Automated DEM/DSM accuracy estimates towards land change detection Jasmee Jaafar & Gary Priestnall Department of Geography, University ofnottingham, Nottingham NG7 2RD Email:jaafar@geography. nottingham. ac. uk Abstract Recognition of buildings and other man-made objects from aerial images is an issue of considerable importance to many users of geo-information, including surveyors, geographers and planners. Integrating a priori information into procedures for the automatic recognition and reconstruction of cartographic objects from imagery has been examined by a number of researchers (Locherbach [1], Priestnall and Glover [2]). The general assumption that buildings are higher than their surrounding surface is used by various researchers such as Weidner and Forstner [3] to aid in detecting buildings from imagery. The results suggest that subtraction of a Digital Elevation Model (DEM) from a Digital Surface Model (DSM) is a feasible technique for detecting man-made objects from imagery. However, since DEM and DSM have inherent levels of accuracy, these accuracy levels will eventually be inherited by the residual surface (DSM-DEM). It is important to understand the inherent accuracy of both the DEM and DSM if the building height derived from these sources is to be used profitably in an object recognition strategy. This paper focuses on the extraction of building heights derived from the subtraction of the DEM from the DSM. Global Positioning Systems (GPS) and traditional surveying techniques have been employed to provide ground reference data for the validation of DEM and DSM which comprises an area of the University of Nottingham campus. The reliability of automated validation of DEM and DSM using synthetic ground reference data generated by a mathematical function along a selected profile is proposed. The results suggest that possible building structures are revealed above the residual surface and can also be detected below the residual surface by generating contour plots at various levels. The automatic validation of DEM and DSM proposed is a promising approached to the task of building extraction since it minimises time and cost by eliminating the need to gather ground reference data.
74 GIS Technologies and their Environmental Applications Introduction Several promising techniques of automatic feature extraction from digital imagery have produced significant results in restricted domains (Weidner and Forstner [3]). However, there is no single technique which is adequate to solve all the problems in automatic image recognition. Priestnall and Glover [2] noted that, "the successful recognition of geographic objects from imagery requires a more global understanding of the image than local pixel-based approaches based upon spectral characteristics". The integration of extra cues or knowledge has been suggested by Shufelt and McKeown [4] to overcome the single technique restriction. Co-operative approaches using context information such as maps suggest that off-the-shelf software and available cartographic data may provide a powerful tool for feature recognition as suggested by Kisner [5]. In automated land use change detection, Priestnall and Glover [2] use vector databases to guide object recognition. This work highlighted the potential of height information in providing a useful cue in activities such as change detection using existing vector databases to 'train' a change detection system, and showed how other cues to a building's existence in the image would re-enforce this procedure. The subtraction of a DEM from a DSM shows a promising approach to the identification of man-made objects. However, understanding the accuracy estimates of the digital model (DEM and DSM) used is a crucial factor (Jaafar and Priestnall [6]). In this paper an approach towards building recognition for land use change detection using DSM and DEM subtraction is highlighted. Evaluation of the accuracy estimates of both DSM and DEM at various grid resolutions are carried out with the aid of ground reference data, and an alternative method towards automating DEM and DSM accuracy estimates is proposed. Methodology The study area chosen is the university campus area of Nottingham. The locality allowed the convenient collection of detailed control data for analysis. The area of interest for the study contains buildings of various heights and dimensions and, in common with urban/industrial environments, contains other surface objects such as trees and cars.
GIS Technologies and their Environmental Applications 75 A reference DEM was created within ARC/Info from contour lines derived from Ordnance Survey 1: 10 000 map data and spot elevations from 1:1250 maps. A DSM, which ideally models the man-made objects as well as the terrain, was then constructed using colour aerial photographs provided by the National Remote Sensing Centre (NRSC) and ERDAS Imagine Orthomax software. The scale of the photographs is 1: 10 000 and the photographs were scanned at a resolution of 21 jum. The stereo base:height ratio is 0.6. Figure 1 shows the procedure of extracting 'blobs' which are potential candidates for the detection of man-made objects and outlines the automated procedures for evaluating the constructed model. The automated approach for analysing the model is based on synthetic ground reference data generated along selected profiles. In this study various grid resolutions of the constructed model are analysed. The grid format is currently the most common approach to store and analyse elevation data in a Geographic Information System (GIS) (Carrara et al [7]). For comparison, evaluation of the models was also carried out based on check points established using GPS and terrestrial surveying techniques with known orders of accuracy Root Mean Square Error The measures of accuracy adopted in this study is the Root Mean Square Error (RMSE) of the elevational residuals at randomly selected check points. RMSE is calculated from the following equation (Gao, [8]). RMSE = -z, where n is the number of check points; Z* i is the 'true' elevation at position i and Z; is the derived elevation at check point i. One hundred and seventy nine randomly selected points were established. Table 1 shows the corresponding estimated accuracy of the derived check points including the determination of the building's height.
