The general interest in sustainable development models has grown enormously over the last 50 years, and architecture and urban planning are certainly two areas in which research on the topic is most advanced. At the same time, the contribution of computer science for a systematic analysis of the territory, both from a morphological point of view and as regards performance, seems to have been underestimated in today's research. In this context, our research aims to joining the two - until now separate - worlds of computer science, and architecture and urban planning. In particular, in this work we present SIMBA: Systematic clusterIng-based Methodology to support Built environment Analysis. SIMBA aims to enhance a consolidated analysis methodology, the Integrated Modification Methodology (IMM), through the integration of advanced analysis methods for the extraction of relevant patterns from built environment data. Using the city of Milan as a case study, we will demonstrate the possibility for SIMBA to be generalised to the analysis of any built environment.
One of the biggest problems of our century is the impact that the behaviour of human beings
is having on the environment. This impact, under the same conditions for the eficiency in
resource management, is obviously greater in more densely populated areas. According to [
In this paper we present SIMBA: a systematic clustering-based methodology to support built
environment analysis. SIMBA has been conceived as a framework extending and enhancing
IMM (Integrated Modification Methodology)[
The IMM methodology aims at improving the environmental performance of the city by
modifying its structural characteristics. The main elements of the methodology related to the structures
are Attributes (e.g., the height of a building) and Metrics (e.g., Building Density (BD) equal to
the total number of buildings in an area divided by the total sample area). Attributes represent
immediately measurable characteristics while Metrics result from calculations. Moreover, from
Metrics, IMM defines the Key Categories, used to represent the products of the synergy between
elementary parts of the city. At the moment, the IMM group has defined 7 Key Categories, but
for the sake of simplicity in our experiments, we will consider only Permeability (the spatial
relationship between urban built-ups and voids) and Porosity (the relationship between the
street network and the spatial components that influence the overall connectivity). The Key
Categories are used to express morphological characteristics, while the tools for performance
evaluation are called Indicators (e.g., Public Transportation Stop Density) organized into Design
Ordering Principle (DOP) families, i.e., families of actions that designers can perform to improve
the current system behaviour [
2. FIRST LEVEL CLUSTERING (ONE FOR EACH DATASET) 3. SECOND LEVEL CLUSTERING 2.3 CLUSTERING
CLUSTERING
CLUSTERING
EVALUATION 3.1 IMPORTANT
DIMENSION CHOICE 3.2
CLUSTERING ON
SELECTED COMBINATION
In this section we present the SIMBA methodology together with the experiments we carried out. Fig. 1 represents the methodological workflow. As you can see, the process takes as input a built environment of any kind: a town, a city or even a country, since it can be generalized to any input, as long as it is possible to identify comparable units in it. The case study discussed in this paper focuses on Milan. The methodology is composed of three main phases: (i) Built Environment Decomposition (BED phase); (ii) First-Level Clustering (FLC phase, one for each dataset); (iii) Second-Lelvel Custering (SLC phase). The following sections will describe each phase together with its application to our case study.
The Built Environment Decomposition phase is useful to make the analysis manageable through Data Mining algorithms, since it aims at reducing the number of features in the single analysis by splitting all the available data into distinct datasets according to diferent Dimensions.
In this phase, we first need to define the granularity at which the analysis has to be conducted (step 1.1). In other words, we need to choose which are the samples we want to cluster, and in our case study these samples are the 88 NILs. Although the definition of NIL is specific to the city of Milan, this does not afect the generalization potential of the methodology, since similar divisions (possibly at diferent scales) can be found in other cities, and therefore step 1.1 can be carried out. After the granularity, we need to define the Dimensions (step 1.2) and retrieve the data according to them (step 1.3). The Dimensions represent the diferent aspects we want to analyse. They can be of any type and category, e.g., environmental performances, morphological characteristics, demographic data and so on, but they remain abstract concepts for us: they have to work as simple guidelines to be used while looking for data to retrieve. In fact, it is in the datasets that the Dimensions are represented. As far as our experiments are concerned, we chose to focus on all the Dimensions at our disposal, as our four datasets include both performance data and structural characteristics. Consequently, the datasets we create for step 1.3 correspond to the ones we already defined in Section 2. Note that in some experiments we have considered all Metrics together, in others Permeability and Porosity separately.
