Keywords

1 Introduction

We live in an age where the demand for accurate and up-to-date spatial data has never been greater. However, obtaining and maintaining such spatial databases is a challenging and expensive task. In this context, crowdsourced maps such as OpenStreetMap (OSM)Footnote 1 can be a viable solution for obtaining a free and up-to-date representation of our world. While an extensive number of applications have been developed around OSM, concerns exist regarding the quality of OSM data. The predominant method for assessing OSM data quality is based on comparing the OSM data with some form of authoritative maps such as the Ordnance Survey UK [2], Google Maps, etc. However, we argue that this process of comparing a crowdsourced (heterogenous) database with authoritative maps is ineffective. Instead we propose the use of machine learning techniques for assessing and possibly improving the data quality of crowdsourced maps without referencing to external repositories.

Specifically, in this paper we focus on the semantic type quality of streets in the OSM where semantic type refers to the class of a street such as motorway, pedestrian, etc. We hypothesize that the semantic types of streets are a function of their geometrical and topological features and develop structured and unstructured machine learning models that can learn the semantic types of streets given such features. Interestingly, the structured learning models can also exploit the inherent spatial relationships within a street network.

2 Methodology

2.1 Data Representation

Appropriate data representation is a fundamental step toward useful knowledge discovery. Therefore, as a first step a novel multi-granular graph-based street network representation system is developed. All streets having same name and same semantic type correspond to a single node in a multi-granular graph. Such a representation makes the various features of a street explicit as opposed to implicit. More details of the multi-granular representation system can be found in [3].

2.2 Feature Extraction

Several topological and geometrical features of streets were extracted using the multi-granular street network representation obtained above. These include length, linearity, number of dead-ends, number of intersections, semantic types of adjacent streets (using a BoW model), node degree, and betweenness centrality.

2.3 Unstructured Learning

Next, we develop an unstructured (or classical) supervised machine learning model to learn the various semantics types of streets in the OSM database. The development of this model involves assessing the performance of the commonly used machine learning classifiers such as naive bayes, SVM, neural networks, and random forests in terms of their generalization performance on test data. More details on the implementation of the unstructured learning of the problem can be found in [4].

2.4 Structured Learning

A street network is a structured input as it consists of several streets, where not only the streets themselves contain information such as geometry, but also the way in which the streets are connected to each other is important. For such a structured input, we obtain a structured output of semantic types of streets over all the streets in the network. We exploit the Conditional Random Field (CRF) framework for performing structured prediction. The CRF framework allows us to leverage prior knowledge available to us in the form of crowdsourced semantics, the geometrical and topological features of individual streets, and the contextual (structural) relationships between various streets into a single unified model.

Fig. 1.
figure 1

Street network represented as a graphical model. x are the observed variables corresponding to the streets in the network and y are the labelling we want to infer. The green lines correspond to the unary potentials in the model and the red line to the pairwise potential.

Full size image

Suppose we have a street network consisting of N streets \(\mathbf x = \lbrace x_1, x_2, \dots , x_N \rbrace \in \mathcal {X}\) and our goal is to predict the semantic type labellings \(\mathbf y = \lbrace y_1, y_2, \dots , y_N \rbrace \in \mathcal {Y}\) for these streets. Figure 1 shows our representation of such a street network as a graphical model where \(x_{ip}\) corresponds to the initial crowd sourced labels or priors and \(x_{if}\) corresponds to the geometric and topological features. Toward the goal of jointly learning the semantic type labelling y, our model maximizes the conditional probability of y given x [6]:

$$\begin{aligned} \hat{\mathbf{y }}= \arg \max _\mathbf{y \in \mathcal {Y}} P(\mathbf y |\mathbf x ;\mathbf w ) \end{aligned}$$
(1)

We use a max-margin approach for determining the model parameters w and a fusion moves approach for inferring the street labellings. More details on the structured learning of the problem can be found in [1, 5].

3 Results and Discussion

We trained and tested our models on two non-overalapping regions from OSM London database. All 19 popular semantic types of streets used in OSM database for classifying a street were considered. An overall classification accuracy of 55.95 % was obtained using the unstructured learning model (random forest). This accuracy increased to 84.75 % when structured learning framework was used. Clearly, and naturally the structured learning framework outperforms the unstructured learning performance as it exploits the inherent structure in street networks. To the best of our knowledge, this is the first time that a structured learning framework has been used in the context of crowdsourced spatial data.

In this work, we considered all the 19 popular semantic types of streets used for classifying a street network. However, such a classification of street network is too fine-grained when compared with the commonly used and understood street network classifications where a street network is usually classified into 4–10 semantic types. In future we propose the development of a multi-layer conditional random field based model for simultaneously learning both the fine-grained (19) and coarse-grained (4–10) semantic types of streets. In addition, the models developed in this paper will also be extended to other map objects such as buildings, Points of Interests (PoIs), etc.

4 Dual Submissions

The work presented in this paper is a summary of the work already published at the following venues:

  1. 1.

    23rd ACM SIGSPATIAL Conference, USA, 2015

  2. 2.

    22nd ACM SIGSPATIAL Conference, USA, 2014

  3. 3.

    Intelligent Systems, Technologies, and Applications, Springer, 2016

  4. 4.

    UL-NUIG Research Day, 2016.

  5. 5.

    Related version submitted to the Indian Workshop on Machine Learning, 2016.