Height Determination of Extended Objects Using Shadows in SPOT Images

Building heights were estimated with relatively high accuracy using shadows in a set of single-look SPOT panchromatic and multispectral images taken from the same satellite simultaneously. Shadows cast b y rows of trees in SPOT images were first used to estimate mean heights of trees. Calibration lines were then constructed to relate the actual mean heights of rows of trees to the estimated heights. Using these calibration lines, heights of some industrial buildings in the image were estimated using their shadows with sub-pixel accuracy. The accuracy achieved is better than$hree metres, or one-third the pixel size of the SPOT panchromatic image. One of the important challenges involved in the process was to determine an appropriate threshold for delineating shadow zones in the images. A technique is provided for this problem. The technique is useful for estimating heights of extended objects situated in pat terrains. The type of resampling used for overlaying a multispectral image over a panchromatic image changes the accuracy of height estimation. However, the change is tolerable if the heights to be estimated ore within the ground-truth data range used for deriving calibration lines. Introduction In the last two decades enormous gains have been made in processing civilian satellite images. This paper presents a quantitative application of remote sensing in building height estimation in the sub-pixel range using SPOT images. The procedure has significant applications in defense for wide-area surveillance, urban monitoring, and environmental studies. Photogrammetrists have been able to extract heights of objects from aerial photographs using parallax in stereo-pair photographs. The lengths of the shadows cast by the objects are also used to determine heights (Huertas and Nevatia, 1988; Liow and Pavlidis, 1990). If the sun and sensor geometry are known, it is fairly simple to establish a relationship between shadow lengths and the heights of objects. The above usages, however, are confined to high resolution photographs, with pixel resolutions much better than the object heights being measured. In a related area, coarse resolution Synthetic Aperture Radar (SAR) images have been used for mapping terrain elevations using shape from shading techniques (Wildey, 1986; Frankot and Chellappa, 1987; Guindon, 1989). These are extensions of shape from shade studies in optical images. In these techniques, the relationship between local shading variations and local slope variations is used to determine terrain shape. The technique assumes that the scattering coefficient is a constant for the area and also that the scattering cross section does not vary with the variation of local incrdence angle. Defence Science and Technology Organisation, Salisbury, SA 5082, Australia. G.M. Sumerling is presently with the Intergraph Corporation, Kew, Victoria 3101, Australia. Although shadows have been used successfully for estimating heights of objects in high resolution aerial photographs, the use of civilian satellite images, such as SPOT, for height estimation of buildings reported in the literature is limited. There have been some reports in this direction lately (Cheng and Thie1,1995; Hartl and Cheng, 1995; Meng and Davenport, 1996). The main reason for the scarcity in applications is that the resolution of the civilian satellite images is much coarser than that of the aerial photographs and shadows are not well defined for short and commonly occurring objects. This causes problems in determining shadow widths that are needed for estimating heights. The solution to this problem is a major issue considered in this paper. A Summary of Related Reports The papers of Cheng and Thiel (1995), Hartl and Cheng (1995), and Meng and Davenport (1996) are closely related to the contents of this paper. Cheng and Thiel (1995) have used a SPOT panchromatic image to estimate building heights using shadows. They have achieved a root-mean-square error of 3.69 m in height estimation of 42 buildings. It should be noted that the mean of their building shadow widths is on the order of 7.7 uixels and that building heights varied from 8 to 66 m. In the present study, the av&age <hadow width is less than 2 pixels. Hence, the challenges involved in achieving sub-pixel accuracy are greater in our study. However, we have achieved an accuracy better than that achieved by Cheng and Thiel (1995). Our estimation of height is indirect, using calibration curves, while Cheng and Thiel (1995) have computed heights directly. The number of buildings in our study is much smaller than theirs. In this paper we have discussed the role of thresholding and have developed a procedure for selecting an optimal threshold for measuring shadow width with sub-pixel accuracy. After the selection of the threshold, the shadows are delineated by segmenting the whole image. Cheng and Thiel (1995) have used an adaptive procedure for thresholding each shadow. Hartl and Cheng (1995) is the continuation of Cheng and Thiel (1995). Here they have reported a root-mean-square error of 6.13 m in height estimation. The study involved height measurement of 77 buildings. Meng and Davenport (1996) have developed a procedure to fix shadow boundaries, which is crucial for height estimation, with sub-pixel accuracy. They claim that the edges can be fixed to 11100th of a pixel. Unfortunately, they have not provided any ground-truth data for accuracy assessment. For accurately locating the shadow boundaries, they create an edge-image template using the point-spread function of the Photogrammetric Engineering & Remote Sensing, Vol. 64, No. 1, January 1998, pp. 3544. 0099-1112/98/6401-0035$3.00/0