Ecological Indicators 125 (2021) 107540

Contents lists available at ScienceDirect

Ecological Indicators
journal homepage: www.elsevier.com/locate/ecolind

The multi-objective spatial optimization of urban land use based on

low-carbon city planning
Gengzhe Wang *, Qi Han , Bauke de vries
Department of Built Environment, Eindhoven University of Technology, Eindhoven, North Brabant 5612AZ, the Netherlands

A R T I C L E I N F O A B S T R A C T

Keywords: As the spatial carrier of carbon emission from land ecosystems and human activities, land use plays an important
Land use classification role in representing the spatial distribution of carbon emissions and carbon sinks. The intension of CO2 emission
Land use optimization is closely attached to the fundamental unit of urban form—land use patterns. The exploration of the low-carbon
Spatial attribute
oriented spatial optimization provides new scope for solving the high carbon emission in the urban area through
Regression tree
Genetic algorithm
spatial planning. In this paper, Eindhoven in the Netherlands was selected as the case study to implement spatial
optimization. A set of parameters containing the spatial attributes of buildings and vegetation were introduced to
classify the land use patterns into six categories through a combined random forest algorithm and regression tree
approach. The results show the geographic features of vegetation are the crucial factors for the carbon emission.
The multi-objective spatial optimization model integrated carbon emission, population, and constraint condi­
tions. It is solved with non-dominated sorting genetic algorithm-II (NSGA-II), in which each gene represents a
specific type of land use category. The optimal solutions were incorporated with a regression model to analyze
the impact of the variation in each land use category. Three categories were proved to be more influential on
carbon emission performance. The optimized land use structure shows the potential to reduce the carbon
emission and offer valuable consults to low carbon land use plan.

1. Introduction
In recent decades, global warming has become the most critical
environmental issue which threatening human society and ecosystems
worldwide (Cramer et al., 2018). According to the report from Inter­
governmental Panel on Climate Change (IPCC), it is believed that the
primary cause of global warming is highly related to the CO2 emission
produced by the energy consumption from fossil fuels (Intergovern­
mental Panel on Climate Change, 2014). Urban areas concentrate most
of the population and human activities. They are the major sources of
CO2 emissions and playing an increasingly important role in global
warming (Pachauri, 2014). Urban CO2 reduction has become the most
urgent task for municipalities internationally to address the adverse
impacts of global warming.
The contribution of carbon emission is highly affected by carbon
sources and carbon sinks. Carbon sources in an urban area mainly refer
to the anthropogenic factors that releasing CO2 to the atmosphere such
as electricity supply, heating, and transportation. In contrast, carbon
sinks are the natural factors that can absorb CO2 such as vegetation and
soil. As the spatial carrier of carbon sources and carbon sinks, land use
represents the carbon emission intensity on a spatial scale. Since the
structure and distribution of land use can highly define the spatial
configuration of the built environment and related human activities, the
spatial layout of urban land use is highlighted as one of the key factors
which have a great impact on carbon emission (Lu and Guldmann,
2012). There are many studies have demonstrated that a reasonable land
use plan could make a significant contribution to mitigating carbon
emission (Gao et al., 2014; Zhang et al., 2018; Shu and Xiong, 2019;
Penazzi et al., 2019).
With the support of geographic information system (GIS) technology
and numerous spatial optimization approaches, many spatial decision
support systems (SDSS) were developed to provide a guide for future
land use planning. Some researchers used a mixed approach to define
the urban energy-saving scenarios integrating GIS Urban Energy Map­
ping, Stakeholders Analysis, and Multi-Criteria Decision Analysis
(Moghadam et al., 2017). Others established a web-based SDSS model to
test existing crop change policies or produce effective crop change de­
cisions using tradeoff analysis (Tayyebi et al., 2016). Some scholars have

* Corresponding author.
E-mail address: G.wang@tue.nl (G. Wang).

