 |  |
| Bildiri Özetleri | |
| Ana Sayfaya Dönüş |
| Back To The Main List |
|
Salih AYDEMIR, M. Ali Çullu Harran University, Faculty of Agriculture, Department of Soil Science, Şanlıurfa, Turkey Abstract Four digital images acquired during the 1998-growing season of a cornfield were used to test and evaluate the radiometric correction process of histogram matching. The digital images were acquired with a new digital aerial system (ADAR 5500). The study focused on determining if radiometric correction was necessary for these digital images, which were acquired in four different times during the growing season (June 1, July 2, September 11, and September 25). Point samples of pseudoinvariant features (road surfaces) were taken to determine average spectral responses in each band of the four images. Plots of average spectral reflectance versus the date of the aerial image yielded multi-temporal classifications that demonstrated that radiometric correction was necessary. Due to the simplicity of the process and the lack of detailed measurements required in other physical based models, histogram matching was chosen as the most appropriate method. Three of the images were referenced to the July 2 scene using this technique. Multi-temporal classifications demonstrated that radiometric correction was accurate in delivering an invariant radiometric reflectance for the road surfaces. The usefulness of the matched (i.e. radiometrically corrected) images was then investigated by attempting to distinguish variations in yield for several ground control points. The matched images displayed improved separation between different yield values showing definite distinctions between areas of high and low yield. Results The first step in the procedure required locating possible PIF features that could be used to investigate changes in reflectance values. The roads that are located around the cornfield are PIF features that should exhibit a common reflectance for the various time periods. These roads are common in all four images in terms of the extent to which they are seen in the image. In addition, the roads form a nice boundary around the field of interest. Therefore, by investigating changes in the reflectance values of this invariant feature, any changes that occur should be related to non-scene-dependent changes. Point samples were taken along the stretch of the roads at particular latitude and longitude points. A total of 100 point samples along the stretch of the road yielded different average reflectance values in each band throughout the growing season. The results are shown in Figure 1, with the vertical axis in terms of the average reflectance values for three of the four bands. From this figure, it is easily seen that radiometric correction is necessary in order to adjust the images to a normalized scene more appropriate for direct comparison. Therefore, radiometric correction was determined to be necessary, and the next question then shifted to the appropriate method for correction. However, before moving directly into the next step, the ancillary data for the area, especially in terms of the rainfall and moisture content data, was investigated to determine if the differences in reflection values for the roads were due to these variables. Daily rainfall amounts were obtained and investigated for the area, and from this investigation, it was concluded that no rainfall had fallen on or at least three days before the time in which the images were taken. In addition, the times in which the images were taken were relatively dry periods in an otherwise wet growing season. Therefore, no rainfall interactions were seen between changes in reflectance values for the roads in the four images. Thus, the images had approximately equivalent moisture contents, and no water was standing on the paved-road surfaces. Two main correction procedures were identified in order to perform simple radiometric corrections on images. Physical based models were not considered possible alternatives due to a lack in atmospheric, sensor, and scene data required by these models. The first technique is a function available in the ERDAS image processing software called histogram matching. Histogram matching attempts to correct for atmospheric and scene characteristics independent of any relationship between the images. In fact, histogram matching is a purely statistical technique that relates the cumulative density function of one image to the density function of another. Because the July 2 image was determined as the most valuable in predicting crop yield and other vegetation properties within the growing season, this image was selected as the reference image. Histogram matching was performed on the other images, referencing each back to the July 2 image. Results of the correction process are shown in Figure 2. Exactly the same point samples were taken along the road, and the 100 samples were averaged to determine the overall reflectance values shown in Figure 2. As can be seen in the figure, radiometric correction performed well, adjusting the reflectance values of the selected PIF feature. The final step in the study involved investigating the usefulness of the matched images in possibly delineating yield or determining coarse-level management zones within the field. By investigating the original spectral profiles of the corn, the difficulty of using unsupervised and supervised classifications was occurred. In addition, an attempt to calculate the NDVI (Normalized Difference Vegetation Index) for both the original and matched images was performed. However, the calculated NDVI did not relate to yield within the field. Therefore, the question that remained involved whether subtle differences were lost in converting the original images to match the reference image, or if the matched images provided greater separation between the spectral reflectance values of the cornfield. Several ground control points were used in this part of the research. These points were used due to detailed ground-truth and yield data for the particular locations within the field. Results indicated that by considering the multi-temporal response of the corn according to the infrared (band 4), differences in yield could be detectable with the matched images. In terms of the original images, no patterns could be seen in the multi-temporal response that indicated any difference in yield for the ground control points (Figure 3a). However, for the matched images, clear distinctions between the areas with good yield and areas with poor yield were observable in the results, as shown in Figure 3b. Conclusions Radiometric correction with histogram matching has proven to be a useful tool in correcting differences between atmospheric and scene characteristics with multi-temporal images taken throughout the growing season of a particular crop. The atmospheric, sensor, and scene characteristics are sure to differ between remotely sensed images acquired throughout a growing season, suggesting the usefulness of a simple radiometric correction scheme. More specifically, reflectance values for PIF features remained relatively invariant throughout a four-month period after histogram matching was performed to relate the images to a reference scene. In addition, the matched images proved useful in distinguishing between different areas of crop yield within the field, especially when compared to the original images. Many factors may be contributed to these results, all of which may not be completely known. Therefore, further investigation into the techniques, principles, and processes of both histogram matching and multi-temporal classification needs to be performed in order to make more generalized conclusions, not dependent on the particular area under research. In addition, further work needs to be carried out to understand more detailed and promising ways of calculating yield, especially in figuring out the problems with the NDVI calculations for this particular field. |