76 GIS Technologies and their Environmental Applications DSM Construction and Evaluation DEM Construction and Evaluation Aerial Photos (Scanned Digital Data ) Profile Selection Orlbomax Jflcopy Dgtl. P'rutryl X y,», j»*», % X y, "21 *M» \, m puled coordinate e ng selected profile i xcluding building ling ) j z. *j y ~^3 f *3 *«yj Z^ Coo r< > i^o selectedp inu o files X»» * X * y y Computed coordio along selected prof Digital Elevation Model (DEM) Digital Surface Model ( Various Odd Size ) AMI, Programmiog kg gu«8 -I i i Zurverf Xpert Package Hf. ^X^v/ Distance His. at coordinated pi. from DSM specific interval on profile DSM - DEM (Residual Surface) Figure 1: Automated evaluation of the digital model. Method GPS Traversing Trig. Levelling (R.L. of Roof) Accuracy Estimates Y 0.02m 0.01 m 0.10m A 0.02m 0.18m 0.10m L 0.01 m 0.03m 0.10m Table 1: Accuracy estimates with respect to method employed..
GIS Technologies and their Environmental Applications 77 Synthetic Ground Reference Data Tempfli [9] regards a terrain profile as a continuous space signal denoted by/(x). If the constructed curve is defined by/%), then the construction error can be shown as; (2) and the accuracy is given by; L Mean square error = J e^ (x) dx/l o (3) where [0,L] is the interval of fix). In reality, a true representation of the profile can be replicated if only critical points on the profile which represent its shape can be determined by terrestrial surveying or by photogrammetric methods. An alternative approach proposed in this study is to generate a mathematical function f(x) using the elevation values available from maps along a selected profile. Heights at 0.5m intervals along the profile (Synthetic ground reference data) are then determined. Using equation (2), the elevational residuals will then be the difference between the Synthetic ground reference data and the height derived from the digital model at every 0.5m as shown in figure 2. H&t Bstenoe Figure 2: The Digital Model (DSM or DEM) mathematical function profile (MF) contrived using point A,B,C,D,E and F elevational residual (d\,di,d$ &W at check point 1,2,3...n ( 0.5m interval).
78 GIS Technologies and their Environmental Applications Profile Nine profiles within the study area were constructed. Figure 3 shows the profiles constructed to generate the Synthetic ground reference data. Figure 3: Profiles chosen to generate synthetic ground reference data. The criteria for the construction of profiles are as follows : (i) profiles are constructedfrom two established known positions, where possible the profile begins and ends at known benchmark locations. (ii) profiles represent all possible directions in the four azimuthal quadrants; and (iii) the profiles represent various degrees of elevation. In order to determine the magnitude and direction of the profiles, the profiles were constructed from two distinct positions. Related elevations along the profile can then be identified clearly and compared with the digital model. Selection of profiles running in various directions is required to conduct the investigation globally. The final criteria ensures that the full range of terrain variability is represented. Mathematical Function The mathematical function representing the profile is derived using spot heights and contour values along the selected profile. The choices of mathematical function (cubic spline, tension spline, linear, etc) to represent the profiles will be a major factor in determining its adequacy of fit. Furthermore, the spatial arrangement of the elevational data on the
GIS Technologies and their Environmental Applications 79 selected profile can have a significant effect in the construction of the profile (Mather [10]). In this study the linear spline is adopted to represent the selected profile. The elevation value derived from fix) at 0.5m intervals are retrieved. RMSEofDEMandDSM Figure 4 shows the RMSE of the DEM and DSM at various grid resolutions using ground reference data and Synthetic ground reference data. 2.5 2.9 1.9 1.7 10m Gnd Rtsolutto Figure 4: RMSE plot of the DEM Oround. SyHthtttt at various grid resolutions. The RMSE of the DEM is seen to increase gradually at resolutions greater than 15m using ground reference data. This tendency has been noted by others such as Gao [8]. Surprisingly, the RMSE using Synthetic ground reference data decreases gradually from 1m to 30 m grid resolution. This effect could be due to the generalisation effect as the grid resolution increases. Since the Synthetic ground reference data are retrieved from the mathematical function based on selected height values from the same source as used in constructing the DEM, this will eventually result in small residuals. This factor might also be the reason for the smaller RMSE computed as compared to ground reference data. Figure 5 shows the RMSE of the DSM using ground reference data and Synthetic ground reference data for grid resolutions between 1 and 5m. There are no significant trends in the RMSE plot with varying grid resolutions for the DSM using ground reference data. Further studies relating to this effect will be carried out in the near future. The RMSE of the DSM using Synthetic ground reference data, however, increases gradually between 1m and 5m in grid resolution. This effect is the reverse of what is experienced using Synthetic ground reference data on DEM evaluation.
