology to monitor most of the parameters associated with these factors. A challenge in landslide monitoring through WSN is that each sensed data item might be critical whereas the underlying wireless communication is often unreliable. In case of landslides, the terrains have irregular shapes, providing harsh conditions for wireless communication thereby more data loss may be expected. This study focuses on the effect of lossy communication in WSN on the efficiency and accuracy of landslide monitoring systems. To this end, collaborative local data analysis is used to enable each node to decide locally whether its sensed data corresponds to a potential event-of-interest. Through extensive simulations, the performance of various landslide prediction and detection models has been evaluated. By and large, the study lets a significant insight into landslide monitoring before implementing a parameterized application for real-world deployment.
Article
Full-text available
The city of Ouro Preto, which is located in the state of Minas Gerais, Brazil, has a long history of mass movements influenced by the regional geology, geomorphology, and anthropic activities, which have resulted in harmful consequences to the population. However, most of the studies conducted in the region are qualitative and are directly dependent on the experience specialists. The aim of this study was to analyse the landslide susceptibility in the urban region of Ouro Preto quantitatively by using discriminant analysis. The landslide inventory was obtained by using unmanned aerial vehicle images and fieldwork. ArcGIS 10.6 and R 3.5.1 software were used, and the following landslide predisposing factors were considered: slope angle, slope aspect, profile curvature, and topographic wetness index (TWI). As geological and geotechnical data are still scarce in the interior of Brazil, we only used data derived from topography to determine the effectiveness of these factors for analysing landslide susceptibility. The slope angle proved to be the factor that most differentiated unstable from stable terrain, followed by TWI. The other parameters were not as effective in differentiating the stability conditions. The model efficiency was 88.6%, the specificity was 93.3%, and the sensitivity was 85.0%. Also, the prediction and success curve were used to evaluate the accuracy of the proposed landslides model, by using the area under the curve (AUC) criteria. It was shown that the AUC values 0.851 for testing and 0.838 for training indicate that the developed model provides an excellent prediction. The main contribution of this work is the demonstration of the effectiveness of using easily accessible data (derived from topography) for analysing landslide susceptibility with a multivariate statistical method. This method can contribute valuable information to urban planning efforts in cities without the need for robust data.
Article
The selection of nonlandslide samples is a key issue in landslide susceptibility modeling (LSM). In view of the potential subjectivity and randomness in random sampling, this paper considers LSM as a positive-unlabeled (PU) learning problem and proposes a two-step deep neural network framework (T-DNN). Through the Spy technique and iteratively training binary classifiers, negative samples with high confidence were identified from the random subsamples with unlabeled sets. Based on the framework and traditional random sampling, we used logistic regression (LR), support vector machine (SVM), and deep neural network (DNN) models for testing and validation. Taking the Changbai Mountain Area in Jilin Province, China, as an example, according to the regional landslide list and the metrological, geographical, and human factors of frequent disasters, landslide susceptibility was evaluated. Results show that the proposed T-DNN method can enhance the selection of negative samples and make the results of landslide susceptibility assessment more reliable and accurate; the area under the receiver operating characteristic curve (AUC) reaches 0.953. In addition, compared with traditional random negative sample sampling, the optimized sample set shows more stable and superior prediction performance in different classifiers.
Article
2018 Curran Associates Inc.All rights reserved. Most recent work on interpretability of complex machine learning models has focused on estimating a posteriori explanations for previously trained models around specific predictions. Self-explaining models where interpretability plays a key role already during learning have received much less attention. We propose three desiderata for explanations in general - explicitness, faithfulness, and stability - and show that existing methods do not satisfy them. In response, we design self-explaining models in stages, progressively generalizing linear classifiers to complex yet architecturally explicit models. Faithfulness and stability are enforced via regularization specifically tailored to such models. Experimental results across various benchmark datasets show that our framework offers a promising direction for reconciling model complexity and interpretability.
Article
Unreliable transit services can negatively impact transit ridership and discourage passengers from regularly choosing public transport. As the most important content of bus service reliability, accurate bus arrival prediction can improve travel efficiency for enabling a reliable and convenient transportation system. Accordingly, this paper proposes a novel deep learning method, i.e. variational mode decomposition long short-term memory (VMD-LSTM), for bus travel speed prediction in urban traffic networks using a forecast of bus arrival information on variable time horizons. The method uses the temporal and spatial patterns of the average bus speed series. The results show that the VMD-LSTM model outperforms other models on forecasting bus link speed series in future time intervals, whereas the artificial neural network model achieves the worst prediction. In conclusion, the VMD-LSTM method can detect irregular peaks of transit samples from a series of temporal or spatial variations and performs better on major and auxiliary corridors.
Article
Along with the great success of deep neural networks, there is also growing concern about their black-box nature. The interpretability issue affects people’s trust on deep learning systems. It is also related to many ethical problems, e.g., algorithmic discrimination. Moreover, interpretability is a desired property for deep networks to become powerful tools in other research fields, e.g., drug discovery and genomics. In this survey, we conduct a comprehensive review of the neural network interpretability research. We first clarify the definition of interpretability as it has been used in many different contexts. Then we elaborate on the importance of interpretability and propose a novel taxonomy organized along three dimensions: type of engagement (passive vs. active interpretation approaches), the type of explanation, and the focus (from local to global interpretability). This taxonomy provides a meaningful 3D view of distribution of papers from the relevant literature as two of the dimensions are not simply categorical but allow ordinal subcategories. Finally, we summarize the existing interpretability evaluation methods and suggest possible research directions inspired by our new taxonomy.
Article
The recent series of innovations in deep learning (DL) have shown enormous potential to impact individuals and society, both positively and negatively. DL models utilizing massive computing power and enormous datasets have significantly outperformed prior historical benchmarks on increasingly difficult, well-defined research tasks across technology domains such as computer vision, natural language processing, and human-computer interactions. However, DL's black-box nature and over-reliance on massive amounts of data condensed into labels and dense representations pose challenges for interpretability and explainability. Furthermore, DLs have not proven their ability to effectively utilize relevant domain knowledge critical to human understanding. This aspect was missing in early data-focused approaches and necessitated knowledge-infused learning (K-iL) to incorporate computational knowledge. This article demonstrates how knowledge, provided as a knowledge graph, is incorporated into DL using K-iL. Through examples from natural language processing applications in healthcare and education, we discuss the utility of K-iL towards interpretability and explainability.