https://doi.org/10.1016/j.ecolind.2021.107540
Received 29 March 2020; Received in revised form 23 September 2020; Accepted 9 February 2021
Available online 5 March 2021
1470-160X/© 2021 The Author(s). Published by Elsevier Ltd. This is an open access article under the CC BY license (http://creativecommons.org/licenses/by/4.0/).
G. Wang et al.
proposed a spatial optimization framework which optimizes the location
of future residential development against several sustainability objec­
tives, the resulting spatial strategies have proved to outperform signif­
icantly than the current local authority strategy (Caparros-Midwood
et al., 2015). Most of decision support tools mainly rely on the raster
data to structure the map of case study, largely due to its simplicity on
data collection and spatial attributes assessment. As for vector data, the
process of spatial query and topological evaluation is relatively
complicated. However, the use of vector data can better fit with the
realistic situation, and it is more suitable to reflect planning practice
(Yao et al., 2019).
Vector data usually deploy on land use patterns to describe the
geometric information and landscape configuration. As the basic unit for
spatial mapping, land use patterns can be illustrated at a different spatial
level from a cadastral plot to an entire neighborhood. Many studies have
proven street block level provides the most sufficiently fundamental
detail on describing the urban structure, and it is an ideal scale to define
units and extract correspondent spatial attributes (Bochow et al., 2010).
The spatial attributes of land use patterns can be measured through a
series of metrics that is quantifying the characteristics such as shape,
composition, and configuration. Remote sensing technology plays a
crucial role in collecting relevant spatial data. With the support of sat­
ellites and airborne sensors that can acquire multi-spectral information
with high resolution, it is possible to describe the urban area in a
detailed characterization. LiDAR system represents the vertical structure
of terrain and above-ground elements with the point cloud, and it can
capture the accurate three-dimensional information at a fine-scale (Yan
et al., 2015). Usually, the basic above-ground elements will be identified
as buildings or vegetations according to their shape and spectral char­
acteristics. Then, the remote sensing data can be further used on
defining the spatial attributes that describing geometry or area coverage
(Hermosilla et al., 2014).
Due to the high complexity and heterogeneous in urban land cover,
the description of spatial characteristics always relying on a set of high
dimensional geometrics datasets. To obtain the spatial features which
have a significant impact on the target variable, it is necessary to reduce
the correlation among the high-dimensional dataset (Li et al., 2011).
Machine learning algorithms are commonly associated with feature se­
lection methods to reduce the redundancy caused by autocorrelation
between parameters (Prendinger et al., 2013). As one of the most
common feature selection strategies, random forest (RF) algorithm was
widely applied to improve the accuracy of selected results. As an
ensemble learning method for classification and regression, RF has high
performance on accuracy and robustness, and it corrects the habit of
overfitting to their training set. The importance of different variables
can be clearly illustrated by the calculated feature scores (Lebourgeois
et al., 2017).
In order to investigate the geographical characteristics among land
use patterns, a variety of classification methods have been widely
applied for determining land use categories from the remote sensing
data, including maximum likelihood classification (MLC), multinomial
logistic regression (MLR), decision tree classification (DTC), classifica­
tion and regression tree (CART), etc. Among these methods, the most
common approach is MLC. It has difficulties in extracting objects with
multiple variables, and it leads to the relatively low accuracy on clas­
sification (Xie et al., 2019). MLR is widely applied to linear classification
issues. However, it mainly focuses on the dynamic process, such as land
use change and urban expansion. In most cases, MLR aims to predict the
probabilities of possible outcomes of the target variable, and the out­
comes are not usually fit with the requirement of the static classification
research (Leng et al., 2017; Lyu and Zhang, 2019). Since Support Vector
Machine (SVM) can get the optimal separating hyperplane during the
training process, it usually has better performance on accuracy
compared with other classification algorithms. Although SVM can deal
with the high dimensional and nonlinear dataset, its sensitivity to the
initial parameters may cause uncertainty during the training process
2
Ecological Indicators 125 (2021) 107540
(Cao et al., 2019). DTC can provide an intuitive structure of the classi­
fication, and it has no requirements for the assumptions of the input data
distribution. Therefore, it can easily integrate with other methods to get
wider applicability (Hua et al., 2017). However, DTC is usually applied
to deal with the data which as a discrete structure. It has great diffi­
culties in solving the vector map with fragmented landscapes and
complex terrain (Yang et al., 2017). Comparing with DTC, CART inherits
all advantages of general decision trees, it not only carry out the capacity
for classification but also can be applied to predict the regression of
continuous variables (Olfaz et al., 2019). It has high operational effi­
ciency in estimating the non-linear relationship between independent
and dependent variables. The regression tree model is considered as a
higher accuracy and predictability approach comparing with other
linear regression models (Wang et al., 2018a; 2018b). Due to the ad­
vantages of CART on the land use classification, it has been widely used
on the fine-scale urban land use studies. Many researchers implemented
CART to estimate the impervious surface in the urban area (Wang et al.,
2018a; 2018b). Others used it to classify the vegetation types in the
urban area based on the land cover map (Verhulp and Van Niekerk,
2017).
To optimize the spatial structure of urban land use, many advanced
algorithms are conducted in the spatial optimization model to find
alternative solutions with high-performance. Following the significant
progress in computational capacity, the majority of the optimization
model has been shifted from basic linear programming to the usage of
heuristic algorithms (Cao et al., 2012). The heuristic algorithms such as
simulated annealing (SA), particle swarm optimization (PSO), and ge­
netic algorithms (GA) allow a trade-off between solution quality and
computation burden, hence they can be used to find near-optimal so­
lutions. Current research regarding spatial optimization usually needs to
consider multiple objectives such as maximizing ecological benefit,
minimizing financial cost, and optimizing the number of urban facilities,
etc. (Jing et al., 2013; Sharmin et al., 2019). These objectives are often
conflicting with each other, the evaluating of trade-offs is required be­
tween the conflicting results to explore the more compromising solu­
tions. Non-Dominated sorting genetic algorithm II (NSGA-II) is widely
applied to generate a series of Pareto optimal solutions which can realize
a reasonable balance among different optimizing objectives. As one of
the variations of multi-objective GA, NSGA-II is based on the mechanics
of fast sorting and elitist strategy. It has great robustness for identifying
the global optimum in a large and complex search space (Corbera et al.,
2016).
In this paper, we established a multi-objective spatial optimization
model based on the land use classification results. It intends to provide a
land use zoning plan for the municipality under the background of low-
carbon city development. Due to the easy accessibility of open data
sources and carbon emission estimation results from a previous study
(Wang et al., 2019), Eindhoven in the Netherlands was selected as the
case study to present the effects of land use classification and spatial
optimization. According to the spatial attributes extracting from remote
sensing data, the land use patterns in the case study can be classified into
different categories with carbon emission value as the target variable.
Furthermore, the spatial optimization will be demonstrated for the land
use allocation problems, in which the carbon emission and population
were chosen as objectives. With the carbon emission and population
boundaries provided by Bureau for Economic Policy Analysis (CPB), the
Pareto fronts of spatial optimization model will be given to offering
alternative land use plans under different scenarios in the future.
2. Materials and methods
2.1. Study area and data sources
Eindhoven is the center of the high-tech industry in the Netherlands.
It locates in the province of North Brabant—the southeast region of the
Netherlands. As the 5th largest city in the Netherlands, Eindhoven is
G. Wang et al.
composed of seven administrative districts with an area of 88.87 km2
and the population around 224740. Due to the successful consequence
of the urban transformation, the economic growth in Eindhoven has
surpassed G4 cities (the four largest Dutch municipalities: Amsterdam,
Rotterdam, The Hague, and Utrecht) in recent years. It has attracted
many migrant workers, experts, and international students from else­
where in the world to inhabit, and the population has increased by 9%
since 2008. However, not enough housing units were built during the
last ten years to accommodate the increasing new inhabitants, and this
has inevitably led to a housing crisis. Meanwhile, as the most important
industrial city in the Netherlands, Eindhoven is also facing the severe
challenge of achieving sustainable development. The municipal council
of Eindhoven has set an objective to reduce 55% of carbon emission in
2030 relative to 1990. The balancing of expanding housing and carbon
reduction has become the most urgent topic for the urban planner of the
municipality. The location and summary information of Eindhoven is
presented in Fig. 1.
The following database was carried out on the case study for the
preprocessing and further analyzing: (1) Land use function map for 2012
is provided by the Central Agency of Statistics (CBS) database, it dem­
onstrates the spatial distribution of different land use function in Eind­
hoven. The original land use function was reclassified into nine
categories: agriculture, company premises, forest, others, recreation,
residential area, retail trade, social-cultural area, and transport. (Fig. 2a)
(2) Digital surface model (DSM) is acquired from the Current Height File
in the Netherlands (AHN2) in 2014. It represents the altitude informa­
tion of Eindhoven on 0.5*0.5 m resolution. (Fig. 2b) (3) The land cover
map is obtained from BAG (Basic registration Addresses and Buildings)
and BGT (Basic registration Large-scale Topography) database. BAG
includes the building information ranged from height and area to spatial
layout. BGT database on the other hand, mainly contains vegetation
information such as the classification of green space and location of
trees. Original vegetation types were all unified as greenspace and tree
to illustrate the spatial layout of vegetation in Eindhoven. (Fig. 2c) (4)
The distribution of carbon emission in Eindhoven is estimated by a
framework that estimating the carbon emission from building, transport,
vegetation, and resident sectors. The details of the estimation process
Fig. 1. Location and summary information of study area.
3
Ecological Indicators 125 (2021) 107540
could be seen in the previous study (Wang et al., 2019). The estimation
results were integrated and presented through ArcGIS. (Fig. 2d)
According to the road network, Eindhoven can be divided into 6754
irregular patterns. Their average area is about 1.213 ha. Among 6754
land use patterns, 3679 of them were identified as residential area, while
799 were recognized as forest or recreation. Since the contradiction
between house shortage and carbon reduction is mainly depends on the
reallocation of residential areas and green spaces. This study only put
residential area, forest, and recreation land use categories into consid­
eration. Other land use categories are assumed to be unchangeable. The
research purpose of this study can be achieved through three steps:
quantification of spatial attributes, classification of urban land use, and
multi-objective spatial optimization. The general flowchart is presented
as Fig. 3. It shows the data sources, corresponding key elements, and
methodologies we applied during the process.
2.2. Quantification of spatial attributes
Buildings and vegetations are expected as the key elements in the
study of low carbon city planning; their geographic structure and spatial
layout have a significant impact on the carbon emission in the urban
area. In order to reflect the geometrics of building and vegetation on a
fine scale, the quantification of spatial attributes is conducted on each
individual land use pattern (Lai and Kontokosta, 2018). Numerous
spatial attributes are extracted based on building and vegetation covers,
these covers were obtained from BAG and BGT database which consist of
the distribution of building patches and vegetation patches respectively.
Combining with the high-spatial-resolution DSM image obtained using
LiDAR technology, the pixels on the DSM image were classified into
building class and vegetation class based on their coverage.
In general, the spatial attributes used to describe the land use pat­
terns are defined as three groups: (1) descriptors of the landscape and
shape properties of building patches; (2) geometric and volumetric at­
tributes regarding of building patches; (3) geometric and volumetric
attributes regarding of vegetation patches. The complexity of building
patches contour is mainly quantified through shape metrics (shape
index, fractal dimension, and compactness). Shape index (SI) measures
G. Wang et al. Ecological Indicators 125 (2021) 107540

Fig. 2. Data sources map in Eindhoven.

