You are on page 1of 13

Geomorphology 354 (2020) 107045

Contents lists available at ScienceDirect

Geomorphology

journal homepage: www.elsevier.com/locate/geomorph

Deep learning-based approach for landform classification from


integrated data sources of digital elevation model and imagery
Sijin Li a,b,c, Liyang Xiong a,b,c,⁎, Guoan Tang a,b,c, Josef Strobl b,d
a
Key Laboratory of Virtual Geographic Environment (Nanjing Normal University), Ministry of Education, Nanjing 210023, China
b
School of Geography, Nanjing Normal University, Nanjing 210023, China
c
Jiangsu Center for Collaborative Innovation in Geographical Information Resource Development and Application, Nanjing 210023, China
d
Department of Geoinformatics – Z_GIS, University of Salzburg, Salzburg 5020, Austria

a r t i c l e i n f o a b s t r a c t

Article history: Landform classification is one of the most important aspects in geomorphological research, dividing the Earth's
Received 28 September 2019 surface into diverse geomorphological types. Thus, an accurate classification of landforms is a key procedure in
Received in revised form 15 January 2020 describing the topographic characteristics of a given area and understanding their inner geomorphological for-
Accepted 15 January 2020
mation processes. However, landform types are not always independent of one another due to the complexity
Available online 16 January 2020
and dynamics of interior and external forces. Furthermore, transitional landforms with gradually changing sur-
Keywords:
face morphologies are widely distributed on the Earth's surface. With this situation, classifying these complex
Landform classification and transitional landforms with traditional landform classification methods is hard. In this study, a deep learning
Deep learning (DL) algorithm was introduced, aiming at automatically classifying complex and transitional landforms. This
DEM algorithm was trained to learn and extract landform features from integrated data sources. These integrated
Imagery data sources contain different combinations of imagery, digital elevation models (DEMs), and terrain derivatives.
The Loess Plateau in China, which contains complex and transitional loess landforms, was selected as the study
area for data training. In addition, two sample areas in the Loess Plateau with complex and transitional loess
hill and ridge landforms were used to validate the classified landform types by using the proposed DL method.
Meanwhile, a comparative analysis between the proposed DL and random forest (RF) methods was also con-
ducted to investigate their capabilities in landform classification. The proposed DL approach can achieve the
highest landform classification accuracy of 87% in the transitional area with data combination of DEMs and im-
ages. In addition, the proposed DL method can achieve a higher accuracy of landform classification with better
defined landform boundaries compared to the RF method. The classified loess landforms indicate the different
landform development stages in this area. Finally, the proposed DL method can be extended to other landform
areas for classifying their complex and transitional landforms.
© 2020 Elsevier B.V. All rights reserved.

1. Introduction the landform evolution process (Drăguţ and Blaschke, 2006; Hiller and
Smith, 2008; Wang et al., 2010).
The Earth's surface consists of various landforms, and different land- In the past century, the visual interpretation of topographic maps
form types contribute to the surface appearance through certain interior and aerial photographs has been one of the most commonly used ap-
and exterior forces of geomorphological, geological, hydrological, and proaches to landform classification (Hammond, 1964; Dikau et al.,
ecological processes (Drăguţ and Eisank, 2012; Evans, 2012; Song 1991). However, the visual interpretation method is time-consuming
et al., 2016). Thus, landform classification is one of the most important and labor intensive, and it relies on the experience of geomorphologists
procedures in geomorphological mapping and further understanding to generate reproducible results (Galli et al., 2008). Automated ap-
of the Earth surface process (Bocco et al., 2001; Wang et al., 2010). An proaches to landform classification can be proposed and conducted
accurate classification of landform in a given area plays an important with the development of data acquisition and analysis methods in geo-
role in expressing the corresponding topographical characteristics, re- science. Hence, such approaches are focused research aspects in geo-
vealing the geomorphological formation mechanisms, and indicating morphology and geomorphometry studies (Drăguţ and Eisank, 2011).
Currently, two main approaches to landform classification can be
⁎ Corresponding author at: School of Geography, Nanjing Normal University, Nanjing
found from previous studies, that is, pixel- and object-based methods
210023, China. (Xiong et al., 2018). The pixel-based method can automatically classify
E-mail address: xiongliyang@njnu.edu.cn (L. Xiong). geomorphological entities on the basis of a series of terrain features,

https://doi.org/10.1016/j.geomorph.2020.107045
0169-555X/© 2020 Elsevier B.V. All rights reserved.
S. Li et al. / Geomorphology 354 (2020) 107045 3

such as ridge and peak. These features were commonly calculated on suited to extracting information that can support the classification of
the basis of the neighborhood analysis method to define the local differ- similar landforms. The homogeneity of landforms requires powerful
ences (Wood, 1996). The classification methods, such as ISODATA, were learning ability and effective classification methods (Dikau, 1989;
used to distinguish landform types (Minar and Evans, 2008; Evans, Verhagen and Drăguţ, 2012).
2012; Jasiewicz and Stepinski, 2013). However, the pixel-based ap- Machine learning (ML) approaches have become one of the most
proach does not show good performance in non-uniform surfaces. popular methods for landform recognition and segmentation in the
Such an approach cannot effectively implement the classification of field of geomorphological research based on the development of
landform types, which cover multiple pixels (Saha et al., 2011). image segmentation methods (Heung et al., 2016). Machine learning al-
The object-based method is another commonly used approach in the gorithms have the capability of information extraction from multiple
field of landform classification. In the object-based method, the defini- and big data sources (Berral et al., 2013). The surface spectra extracted
tion of topographic objects is necessary to meet the requirement of from imagery and the morphological features derived from digital ele-
the conceptual model of landform entities (Blaschke, 2010). The adja- vation models (DEMs) can be used and fused for data training by ML
cent growing method is adopted in object-based approaches, and approaches. Such approaches can be applied in geomorphological re-
neighboring cells with similar features are merged into a polygonal search. Recently, among the ML methods, deep learning (DL) ap-
objectrepresenting a meaningful spatial entity (Drăguţ and Blaschke, proaches receive particular attention due to their active learning
2006; Romstad and Etzelmüller, 2012; Verhagen and Drăguţ, 2012). capability (LeCun et al., 2015; Ma et al., 2019). DL algorithms have
The results of object-based methods include topological relationships been widely used to identify terrain features, that is, landform recogni-
and morphological characteristics of different objects, and these tion, due to the powerful ability of information recognition and extrac-
object-based methods have been widely used in the classification of tion (Li and Hsu, 2018). However, in geomorphological studies,
landform elements (Verhagen and Drăguţ, 2012). However, previous landform recognition or extraction processes are employed to classify
applications have also demonstrated that the results based on the landforms into different types. Currently, most applications of DL
object-based method are sensitive to morphological discontinuities methods in landform classification research focus on the analysis of
within study areas and data sources (Drăguţ et al., 2009; Anders et al., macro-landforms in a broad spatial scale range with clear landform
2011). Moreover, transitional landforms with gradually changing sur- boundaries (Ehsani and Quiel, 2008; Li et al., 2017; Palafox et al.,
face morphology (i.e., morphological continuity) are widely distributed 2017). However, gradually changing surfaces and homogeneous land-
on the Earth's surface (Fig. 1). forms increase the difficulty of accurate landform classification pro-
Different landforms reflect various geomorphologic processes and cesses (Drăguţ and Blaschke, 2006). For instance, the Loess Plateau,
varied influence factors, whereas the surface morphology over the which is located in the northwestern part of China, has three typical
space consists of non-, sudden-, and gradual-change areas (Murray loess landforms (Fig. 1), which represent the different developmental
et al., 2009). In the non- and sudden-change areas, spatial boundaries stages of loess landscapes. The similar forces and environments cause
and terrain features between different landforms appear well defined transition landforms with large areas, which have similar characteristics
and can be detected easily. However, in transitional areas, landforms that cannot be simply distinguished (Xu et al., 2007; Zhao et al., 2017).
are gradually changing from one type to another, and the terrain fea- Dividing landforms in these transitional areas into specific types using
tures of these landforms change continuously (Phillips, 2003; Drăguţ traditional methods is still difficult (Fig. 1d, e). Thus, a suitable landform
and Blaschke, 2006; Xu et al., 2007). For instance, in the transitional classification method must be proposed, and constructing an approach
landforms of loess hill and ridge, their shapes are gradually changing on the basis of DL algorithms to classify landforms is feasible.
with similar terrain features. In this transitional area, defining criteria This study aims to classify landforms in transitional areas by using a
to distinguish different landforms in transition regions with traditional DL approach. First, within the study area in the Loess Plateau, a dataset
methods is difficult. The above outlined traditional methods are less of landform features from images, DEMs, and terrain derivatives in the

