Elsevier

Pattern Recognition Letters

Volume 26, Issue 9, 1 July 2005, Pages 1201-1220
Pattern Recognition Letters

An automatic method for road extraction in rural and semi-urban areas starting from high resolution satellite imagery

https://doi.org/10.1016/j.patrec.2004.11.005Get rights and content

Abstract

In this paper an efficient method for automatic road extraction in rural and semi-urban areas is presented. This work seeks the GIS update starting from color images and using preexisting vectorial information. As input data only the RGB bands of a satellite or aerial color image of high resolution is required. The system includes four different modules: data preprocessing; binary segmentation based on three levels of texture statistical evaluation; automatic vectorization by means of skeletal extraction; and finally a module for system evaluation. In the first module the color image is rectified and geo-referenced. The second module uses a new technique, named Texture Progressive Analysis (TPA), in order to obtain the segmented binary image. The TPA technique is developed in the evidence theory framework, and it consists in fusing information streaming from three different sources for the image. In the third module the obtained binary image is vectorized using an algorithm based on skeleton extraction techniques and morphological methods. The result is an extracted road network which is defined as a structural set of elements geometrically and topologically corrects. The fourth module is an evaluation of the procedure using a popular method. Experimental results show that this method is efficient in extracting and defining road networks from high resolution satellite and aerial imagery.

Introduction

Cartographic object extraction from digital imagery is a fundamental operation for GIS update. However the complete automation of extraction processes is still an unsolved problem. In fact, many works on this topic have been presented (Mena, 2003), but the manual intervention of the operator in extracting, defining and validating cartographic objects for GIS update is still needed.

Nevertheless, important advances have been achieved. In (Mayer, 1999 and Baltsavias et al., 2001) many of these approaches can be found. The first paper presents the state of the art on automatic object extraction techniques, where the majority of the work in this area is discussed. In the second very interesting papers on cartographic object extraction and GIS update can be consulted. Sowmya and Trinder (2000) is also very important for us because it presents a review about some of the approaches used in knowledge representation and modeling for machine vision, which is a fundamental topic in automatic extraction matter.

Recently, some automatic systems for serve as support in GIS update tasks have been developed. Between them we can find Heipke et al. (2000), where different aspects of image analysis are discussed and a framework is provided for scene interpretation which is based on the integration of image analysis and a GIS data model. This work includes two examples concerned with the combined extraction of roads and trees, and with the multitemporal interpretation and monitoring of moorland. Likewise, Eidenbenz et al. (2000) presents the project ATOMI which aims to update vector data of road centerlines and building roof outlines from 1:25000 maps, fitting in to the real landscape, improve the planimetric accuracy to 1 m and derive height information with 1–2 m accuracy. Also important for us is the work Wallace et al. (2001). It presents the research project Automatic Linear Feature Identification and Extraction (ALFIE), which uses a geographical information system built around an object oriented geospatial database. Likewise, Bückner et al. (2002) and Gerke (2002) describe the system named Geo Automatic Image Data Analyser (GeoAIDA), which have been developed at the Institute of Communication Theory and Signal Processing of Hannover, and it allows an intelligent, concise and flexible control of a scene interpretation by utilizing a semantic scene description. The system produces a hierarchic, pictorial description of the results as well as the structural context of the identified objects including the associated attributes.

However, any of previous works achieves the complete automation in the cartographic process. In fact, this problem is bigger that many researches have opted for semi-automatic extraction methods. Between them Baumgartner et al. (2002) presents a prototype system for semi-automatic extraction of road axes and a study on its efficiency for operational use. The core of this system is a road tracker based on profile matching, which is enhanced with a graphical user interface that guides the operator through the whole data acquisition process. Likewise, in Zhao et al. (2002) a semi-automatic method to create and update road maps in urban and suburban area using high resolution satellite images is proposed. In this research road mask is defined as a mask of road pixels, which are discriminated from others using a commercial remote sensing software.

Mainly focusing on GIS update is the framework presented in Ohlof et al. (2000) which is applied in many cartographic military systems of OTAN countries. This paper reports on the results of two projects conducted for the AmilGeo of the German Federal Armed Forces. The first project consists of to establish an operational workflow to update existing Vmap Level 1 data using commercially available satellite imagery. The second project is a study focused on the generation and update of Vmap Level 2 data using both satellite and airborne imagery. Also very interesting, although outside of the military environment, Bonnefon et al. (2002) presents a complete process to update and upgrade geographic linear features in GIS, such as roads, railways and little rivers, with methods as automatic as possible and with a quality evaluation.