Fig. 3. Flowchart of major steps in research.

the patch shape complexity comparing with the same size patch with a in scale. Compactness (C) measures the degree of a patch which close to
standard shape. Fractal dimension (FD) offers a numerical description a circle. The geometric and volumetric attributes of buildings are
for patch shape by computing fractal complexity as a ratio of the change described in terms of number, height, area, and volume. The number of

4
G. Wang et al. Ecological Indicators 125 (2021) 107540

buildings (N) in each land use pattern can be directly accounted for from show the importance of differences among selected spatial attributes.
the BAG database. The building height is characterized by maximum For the classification of urban land use driven by the carbon emis­
(maxh), mean (h), and standard deviation (SDh) value which computed sion, the regression tree algorithm, based on the classification and
from the BAG and DSM data. Building coverage ratio (BCR) is used to regression tree (CART) theory was applied during the process (Breiman
describe the geometric features of buildings, and it computed the pro­ et al., 2017). It provides a rule-based model to predict a continuous
portion of the area covered by building class, the result was expressed as response variable from various explanatory variables. In this study,
a percentage. The volumetric attributes of buildings were derived by CART is used to predict the values of carbon emission variables in
combining the height information with geometric features. The sum different spatial attributes contexts. CART is achieved by an iterative
volume of buildings (VolumeB ) presented the volume of all building process known as binary recursive partitioning. The partition is based on
patches in the individual land use pattern while the mean volume of the principle of maximizing the measure of node homogeneity, the
buildings (mVolumeB ) expressed the average value. Analogously, vege­ explanatory variable that can maximize the reduction in the residual
tation coverage ratio (VCR) computed the proportion of the area sum of squares or deviance is selected to split available samples. The
covered by vegetation class. The sum volume of vegetations (VolumeV ) split result is presented as a binary decision tree. This process will keep
can be derived by combining the height information with vegetation repeating on the descendant nodes until all samples fall into their cor­
coverage area. Since the shape and area of each land use pattern is responded nodes. The regression tree model starts with a full dataset,
different from each other, many non-proportional spatial attributes proceeds through a series of nodes and ends up with a set of terminal
(perimeter of buildings, buildings volume, mean buildings volume, nodes that contain predicted values. A calibration dataset is used to
buildings number, building coverage area, vegetations volume, vegeta­ assess the relationship between the response variable and explanatory
tion coverage area) were normalized to fit the requirement of vector variables. The regression tree model could be too elaborate and over­
study. The spatial attributes and their corresponding equations can be fitting with many noisy features, which leads to the arbitrary and in­
presented in Table 1. accuracy for generalization and interpretation. To avoid overfitting,
pruning is normally conducted to simplify the regression tree by
2.3. Classification of urban land use removing excessive nodes. It can improve the predictive accuracy of the
regression tree model through cross-validation. First, a perfectly fit tree
To select the main impact factors of carbon emission, it is necessary is grown based on one portion of the calibration dataset. Then, the
to measure the importance of each spatial attribute on carbon emission. remaining calibration data is passed down through the tree. The tree size
As a wrapper feature selection algorithm, random forest use a mass of is optimized by minimizing the cross-validated error, the pruning pro­
trees to ensemble the high dimensional data set (Ma et al., 2017). It aims cess will end when the deviance has reached a minimum, or the corre­
to minimize the correlation among the ensembled trees. The conditional sponding decrease in deviance can justify no more additional nodes. The
variable importance evaluation of the spatial attributes can be calcu­ establishment and analysis of CART can be performed through the
lated by “Boruta” package in R. It creates shuffled copies of all features “rpart” package in R. The analysis results of CART not only illustrate the
to add randomness in the given dataset, the copies of the features can carbon emission of the land use pattern under different contexts.
also be named as shadow features (Kursa and Rudnicki, 2015). Then, a Objectively, all land use patterns are also classified into different cate­
classifier is trained with an extended dataset to measures the importance gories based on their spatial attributes.
of each feature. It will check whether the importance of real features is
higher than the best of its shadow feature in each iteration. The features
which are deemed unimportant will be rejected. The algorithm will keep 2.4. Multi-objective spatial optimization
running until all features are confirmed or rejected. The filtered results
The spatial optimization model is established based on the require­
ment for low carbon emission and maximum population capacity. The
Table 1
objectives of carbon emission and population capacity are all integrated
Spatial attributes and equations extracted from land use patterns.
into the multi-objective optimization model with other constraint con­
Attribute (Unit) Equation
ditions. The Pareto fronts consist of a series of optimized solutions that
Perimeter of Buildings (m) PerimeterB none of the objective functions can be further optimized without
Normalized Perimeter of Buildings (m)
nPerimeterB =
PerimeterB degrading other objective values. It was used to describe alternative land
Normalized Buildings Volume (m3/m2) ∑b Area use plans in the case study, where the optimized Pareto fronts provide
hi r2
nVolumeB = i
Area the trade-offs between carbon emission and population capacity. The
Normalized Mean Buildings Volume (m3/m2)
nmVolumeB =
sVolumeB multi-objective spatial optimization of land use is based on two aspects
N∙Area which are land use carbon emission C and land use population capacity
Normalized Buildings Number (/m2) N
nN =
Area
∑b 2
P. It was designed to minimize the carbon emission under maximum
Building Coverage Ratio (%)
BCR = i ∙100
r population capacity. The objective functions and constrain condition
Area
Max Building Height (m) maxh = max(hi ) can be illustrated as Eq. (1)
∑n
Mean Building Height (m) hi ⎧
h = i ⎪
N√̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅̅ ⎪
⎪
⎪
SD of Building Height (m) ∑n 2 ⎪
⎪
i (hi − h) ⎪ minC = ∑n X ∙K
⎪
SDh = ⎪
n− 1 ⎪
⎪
⎪ i=1
i i
Compactness of Buildings 4∙π∙AreaB ⎪
⎪ ∑n
C = ⎪
⎪
PerimeterB ⎪ maxP =
⎪ Xi ∙Si
Shape Index of Buildings PerimeterB ⎪
⎨ i=1
SI = √̅̅̅̅̅̅̅̅̅̅̅̅̅ ∑n
4∙ AreaB s.t. i=1 Xi = T (1)
Fractal Dimension of Buildings log(PerimeterB /4) ⎪
FD = ⎪
⎪
⎪ ∑n
log(Area
∑v B 2) ⎪
⎪ Xi ∙Ki < Cbase
Normalized Vegetations Volume (m3/m2) hi r ⎪
⎪ i=1
nVolumeV = i ⎪
⎪ ∑n
∑v 2Area ⎪
⎪
⎪ Xi ∙Si ≥ Pbase
Vegetation Coverage Ratio (%) hi r ⎪
⎪
VCR = i
i=1
∙100 ⎪
⎪
Area ⎪
⎪
⎩
Note: b: the number of the pixel covered by building class; v: the number of the
pixel covered by vegetation class; hi : height obtaining from AHN2 for the pixel i;
r: spatial resolution; Area: area of land use pattern. where Xi is the area of land use category i; Ki is the carbon emission

