Laserfiche WebLink
MEMO 89.2 <br />2.2.2 Geometric Correction <br />The geometric correction consisted of georeferencing the Landsat images to Digital Ortho Quadrangles <br />(DOQ) projected to the UTM coordinate system, Datum NAD27, Spheroid Clarke 1866. This procedure <br />allowed us to combine the high spatial accuracy of the DOQ with the multi-spectral information of the <br />Landsat imagery. For this purpose, tiled DOQs were displayed side-by-side with the frames using the <br />Image Geometric Correction Tool in ERDAS Imagine. The geometric correction process includes initial <br />identification of 3-5 points manually, which are then used by ERDAS Imagine to auto-locate additional <br />points between the uncorrected Landsat image and reference DOQ. These additional points are all <br />analyst inspected for accuracy and individually accepted or modified. Adequate numbers (17-22) of <br />Ground Control Points (GCPs) were then manually obtained on-screen for each Landsat frame to generate <br />a first order polynomial relationship between the Landsat and the DOQ. The GCPs were selected using <br />highly visible features on the Landsat frame that could be positively identified on both the Landsat and <br />the DOQ (Figure 2), and that are stable over time (e.g. road intersections or buildings). <br />Figure 2. Example of a Ground Control Point (GCP) Selected on (a) the Landsat Image and (b) on the <br />DOQQ <br />In the case of Landsat scenes 034/032, 034/033, located over mountainous terrain, a Landsat <br />orthorectification model included in ERDAS Imagine was used. This model is derived by space resection <br />based on colinearity equations and requires elevation information for removing relief displacement <br />(Imagine 8.7 Online Documentation, 2003). Consequently, the Landsat images were orthorectified using <br />DOQs, and a 30-meter Digital Elevation Model (DEM). At one pixel of RMS for Landsat data, any given <br />feature within the corrected imagery could potentially be up to 30 meters away from its actual location <br />(Landsat pixel = 30m). The Root Mean Square errors (RMS) achieved in the georeferencing process <br />were under 0.5 pixels or less than 15 meters in all cases, and thus deemed high quality results well within <br />acceptable limits. Table 1 shows the number of GCPs selected for each Landsat frame in the SPDSS <br />study area, and their corresponding root mean square (RMS) errors. RMS errors summarized in Table 1 <br />are reported directly from the ERDAS Imagine Geometric Correction Tool. This level of RMSE should <br />be replicable by awell-trained careful analyst using similar quality datasets. <br />Page 3 of 45 ~RFversfde TecAnotogy, fnc. <br />4'JaYCr Resources Errgi~ecr:np an~i CansaFlrnp <br />(b) noQ wAn27, Clarke Is66~ <br />(a) Non-rectified Landsat TM <br />