Peter Doucette

Self-organised clustering for road extraction in classified imagery

Abstract The extraction of road networks from digital imagery is a fundamental image analysis operation. Common problems encountered in automated road extraction include high sensitivity to typical scene clutter in high-resolution imagery, and inefficiency to meaningfully exploit multispectral imagery (MSI). With a ground sample distance (GSD) of less than 2 m per pixel, roads can be broadly described as elongated regions. We propose an approach of elongated region-based analysis for 2D road extraction from high-resolution imagery, which is suitable for MSI, and is insensitive to conventional edge definition. A self-organising road map (SORM) algorithm is presented, inspired from a specialised variation of Kohonens self-organising map (SOM) neural network algorithm. A spectrally classified high-resolution image is assumed to be the input for our analysis. Our approach proceeds by performing spatial cluster analysis as a mid-level processing technique. This allows us to improve tolerance to road clutter in high-resolution images, and to minimise the effect on road extraction of common classification errors. This approach is designed in consideration of the emerging trend towards high-resolution multispectral sensors. Preliminary results demonstrate robust road extraction ability due to the non-local approach, when presented with noisy input.

Automated Road Extraction from High Resolution Multispectral Imagery

This work presents a novel methodology for fully automated road centerline extraction that exploits spectral content from high resolution multispectral images. Preliminary detection of candidate road centerline components is performed with Anti-parallel-edge Centerline Extraction (ACE). This is followed by constructing a road vector topology with a fuzzy grouping model that links nodes from a self-organized mapping of the ACE components. Following topology construction, a Self-Supervised Road Classification (SSRC) feedback loop is implemented to automate the process of training sample selection and refinement for a road class, as well as deriving practical spectral definitions for non-road classes. SSRC demonstrates a potential to provide dramatic improvement in road extraction results by exploiting spectral content. Road centerline extraction results are presented for three 1 m colorinfrared suburban scenes which show significant improvement following SSRC.

Automated Extraction of Linear Features from Aerial Imagery Using Kohonen Learning and GIS Data

An approach to semi-automated linear feature extraction from aerial imagery is introduced in which Kohonens self-organizing map (SOM) algorithm is integrated with existing GIS data. The SOM belongs to a distinct class of neural networks which is characterized by competitive and unsupervised learning. Using radiometrically classified image pixels as input, appropriate SOM network topologies are modeled to extract underlying spatial structures contained in the input patterns. Coarse-resolution GIS vector data is used for network weight and topology initialization when extracting specific feature components. The Kohonen learning rule updates the synaptic weight vectors of winning neural units that represent 2-D vector shape vertices. Experiments with high-resolution hyperspectral imagery demonstrate a robust ability to extract centerline information when presented with coarse input.

Spatiospectral cluster analysis of elongated regions in aerial imagery

The extraction of road networks from digital imagery is a fundamental operation in geospatial applications. In images captured by new satellite sensors with a ground sample distance of less than 2 meters per pixel, roads can be broadly described as elongated regions. We introduce a novel technique of spatiospectral cluster analysis in which the spatial properties of elongated regions are identified from unsupervised analysis of their corresponding spectral properties. Preliminary results demonstrate a fully automated process in which road centerline topology can be identified in high-resolution aerial imagery in the presence of typical clutter.

An evaluation methodology for vector data updating

The methods used to evaluate automation tools are a critical part of the development process. In general, the most meaningful measure of an automation method from an operational standpoint is its effect on productivity. Both timed comparison between manual and automation based-extraction, as well as measures of spatial accuracy are needed. In this paper, we introduce the notion of correspondence to evaluate spatial accuracy of an automated update method. Over time, existing vector data becomes outdated because 1) land cover changes occur, or 2) more accurate overhead images are acquired, and/or vector data resolution requirements by the user may increase. Therefore, an automated vector data updating process has the potential to significantly increase productivity, particularly as existing worldwide vector database holdings increase in size, and become outdated more quickly. In this paper we apply the proposed evaluation methodology specifically to the process of automated updating of existing road centerline vectors. The operational scenario assumes that the accuracy of the existing vector data is in effect outdated with respect to newly acquired imagery. Whether the particular approach used is referred to as 1) vector-to-image registration, or 2) vector data updating-based automated feature extraction (AFE), it is open to interpretation of the application and bias of the developer or user. The objective of this paper is to present a quantitative and meaningful evaluation methodology of spatial accuracy for automated vector data updating methods.