About the phases members of the general automatic object extraction process, Markov and Napryushkin (2000) proposes a detailed and efficient sequence based on segmentation and classification for solving the remote sensing data interpretation problem in the GIS framework. Although this sequence requires manual operation in some steps, it is very important for our work. It is the following: (1) designing the list of geometrical objects to be extracted, conditional splitting the initial raster image into n classes. (2) Selecting on the initial image the samples determining accordingly each class and specifying its features. Each sample contains the information on the certain type of objects. The set of geometric figures of each sample is the basis for marking a separate vector layer of digital thematic map in vector GIS. (3) Classification of initial raster image with use of the supervised classifier based on Bayesian decision rule. The subsequent actions are carried out for every ith sample, where i = 1, 2,  , n. (4) Transformation of the classified image in two level raster one describing the ith class. In the obtained two level image every pixel of selected class accepts value of unity, others pixels of the image are treated as background and accept zero value. (5) Vectorization of the binary image of ith class. This phase is carried out on the basis of plane pass algorithms. As result the set of geometrical objects is generated. (6) Correction of the errors and exporting the obtained vector data into vector GIS.

In the automatic road extraction topic, another important reference also is Amini et al. (2002). In this work a new approach for automatic extraction of main roads in large scale imagemaps is proposed. The paper describes how the gray scale imagemap is converted to a simplified imagemap using morphological algorithms. The proposed method consists of two stages. In the first stage, the simplified imagemap is segmented and converted to a binary image. In the second stage, the resolution of the simplified image is reduced through the Wavelet transform and the skeleton of roads is extracted.

The support on GIS information and the use of context information are also common characteristics in many works on road extraction, including our system. In this field Ruskoné and Airault (1997) and Mayer (1999) are two of the most indicative works; and other important works also are: Agouris et al. (1998) where a road extraction method which is governed for a fuzzy system is proposed; Jeon et al. (2000) where curvilinear structures in SAR images together with digital maps are analyzed; McKeown et al. (1999a) and Fabre et al. (2001) where context information into the study of data fusion in hyperspectral image processing is used; and Hinz et al. (2001) which proposes the generation of models in the context information analysis.

Data fusion is other important aspect in our work on automatic extraction. In this matter Hellwich and Wiedemann (2000) offers an approach to the combined extraction of linear as well as surface objects from multisensor image data based on a feature and object level fusion. In this case, data sources are high resolution panchromatic digital orthoimages, multispectral image data, and interferometric SAR data. Very interesting also are McKeown et al. (1999b) and Peddle and Ferguson (2002). The last work proposes three methods for optimizing the process of data fusion, relative to the specification of user defined inputs, based on different levels of empirical testing and computational efficiency. These methods are: the exhaustive search by recursion, the isolated independent search, and the sequential dependent search. Likewise, Fabre et al. (2001) uses pixel fusion in order to elaborate a classification method at level pixel. This paper proposes a formalism of modeling of the sensor reliability to the context that leads to two methods of integration: the first one amount to integrate this further information in the fusion rule as degrees of trust, and the second models the sensor reliability directly as mass function according to evidence theory, which can be studied in Kohlas (1995) Kohlas and Besnard, 1995a, Kohlas and Besnard, 1995b, Kohlas (1997) and Bauer (1997).

Finally, a fundamental topic in automatic object extraction also is the knowledge representation and modeling for computer vision. In this field, Mayer (1999) presents an exhaustive study and detailed analysis on automatic object extraction. This paper defines criteria in order to establish a model of knowledge. The model comprises: the derivation of characteristic properties from the function of objects, three dimensional geometry and material properties, scales and levels of abstraction/aggregation, and local and global context. The strategy consists of grouping, focusing on different scales, context based control and generation of evidence from structures of parts, and fusion of data and algorithms. Likewise Sowmya and Trinder (2000) presents a review of the approaches used in knowledge representation and modeling for machine vision, and give examples of their applications in research for image understanding of aerial and satellite imagery. More recently Dell’aqua and Gamba (2001) offers a fuzzy approach to the analysis of airborne synthetic aperture radar images of urban environments. In particular, it shows how to implement structure extraction algorithms based on fuzzy clustering unsupervised approaches. Another reference is Andersen et al. (2002) where LIDAR sensing geometry is explicitly modeled in the domain of scan space three dimensional, analogue to two dimensional image space. Here, prior models for object configurations take the form of Markov marked point processes, where pair wise object interactions depend upon object attributes. Given the complexity of the distribution used, inferences are based upon dependent samples generated via Markov chain Monte Carlo simulation. Also in Zhang and Baltsavias, 2002a, Zhang and Baltsavias, 2002b a concept for road network reconstruction from aerial images using knowledge based image analysis is presented. In contrast to other approach, the proposed approach uses multiple cues about the object existence, employs existing knowledge, rules an models, and treats each road subclass differently to increase success rate and reliability of the results. Finding 3D edges on the road and specially the road borders is a crucial component of their procedure and is the focus of theses papers.