Fig. 1. Typical loess landforms. (a) loess tableland; (b) loess ridge; (c) loess hill; (d) transition area of mixed landforms of loess ridge and hill; (e) loess hill-dominated area (landform
boundaries are delineated by visual interpretation).
4 S. Li et al. / Geomorphology 354 (2020) 107045

Fig. 2. Study areas. (a) Location of the Loess Plateau in China; (b) DEM of the Loess Plateau; (c) image of Wuqi; (d) DEM of Wuqi; (e) image of Yulin; (f) DEM of Yulin.

gradually changing landforms was established. Thereafter, we proposed the formation process of loess landforms and are crucial in studies on
a network structure on the basis of DL, which can be trained to learn the landform evolution (Cao et al., 2013; Song et al., 2013). In addition,
landform features. We then discuss the performance of the DL algorithm most anthropogenic terraces were constructed in this convex loess land-
in landform classification and investigate the performance of experi- form area (Dai et al., 2019). Different loess landforms in this convex area
mental results in illustrating the geomorphological process. After that, also demonstrate different soil erosion processes (Fu, 1989; Fu et al.,
we compare the accuracies of the proposed DL and random forest 2000; Liu et al., 2016). Thus, accurate loess landform classification can
methods. Finally, the development stages of landforms in the study support the planning of agricultural production and soil and water con-
areas are assessed. servation (Fu, 1989; Zhou et al., 2010).
We observed through the practical experiments that the differences
between loess tableland and other two landforms are clear. However,
2. Study area and data distinguishing loess ridge from loess hill is difficult. The forms of these
two types are similar, and their definitions are ambiguous, making the
In this study, the Loess Plateau in China was chosen as a study area to classification difficult. In this study, we aim to implement a classifying
classify landforms by using the DL approach due to the complex terrain network with the prime performance objective of distinguishing loess
combination and transitional landform types. The Loess Plateau is ridge and hill.
located in the middle and upper reaches of the Yellow River (Fig. 2a, b) The areas of basic data, which include training, validation, and test
and covers an area of 640,000 km2. The typical landforms in this area data, were selected on the basis of the landscape evolution and morpho-
can be used to reflect various geomorphic processes including gully logical characteristics. Fig. 2b shows the five typical areas chosen in this
erosion (Sun, 2002). Currently, loess landforms in convex landform study, marked by circles with a number. These areas are located in the
(i.e., positive landform) areas are recognized as three basic types (Zhou transitional areas of loess hill-ridge, providing the required information
et al., 2010, Xiong et al., 2014a, 2014b, 2014c). These types are loess table- for landform classification. In addition, the two sample areas, Wuqi
land, ridge, and hill (Table 1). Most previous research has focused on con- (Fig. 2c, d) and Yulin (Fig. 2e, f), were selected to discuss the network's
cave landform (i.e., negative landform) objects, such as gullies and vallies. results. These two areas represent different development stages of loess
Loess landforms on the convex terrain surface are representative geo- landforms. Wuqi has more loess ridges, whereas Yulin suffered more in-
graphical objects (Zhou et al., 2010). Such geographical objects reflect tense erosion and has more loess hills. No training data exist in these
two areas to enhance the accuracy of the experiment. The data of
Table 1 these two sample areas are only used in the experiments after the opti-
Characteristics of three basic loess landforms.
mization and assessment of the network.
Type Development stage Feature The basic data we used in this study are DEM, terrain derivatives, and
Loess Young stage The top of tableland is flat and it contains
remote sensing images, (Table 2). The DEM data were collected from
Tableland villages and croplands. version 2 of the ASTER Global Digital Elevation Model (GDEMV2). The
Loess Ridge Middle stage They have the long strip shape and connect GDEMV2 is one of the most complete high-resolution digital topo-
with other ridges and hills. graphic datasets of the world at a spatial resolution of 1 arc-second (ap-
Loess Hill Mature stage Their visual features are similar to mounds.
proximately 30 m) (Satgé et al., 2015; Diani et al., 2017; Liu et al., 2019).
S. Li et al. / Geomorphology 354 (2020) 107045 5