80 GIS Technologies and their Environmental Applications Ground Synthetic Figure 5: RMSE plot of DSM at various grid resolutions. This might be due to the limited number of control points used for the matching process in the DSM creation. As a result, residuals decrease in magnitude as the grid resolution increases. It is important to clarify that the use of ground reference data evaluates the DEM/DSM with respect to the 'real world', whereas Synthetic ground reference data will eventually compare the interpolation outcome along the selected profiles. However, both give an indication of the DEM/DSM accuracy estimates in their own 'way'. RMSE of residual surface The RMSE of the residual surface is computed based on the measured height of existing buildings at thirty five known positions. The RMSE plot of the residual surface (DSM-DEM) is shown in Figure 6. Referring to Figure 6, the RMSE of the residual surface are between 6m and 7m. The RMSE for the residual surface does adhere with the sum of the accuracy estimates of the DEM and DSM using ground reference data. It is worth pointing out that due to the uncertainty level (± 7m), man-made objects of a certain elevation might not be shown above the residual surface. To illustrate this concept, Figure 7 shows the contour plot of the residual surface at -1m, Om and 1m. It is shown that the building edges can be detected below the residual surface at -1m..25m 1 in 3m 5m Grid Resolution Figure 6: RMSE of the residual surface.
GIS Technologies and their Environmental Applications 81-1 m 0 m 1m Figure 7: Contour plot of the residual surface at -1m, Om and 1m. Conclusion In this paper the possibility of detecting building structures using subtraction between DSM and DEM is demonstrated. It is shown that building structure might not only be shown above the residual surface but below the surface due to errors in digital model construction. Therefore, it is possible that buildings which are a few meters in height might not be revealed as 'blobs' in the residual surface but as 'submerged blobs'. It is also shown that, using Synthetic ground reference data the accuracy estimates of the digital model can be determined. Even though, it will not relate to the 'real world', it will give a good approximation of the uncertainty level to which the residual surface should be interrogated towards an automated extraction procedure. Acknowledgements The authors wish to thank Prof. P.M. Mather for his valuable suggestions, supervision and critical review of the manuscript. The authors wishes to acknowledge Mr. Kenny Gibson, from the Institute of Engineering Surveying and Space Geodesy, University of Nottingham for his assistance in the GPS survey. Imagery data from the National Remote Sensing Center. References [1] Locherbach, T, Reconstruction of land-use units for the integration of gis and remote sensing data, in H Ebner, C. Heipke & K. Eder, eds, * Spatial Information from Digital Photogrammetry and Computer Vision', Vol. 30/3, ISPRS, Munich Germany, 1994.
82 GIS Technologies and their Environmental Applications [2] Priestnall G. And Glover R., A control strategy for automated land use detection: An integration of vector-based GIS, remote sensing and pattern recognition. Proceedings of the Geographical Information Systems Research - UK Conference (GISRUK 97), 9-11 April, 1997, University of Leeds. Leeds UK [3] Weidner U.,and Forstner W, Towards automatic building extraction from high resolution digital elevation models, ISPRS Journal of Photogrammetry & Remote Sensing, 50 (4), pp. 38-49, 1995. [4] Shufelt J. and Mckeown D.M., Fusion of monocular cues to detect man-made structure in aerial imagery, Computer Vision, Graphics, and Image Processing: Image Understanding, 57(3);307-330, May, 1993. [5] Kisner D, An integrated system to evaluate the potential of remotely sensed building height information as input into an automated change detection strategy, Msc. Dissertation in GIS, Department of Geography, University of Nottingham, September 1997. [6] Jaafar J. and Priestnall G, A critical evaluation of the potential of automated building height extraction from stereo imagery for land use change detection, to be presented at the Geographical Information Systems Research - UK Conference (GISRUK 98), 31-2 April, 1998, University of Edinburgh, Edinburgh. [7] Carara, A., Bitelli, G and Carla',R., Comparison of techniques for generating digital terrain models from contour lines. International Journal of Geographical Information Science, Vol 11, No. 5,451-473, 1997. [8] Gao, Jay, Resolution and accuracy of terrain representation by grid DEMs at a micro scale. International Journal of Geographical Information Science, Vol 11, No. 2,199-212, 1997. [9] Tempfli, K, Spectral analysis of terrain relief for the accuracy estimation of Digital Terrain Models. LT.C Journal, 1980-3: 478-510, 1980. [10] Mather, P.M., Computational methods of multivariate analysis in physical geography. (John Wiley & Sons), 1976.