5
G. Wang et al. Ecological Indicators 125 (2021) 107540

coefficient of the relative land use category i; Si is the population ca­ were set as dependent variables. The structure of the resultant MLR
pacity coefficient of the relative land use category i; T is the area of all models of the carbon emission reduction and carbon emission per person
land use patterns in the case study; Cbase is the carbon emission value in reduction can be illustrated as the Eq. (2) and Eq. (3).
the baseline year; Pbase is the population capacity in the baseline year; n
CER = a1 Xc1 + a2 Xc2 + ⋯ai Xci + β + ε (2)
is the number of classified land use categories. Since spatial optimization
is only conducted in the current land use, urban expansion and urban
CERper = a1 Xc1 + a2 Xc2 + ⋯ai Xci + β + ε (3)
shrinkage are not be considered in this case. The total area of different
land use categories remains unchanged. The total carbon emission where CER, CERper are the amounts of carbon emission reduction and
amount of optimized land use is supposed to be lower than the value in carbon emission reduction per person in the case study, respectively;
the baseline year. It is also required that the population capacity won’t Xc1 , Xc2 , …, Xci are the variation of land use area in land use category 1,
reduce after the optimization. 2,…, i; a1 , a2 , …, ai are the estimated coefficients of explanatory vari­
The algorithm applied in the optimization model must be able to ables; β is the model intercept; ε is the model residual.
identify which sets of control variables result in better objectives out­
comes and exploring if better outcomes could be achieved. Non- 3. Results
dominated sorting genetic algorithm-II (NSGA-II) is selected consid­
ering the complexities caused by two competing objective functions. 3.1. Spatial attributes selection
NSGA-II is an improved version of NSGA which worked as an extension
of the genetic algorithm (GA) for multiple objective function optimiza­ There are a total of 13 spatial attributes are extracted from land use
tion. It involves the elitist approach which consists of non-dominated patterns in Eindhoven, and feature selection was conducted on the
ranking and crowding distance sorting methods. The chromosomes in initial spatial attribute dataset through the random forest algorithm. The
the GA model are encoded as land use categories, and every chromo­ selection results are presented as Fig. 4, and it proves that all selected
some represents a scheme of the combination of different land use cat­ spatial attributes have a significant impact on carbon emission, all of
egories. The value of each gene in the chromosome refers to the area of them need to remain in further analysis. Among these spatial attributes,
the specific land use category in that land use combination scheme. In normalized vegetation volume has the highest mean importance with
order to sort a population with N solutions according to the level of non- the value of 44.41. It is followed by vegetation coverage ratio with the
domination, each solution is compared with every other to find if it is mean importance of around 22.97. Other spatial attributes such as
dominated. This requires O(mN) comparisons, where m is the number of compactness and normalized building number also have relatively high
objectives. This process will find out the members of the first non- importance on the carbon emission, the values of compactness and
dominated class for all populations, and the total complexity is O normalized buildings number are 19.08 and 15.84, respectively.
(mN2). In order to find the individuals in the next front, the solutions of
the first front are stored to avoid duplicate comparisons and the pro­
3.2. Land use category classification
cedure above will be repeated (Chan et al., 2016). This procedure will be
repeated until all subsequent fronts are found. With the help of a
The simplest and most accurate regression tree is generated from the
bookkeeping strategy, the computation of all fronts can be decreased to
initial land use pattern dataset using their selected spatial attributes to
O(mN2) at most (Bao et al., 2017). In order to preserve the distributivity
predict the carbon emission in different contexts. The growing and
and diversity of solutions, crowding distance is introduced to estimate
pruning process of the regression tree can be achieved through the rpart
the density of solutions in the population. It calculates the average
() function in R. As it is presented in Fig. 5, the generated regression tree
distance between two points alongside the objective axes. Following this
has three layers and six terminal nodes. It has classified all land use
process, a loop is generated to respond to the computation of all
patterns into six different land use categories, and they were named
crowding distances for each solution (Lamamra et al., 2014).
from class A to class F. The value in each terminal node represents the
predicted carbon emission of the corresponding land use category.
2.5. Regression modeling
Further, the classification results were input into the initial dataset of
land use patterns. Combining with the data of per area population in
In order to further explore the impact of land use variation on carbon
individual land use patterns, the average population capacity in each
emission, regression modeling is needed to formalize the relationship
land use category could be calculated. Table 2 has shown the land use
between classified land use categories and carbon emission, as well as
classification results along with their carbon emission and population
the degree to which this is modified by the land use area change. In this
study, a stepwise multiple linear regression (MLR) was developed to
explain the main impact factor of land use categories for low-carbon city
planning. The stepwise regression has been widely used for screening
predictor variables during the multivariate analysis (Khaledian et al.,
2017). This method can automatically choose the independent variables
based on the prespecified criterion in the fitting regression model, a
variable is considered for addition to or subtraction from the explana­
tory variables set in each step. It aims to maximize the prediction ca­
pacity of the regression model with minimum explanatory variables (Al
Rifat and Liu, 2019). For each MLR model, the adjusted R2 values were
tested to assess the performance of prediction. In each regression model,
multicollinearity between the independent variables was also tested
using Tolerance and Variance Inflation Factors (VIF) in the correlation
matrix. Generally, the independent variables were considered a low
degree of multicollinearity when the tolerance value is close to 1, while
the VIF value is between 1 and 4 (Hair et al., 2010). The regression
analysis was achieved through SPSS, the variation of the land use area
within different land use categories were set as independent variables,
while the reduction of carbon emission and carbon emission per person Fig. 4. Features selection result of extracted spatial attributes.

6
G. Wang et al. Ecological Indicators 125 (2021) 107540

Fig. 5. Land use classification result achieved by regression tree.

Table 2
The statistics of carbon emission and population capacity of land use categories.
Land use categories Carbon emission (kg/m2) Population capacity (persons/ha) No.

Mean Max Min Median Mean Max Min Median

Class A 5.32 52.99 − 0.35 1.92 13.96 122.10 0.00 3.26 278
Class B 11.66 55.75 − 0.46 9.16 32.35 172.88 0.00 25.11 921
Class C 18.48 93.36 − 0.39 17.46 50.04 475.05 0.00 46.82 1409
Class D 22.95 63.42 2.05 22.48 61.51 157.83 8.34 60.55 660
Class E 29.80 114.40 3.35 28.28 76.94 326.43 16.28 74.30 359
Class F 33.03 95.31 4.34 32.09 86.08 152.13 18.63 85.43 851

capacity values. Among the classified land use categories, for the land
use pattern in which the normalized buildings volume value is lower
than 1.5 m3/m2 were classified into the left section at the initial node, it
contains Class A, Class B, and Class C land use category. 278 patterns
were identified as Class A since their normalized vegetation value is
higher than 4.8 m3/m2. Class A land use category has the lowest mean
carbon emission and population capacity, and the mean values are 5.32
kg/m2 and 13.96 persons/ha, respectively. Class B represents the land
use patterns in which the normalized vegetation value is lower than 4.8
m3/m2 and higher than 1.3 m3/m2, while Class C represents the land use
patterns have the normalized vegetation value lower than 1.3 m3/m2.
Class C takes the largest proportion among all land use categories, the
number is 1409, and the percentage value is 31.46%. It is followed by
Class B, which has 921 patterns in total, and it takes 20.57% among all
land use patterns. For the land use patterns in which the normalized
building’s volume is larger than 1.5 m3/m2, they were classified into the
right section at the initial node, it includes the land use categories such
as Class D, Class E, and Class F. More specifically, 660 patterns were
recognized as Class D, and 359 of them were considered as Class E. The
normalized buildings number value in Both Class D and Class E is higher
than 0.0038/m2. They are further distinguished by the value of vege­
tation coverage ratio. In Class D, the vegetation coverage ratio is higher
than 0.62, while in Class E, their vegetation coverage ratio is lower than
0.62. The rest 851 land use patterns were considered as Class F, the
normalized building’s number in these patterns is higher than 0.0038/
m2. Class F also has the highest mean carbon emission and population
capacity, and their values are 33.03 kg/m2 and 86.08persons/ha,
respectively. The mean values of carbon emission in Class B, Class C,
Class D, and Class E are 11.66 kg/m2, 18.48 kg/m2, 22.95 kg/m2, and
29.80 kg/m2 respectively. While the mean values of population capacity
in Class B, Class C, Class D, and Class E are 32.35persons/ha,
50.04persons/ha, 61.51persons/ha, 76.94persons/ha, respectively.
With the rising values in buildings and vegetation volume from Class A
to Class F, the carbon emission and population capacity are also grad­
ually increasing. In order to reflect the realistic image of classified land
use categories, the typical land use pattern in each land use category was
selected and visualized through ArcScene. Fig. 6 presents the 3D
geographical model of the sample land use pattern from Class A to Class
F. Class A usually represents the greenspace area covered by high
vegetation volume and vegetation coverage ratios, such as park, forest,
and recreation area. Class B and Class C are the land use patterns
distributed with separated buildings. Comparing with Class C, the land
use patterns identified in Class B usually have more vegetation. For the
Class D and Class E, they are the land use patterns which equipped with
higher buildings volume and building density. The vegetation coverage
in Class D is higher than Class E. As the last land use category, Class F
refers to the land use patterns which have the most compact building
distribution and the highest buildings density.
3.3. Multi-objective spatial optimization of land use
According to the land use classification results, six variables were set
in the multi-objective optimization model. They are: Class A(X1 ), Class B
(X2 ), Class C(X3 ), Class D(X4 ), Class E(X5 ), Class F(X6 ). The mean values
of carbon emission and population capacity in each land use category
were assigned on their corresponding coefficient (Ki and Si ) respectively.
In this study, the baseline year is set as 2014. According to the classi­
fication results, the land use area of each land use categories in the
baseline year is 5.36 km2(X1 ), 8.21 km2(X2 ), 6.32 km2(X3 ), 6.96
km2(X4 ), 3.94 km2(X5 ), 6.49 km2(X6 ) respectively, the total area is
37.28 km2. Combining with the mean value of carbon emission and
population capacity in different land use categories, the total carbon
emission amounts in the baseline year is calculated as 732.55kt, while
the population capacity in the residential area is 194659. From the
constrained conditions in equation (1), it can be conducted that the
boundaries of the maximum carbon emission and minimum population