Table 2 structure on the basis of previous studies (Bu et al., 2014). The network
Information on the selected data. was also trained in this stage. After obtaining the trained network with
Data Resolution Function optimal training performance, the test data were used to calculate the
DEM 30 m Providing information about surface relief and elevation.
accuracy of the network. The classification performances based on dif-
Slope 30 m Reflecting the changing rate of surface relief. ferent input datasets were also compared, and an optimal data combi-
Aspect 30 m Reflecting development direction of landforms. nation with the highest accuracy was chosen in this process. The last
Image 8.96 m Providing information about optical spectrum and texture. step focused on the analysis of landform development. We tried to as-
sess the development stage in the study areas on the basis of the results
we obtained from our approach.
In the Loess Plateau, the length or radius of loess landforms is larger
than 30 m, and the features of loess landforms can be observed clearly 3.1. Training data collection
in the DEM at a resolution of 30 m. Moreover, terrain derivatives
(i.e., slope and aspect), which were derived from the DEM data, were DL algorithms commonly need a large amount of data to train a net-
also used in this study to further represent the differences between var- work, whereas test data are also necessary to investigate the feasibility
ious loess landforms (Kienzle, 2004). A slope reflects the changing rate and accuracy of algorithms (Gevaert et al., 2018). First, we collected im-
of surface elevation, which can help detect the boundary of loess land- ages and DEMs in loess landform transition areas and calculated several
forms (Iwahashi et al., 2001; Weiss, 2001). An aspect also provides in- terrain derivatives on the basis of DEMs. In our study, we resampled the
formation about the development direction of landforms (Chang and cell size of DEM to an equal level of imagery. In this process, DEM data
Tsai, 1991). For instance, the characteristics of aspect in loess ridge were resampled using the nearest neighbor method to obtain data
areas include several strip-type polygons, which should be differed to with the same cell size as remote sensing imagery. This resampling pro-
that in loess hill areas. Remote sensing images also have a great signifi- cess in DL algorithms has been proven to exert a limited influence on the
cance in landform classification (Drăguţ and Blaschke, 2006; Wan et al., accuracy of the landform classification process (Latifovic et al., 2018;
2012). Images reflect the landform features of spectrum and texture. Metelka et al., 2018).This is because the resolution of resampled DEM
Google Earth images were used in our study because of the similarity is finer than that of the original DEM. In addition, resampling is neces-
of loess landforms. We selected remote sensing images at a spatial res- sary when we want to combined multi-source data with different reso-
olution of 8.96 m as one of the basic data. lutions. The network requires the same cell size as input data.
Meanwhile, annotated data were also prepared. Annotated data can
3. Methodology be viewed as “labeled data” which represent different polygons. These
polygons should be different landform units with corresponding attri-
The proposed approach involves these following four steps: (1) data bute values. We interpreted all landforms existing on the images and
preparation, (2) DL network building and training, (3) accuracy test and manually drew their boundaries. These annotated data were also
validation, and (4) landform analysis of study area. First, remote sensing added to the total dataset. The class balancing problem was also consid-
imagery, DEMs, and their derivatives were collected as the input ered in this step, and we randomly annotated similar numbers of poly-
dataset. Second, a DL approach was established for landform classifica- gons for each type (W. Zhang et al., 2018). In our study, the number of
tion on the basis of the prepared input dataset. We built a network loess landform polygons (i.e., loess ridge and loess hill) was more than
with multi-channel geomorphological feature extraction and fusion 1500. They were collected from visually interpreting the original data,

Fig. 3. U-net architecture (example for 32 × 32 pixels in the lowest resolution) (Ronneberger et al., 2015).
6 S. Li et al. / Geomorphology 354 (2020) 107045

and a dataset was built on the basis of these data. In the prepared external geomorphological processes, which lead to the deviation be-
dataset, each type of data has 150 samples. tween the natural and theoretical features of landforms. The elastic de-
The complete dataset was randomly divided into the following three formation could simulate a part of landform change and enhance the
parts: (1) training data for analyzing the landform features, (2) valida- classification ability of the network. Moreover, this strategy could re-
tion data for minimizing the over fitting, and (3) test data for evaluating duce the required amount of available training samples (Norman
the performance of the trained network. The percentages of these three et al., 2018). An enormous amount of typical samples that cannot be dis-
parts are 80%, 10%, and 10%, respectively. tinguished by manual visual interpretation exists. Hence, such a strat-
In this research, all samples sizes are 224 × 224 pixels (approxi- egy is an improvement for landform classification.
mately 2 km × 2 km). The selection of this sample size was based on
the following reasons. First, the features of loess landforms in the area
can be reflected clearly in the images, and the information that will be 3.2.2. Modification based on U-Net
used for the classification is preserved. In other words, the sample size The traditional U-Net structure has only one input channel. How-
is suitable for the purpose of our study to classify the loess landforms. ever, in geomorphological research, landform is a complex combination
Furthermore, the specific sample size affects the computational time of characteristics, and its features cannot be expressed completely on
on the basis of a satisfactory precision. the basis of one channel of data. For instance, images can describe the
texture of landforms, whereas they cannot show information about sur-
3.2. Deep-learning-based algorithm face elevation. The classification of landforms requires various data sets
to represent different information and the completeness of features in-
In this research, we chose a typical structure for a deep convolution fluences the accuracy of the network. Hence, the architecture with one
network, U-Net (Ronneberger et al., 2015), as the basic network struc- input channel limited the application of U-Net or other similar networks
ture to implement the classification of landforms because of its effec- in the field of landform studies. In this study, a multi-channel feature fu-
tiveness and simplicity. We changed certain layers and adjusted them sion structure (Fig. 4) was used as a replacement of the U-Net's
to geomorphological research on the basis of the existing structure of contracting path. In our network, the number of input channels is
U-Net. changeable. Each type of the previous dataset we established, such as
images, DEM, and slope, can be used as an input of one channel. In
3.2.1. U-Net each channel, the first three layers are similar to the original U-Net. A
Fig. 3 shows the network structure, which can be divided into two concatenation operation was established between the third and fourth
paths, contracting and expanding paths (Ronneberger et al., 2015). convolutional layers to connect different feature maps. The different in-
The function of the contracting path is to extract valuable information formation extracted from divergent data can be fused and combined
from input data and compress images. The clear difference between into one powerful feature map through the concatenation layer. Each
U-Net and traditional CNN is the architecture of the expanding path. training sample will pass 16 convolution layers to extract features and
The up-convolution for upsampling replaces the fully connected layers. reduce the size of the sample. Various features can be extracted by the
U-Net uses extensive data augmentation by applying elastic deforma- proposed network from different data, and Fig. 5 shows the feature
tion to the available training images (Ronneberger et al., 2015). The maps of four convolution layers.
strategy of data augmentation was used to meet the segmentation Another modification of our network is about the concatenation be-
needs of biomedical images (Norman et al., 2018). Deformation is a tween the contracting and expanding paths. In the process of concatena-
common phenomenon for biological cells, and elastic deformation can tion, choosing the sequence of channels will also influence the result of
simulate it efficiently. Similar principles exist in the research of land- classification. In this study, the channel of remote sensing images was
forms. Landforms suffered the various influences of the inner and chosen to connect with corresponding layers of the expanding path.

