Keywords: AVHRR, geometric correction, registration, GCP, matching, edge line
Abstract This paper describes a geometric registration method for compositing AVHRR images. In the method, we consider two problems to improve geometric registration accuracy. The first is a development of automated detection procedure for GCPs locations in AVHRR imagery. The second is to develop image to map registration procedure for fine correction. Using the existing and developed methods, the geometric registration is carried out in two stages. In the first stage-correction, an accuracy of AVHRR image navigation is increased by automated technique of GCPs identification and correction of terrain elevation. In the second, the navigated AVHRR image is overlaid on reference coastline image for determination offsets and a final correction is carried out by registration of the navigated image to fixed reference image. This method was implemented by using "C" under UNIX operating system in HP workstation and has been operationally used for geometric registration of AVHRR composite data without operator intervention. The registration software package is applicable to any pre-processed AVHRR data for fine geometric correction.
1. Introduction
The availability of daily Advanced Very High Resolution Radiometer (AVHRR) imagery which covers very large regions has been shown to be useful for monitoring global and continental environmental conditions. AVHRR imagery has major advantages such as high frequent overpass and wide scanning provide the chance of producing cloud free images by compositing AVHRR images collected in regular intervals (3). However, for users of the data, it has been often showed difficult to pre-process and produce multi-temporal AVHRR composite data set. One of the most difficult problems in the compositing process is to provide high accuracy of geometric correction for each individual scene to be composited.
This paper considers improvements on geometric correction in two ways and designing full registration system for AVHRR composite data. The first improvement is the increase of systematic correction accuracy using automated detection procedure of GCPs locations in AVHRR imagery and elevation data. The second is to develop an automated image registration of navigated AVHRR images to fixed map image.
2. Method
The approach of accurate geometric registration used in the compositing AVHRR data consists of two-stage correction procedures. In the first stage of correction is to calculate ground location of the image pixels using an orbital model and it includes two steps: identification of GCPs and AVHRR image navigation. The second stage correction is motivated by the need to increase multitemporal positional accuracy of the navigated images for compositing requirement and it includes estimation of navigation accuracy, determination of mapping function and resampling. A flow of geometric registration method for composite AVHRR data set is shown in Figure 1. In the registration approach, we use PaNDA package for image navigation procedure and the rest of all steps were developed.
Figure 1. Processing steps for geometric registration of AVHRR data
3. Development Of Geographical Database
The registration procedure developed requires the use of geographical database consisting of GCPs data set, ground control matching pattern (GCMP), reference overlay image of geographic features such as coastline and rivers and elevation data. Therefore, these geographical databases were created in the software development stage.
The geographical database used in the geometric registration includes coastline vector data, elevation data, and library of GCPs and GCMPs. These data are used to precise identification of GCPs positions in satellite and ground projections, and accurate matching of shape similarity. The coastline vector data were created from Digital Chart of the World (DCW). However, the vector data created from only DCW does not fully satisfy our requirement, because data for rivers are not available in DCW. When image to map registration is used, the digital data for rivers are extremely important where lakes or coastal data are not available. Therefore, the hydrological feature of rivers was generated from water mask data (EROS DATA Center) and added to the vector data. Each of the points included in the vector data, has latitude, longitude and elevation values.
The coastline vector data was converted to reference overlay image of coastline. GCPs data set was created from the reference overlay image by selecting easily detectable and most suitable locations for matching procedure. Using the reference overlay image, land and water mask image had been created and then it was used to generate bitmap image chips with 32 by 32 pixels size for pattern matching of shape similarity. Each of these chips is assigned GCP location, which we are referred to as GCMP.
The digital elevation data used for correction for terrain elevation were generated from GTOPO30 according to whole coverage area of the final composite data set.
4. First-Stage Correction
4.1. Identification of GCPs locations in AVHRR imagery
The shifts of GCPs locations in AVHRR image from the corresponding locations in the reference overlay image is determined by matching image window of edge line derived from AVHRR image with binary template image extracted from the reference overlay image. We refer these images as to searching window and reference window respectively and they are created using same ground point selected from the GCP data set.
The reference window of edge line is created in the following way. Before the matching procedure, once the reference overlay image had been created in satellite projection from the coastline vector data described in section 3. The reference overlay image includes coast and shorelines of the area, which is exactly same as whole coverage area of the raw AVHRR imagery. Then the reference window of 32 by 32 binary data is extracted from the reference overlay image centered GCP position.
The searching window from raw AVHRR imagery is created in the following way. At first, an image window of 64 by 64 data is extracted from raw AVHRR channel 1, 2 and corresponding pixel numbers centered at selected GCP position (Figure 2, 2nd and 3rd columns). The channel 1 and 2 were calibrated and a suitability of the image window for the subsequent matching process is tested by cloudiness and sensor scan angle.
A searching window of edge line for the matching is created from the image window by edge extraction using NDVI value, an adaptive thresholding and filtering. Thus three searching windows of edge line for each GCP were produced: first one by NDVI value (Figure 2, 4-th column) and second one by spatial filtering method (Figure 2, 5-th column) and third one by optimal thresholding method (Figure 2, 6-th column). In the searching window, cloudy pixels belonging to edge line are not used for matching.
After the reference window and searching windows had been generated, a similarity between the two windows is determined by image correlation procedure. In the result of image correlation procedure, the location of selected GCP was determined simultaneously in the three searching windows and the searching window with maximum correlation coefficient was selected (Figure 2, the area where best match occurs is marked by small window in the searching windows).