7
G. Wang et al.
Fig. 6. 3D geographical model of land use categories.
capacity are 732.55kt and 194659, respectively. The total area of all
land use categories remains the same as the value in the baseline year
(37.28 km2). Since the results could be trapped in a locally optimal so­
lution, the multi-objective optimization model was run 100 times to
make the final solution getting close to the global optimal solution as
much as possible. Fig. 7 has presented the scatter plot of the baseline
year scenario and all optimal solutions after 100 times simulation.
Comparing with the baseline year scenario, the optimal solutions have
either lower carbon emission or higher population capacity. In order to
assess the comprehensive performance of the optimal solutions, the level
of carbon emission per person was used to evaluate the optimized re­
sults. The top 4 solutions with the best performance on the carbon
emission per person were highlighted as the planning scenarios (S1, S2,
S3, S4), these solutions are scattered among the edge of all optimal so­
lutions. Table 3 has described the land use area of each land use category
in different planning scenarios, as well as their values of carbon emis­
sion, population, and carbon emission per person.
Fig. 7. Scatter plot of optimal solutions.
8
Ecological Indicators 125 (2021) 107540
3.4. Regression modeling of land use category and carbon emission
Based on the simulation results of the multi-objective spatial opti­
mization model, a series of scenarios with different land use categories
proportions were calculated. They were conducted in the MLR analysis
to get the impact of land use category on carbon emission in the case
study. Table 4 presents the results of MLR models linking carbon emis­
sion with the variation of land use area in different land use categories.
The regression results in model 5 show that the land use categories Class
A, Class E, and Class F have a significant impact on carbon emission,
their coefficients are − 0.383, − 0.447, and − 0.332, respectively.
Therefore, the variation of land use category Class E has the most sig­
nificant impact on carbon emission, followed by Class A and Class F,
respectively. The collinearity statistics show the tolerance values of all
parameters are close to 1, while VIF values are all slightly higher than 1.
It proves the independent variables in model 5 have a low degree of
multicollinearity. Table 5 presents the results of MLR models linking
carbon emission per person with the variation of land use area in
different land use categories. The tolerance values and VIF values in
G. Wang et al. Ecological Indicators 125 (2021) 107540

Table 3 model, the tolerance values and VIF values in model 3 show the inde­
Optimization planning scenarios by the multi-objective spatial optimization pendent variables have a low degree of multicollinearity. The co­
model. efficients of land use categories Class A, Class B, and Class E are
Scenarios S0* S1 S2 S3 S4 − 7.734E-6, − 2.164E-6, and − 2.467E-6, respectively. It explains the
Land use Class 5.36 4.45 4.42 4.41 4.53
variation of land use category Class A also has the greatest impact on
area (km2) A carbon emission per person, while land use category Class E has a
Class 8.21 8.44 8.72 8.42 8.54 greater impact on carbon emission per person than land use category
B Class B. Fig. 8a and Fig. 8b presents the histogram and P-P plot of carbon
Class 6.32 7.54 7.37 7.15 6.88
emission reduction (CER) MLR model and carbon emission per person
C
Class 6.96 7.43 7.12 7.93 7.88 reduction (CERper) MLR model, respectively. The histograms in both
D Fig. 8a and Fig. 8b show the standardized residual generally follows the
Class 3.94 3.57 3.47 4.16 3.78 0 means value and 1 standard deviation normal distribution. Mean­
E while, the P-P plots are all linear over much of the range of residuals. The
Class 6.49 5.86 6.18 5.21 5.67
F
histogram and P-P plot in Fig. 8 prove that the standardized residuals in
Total 37.28 37.28 37.28 37.28 37.28 both models are normally distributed. It means the regression model
Carbon emission(kt) 732.55 731.60 732.40 731.81 731.49 linking carbon emission with different land use categories and the
Population(persons) 194,659 194,785 194,970 194,805 194,714 regression model linking carbon emission per person with different land
Carbon emission per 3763.25 3755.95 3756.50 3756.65 3756.72
use categories both fit the basic application conditions.
person(kg/person)

*S0 represents the baseline year scenario. 4. Discussion

4.1. Land use classification based on spatial attributes

Table 4
MLR models linking carbon emission with different land use categories.
The combination methodologies of the random forest algorithm and
Parameters Estimate S.E. p Collinearity statistics Adjusted regression tree that was proposed in this study proved to be effective in
R2
Tolerance VIF classifying land use patterns according to the carbon emission amounts.
1 Class − 0.555 0.045 0.000 1.000 1.000 0.604 The configuration of buildings and vegetation plays an important role in
A mitigating carbon emission, and the classified land use categories are
2 Class − 0.852 0.064 0.000 0.368 2.717 0.704 distinguished by the spatial distribution of the built environment (Khan
A and Pinter, 2016). Comparing with the previous research which only
Class B − 0.261 0.045 0.000 0.368 2.717
3 Class − 0.813 0.059 0.000 0.360 2.781 0.750
focusing on the classification based on the landscape properties (Zhou
A et al., 2017; Wang et al., 2018a), this study has fully illustrated the
Class B − 0.246 0.041 0.000 0.365 2.736 relation between carbon emission and spatial attributes during the
Class E − 0.217 0.050 0.000 0.973 1.028 process of land use classification. Combining with the population dis­
4 Class − 0.466 0.134 0.001 0.066 15.170 0.767
tribution data, the carbon emission and population capacity within
A
Class B − 0.051 0.079 0.517 0.093 10.738 different land use categories were clearly presented through Table 2 and
Class E − 0.408 0.082 0.000 0.333 3.001 Fig. 6. The results show they are highly related to the building density
Class F − 0.278 0.097 0.005 0.151 6.630 and vegetation coverage in each land use pattern.
5 Class − 0.383 0.041 0.000 0.705 1.419 0.769 The CART model presented in this research shows the classification
A
Class E − 0.447 0.056 0.000 0.702 1.425
results are significantly related to the urban structure and landscape
Class F − 0.332 0.049 0.000 0.592 1.689 features. It should be noted that the CART model may fail to estimate all
appropriate spatial attributes. The samples used to generate the fitted
model are usually limited; it is difficult to account for all the variability
model 5 significantly exceed the boundary. In model 4, the tolerance encountered in reality (Johnson and Jozdani, 2018). Thus, the classified
values of Class A and Class B are lower than 0.1 and the VIF values are land use categories only partly reflect the urban spatial layout in a
higher than 10. It means multicollinearity among the independent var­ specific case study. Considering the land use patterns in Eindhoven
iables is existed in model 4. When Class F is excluded from the MLR presented as the low-rise buildings dominated, the results of the