Fig. 4. Modified structure in this study (replacing the structure which is labeled by the red box in Fig. 3).
S. Li et al. / Geomorphology 354 (2020) 107045 7

Fig. 5. Features extracted by the proposed method from different data.

3.3. Network validation and accuracy test application in the case studies. We conducted a three-step assessment
to evaluate the fine-tuned network. First, we recorded the learning
The novel network was optimized during the training process. In this curve of the network, which could help to minimize over-fitting with
step, we evaluated the accuracy of the trained network before its the validation dataset. We adopted an early stopping strategy
8 S. Li et al. / Geomorphology 354 (2020) 107045

Table 3 4. Results
Data combinations of case studies.

Case number Data type 4.1. Network optimization and accuracy


Case 1 Image
Case 2 Image + DEM 4.1.1. Optimization
Case 3 Image + DEM + aspect We optimized the network during the training process and assessed
Case 4 Image + DEM + slope the accuracy of the trained network before the application in the case
Case 5 All indices
studies. The validation dataset was used to analyze the over-fitting of
the network. Fig. 6 depicts the loss curves of the training process
based on the different data combination.
We should decrease the influence of over-fitting in the training pro-
(Loughrey and Cunningham, 2005) and trained the network with cess to obtain the optimal result of a network (Q. Zhang et al., 2018). A
predefined 1000 epochs so that we could record the full learning significant feature of over fitting is that the values of validation loss
curve of our network. The curves of training and validation losses can reach their lowest and then rebound, whereas the training loss values
be used to ascertain the optimal value of network epochs. Moreover, still descend. For instance, Fig. 6b shows that the curve indicates that
we assessed the accuracies of classification on the basis of the reference the network tends to “memorize” the training data rather than “learn”
data, which were conducted by human interpretation. We assessed the to extract features after the 140th epoch. We ascertained the optimal
accuracy of classification at two levels, object and pixel levels. For the epoch values of five data combinations (Table 3), and Fig. 6 shows
object level, first, we automatically calculated the intersection over these epoch values in blue. We found that the threshold values are not
union (IoU) value of each polygon produced by network. The IoU was high, and these values decrease gradually with the increase of input
calculated as follows: channel number based on the learning curves. These two phenomena
can be explained by two reasons. First, the size of data remarkably influ-
ences the necessary epoch of training process. In our study, the rela-
IoU ðA; BÞ ¼ areaðA∩BÞ=areaðA∪BÞ; ð1Þ tively small size of training sample directly causes the early
convergence with small epoch values. Second, the use of additional
data can supply extra information about landform features, and it also
where A is a polygon in the classification result and B is a polygon in the can accelerate the process of learning, which results to the gradual de-
reference data representing the same landform object. A high value of crease of epoch values.
IoU means a high overlap accuracy of the classifying result. In this
study, we set a threshold of IoU as 0.45, which is slightly lower than a
traditional value of 0.5 because of the similarity of loess landforms. A 4.1.2. Accuracy assessment
polygon with an IoU value that was greater than 0.45 was counted as We completed the training of network with the required epoch on
a true positive (TP). The polygons which have the IoU value of 0–0.45 the basis of the former analysis. Thereafter, we assessed the accuracy
were counted as false positives (FP). Moreover, the polygons that of the network with the hold-out test dataset. First, we calculated the
were not detected by networks in the classification result but exist in accuracy of classification in terms of landform objects. Tables 4 and 5
the reference data were regarded as true negatives (TN). The false neg- show the results. Precision and recall show a negative relationship,
atives (FN) represent polygons, which should be classified as back- which means that an increase in one indicator correspond with a de-
ground, but were classified as a part of loess landforms. In the field of crease in the other indicator. This phenomenon is common in the field
DL, precision, recall, and over-accuracy are three common indicators of DL research (Davis and Goadrich, 2006). Tables 4 and 5 show that
that reflect the performance of networks (Davis and Goadrich, 2006). the precision of the loess hill identification in all experiments is greater
On the basis of these indicators, we calculated precision, recall, and
than 0.7, except for case 4. For the classification of loess ridge, the values
overall accuracy (OA): of recall are higher than the results of loess hill. All recall values of loess
ridge classification are over 0.75. For the values of overall accuracy, cases
Precision ¼ TP=ðTP þ FPÞ; ð2Þ 2 and 5 exhibit excellent performance in the classification of loess hill
and loess ridge.
We can see that the factors of slope and aspect did not have the pos-
Recall ¼ TP=ðTP þ FNÞ; ð3Þ itive effect on the classification of loess landforms. For aspect, features of
loess hill are similar to several parts of loess ridge. A part of loess hill re-
mains at the transition stage between loess ridge and hill. Slope can im-
OA ¼ TP=ðTP þ FP þ TN þ FNÞ; ð4Þ prove the performance of boundary detection. However, for the
classification of loess landforms, the result based on slope is not satisfac-
tory. The reason is that loess ridge and hill are located in convex terrain.
where TP, FP, TN and FN are the polygon's accounts of true positive, false The difference of slope between these two entities is insignificant,
positive, true negative, and false negative, respectively. For the pixel though, leading to worse results of loess hill classification.
level, we calculated the percentage of pixels with the same distribution Table 6 shows the percentage of correctly classified pixels. The re-
of reference data. Finally, in this study, we compared the accuracies of sults are satisfactory and all over 70%, which reflect that the network
the DL and RF approaches, and we also analyzed the differences of has a good performance at the pixel level. The result based on case 2 (in-
two methods. cluding image and DEM) has the highest accuracy, corresponding with
In addition, we conducted five independent experiments using dif- the assessment of object level. Another phenomenon is that the accura-
ferent data combinations. The performance of the fine-tuned network cies of loess hill are higher than for loess ridge, due to the boundaries of
was also tested on the basis of these five data combinations. Image, loess ridge being more complex than for loess hill. The omission errors
DEM, and terrain derivatives further reflect a part of landform features, are predominantly located in the area of loess ridge boundaries. The
and the combination of different data can enhance the expression of similarity between loess ridge boundary and gully leads to the wrong
landform features. Table 3 shows the information about the data combi- classification at the pixel level. Fig. 7 shows that all polygons in pre-
nations. The accuracy and the performance on the basis of different data dicted results have a similar shape as the actual landforms, but certain
combinations was analyzed to choose the best data combination. redundant parts also exist near the margin. We consider the position
S. Li et al. / Geomorphology 354 (2020) 107045 9

Fig. 6. Loss graphs for network optimization of different data combinations. (a) Case 1; (b) case 2; (c) case 3; (d) case 4; (e) case 5.

and the area of predicted results as satisfactory, this network therefore landform classification results on the basis of case 2 (i.e., images and
can be used in the practical case study based on the former analysis. DEMs). The number of loess hill units in Yulin (Fig. 8a) is clearly more
than that in Wuqi (Fig. 8b). This demonstrates that the Yulin area has
4.2. Loess landform classification reached the medium stage of loess ridge-hill development due to in-
tense erosion. However, the Wuqi area remains at the earlier stage
After the training and test process, the network is qualified and can and mainly involves loess ridges. The transition from loess ridge to
be used to classify the two loess case areas (Fig. 2). Fig. 8 shows the loess hill in Wuqi is obvious in the next period. In addition, results of
10 S. Li et al. / Geomorphology 354 (2020) 107045

Table 4 5. Discussion
Accuracy assessment of loess hill landform classification (object level).