Once we have our datasets ready, we perform First-Level Clustering for each of them. This phase is needed to:(i) prepare the datasets for clustering; (ii) select the important features; (iii) perform clustering; (iiii) evaluate each obtained cluster. Outputs of this phase are:(i) diferent clusterizations for each dataset; (ii) distances between the NILs for each dataset.
These outputs are useful to investigate patterns in each dataset and so for each Dimension. For
what concerns the setting for this phase, note that the final choice of the techniques used for
preprocessig, the algorithm chosen for feature selection, and the clustering algorithm itself, were
dictated by the nature of the data at our disposal. On the other hand, what is also important to
underline is that, once the process has been formalised, even if the specific setting is dependent
on the input data, the procedure remains unchanged. For what concerns the pre-processing
phase, we mostly had to manage missing values and perform some feature engineering. We
strictly collaborated with the experts in this step (2.1) to preserve the data semantics as much
as possible. As far as step 2.2 is concerned, for the manual feature selection we relied totally on
the experts, asking them to select no more than 8 features for each dataset; for the automated
case instead, the selection was conducted using the entropy-based algorithm presented in [
The last step of this phase is dedicated to the evaluation of the clusters, which is still a tricky
part in this field and most of the times is strongly application-dependent. Firstly, we tried to
have an absolute evaluation of each cluster result using internal metrics such us the Dunn and
Davies-Bouldin [
To do so, we first defined the comparison_matrix, a matrix used to store the number of common elements among the clusters obtained by selecting diferent set of features. For each experiment (dataset) we computed the comparison_matrix between the two approaches (the manual and the automated one). The pseudo code for the computation is shown in Algorithm 1 below.
Algorithm 1: Comparison_matrix computation Input: _ Output: _ _ = _.[0] for = 0, < (_), + + do for = 0, < (_), + + do _[][] = (_[] == _[]).() return _
In 1, the comparison_matrix CM is a matrix of dimension , where is the number of NILs for the dataset. For two NILs i and j, CM [i][j] = 2 if the two methods put the NILs i and j in the same cluster, CM [i][j] = 1 if only one method puts the two NILs in the same cluster, and CM [i][j] = 0 otherwise. This means that the positive cases correspond to the values 0 and 2 since they mean that, in the two approaches, the clustering has produced the same result for that pair of NILs.
Finally, to evaluate each experiment, we defined the measure score, whose computation is shown in Algorithm 2.
Algorithm 2: Score computation Input: _, _, __ Output: for = 0, _., + + do for = 0, _., + + do if _[][] == 0 _[][] == _ then = + 1
return = __
In a given experiment, for each pair of NILs, score counts how many values 0 or 2 occur in the corresponding comparison_matrix, and normalizes this number w.r.t. the number of NILs present in that experiment (88 or 86). Assuming that a good result for our experiment is that the two clustering runs group all the NILs in the same way, in the Manual and in the Automated case, score can be seen as an accuracy measure for the procedure. In addition, this measure provides a direct and simple comparison between Manual and Automated cases.
Therefore, the score provides a numerical evaluation of the clusters obtained, and allows the procedure to be repeated in any proposed scenario. However, given the nature of the experiments and the absence of any real ground truth, a qualitative, although less formal, evaluation by the experts cannot be ignored. For this reason, in Figure 2 we report the most relevant results directly on the map of Milan. Each map is divided into NILs coloured according to the cluster they belong to. The colours have also been chosen to trace which clusters are (a) Indicators Manual n_clusters = 3 (b) Indicators
Automated n_clusters = 4 (c) Metrics Manual (d) Metrics Automated
n_clusters = 4 n_clusters = 5 (e) Porosity Manual (f) Porosity Automated (g) Permeability all
n_clusters = 2 n_clusters = 2 n_clusters = 4 closest to each other; the outliers are always coloured red, therefore, in the figure, they represent a single cluster, but they contribute individually to increase the n_cluster parameter which, as we note, varies among the diferent experiments. In addition, black is used to highlight the NILs that are missing in some of the datasets. In table 2 we also report the scores obtained for all the datasets.