Table 5
MLR models linking carbon emission per person with different land use categories.
Parameters Estimate S.E. p Collinearity Statistics Adjusted R2

Tolerance VIF

1 Class A − 5.524E-6 0.000 0.000 1.000 1.000 0.789

2 Class A − 8.181E-6 0.000 0.000 0.368 2.717 0.896
Class B − 2.335E-6 0.000 0.000 0.368 2.717
3 Class A − 7.734E-6 0.000 0.000 0.360 2.781 0.978
Class B − 2.164E-6 0.000 0.000 0.365 2.736
Class E − 2.467E-6 0.000 0.000 0.973 1.028
4 Class A − 4.548E-6 0.000 0.000 0.066 15.170 1.000
Class B − 3.759E-7 0.000 0.000 0.093 10.738
Class E − 4.220E-6 0.000 0.000 0.333 3.001
Class F − 2.552E-6 0.000 0.000 0.151 6.630
5 Class A − 2.058E-6 0.000 0.000 0.001 1655.105 1.000
Class B 8.799E-7 0.000 0.000 0.001 865.477
Class E − 6.215E-6 0.000 0.000 0.002 556.414
Class F − 5.160E-6 0.000 0.000 0.001 1507.622
Class D − 8.170E-7 0.000 0.000 0.003 306.434