Loess Hill Predicted class Precision Recall Overall 5.1. Comparison between DL and RF methods
accuracy
Positive Negative
We compared the performance of our network with that of the RF
Actual Case True 122 85 0.76 0.59 0.50
method for automated landform classification. The RF method has
class 1 False 39 0
Case True 125 36 0.86 0.78 0.69 been widely used in image classification on the basis of multiple deci-
2 False 20 0 sion trees (Belgiu and Drăguţ, 2016; Breiman, 2001). In this RF experi-
Case True 100 88 0.74 0.53 0.45 ment, we first constructed 300 trees to conduct the experiment of
3 False 35 0 landform classification of loess ridge and hill. To compare the classifica-
Case True 90 80 0.66 0.53 0.42
4 False 46 0
tion results of RF and DL, the experiment on the RF method was proc-
Case True 136 45 0.77 0.75 0.62 essed based on same dataset of the DL method. Second, the annotated
5 False 40 0 data (labeled data) were transformed to the required data format of
the RF method. Finally, after the training process, the landforms of
loess ridge and hill were classified. The results of the classification
using the two methods were compared at two different levels. At first,
Table 5
Accuracy assessment of loess ridge landform classification (object level). at the object level, as seen in Fig. 9, the predicted polygons are broken
and have many holes. Besides, polygons in the result of the RF approach
Loess Ridge Predicted Class Precision Recall Overall
include some areas of negative terrain, considered erroneous results.
accuracy
Positive Negative Omission errors are common in the results of RF. In addition, the RF ap-
Actual Case True 125 39 0.63 0.76 0.52 proach has difficulties to distinguish the type of landform. In the results
class 1 False 75 0 of RF, almost all polygons are classified into loess ridge. Finally, Fig. 9
Case True 133 24 0.82 0.85 0.71 shows that the boundaries of the RF approach as very irregular and bro-
2 False 30 0
ken. These results show that the performance of the RF approach is
Case True 120 35 0.59 0.77 0.50
3 False 85 0 worse than our DL network.
Case True 124 29 0.66 0.81 0.57 We also assessed the accuracy of the RF approach at the pixel level.
4 False 65 0 Table 7 compares the accuracy of the DL network and the RF approach.
Case True 138 30 0.77 0.82 0.66 Clearly, the percentage of the correct classification of the RF approach is
5 False 41 0
much less than the percentage of the DL approach. Overall, the perfor-
mance of the DL network is better than the RF approach. In transition
areas, distinguishing two similar landforms on the basis of a simple
Table 6 dataset is difficult in the RF approach. The RF approach requires addi-
Accuracy assessment for landform classification (pixel level). tional representative datasets to improve the accuracy of classification.
However, compared with the RF approach, the DL approach offers a bet-
Case 1 Case 2 Case 3 Case 4 Case 5
ter performance based on the readily accessible data with easy pre-
loess ridge 0.73 0.83 0.73 0.77 0.73 processing. The accuracy of networks for the DL approach shows
loess hill 0.84 0.91 0.85 0.88 0.83
improvement.

5.2. Indication of loess landform development stages

case 2 show that DEMs can enhance the discriminability of each class, The creation of landforms is a complex process, and the changing of
and the area of overlap of two entities decreased obviously. The joining landform units is a process of geomorphic evolution at the micro level
of DEMs provides the information of terrain relief and improves the per- (Zhou et al., 2010). The number of landform units directly reflects the
formance of the network. Compared with other results, loess hills de- stage of landscape development and the intensity of influential processes.
tected in case 2 have more regular circular shapes. Moreover, the Fig. 8 shows that the difference of results in two study areas is clear. The
visual effect is also better than others, and the boundaries of polygons number of polygons representing the loess hill in Yulin, is significantly
are more similar to actual landform entities. Hence, we consider that higher than in Wuqi, and the area of spatial distribution in Yulin is also
DEM offers the information about elevation and terrain and improved larger than area in Wuqi. In addition, in our study, the number of poly-
the performance of landform classification. gons and other derivatives, which can depict surfaces, are computable.

Fig. 7. Boundary differences between predicted results and ground truth. (a) Case 1; (b) case 2; (c) case 3; (d) case 4; (e) case 5.
S. Li et al. / Geomorphology 354 (2020) 107045 11

Fig. 8. Landform classification results of Yulin and Wuqi. (a) Yulin. (b) Wuqi.

Fig. 9. Comparison between DL and RF approaches. (a) Case 1 of DL; (b) case 2 of DL; (c) case 3 of DL; (d) case 4 of DL; (e) case 5 of DL; (f) case 1 of RF; (g) case 2 of RF; (h) case 3 of RF;
(i) case 4 of RF; (j) case 5 of RF.

Fig. 10 shows that the differences of these factors calculated in two stud- Wuqi. Loess hill represents more intense erosion than loess ridge on the
ies. The number of loess ridge in the Yulin area is higher than those in the basis of the prior knowledge. Yulin underwent more intense erosion
Wuqi area, whereas the areas of loess ridge in the Yulin area are smaller and the period of erosion in Yulin is also longer than Wuqi. The method
than those in the Wuqi area. This phenomenon shows that the surface we propose improves the performance of landform classification. The
of the Yulin area is more broken than the surface of the Wuqi area, and landform development in the study area can be judged and analyzed
Yulin has a more mature development stage than Wuqi. In addition, through the relative calculation of landforms' factors.
Yulin has higher area and polygon number of loess hill than that of Moreover, in a specific area such as the Loess Plateau, the landforms
in convex terrain have the phenomenon of geomorphological inheri-
Table 7
tance (Xiong et al., 2014a, 2014b, 2014c). Extending the research of
Percentage of correctly classified pixels. paleotopography to the micro level and constructing the conversion
model of the modern surface to paleotopography could be possible
Case 1 Case 2 Case 3 Case 4 Case 5
based on the analysis of landforms. The automated classification of land-
RF approach 35% 59% 59% 60% 61% forms can support the research on geomorphological inheritance and
DL approach 78% 87% 79% 83% 78%
improve the efficiency and accuracy of the results.
12 S. Li et al. / Geomorphology 354 (2020) 107045

Fig. 10. Statistics results of polygon number, perimeter, and area. (a) Polygon number; (b) perimeter; (c) area.