Restricting the input image type as well as the cartographic object to extract, we limit the automatic extraction problem and, subsequently, good results could be obtained. In this line, following hypotheses and input requirements have been selected for our system on automatic extraction and GIS update:

  • 1.

    Linear cartographic objects. Linear objects like roads or river have been selected. Therefore the system is not valid for building, trees or superficial shapes, though the segmentation phase could be applied for these objects. Specifically, our objective is the automatic road extraction, including the geometrical and topological definition of the road network.

  • 2.

    Aerial or satellite color images. RGB bands of an aerial or satellite color image have been only selected as input data. Therefore, in our system the multitemporal, multispectral or stereoscopic analysis are excluded. Although these techniques improving the quality of results, we want to offer a simple system on road extraction which provides reasonable outputs starting from a minimum of input requirements.

  • 3.

    High resolution imagery. An appreciable width of the roads in the input image is needed for our method, since the texture analysis is used in order to obtain a binary segmented image. Therefore Ikonos, Quick Bird and high resolution aerial imagery can be used. Landsat, Spot and low resolution aerial imagery are excluded.

  • 4.

    Rural and semi-urban areas. In this work only rural and semi-urban areas are considered, because in urban areas frequently the texture analysis of the ground can not be applied.

  • 5.

    Automatic system. By means of our system reasonable outputs are obtained when the algorithm is applied starting from images of mentioned characteristics.

For solving the automatic road extraction problem, the following flow of work has been accepted for us based on Ohlhof et al. (2000); Markov and Napryushkin (2000); Wallace et al. (2001) and Amini et al. (2002):

  • 1.

    Data preprocessing.

  • 2.

    Object extraction and classification.

  • 3.

    Raster vector conversion.

  • 4.

    Validation of results.

  • 5.

    Assignment of attributes.

  • 6.

    Update of existing vector data.

In this line, we proposes a new method for automatic road extraction and GIS update which starting from a minimum of input data provides the geometrical and topological definition of the road network from high resolution imagery.

Section snippets

Objectives and modules of the automatic road extraction system

In order to expedite the integration of our algorithm in other object extraction existing systems, the following goals have been chosen:

  • 1.

    Acquisition of initial GIS information and image preprocessing.

  • 2.

    Automatic road network detection by means of segmentation based on texture progressive analysis.

  • 3.

    Geometrical and topological definition of road network based on skeleton extraction methods.

  • 4.

    Evaluation, validation and storage in GIS of graphic elements which have been obtained.

These objectives are

First module: Data preprocessing

As external input data, only the RGB bands of a high resolution image is needed for our algorithm. In this paper an Ikonos image in order to clarify the system description have been chosen. In (Dial et al., 2001) several aspects about the use of Ikonos imagery in automated road extraction are analyzed.

Second module: Segmentation process (low level of knowledge)

Images segmentation represents a first step in many tasks that pattern recognition or computer vision has to deal with. There are many papers about segmentation of images using color, see Skarbek and Koschan (1994) for an early survey and Cheng et al. (2001) for a more recent one. Several authors are applying different techniques for color in order to improve the final result of the segmentation, for example, Park et al. (1998) presents a new algorithm based in mathematical morphology which

Third module: Geometrical and topological vectorization process (mid-level)

Starting from the segmented image, now our system seeks the geometrical and topological definition of extracted road network. For this goal an algorithm based on skeleton extraction and graph theory is applied. This algorithm comprises the following steps.

Other obtained results

Below other results are shown. In each figure only the original image, the training area and the automatic vectorization are presented. The numerical results are omitted in order to abbreviate this paper.

Evaluation

According to the method developed in Wiedemann et al. (1998) for evaluating automatic road extraction systems, and by comparison between automatic results and manual results, the following data have been obtained for our system (Fig. 8, Fig. 9, Fig. 11, Fig. 12).

VariablesFig. 8Fig. 9Fig. 10Fig.

Conclusions

In this paper a solid system for automatic road extraction has been presented. This system includes the geometrical and topological definition of the graphic elements, and it seeks decreasing the work which is manually realized for the operators in their tasks for GIS update.

Our system constitutes a proposal of high efficiency, since admitting imperfections in the results, only GIS information and the RGB bands of a satellite or aerial color image are required for starting the automatic process.

References (73)

  • A. Moghaddamzadeh et al.

    A Fuzzy region growing approach for segmentation of color images

    Pattern Recognition

    (1997)
  • S.H.O. Park et al.