9
G. Wang et al.
Fig. 8. Histogram and P-P plot of the standardized residual in MLR models.
regression tree show the land use classification is mainly determined by
the spatial attributes such as normalized buildings volume, normalized
vegetation volume, normalized buildings number, and vegetation
coverage. The classification process only distinguished the differences in
building layout among land use patterns based on their building density.
The vertical structure of buildings among different land use patterns are
hard to be recognized in Eindhoven. Comparing with the spatial attri­
butes of buildings patches, the classified land use categories show more
significant differences in the geometric and volumetric attributes of
vegetation. The distinction among the land use categories Class A, Class
B, and Class C mainly depend on the values of normalized volume of
vegetation, while the distinction between Class D and Class E is relying
on the vegetation coverage ratio. Therefore, the classification results in
Eindhoven prove the geographic features of vegetation are the relatively
more important factors for the carbon emission. From the visualization
results in Fig. 6, it can be briefly summed up that carbon emission and
population capacity are positively correlated with building density, and
negatively correlated with vegetation volume.
10
Ecological Indicators 125 (2021) 107540
There are still several limitations within the current land use clas­
sification. Due to the conversion rules raster data and vector data in the
carbon emission estimation model, some land cells were assigned with
the carbon emissions which have a large physical distance (Wang et al.,
2019). These carbon emissions have been further assigned to each land
use category during the classification process. It leads to the result that
the carbon emission value in the pure greenspace land use category
(Class A) is greater than its carbon sequestration value. Since land use
categories Class A, Class B, and Class C are only distinguished by the
values of normalized vegetation volume. They lack the connection with
the geographic features of buildings, and it weakens their accuracy on
reflecting the population capacity in each land use category. To further
generalize more accurate results and better adapt to more land cover
types, one of the most viable ways is increasing the diversity of the data
collection which can provide more detailed information for the land use
classification.
G. Wang et al.
4.2. Land use plan projection from spatial optimization
According to the results of land use classification, the multi-objective
spatial optimization model proposed in this research using the evolu­
tionary algorithm to generalize a series of possible future land use plans.
The projected land use solutions in Fig. 7 and Table 3 show the land use
area variation in different land use categories may lead to the reduction
of carbon emission and increment on population capacity. It proves that
even without introducing renewable energy or mitigation technologies,
the aims of reducing carbon emission can still be achieved by the opti­
mization of the urban spatial structure. It also provides the possibility of
increasing the population capacity without any land use expansion.
Unlike other land use spatial optimization models that set the policies
regarding energy and social-economic development as the optimal
conditions, this study directly linking the carbon emission with spatial
planning policies. It provides a more straightforward perspective for
analyzing the possible land use plans, the projected scenarios from the
spatial optimization model have proved its great potential on reaching
the requirement of the low-carbon city by reasonable spatial planning.
With the support of the specific future policy, the multi-objective
spatial optimization model could have many extensions on its imple­
mentation. For instance, by introducing extra limitations on land use
categories, the spatial optimization model can be used to predict the
urban land use structure under different land use policies (Wang et al.,
2018b). In order to predict the effect of new energy policy on land use
plan, the carbon emission coefficient in the carbon emission estimation
framework can be adjusted following the available reference for CO2
emission factor in the Netherlands (Status, 2011). Mostly, the new en­
ergy policy has occurred in the building or transport sector. According to
the requirements of specific energy policy, the relevant carbon emission
distribution map can be obtained by updating the carbon emission co­
efficient in building or transport carbon emission estimation model. The
estimated results can be further applied in a comparative study. To sum
up, the developed multi-objective spatial optimization model can be
widely applied to support land use planning in various situations.
However, due to the reason that limitation conditions in the current
study only include basic limits for carbon emission and population ca­
pacity, the range for the model fitness is relatively wide. It leads to the
consequence that the results are easily trapped in the local optimum.
This issue could be solved by involving more precise land use limitation
conditions in the spatial optimization model.
Although the setting in the current model has some limitations, the
massive number of optimal solutions still provide a possibility to identify
the impact of land use variation on the carbon emission. The regression
results presented in Table 4 illustrate the increase of land use area in
Class A, Class E, and Class F categories have a negative impact on the
carbon emission reduction in the case study. Table 5 shows the increase
of land use area on Class A, Class B, and Class E categories have a
negative impact on the reduction in carbon emission per person. Usu­
ally, the greenspace plays an important role in balancing the carbon
emission in an urban area. Since the objectives of the optimization
model are maximum population capacity and minimizing carbon
emission, the increasing of land use area in pure greenspace land use
category like Class A may lead to the reduction in inhabitants, which
needs to be compensated by increasing the high carbon emission land
use categories such as Class E and Class F. This leads to the opposite
effect between increasing pure greenspace and reducing carbon emis­
sion. To reduce carbon emissions and to maintain population capacity
above its level in the baseline scenario, it is not suggested to develop
more pure greenspace area in Eindhoven. However, a certain proportion
of greenspace is still helpful in reducing carbon emissions in neighbor­
hoods. It not only directly absorbs the CO2 in the process of photosyn­
thesis but also decreases the cooling demands by shading and
evapotranspiration, and heating demands by windspeed reduction
(Zhang et al., 2017). This leads to the results that the increased land use
area in the Class C and Class D with well-balanced green space and
11
Ecological Indicators 125 (2021) 107540
housing did not show a negative impact on the reduction of carbon
emission in the regression model. To maintain the balance between
maximizing population capacity and minimizing carbon emission. The
regression results suggested to decrease land use area in pure green­
space, and high-density housing with low greening, especially Class A
and Class E. Because the increasing land use area in these two land use
categories have a negative impact on both carbon emission and carbon
emission per person.
5. Conclusion
With the increasing amounts of carbon emission in an urban area, the
optimization of land use structure holds great potential in reducing
carbon emission rates and providing sustainable land use plans. This
study proposed an integrated framework for urban land use classifica­
tion and spatial optimization. First, with the support of remote sensing
and land use and land cover map extracted from the local database.
Thirteen spatial attributes were selected to capture the configuration of
buildings and vegetation in each land use pattern. Then, the constructed
spatial attributes were combined with the results from the carbon
emission estimation framework to conduct land use classification using
random forest algorithm and CART. The land use patterns in Eindhoven
were classified into six categories according to the spatial attributes such
as normalized buildings volume, normalized vegetation volume, vege­
tation coverage, and building density. The classification results clearly
revealed the carbon emission and population capacity under a different
spatial context. Finally, a multi-objective spatial optimization model
based on NSGA-II was developed to minimize carbon emission and
maximum population capacity. The results based on the experimental
data illustrated the optimization results could easily be trapped in the
local optimum solutions due to the lack of sufficiency limitation con­
ditions. Despite the local optimal problems, multi-objective spatial
optimization still able to propose a series of land use solutions that can
meet the requirement of low-carbon city planning. The MLR analysis
also presents the impact of each land use category on carbon emission
and carbon emission per person, and the results provide a valuable
reference for the municipality about future land use zoning plan in
Eindhoven.
Based on the current results, the future investigation on the land use
classification and spatial optimization could focus on two aspects. On
the one hand, regarding the classification of land use categories, more
precise geographical information could be introduced to improve the
accuracy of capturing effective spatial attributes. On the other hand, to
reduce the locally optimum problems in the current model, the detail
regulations about building density and vegetation coverage could be
considered. It not only provides more limitation conditions in the opti­
mization process but also gives the possibilities to predict land use
zoning plans under specific urban development scenarios. With further
exploration of the current system, the spatial optimization model may
have better flexibility in adopting the requirements of urban planners.
CRediT authorship contribution statement
Gengzhe Wang: Conceptualization, Methodology, Software, Vali­
dation, Formal analysis, Data curation, Writing - original draft, Visual­
ization. Qi Han: Writing - review & editing, Supervision. Bauke de
vries: Writing - review & editing, Supervision.
Declaration of Competing Interest
The authors declare that they have no known competing financial
interests or personal relationships that could have appeared to influence
the work reported in this paper.
G. Wang et al.
Acknowledgements
This research is supported by the China Scholarship Council (grant
number 201606170088). Here we would like to acknowledge their
financial support.
References