5.3. Implication of landform classification in other areas areas (Dian, 1996; Tang and Day, 2000). Fenglin and Fengcong can be
viewed as important signs of karst landform developmental degree
Research on specific landforms helps describe the development and (Zhu, 2009), which have a similar role as loess ridge and hill (Yang
formation process in a given area (Hammond, 1964; Evans, 2012). In the et al., 2019). However, due to the morphological similarity of these
typical research of geomorphology, a common problem is the complex- two landforms, the classification of Fenglin and Fengcong still has
ity and similarity of landforms. An advanced structure of a classification many problems (Smart et al., 1986; Waltham, 2008). The classification
model is necessary for accurate results. However, for other methods in of loess ridge and loess hill that we conducted in this research is
different study, it is difficult to apply an algorithm to other tasks areas under a similar situation as the classification of Fenglin and Fengcong.
and acquire a satisfied result. A special dataset should be used to realize network training based on
A significant advantage of DL algorithms is the transferability of the our method, and improve the accuracy of karst landform classification.
network (Long et al., 2015). The fine-tuned network could be applied to
similar tasks and perform well through the replacement of training data 5.4. Possible limitation of the proposed method
(Marmanis et al., 2015). This process is called transfer learning, which is
the ability of a system to recognize and apply knowledge and skills The method is feasible and can support geomorphological research
learned in previous tasks to novel domains. Transfer learning can im- to realize the classification of landforms and analyze the development
prove the efficiency of landform research (Shin et al., 2016). In this pro- stage of landscapes based on DL (W. Zhang et al., 2018; Ma et al.,
cess, a relatively small dataset can be used to adjust a network to an 2019). The network used in this study successfully utilizes multiple
optimal stage. The formation of different landforms is based on similar data sources to extract landform features from training datasets and
processes, and the similarity of features could also be discovered from classifies the targets of interest into correct categories. However, certain
the appearance of landforms. Hence, transfer learning is a reasonable problems still exist in our study. We envision two possible ways to im-
process for a geomorphological study. Moreover, the transferability of prove the performance of our network: (1) find the optimal definition
networks decreases time and specifically efforts to collect basic data of landform boundary: in our research, we have proven that our net-
and train the classification network. work can detect the correct position of landforms with a comparatively
Karst landforms have similar features as the loess landforms that we precise boundary result. However, the boundaries of some landform en-
tackled in this study. For instance, Fenglin and Fengcong are two typical tities are ambiguous. Boundaries with a clear definition could improve
karst landforms with prominent features (Fig. 11). Fenglin areas have the accuracy and performance of landform classification. (2) Provide
many pinnacles with large slopes and less dense peaks, and most pinna- different terrain derivatives and integrate multi-level features of land-
cles are tower-shaped. However, the pinnacles in Fengcong areas are forms. Different data could depict the various features of landforms
mostly cone-shaped, and they have less slopes than pinnacles in Fenglin from different perspectives (Zhao et al., 2017). We expect to find several

Fig. 11. Images of Karst landforms. (a) Fenglin and (b) Fengcong (Meng et al., 2018).
S. Li et al. / Geomorphology 354 (2020) 107045 13