Figures 2a and 2b represent the clusters obtained by applying the Manual and the Automated algorithms to the Indicators dataset. What emerges from these maps is that the two algorithms provide almost the same results. Indeed, one might think that indicators only highlight macrodiferences among NILs, but the results have been judged by the experts to be perfectly coherent with the way DOP families describe NILs. Moreover, as shown in 2, the score is rather close to the number of NILs (88).
Looking at the results related to the Metrics dataset (Figure 2c and 2d) we note that, in this case, the two algorithms (Manual and Automated) perform really diferently. What emerges is that the features selected manually can express finer diferences, while the automated algorithm only separates the central NILs from the more peripheral ones, with some exceptions in the very peripheral areas. This is probably due to the fact that in the Manual case experts selected both Porosity and Permeability metrics to balance their efects, while the entropy-based algorithm mostly selected porosity-related ones. This is probably due to the imbalance between the number of metrics related to Porosity (52) and those related to Permeability (7) that afected the automated selection of the features.
This is confirmed by looking at the remaining maps in Figure 2 where we note that using only Porosity-related metrics we have a sharp separation between centre and periphery NILs, while, using the ones related to Permeability, it is possible to define some clusters also in the peripheral areas. This gives us an idea of how SIMBA can be useful to study the efect of the diferent Key Categories separately or jointly, and this is of huge importance for the IMM group.
Second-Level Clustering is the third and last phase of SIMBA, and it takes two inputs: (i) the First-Level Clustering results, to identify on which dataset the clustering performed better; (ii) the IMM experts’ indications about the Dimensions to focus on to continue the analysis (step 3.1); and produces two outputs:(i) the clustering results on the dataset created by combining the important Dimensions; (ii) a formalization of distances between NILs considering only selected features of the important Dimensions.
In the previous sections we showed the results obtained by clustering NILs keeping the datasets separated. This was an important step, since it allowed us to understand what are the important IMM elements to focus on, both in an automated way and in a more guided one. As we have seen, both looking at the score values and at the produced maps, the Indicators dataset was the one where the procedure performed best. This is probably due to the fact that the indicators themselves are well separated into DOP families and the entropy-based featureselection algorithm can extract at least one indicator for each family. Other interesting results, as we have shown, are obtained on the Metrics dataset. This time the score is considerably worse, but the meaning of the resulting clusters in terms of morphology can’t be ignored. This is actually why we decided to add this last phase to the methodology, applying clustering to a dataset composed both by indicators and metrics (step 3.2). The results are reported in Figure 3.
By comparing maps, three important things emerge:(i) most of the NILs of the city centre are always in the same cluster; (ii) looking at the manual case, this seems to show practically the same results as the Indicator case, while the automated one contains also some of the clusters identified in the Metrics automated case. This might mean that there could be some indicators that guide the whole creation of clusters, but some of the metrics selected in the automated case end up smoothing its efect.
In this work, we presented SIMBA, a systematic clustering-based methodology to support built-environment analysis, and we proved its potential as a support tool for architects and urban planners in understanding and analysing urban environments. The biggest limitation of the work is the nature of the data. On the one hand, the scarcity of the samples (only 88 NILs) makes the analysis sensitive to outliers and dificult to formally generalise; on the other hand, the excessive number of characteristics, which are often redundant and not very informative, afects the quality of the results produced. Two aspects on which future work will therefore focus are the definition of finer granularity and the selection of features. Additional eforts will be devoted to refining the human-in-the-loop approach that SIMBA wants to support, by means of explanations that can guide the domain experts (and eventually, diferent experts) in the selection of relevant analysis Dimensions in the SLC phase. In addition, to assess the actual possibility for SIMBA to be generalised to the analysis of any built environment, diferent datasets with diferent characteristics should be analyzed.
We would like to thank Dott. Hadi Mohammad Zadeh who has contributed to this work and its future development.