|
Adaptive Multi-Image Matching Algorithm for the Airborne Digital Sensor ADS40
Radiometric Aspects and Image Preprocessing
Radiometric Quality Analysis Although ADS data are 14-bit, the effective number of bits is less (ca. 12 according to LHS). Histograms of the Level 0 images that were processed show a strong peak towards the darker grey values (about 310-340 in the panchromatic channels). The effective grey level range was 13 bits for the panchromatic channels and 12 for the multispectral, excluding grey values at the histogram ends with frequency less than ca. 0.0012% (5 times less the frequency occurring if all 14-bit values were equally occupied). A noise analysis has been performed for different datasets in Level 0 and Level 1 images, according to the method presented in (Baltsavias et al, 2001). The noise characteristics were analysed in non-homogeneous areas and the standard deviation, as an indication of noise, was calculated with a 3x3 mask size every 3 x 3 pixels, for various grey value ranges. Table 2 shows results for the Waldkirch data, whereby the Pan results represent the average of the 3 channels. Table 2. Noise estimation (mean standard deviation) of Waldkirch data using non-homogeneous areas.
The standard deviations indicate that noise is generally low and even more for the Level 1 images, due to the resampling of pixels during rectification. The multispectral channels exhibit slightly less noise than the panchromatic. Noise is increasing with increasing grey values. Some grey value ranges have no reliable standard estimation, due to the small number of samples. These results are consistent with the noise analysis that has been performed by DLR (Bφrner et al., 2000), stating that the standard deviation is about 2 grey values. Image Pre-processing In order to optimise the images for subsequent computer processing, various pre-processing methods were developed, implemented and tested. All pre-processing was applied to the 14-bit data. First, two adaptive local filters were developed (Baltsavias et al., 2001). They reduce noise, while sharpening edges and preserving even fine details such as one-pixel wide lines, corners and line end-points. The effect of the two local filters is generally quite similar, although they use a different number and size of masks, and one employs a fuzzy method. They require as input an estimate of the noise, which may be known or estimated by the method mentioned above. Table 3 shows the noise estimation before and after the adaptive filtering. After filtering, noise is reduced by about factors 2 to 3. Table 3. Noise estimation (mean standard deviation) using non-homogeneous areas before and after noise reduction, using the Waldkirch dataset.
Next, a new version of the Wallis filter (Baltsavias, 1991), which estimates automatically some of the filter parameters, is applied, and finally a reduction to 8-bit imagery by histogram equalisation is performed. With Wallis, the contrast and thus also the noise is enhanced, but since the noise has been reduced in the first preprocessing stage, the noise level still remains under the noise level of the original images. Histogram equalisation is used, since it preserves more grey values that are more frequently occurring. The histogram equalisation was iterative with the aim being to occupy all 8 bits with similar frequencies. The histogram equalisation may lead to too strong bright and dark regions for visual interpretation, and thus in such cases it can be replaced by a histogram normalisation (Gaussian form). Figure 1 shows the radiometric improvement of 8-bit Level 1 images after pre-processing. 
Figure 1. Detail of roof (Level 1 data after Wallis filtering and reduction to 8-bit by histogram normalisation) without (left) and with (right) noise reduction. On the right, noise is reduced, edges are sharpened and small details are preserved. Matching General principles Methods for automatic DTM and DSM generation exist already in various commercial digital photogrammetric systems, including SOCET SET of LH Systems. All existing matching algorithms are geared towards frame aerial imagery, with a usual overlap of 60% per stereopair; while in matching only two images are used. The ADS40 however, offers new capabilities, which make necessary the development of new matching algorithms based on another philosophy. ADS40 provides full 100% overlap of 3 to 7 images for each strip. Thus, a multi-image matching approach is followed, leading to substantial reduction of problems caused by occlusions, multiple solutions, image noise, and surface discontinuities and higher measurement accuracy through the intersection of more than two image rays. The known interior and exterior orientation are used to enforce geometric constraints, restricting the search space along quasi-epipolar lines. The sensor model developed by LH Systems is used in all transformations between image and ground coordinates. The main aim of the developed methods is to derive a DSM/DTM from Level 1 images, and a by far secondary one is to measure tie points or derive a coarse DSM (using Level 0 or Level 1 images), results which can be used in the bundle adjustment. In the first matching stages, cross-correlation or other similarity or difference measures are used delivering pixel accurate results, which can be improved by subsequent sub-pixel matching methods, e.g. least squares matching. Level 1 images generally exhibit small scale and rotation differences, especially in the upper pyramid levels, and thus the use of only shifts in matching, without additional shaping parameters, is justified, leading to faster processing than other time consuming matching methods like least squares matching. Apart from normalised zero-mean cross correlation, the sum of squared differences or of the absolute differences can be used and thus further decrease processing time. Although investigations have reported that these difference measures are inferior to normalized zero-mean cross correlation, e.g. sensitive to radiometric differences, after the pre-processing and radiometric equalization that we apply, first tests show that these measures perform equally well. Matching is performed at pixels of extracted features (Sec. 3.2) and approximations are derived by a modified image pyramid concept (Sec. 3.3). In matching, different type of image information can be used: original grey value images, binary images by thresholding the edge magnitude, images where edge magnitude is kept and non-edges are set equal to 0, variations of the two previous image types, by encoding through appropriate values information on the edge pixel orientation or sign. The rationale behind using image types other than the original grey values, and especially binary ones, was a possible matching speed-up, especially in the upper pyramid levels, where not the outmost accuracy is sought. Initial tests showed that coarse matching with binary images was slightly faster and the accuracy was similar compared to the points matched using grey level images. Therefore, binary images may be used in the upper pyramid levels without significant loss in accuracy but not in the finer pyramid levels. |