derivatives that could reflect more representative features of landforms. Dikau, R., 1989. The application of a digital relief model to landform analysis in geomor-
phology. Three dimensional applications in geographical information systems 51–77.
The representative data combination can improve the accuracy of land- Dikau, R., Brabb, E.E., Mark, R.M., 1991. Landform Classification of New Mexico by Com-
form classification. puter. Open File Report 91-634. US Geological Survey. 15.
Drăguţ, L., Blaschke, T., 2006. Automated classification of landform elements using object-
based image analysis. Geomorphology 81 (3-4), 330–344.
6. Conclusions Drăguţ, L., Eisank, C., 2011. Object representations at multiple scales from digital elevation
models. Geomorphology 129 (3–4), 183–189.
Drăguţ, L., Eisank, C., 2012. Automated object-based classification of topography from
(1) The DL method was used for landform classification and demon-
SRTM data. Geomorphology 141, 21–33.
strates good performance. In this study, we constructed a network Drăguţ, L., Schauppenlehner, T., Muhar, A., Strobl, J., Blaschke, T., 2009. Optimization of
based on a modified U-Net. Multiple data layers were collected and scale and parametrization for terrain segmentation: an application to soil-landscape
modeling. Comput. Geosci. 35 (9), 1875–1883.
fused as the input to the network. The classification accuracy is satisfac-
Ehsani, A.H., Quiel, F., 2008. Geomorphometric feature analysis using morphometric pa-
tory, and the results of experiments can be used in further geomorpho- rameterization and artificial neural networks. Geomorphology 99 (1–4), 1–12.
logical research. Evans, I.S., 2012. Geomorphometry and landform mapping: what is a landform? Geomor-
(2) Constructing several data combinations and the accuracies based phology 137 (1), 94–106.
Fu, B., 1989. Soil erosion and its control in the Loess Plateau of China. Soil Use Manag. 5
on the different data layers were assessed. (i) The remote sensing offers (2), 76–82.
immense information about the object boundary and area; (ii) DEM can Fu, B., Chen, L., Ma, K., Zhou, H., Wang, J., 2000. The relationships between land use and
improve the distinction of loess ridge and hill; (iii) Aspect data can im- soil conditions in the hilly area of the loess plateau in northern Shaanxi, China. Catena
39 (1), 69–78.
prove the accuracy of loess hill detection, but its the performance is Galli, M., Ardizzone, F., Cardinali, M., Guzzetti, F., Reichenbach, P., 2008. Comparing land-
worse than that of DEM; (iv) slope data can improve the accuracy of slide inventory maps. Geomorphology 94 (3–4), 268–289.
boundaries slightly; and (v) the best data combination in our study is Gevaert, C., Persello, C., Nex, F., Vosselman, G., 2018. A deep learning approach to DTM ex-
traction from imagery using rule-based training labels. ISPRS J. Photogramm. Remote
image and DEM. Sens. 142, 106–123.
(3) On the basis of the comparison with the RF approach, the perfor- Hammond, E.H., 1964. Analysis of properties in land form geography: an application to
mance of the DL approach is better than the RF approach. The result of broad-scale land form mapping. Ann. Assoc. Am. Geogr. 54 (1), 11–19.
Heung, B., Ho, H.C., Zhang, J., Knudby, A., Bulmer, C.E., Schmidt, M.G., 2016. An overview
the DL approach shows higher accuracy and more reasonable classifying
and comparison of machine-learning techniques for classification purposes in digital
results. In addition, the indication of landform development stage from soil mapping. Geoderma 265, 62–77.
the classification results of the DL approach has also been proven in this Hiller, J., Smith, M., 2008. Residual relief separation: digital elevation model enhancement
for geomorphological mapping. Earth Surf. Process. Landf. 33 (14), 2266–2276.
study. The development stage of loess landforms is reflected by the re-
Iwahashi, J., Watanabe, S., Furuya, T., 2001. Landform analysis of slope movements using
sult of the DL approach. Furthermore, the approach proposed in this DEM in Higashikubiki area, Japan. Comput. Geosci. 27 (7), 851–865.
study can be applied to different landform areas, such as karst landform. Jasiewicz, J., Stepinski, T.F., 2013. Geomorphons—a pattern recognition approach to classi-
fication and mapping of landforms. Geomorphology 182, 147–156.
Kienzle, S., 2004. The effect of DEM raster resolution on first order, second order and com-
Declaration of competing interest pound terrain derivatives. Trans. GIS 8 (1), 83–111.
Latifovic, R., Pouliot, D., Campbell, J., 2018. Assessment of convolution neural networks for
surficial geology mapping in the South Rae geological region, Northwest Territories,
The authors declare no conflict of interest. Canada. Remote Sens. 10 (2), 307.
LeCun, Y., Bengio, Y., Hinton, G., 2015. Deep learning. Nature 521 (7553), 436.
Acknowledgments Li, W., Hsu, C.Y., 2018. Automated terrain feature identification from remote sensing im-
agery: a deep learning approach. Int. J. Geogr. Inf. Sci. 1–24.
Li, W., Zhou, B., Hsu, C.Y., Li, Y., Ren, F., 2017. Recognizing terrain features on terrestrial
This work was supported by the National Natural Science Foundation surface using a deep learning model: an example with crater detection. Proceedings
of China (41930102, 41971333, 41601411, 41671389); and Priority Aca- of the 1st Workshop on Artificial Intelligence and Deep Learning for Geographic
demic Program Development of Jiangsu Higher Education Institutions. Knowledge Discovery, ACM, pp. 33–36.
Liu, K., Ding, H., Tang, G., Na, J., Huang, X., Xue, Z., Li, F., 2016. Detection of catchment-scale
gully-affected areas using unmanned aerial vehicle (UAV) on the Chinese Loess Pla-
References teau. ISPRS Int. J. Geo Inf. 5 (12), 238.
Liu, K., Song, C., Ke, L., Jiang, L., Pan, Y., Ma, R., 2019. Global open-access DEM perfor-
Anders, N.S., Seijmonsbergen, A.C., Bouten, W., 2011. Segmentation optimization and mances in Earth’s most rugged region High Mountain Asia: a multi-level assessment.
stratified object-based analysis for semi-automated geomorphological mapping. Re- Geomorphology 338, 16–26.
mote Sens. Environ. 115 (12), 2976–2985. Long, M., Cao, Y., Wang, J., Jordan, M.I., 2015. Learning transferable features with deep ad-
Belgiu, M., Drăguţ, L., 2016. Random forest in remote sensing: a review of applications and aptation networks. arXiv preprint arXiv 1502.02791.
future directions. ISPRS J. Photogramm. Remote Sens. 114, 24–31. Loughrey, J., Cunningham, P., 2005. Using early-stopping to avoid overfitting in wrapper-
Berral, J.L., Gavalda, R., Torres, J., 2013. Power-aware multi-data center management using based feature subset selection employing stochastic search. 10th UK Workshop on
machine learning. 42nd International Conference on Parallel Processing, IEEE, Case-Based Reasoning, pp. 3–10.
pp. 858–867. Ma, L., Liu, Y., Zhang, X., Ye, Y., Yin, G., Johnson, B.A., 2019. Deep learning in remote sens-
Blaschke, T., 2010. Object based image analysis for remote sensing. ISPRS Journal of Pho- ing applications: a meta-analysis and review. ISPRS J. Photogramm. Remote Sens.
togrammetry and Remote Sensing 65 (1), 2–16. 152, 166–177.
Bocco, G., Mendoza, M., Velázquez, A., 2001. Remote sensing and GIS-based regional geo- Marmanis, D., Datcu, M., Esch, T., Stilla, U., 2015. Deep learning earth observation classification
morphological mapping—a tool for land use planning in developing countries. Geo- using ImageNet pretrained networks. IEEE Geosci. Remote Sens. Lett. 13 (1), 105–109.
morphology 39 (3–4), 211–219. Meng, X., Xiong, L., Yang, X., Yang, B., Tang, G., 2018. A terrain openness index for the ex-
Breiman, L., 2001. Random forests. Mach. Learn. 45 (1), 5–32. traction of karst Fenglin and Fengcong landform units from DEMs. J. Mt. Sci. 15 (4),
Bu, S., Cheng, S., Liu, Z., Han, J., 2014. Multimodal feature fusion for 3D shape recognition 752–764.
and retrieval. IEEE MultiMedia 21 (4), 38–46. Metelka, V., Baratoux, L., Jessell, M.W., Barth, A., Ježek, J., Naba, S., 2018. Automated rego-
Cao, M., Tang, G., Zhang, F., Yang, J., 2013. A cellular automata model for simulating the lith landform mapping using airborne geophysics and remote sensing data, Burkina
evolution of positive–negative terrains in a small loess watershed. Int. J. Geogr. Inf. Faso, West Africa. Remote Sens. Environ. 204, 964–978.
Sci. 27 (7), 1349–1363. Minar, J., Evans, I.S., 2008. Elementary forms for land surface segmentation: the theoreti-
Chang, K.T., Tsai, B.W., 1991. The effect of DEM resolution on slope and aspect mapping. cal basis of terrain analysis and geomorphological mapping. Geomorphology 95 (3–
Cartography and geographic information systems 18 (1), 69–77. 4), 236–259.
Dai, W., Na, J., Huang, N., Hu, G., Yang, X., Tang, G., Li, F., 2019. Integrated edge detection Murray, A.B., Lazarus, E., Ashton, A., Baas, A., Coco, G., Coulthard, T., Paola, C., 2009. Geo-
and terrain analysis for agricultural terrace delineation from remote sensing images. morphology, complexity, and the emerging science of the Earth's surface. Geomor-
Int. J. Geogr. Inf. Sci. 1–20. phology 103 (3), 496–505.
Davis, J., Goadrich, M., 2006. The relationship between Precision-Recall and ROC curves. Norman, B., Pedoia, V., Majumdar, S., 2018. Use of 2D U-Net convolutional neural net-
Proceedings of the 23rd International Conference on Machine Learning. ACM, works for automated cartilage and meniscus segmentation of knee MR imaging
pp. 233–240. data to determine relaxometry and morphometry. Radiology 288 (1), 177–185.
Dian, Z., 1996. A morphological analysis of Tibetan limestone pinnacles: are they rem- Palafox, L.F., Hamilton, C.W., Scheidt, S.P., Alvarez, A.M., 2017. Automated detection of
nants of tropical karst towers and cones? Geomorphology 15 (1), 79–91. geological landforms on Mars using Convolutional Neural Networks. Comput. Geosci.
Diani, K., Tabyaoui, H., Kacimi, I., El Hammichi, F., Nakhcha, C., 2017. Stream network Model- 101, 48–56.
ling from aster GDEM using ArcHydro GIS: application to the upper Moulouya River Phillips, J.D., 2003. Sources of nonlinearity and complexity in geomorphic systems. Prog.
Basin (Eastern, Morocco). Journal of Geoscience and Environment Protection 5(05), 1. Phys. Geogr. 27 (1), 1–23.
14 S. Li et al. / Geomorphology 354 (2020) 107045