    Color image segmentation based on 3D clustering: Morphological approach

    Pattern Recognition

    (1998)
  • D.R. Peddle et al.

    Optimization of multisource data analysis: an example using evidential reasoning for GIS data classification

    Computers and Geosciencies

    (2002)
  • G. Rellier et al.

    Local registration and deformation of a road cartographic database on a SPOT satellite image

    Pattern Recognition

    (2002)
  • A. Sowmya et al.

    Modelling and representation issues in automated feature extraction from aerial and satellite images

    Journal of Photogrammetry and Remote Sensing

    (2000)
  • J.F. Yang et al.

    Color image segmentation using fuzzy C-means and eigenspace projections

    Signal Processing

    (2002)
  • Y. Zhong et al.

    Object location using color texture and shape

    Pattern Recognition

    (2000)
  • D. Zugaj et al.

    A new approach of color images segmentation based on fusing region and edge segmentations outputs

    Pattern Recognition

    (1998)
  • P. Agouris et al.

    Using a fuzzy supervisor for object extraction within an integrated geospatial environment

    International Archives of Photogrammetry and Remote Sensing

    (1998)
  • Andersen, H.E., Reutebuch, S.E., Schreuder, G.F., 2002. Bayesian object recognition for the analysis of complex forest...
  • Baumgartner, A., Hinz, S., Wiedemann, C., 2002. Efficient methods and interfaces for road tracking. In: Proceedings of...
  • Bhattacharya, A.K., Srivastava, P.K., Bhagat, A., 2001. A modified texture filtering technique for satellite images....
  • Belongie, S., Carson, C., Greenspanand, H., Malik, J., 1998. Color and texture based image segmentation using EM and...
  • Borgefors, G., 1988. Hierarchical chamfer matching: a parametric edge matching algorithm. In: IEEE Transactions on...
  • Bückner, J., Müller, S., Pahl, M., Stahlhut, O., 2002. Semantic Interpretation of Remote Sensing Data. ISPRS....
  • T.F. Burks et al.

    Classification of weed species using color texture features and discriminant analysis

    Transactions of the ASAE

    (2000)
  • T.F. Burks et al.

    Backpropagation neural network design and evaluation for classifying weed species using color image texture

    Transaction of the ASAE

    (2000)
  • Conci, A., Proenca, C.B., 1997. A box-counting approach to color segmentation. In: Proceedings of International...
  • F. Dell’aqua et al.

    Detection of urban structures in SAR images by robust fuzzy clustering algorithms: The example of street tracking

    IEEE Transactions on Geoscience and Remote Sensing

    (2001)
  • Dial, G., Gibson, L., Poulsen, R., 2001. IKONOS Satellite Imagery and its Use in Automated Road Extraction. Baltsavias,...
  • Eidenbenz, C., Käser, C., Baltsavias, E., 2000. Atomi-Automated reconstruction of topographic objects from aerial...
  • S. Fabre et al.

    Impact of Contextual Information Integration on Pixel Fusion

    IEEE Transactions on Geoscience and Remote Sensing

    (2001)
  • K. Fukunaga

    Statistical Pattern Recognition

    (1990)
  • Gagalowicz, A., 1987. Texture modelling applications. The Visual Computer, Springer Verlag, vol. 3, pp....
  • Cited by (169)

    • Vectorial and topologically valid segmentation of forestry road networks from ALS data

      2023, International Journal of Applied Earth Observation and Geoinformation
    • Automated road extraction using reinforced road indices for Sentinel-2 data

      2022, Array
      Citation Excerpt :

      A research article presented by Shao et al. [12] unveils an algorithm named as “Fast Line Feature Detector” to detect centerline of roads on a high resolution image as ribbon-like linear features. Another research work done by Mena and Malpica [13] uses high resolution RGB image to perform Texture Progressive Analysis (TPA) for road extraction and vectorization. We found another relevant research work that utilizes multi-tier segmentation on high resolution aerial images to differentiate roads in an urban area [14].

    • Correction, update, and enhancement of vectorial forestry road maps using ALS data, a pathfinder, and seven metrics

      2022, International Journal of Applied Earth Observation and Geoinformation
      Citation Excerpt :

      Most studies used optical passive remote sensing imagery and then used image classification approaches. The vast majority of studies have been both developed and then validated in urban or peri-urban areas (e.g. Mena and Malpica, 2005; Amo et al., 2006; Clode et al., 2007; Miao et al., 2013; see also the review by Wang et al., 2016). In a forest-specific context, Kearney et al. (2020) used RapidEye 5 m satellite imagery to map forestry roads using a Convolutional Neural Network.

    View all citing articles on Scopus
    View full text