Al Rifat, S.A., Liu, W., 2019. Quantifying spatiotemporal patterns and major explanatory
factors of urban expansion in miami metropolitan area during 1992–2016. Remote
Sens. 11 https://doi.org/10.3390/rs11212493.
Bao, C., Xu, L., Goodman, E.D., Cao, L., 2017. A novel non-dominated sorting algorithm
for evolutionary multi-objective optimization. J. Comput. Sci. 23, 31–43. https://
doi.org/10.1016/j.jocs.2017.09.015.
Bochow, M., Taubenböck, H., Segl, K., Kaufmann, H., 2010. An automated and adaptable
approach for characterizing and partitioning cities into urban structure types. Int.
Geosci. Remote Sens. Symp. 7, 1796–1799. https://doi.org/10.1109/
IGARSS.2010.5652972.
Breiman, L., Friedman, J.H., Olshen, R.A., Stone, C.J., 2017. Classification and regression
trees. Class. Regres. Trees. https://doi.org/10.1201/9781315139470.
Cao, K., Huang, B., Wang, S., Lin, H., 2012. Sustainable land use optimization using
Boundary-based Fast Genetic Algorithm. Comput. Environ. Urban Syst. https://doi.
org/10.1016/j.compenvurbsys.2011.08.001.
Cao, K., Guo, H., Zhang, Y., 2019. Comparison of approaches for urban functional zones
classification based on multi-source geospatial data: A case study in Yuzhong
District, Chongqing. China. Sustain. 11 https://doi.org/10.3390/su11030660.
Caparros-Midwood, D., Barr, S., Dawson, R., 2015. Optimised spatial planning to meet
long term urban sustainability objectives. Comput. Environ. Urban Syst. https://doi.
org/10.1016/j.compenvurbsys.2015.08.003.
Chan, T.M., Tang, K.S., Kwong, S., Man, K.F., 2016. Multiobjective optimization
methods, in: Intelligent Systems. https://doi.org/10.1201/9781315218427-24.
Corbera, S., Olazagoitia, J.L., Lozano, J.A., 2016. Multi-objective global optimization of a
butterfly valve using genetic algorithms. ISA Trans. https://doi.org/10.1016/j.
isatra.2016.03.008.
Cramer, W., Guiot, J., Fader, M., Garrabou, J., Gattuso, J.P., Iglesias, A., Lange, M.A.,
Lionello, P., Llasat, M.C., Paz, S., Peñuelas, J., Snoussi, M., Toreti, A., Tsimplis, M.N.,
Xoplaki, E., 2018. Climate change and interconnected risks to sustainable
development in the Mediterranean. Nat. Clim. Chang. https://doi.org/10.1038/
s41558-018-0299-2.
Gao, L., Wen, X., Guo, Y., Gao, T., Wang, Y., Shen, L., 2014. Spatiotemporal variability of
carbon flux from different land use and land cover changes: A case study in Hubei
Province, China. Energies 7, 2298–2316. https://doi.org/10.3390/en7042298.
Hair, J.F., Black, W.C., Babin, B.J., Anderson, R.E., 2010. Multivariate data analysis_
sumario. Analysis 816. https://doi.org/10.1016/j.ijpharm.2011.02.019.
Hermosilla, T., Palomar-Vázquez, J., Balaguer-Beser, Á., Balsa-Barreiro, J., Ruiz, L.A.,
2014. Using street based metrics to characterize urban typologies. Comput. Environ.
Urban Syst. 44, 68–79. https://doi.org/10.1016/j.compenvurbsys.2013.12.002.
Hua, L., Zhang, X., Chen, X., Yin, K., Tang, L., 2017. A feature-based approach of decision
tree classification to map time series urban land use and land cover with landsat 5
TM and landsat 8 OLI in a Coastal City, China. ISPRS Int. J. Geo-Inform. 6, 1–18.
https://doi.org/10.3390/ijgi6110331.
Intergovernmental Panel on Climate Change, 2014. Climate Change 2014: Synthesis
Report. Contribution of Working Groups I, II and III to the Fifth Assessment Report of
the Intergovernmental Panel on Climate Change, IPCC.
Jing, W., Lu, H., Qin, Y., Sun, C., Jiang, L., Zhao, J., 2013. Multi-objective land use
optimization based on low-carbon development using NSGA-II. Int. Conf.
Geoinformatics 1–5. https://doi.org/10.1109/Geoinformatics.2013.6626051.
Johnson, B.A., Jozdani, S.E., 2018. Identifying generalizable image segmentation
parameters for urban land cover mapping through meta-analysis and regression tree
modeling. Remote Sens. 10 https://doi.org/10.3390/rs10010073.
Khaledian, Y., Kiani, F., Ebrahimi, S., Brevik, E.C., Aitkenhead-Peterson, J., 2017.
Assessment and monitoring of soil degradation during land use change using
multivariate analysis. L. Degrad. Dev. https://doi.org/10.1002/ldr.2541.
Khan, F., Pinter, L., 2016. Scaling indicator and planning plane: An indicator and a visual
tool for exploring the relationship between urban form, energy efficiency and carbon
emissions. Ecol. Indic. 67, 183–192. https://doi.org/10.1016/j.
ecolind.2016.02.046.
Kursa, M.B., Rudnicki, W.R., 2015. Feature Selection with the Boruta Package . J. Stat.
Softw. https://doi.org/10.18637/jss.v036.i11.
Lai, Y., Kontokosta, C.E., 2018. Quantifying place: Analyzing the drivers of pedestrian
activity in dense urban environments. Landsc. Urban Plan. https://doi.org/10.1016/
j.landurbplan.2018.08.018.
Lamamra, K., Belarbi, K., Belhani, A., Boukhtini, S., 2014. NSGA2 based of multi-criteria
decision analysis for multi-objective optimization of fuzzy logic controller for non
linear system. J. Next Gener. Inf. Technol. 5, 57–64.
Lebourgeois, V., Dupuy, S., Vintrou, É., Ameline, M., Butler, S., Bégué, A., 2017.
A combined random forest and OBIA classification scheme for mapping smallholder
agriculture at different nomenclature levels using multisource data (simulated
Sentinel-2 time series VHRS and DEM). Remote Sens. https://doi.org/10.3390/
rs9030259.
12
Ecological Indicators 125 (2021) 107540
Leng, X., Wang, J., Ji, H., Wang, Q., Li, H., Qian, X., Li, F., Yang, M., 2017. Prediction of
size-fractionated airborne particle-bound metals using MLR, BP-ANN and SVM
analyses. Chemosphere. https://doi.org/10.1016/j.chemosphere.2017.04.015.
Li, S., Wu, H., Wan, D., Zhu, J., 2011. An effective feature selection method for
hyperspectral image classification based on genetic algorithm and support vector
machine. Knowledge-Based Syst. 24, 40–48. https://doi.org/10.1016/j.
knosys.2010.07.003.
Lu, J., Guldmann, J.M., 2012. Landscape ecology, land-use structure, and population
density: Case study of the Columbus Metropolitan Area. Landsc. Urban Plan. 105,
74–85. https://doi.org/10.1016/j.landurbplan.2011.11.024.
Lyu, F., Zhang, L., 2019. Using multi-source big data to understand the factors affecting
urban park use in Wuhan. Urban For. Urban Green. https://doi.org/10.1016/j.
ufug.2019.126367.
Ma, L., Fu, T., Tiede, D., Blaschke, T., Ma, X., Chen, D., Zhou, Z., Li, M., 2017. Evaluation
of feature selection methods for object-based land cover mapping of unmanned
aerial vehicle imagery using random forest and support vector machine classifiers.
ISPRS Int. J. Geo-Inform. 6, 51. https://doi.org/10.3390/ijgi6020051.
Moghadam, S.T., Lombardi, P., Mutani, G., 2017. A mixed methodology for defining a
new spatial decision analysis towards low carbon cities. Proc. Eng. 198, 375–385.
https://doi.org/10.1016/j.proeng.2017.07.093.
Olfaz, M., Tirink, C., Önder, H., 2019. Use of CART and CHAID algorithms in Karayaka
sheep breeding. Kafkas Univ. Vet. Fak. Derg. https://doi.org/10.9775/
kvfd.2018.20388.
R.K. Pachauri, L.A.M. and C.W.T., 2014. Climate Change 2014 Synthesis Report
Summary Chapter for Policymakers. IPCC 151. https://doi.org/10.1017/
CBO9781107415324.
Penazzi, S., Accorsi, R., Manzini, R., 2019. Planning low carbon urban-rural ecosystems:
An integrated transport land-use model. J. Clean. Prod. 235, 96–111. https://doi.
org/10.1016/j.jclepro.2019.06.252.
Prendinger, H., Gajananan, K., Bayoumy Zaki, A., Fares, A., Molenaar, R., Urbano, D.,
Van Lint, H., Gomaa, W., 2013. Tokyo virtual living lab: Designing smart cities based
on the 3D internet. IEEE Internet Comput. 17, 30–38. https://doi.org/10.1109/
MIC.2013.87.
Sharmin, N., Haque, A., Islam, M.M., 2019. Generating Alternative Land-use Allocation
for Mixed Use Areas: Multi-Objective Optimization Approach. Geogr. Anal. 51,
448–474. https://doi.org/10.1111/gean.12181.
Shu, H., Xiong, P. Ping, 2019. Reallocation planning of urban industrial land for structure
optimization and emission reduction: A practical analysis of urban agglomeration in
China’s Yangtze River Delta. Land Use Policy 81, 604–623. https://doi.org/
10.1016/j.landusepol.2018.11.034.
Status, D., 2011. Nederlandse lijst van energiedragers en standaard CO 2 -
emissiefactoren , versie januari 2011 Colofon 1–20.
Tayyebi, Amin, Tayyebi, Amirhossein, Vaz, E., Arsanjani, J.J., Helbich, M., 2016.
Analyzing crop change scenario with the SmartScapeTM spatial decision support
system. Land Use Policy 51, 41–53. https://doi.org/10.1016/j.
landusepol.2015.11.002.
Verhulp, J., Van Niekerk, A., 2017. Transferability of decision trees for land cover
classification in a heterogeneous area. South Afr. J. Geomatics 6, 30. https://doi.
org/10.4314/sajg.v6i1.3.
Wang, G., Han, Q., de Vries, B., 2019. A geographic carbon emission estimating
framework on the city scale. J. Clean. Prod. 118793 https://doi.org/10.1016/j.
jclepro.2019.118793.
Wang, Y., Li, X., Zhang, Q., Li, J., Zhou, X., 2018b. Projections of future land use changes:
Multiple scenarios-based impacts analysis on ecosystem services for Wuhan city.
China. Ecol. Indic. 94, 430–445. https://doi.org/10.1016/j.ecolind.2018.06.047.
Wang, J., Wu, Z., Wu, C., Cao, Z., Fan, W., Tarolli, P., 2018a. Improving impervious
surface estimation: an integrated method of classification and regression trees
(CART) and linear spectral mixture analysis (LSMA) based on error analysis.
GIScience Remote Sens. 55, 583–603. https://doi.org/10.1080/
15481603.2017.1417690.
Xie, Z., Chen, Y., Lu, D., Li, G., Chen, E., 2019. Classification of land cover, forest, and
tree species classes with Ziyuan-3 multispectral and stereo data. Remote Sens.
https://doi.org/10.3390/rs11020164.
Yan, W.Y., Shaker, A., El-Ashmawy, N., 2015. Urban land cover classification using
airborne LiDAR data: A review. Remote Sens. Environ. 158, 295–310. https://doi.
org/10.1016/j.rse.2014.11.001.
Yang, C., Wu, G., Ding, K., Shi, T., Li, Q., Wang, J., 2017. Improving land use/land cover
classification by integrating pixel unmixing and decision tree methods. Remote Sens.
9 https://doi.org/10.3390/rs9121222.
Yao, J., Murray, A.T., Wang, J., Zhang, X., 2019. Evaluation and development of
sustainable urban land use plans through spatial optimization. Trans. GIS 23,
705–725. https://doi.org/10.1111/tgis.12531.
Zhang, G., Ge, R., Lin, T., Ye, H., Li, X., Huang, N., 2018. Spatial apportionment of urban
greenhouse gas emission inventory and its implications for urban planning: A case
study of Xiamen. China. Ecol. Indic. 85, 644–656. https://doi.org/10.1016/j.
ecolind.2017.10.058.
Zhang, Y., Murray, A.T., Turner, B.L., 2017. Optimizing green space locations to reduce
daytime and nighttime urban heat island effects in Phoenix, Arizona. Landsc. Urban
Plan. https://doi.org/10.1016/j.landurbplan.2017.04.009.
Zhou, W., Song, Y.Q., Pan, Z.K., Liu, Y.L., Hu, Y.M., Cui, X. Sen, 2017. Classification of
Urban Construction Land with Worldview-2 Remote Sensing Image Based on
Classification and Regression Tree Algorithm. Proc. - 2017 IEEE Int. Conf. Comput.
Sci. Eng. IEEE/IFIP Int. Conf. Embed. Ubiquitous Comput. CSE EUC 2017 2,
277–283. https://doi.org/10.1109/CSE-EUC.2017.237.