Romstad, B., Etzelmüller, B., 2012. Mean-curvature watersheds: a simple method for seg- Wang, D., Laffan, S.W., Liu, Y., Wu, L., 2010. Morphometric characterisation of landform
mentation of a digital elevation model into terrain units. Geomorphology 139, from DEMs. Int. J. Geogr. Inf. Sci. 24 (2), 305–326.
293–302. Weiss, A., 2001. Topographic position and landforms analysis. Poster Presentation, ESRI
Ronneberger, O., Fischer, P., Brox, T., 2015. U-net: convolutional networks for biomedical User Conference, San Diego, CA Vol. 200.
image segmentation. International Conference on Medical Image Computing and Wood, J., 1996. The Geomorphological Characterization of Digital Elevation Models. Uni-
Computer-Assisted Intervention. Springer, Cham, pp. 234–241. versity of Leicester, Doctoral dissertation.
Saha, K., Wells, N.A., Munro-Stasiuk, M., 2011. An object-oriented approach to auto- Xiong, L., Tang, G., Li, F., Yuan, B., Lu, Z., 2014a. Modeling the evolution of loess-covered
mated landform mapping: a case study of drumlins. Comput. Geosci. 37 (9), landforms in the Loess Plateau of China using a DEM of underground bedrock surface.
1324–1336. Geomorphology 209, 18–26.
Satgé, F., Bonnet, M.P., Timouk, F., Calmant, S., Pillco, R., Molina, J., Garnier, J., 2015. Accu- Xiong, L., Tang, G., Yan, S., Zhu, S., Sun, Y., 2014b. Landform-oriented flow-routing algo-
racy assessment of SRTM v4 and ASTER GDEM v2 over the Altiplano watershed using rithm for the dual-structure loess terrain based on digital elevation models. Hydrol.
ICESat/GLAS data. Int. J. Remote Sens. 36 (2), 465–488. Process. 28 (4), 1756–1766.
Shin, H.C., Roth, H.R., Gao, M., Lu, L., Xu, Z., Nogues, I., Summers, R.M., 2016. Deep Xiong, L., Tang, G., Yuan, B., Lu, Z., Li, F., Zhang, L., 2014c. Geomorphological inheritance for
convolutional neural networks for computer-aided detection: CNN architectures, loess landform evolution in a severe soil erosion region of Loess Plateau of China
dataset characteristics and transfer learning. IEEE Trans. Med. Imaging 35 (5), based on digital elevation models. Science China Earth Sciences 57 (8), 1944–1952.
1285–1298. Xiong, L., Zhu, A., Zhang, L., Tang, G., 2018. Drainage basin object-based method for re-
Smart, P., Waltham, T., Yang, M., Zhang, Y., 1986. Karst geomorphology of western Gui- gional-scale landform classification: a case study of loess area in China. Phys. Geogr.
zhou, China. Cave Science 13 (3). 39 (6), 523–541.
Song, X., Tang, G., Li, F., Jiang, L., Zhou, Y., Qian, K., 2013. Extraction of loess shoulder-line Xu, Z., Lin, Z., Zhang, M., 2007. Loess in China and loess landslides. Chin. J. Rock Mech. Eng.
based on the parallel GVF snake model in the loess hilly area of China. Comput. 26 (7), 1297–1312.
Geosci. 52, 11–20. Yang, X., Tang, G., Meng, X., Xiong, L., 2019. Classification of Karst Fenglin and Fengcong
Song, X.D., Brus, D.J., Liu, F., Li, D.C., Zhao, Y.G., Yang, J.L., Zhang, G.L., 2016. Mapping soil landform units based on spatial relations of terrain feature points from DEMs. Remote
organic carbon content by geographically weighted regression: a case study in the Sens. 11 (16), 1950.
Heihe River Basin, China. Geoderma 261, 11–22. Zhang, W., Witharana, C., Liljedahl, A., Kanevskiy, M., 2018. Deep convolutional neural
Sun, J., 2002. Provenance of loess material and formation of loess deposits on the Chinese networks for automated characterization of arctic ice-wedge polygons in very high
Loess Plateau. Earth Planet. Sci. Lett. 203 (3–4), 845–859. spatial resolution aerial imagery. Remote Sens. 10 (9), 1487.
Tang, T., Day, M.J., 2000. Field survey and analysis of hillslopes on tower karst in Guilin, Zhang, Q., Yang, L., Chen, Z., Li, P., 2018. A survey on deep learning for big data. Informa-
southern China. Earth Surf. Process. Landf. 25 (11), 1221–1235. tion Fusion 42, 146–157.
Verhagen, P., Drăguţ, L., 2012. Object-based landform delineation and classification from Zhao, W., Xiong, L., Ding, H., Tang, G., 2017. Automatic recognition of loess landforms
DEMs for archaeological predictive mapping. J. Archaeol. Sci. 39 (3), 698–703. using Random Forest method. J. Mt. Sci. 14 (5), 885–897.
Waltham, T., 2008. Fengcong, fenglin, cone karst and tower karst. Cave and karst science Zhou, Y., Tang, G., Yang, X., Xiao, C., Zhang, Y., Luo, M., 2010. Positive and negative terrains
35 (3), 77–88. on northern Shaanxi Loess Plateau. J. Geogr. Sci. 20 (1), 64–76.
Wan, S., Lei, T.C., Chou, T.Y., 2012. A landslide expert system: image classification through Zhu, X., 2009. Discussions on fenglin karst in China. Carsologica Sinica 28, 155–168.
integration of data mining approaches for multi-category analysis. Int. J. Geogr. Inf.
Sci. 26 (4), 747–770.

You might also like