Automated vector-to-raster image registration

The variability of panchromatic and multispectral images, vector data (maps) and DEM models is growing. Accordingly, the requests and challenges are growing to correlate, match, co-register, and fuse them. Data to be integrated may have inaccurate and contradictory geo-references or not have them at all. Alignment of vector (feature) and raster (image) geospatial data is a difficult and time-consuming process when transformational relationships between the two are nonlinear. The robust solutions and commercial software products that address current challenges do not yet exist. In the proposed approach for Vector-to-Raster Registration (VRR) the candidate features are auto-extracted from imagery, vectorized, and compared against existing vector layer(s) to be registered. Given that available automated feature extraction (AFE) methods quite often produce false features and miss some features, we use additional information to improve AFE. This information is the existing vector data, but the vector data are not perfect as well. To deal with this problem the VRR process uses an algebraic structural algorithm (ASA), similarity transformation of local features algorithm (STLF), and a multi-loop process that repeats (AFE-VRR) process several times. The experiments show that it was successful in registering road vectors to commercial panchromatic and multi-spectral imagery.

A Methodology for Automated Vector-to-Image Registration

Registration and alignment of feature (e.g., vector) and raster geospatial data is a difficult and time-consuming process when performed manually. This paper presents an approach for vector-to-raster registration. Candidate features are auto-extracted and vectorized from imagery, which are the basis to compare against existing vector layer(s) to be registered. Given that automated feature extraction (AFE) methods are imperfect, the objective is to determine and gather a sufficient signal-to-noise ratio from AFE upon which to base a registration process between vector data sets. Two vector registration methods were investigated. The first is based on an algebraic structural algorithm (ASA) in which structural components (e.g., angles, lengths and areas) are used as similarity metrics. The second is based on a similarity transformation of local features (STLF) in which a 4-parameter transformation is used to align features on a local basis. Experiments were performed to register road vector data to commercial panchromatic and multispectral QuickBird imagery.

Spatial analysis of image registration methodologies for fusion applications

Data registration is the foundational step for fusion applications such as change detection, data conflation, ATR, and automated feature extraction. The efficacy of data fusion products can be limited by inadequate selection of the transformation model, or characterization of uncertainty in the registration process. In this paper, three components of image-to-image registration are investigated: 1) image correspondence via feature matching, 2) selection of a transformation function, and 3) estimation of uncertainty. Experimental results are presented for photogrammetric versus non-photogrammetric transfer of point features for four different sensor types and imaging geometries. The results demonstrate that a photogrammetric transfer model is generally more accurate at point transfer. Moreover, photogrammetric methods provide a reliable estimation of accuracy through the process of error propagation. Reliable local uncertainty derived from the registration process is particularly desirable information to have for subsequent fusion processes. To that end, uncertainty maps are generated to demonstrate global trends across the test images. Recommendations for extending this methodology to non-image data types are provided.

The Sequential Generation of Gaussian Random Fields for Applications in the Geospatial Sciences

This paper presents practical methods for the sequential generation or simulation of a Gaussian two-dimensional random field. The specific realizations typically correspond to geospatial errors or perturbations over a horizontal plane or grid. The errors are either scalar, such as vertical errors, or multivariate, such as , , and errors. These realizations enable simulation-based performance assessment and tuning of various geospatial applications. Both homogeneous and non-homogeneous random fields are addressed. The sequential generation is very fast and compared to methods based on Cholesky decomposition of an a priori covariance matrix and Sequential Gaussian Simulation. The multi-grid point covariance matrix is also developed for all the above random fields, essential for the optimal performance of many geospatial applications ingesting data with these types of errors.

An approach for evaluating assisted target detection technology

The literature is replete with assisted target recognition (ATR) techniques, including methods for ATR evaluation. Yet, relatively few methods find their way to use in practice. Part of the problem is that the evaluation of an ATR may not go far enough in characterizing its optimal use in practice. For example, a thorough understanding of a methods operating conditions is crucial, e.g., performance across different sensor capabilities, scene context, target occlusions, etc. This paper describes a process for a rigorous evaluation of ATR performance, including a sensitivity analysis. Ultimately, an ATR algorithm is deemed valuable if it is actually utilized in practice by users. Thus, quantitative analysis alone is not necessarily sufficient. Qualitative user assessment derived from user testing, surveys, and questionnaires is often needed to provide a more complete interpretation of an evaluation for a particular method. We demonstrate our ATR evaluation process using methods that perform target detection of civilian vehicles.