an object-based method for contrail detection in avhrr ...gis.geog.uconn.edu/personal/paper1/journal...

412

GIScience & Remote Sensing, 2012, 49, No. 3, p. 412–427. http://dx.doi.org/10.2747/1548-1603.49.3.412Copyright © 2012 by Bellwether Publishing, Ltd. All rights reserved.

An Object-Based Method for Contrail Detection in AVHRR Satellite Images

Daxiang Zhang, Chuanrong Zhang,1 and Robert CromleyDepartment of Geography and Center for Environmental Sciences and Engineering, University of Connecticut, Storrs, Connecticut 06269

David Travis Department of Geography and Geology, University of Wisconsin-Whitewater, Whitewater, Wisconsin 53190

Daniel CivcoDepartment of Natural Resources and the Environment, University of Connecticut, Storrs, Connecticut 06269

Abstract: Contrails are important in understanding how aviation potentially affects climate change. The commonly used methods for detecting contrails usually only consider the spectral information of the image. This study focuses on introducing an object-based method for detecting contrails in AVHRR images. An object-based method utilizes not only the spectral characteristics, but also spatial information to aid in the detection. The results of the experimental study demonstrate that it is prac-tical to use an object-based classification method for contrail detection. Compared with pixel-based classification methods, the accuracy of the results is improved using the object-based method.

INTRODUCTION

Contrails are artificial clouds that are the visible trails of condensed water vapor made by the exhaust of aircraft engines. Contrail detection from satellite images is important for several reasons. For example, contrails can add a significant amount of high-level thin cloudiness over high-traffic areas as summarized in Seaver and Lee (1987), and the additional high thin cloudiness may lead to warmer surface tempera-tures (Liou, 1986); thus, contrails may have an influence on climate change. With the rapid increase in aviation in recent decades, it is important to understand the potential climatic impact of emissions from aircraft as well as the influence of persisting con-trails in lowering atmospheric temperature. Detecting contrails, therefore, is critical in understanding the atmospheric effects of aviation. In addition, contrails may mask other important landscape information in a satellite image. It is, however, difficult to detect contrails in satellite imagery because they are thin and their brightness values

1Corresponding author; email: [email protected]

contrail detection in avhrr images 413

are quite similar to the background landscape portrayed in an image. In addition, their shapes may be changed under the influence of wind and atmospheric dissipation, mak-ing them more difficult to identify.

Due to the aforementioned recent increase in high-altitude jet air traffic, there has been considerable research in the past decades focused on developing methods for detecting contrails. Early work on contrail detection was mainly based on the visual detection of contrails in satellite images. For example, Degrand et al. (1991) applied hard-copy displays of high-resolution Defense Meteorological Satellite Program (DMSP) data to identify contrails. Carleton and Lamb (1986) adopted DMSP Operational Linescan System (DMSP-OLS) data (with a spatial resolution of 600 × 600 m) into their research to detect contrails manually. Bakan et al. (1994) used a visual inspection method from Advanced Very High Resolution Radiometer (AVHRR) images to report contrail coverage over Europe and the North Atlantic. Degrand et al. (1991) manually interpreted contrails on satellite images over North America and in particular the United States. Travis (1996) manually counted pixels to determine the width and length of the contrails. Although the manual inspection methods performed well, these methods are highly subjective and time consuming.

To overcome the limitations of manual detection methods, researchers next attempted to develop automatic algorithms to detect contrails. Lee (1989) initiated a method for the automatic detection of contrails from AVHRR satellite images by using the brightness temperature difference in channels 4 (10.3–11.3 μm) and 5 (11.5–12.5 μm). Based on this concept, Engelstad et al. (1992) worked on pattern recognition algorithms to detect linear contrails. The algorithms included ridge detection and the Hough transform. Ridge detection is a pixel-based method to differentiate ridge pixels from background pixels and the Hough transform is applied to detect straight lines among these ridge pixels. The algorithms were not accurate because of spurious con-trail detection arising from linear streaks of natural cirrus clouds, and due to the fact that the Hough transform algorithm is limited in detecting curved contrails. Forkert et al. (1993) used a similar approach, but their method could sometimes misinter-pret linear features (such as coastlines, valleys, and cloud edges) as contrails. Weiss (1998) improved the ridge detection and Hough transform algorithms to create con-trail-enhanced images to aid in the detection process. When contrails are quite narrow, the search method proved to be efficient and overcame the false detection problems. However, this method is pixel-based and is less effective in processing short line seg-ments that constitute noise.

Later, neural networks were applied to contrail detection by Meinert et al. (1997). However, too much time and effort was needed to acquire well-chosen samples for their method and when substantial amounts of data were involved, the neural network model required a large amount of computational time to achieve acceptable detection results. Mannstein et al. (1999) introduced a method to detect linear contrail features by using a scene-invariant threshold and binary masks. A drawback of this method was that the masks were sometimes insufficient to remove all non–contrail edge features. Recently, other studies incorporated Mannstein’s automated algorithm to detect con-trails. For example, Meyer et al. (2007) detected the contrails in Thailand and Japan, and Palikonda et al. (2001) identified the contrails over some regions of the United States in AVHRR and MODIS images using this approach. More recently, Hetzheim (2007) proposed a very complicated approach using mathematical methods to detect

414 zhang et al.

contrails based on texture or stochastic behaviors of contrails. Although these math-ematical methods may better distinguish contrails from background features in the satellite images, they are still very time consuming.

In this research, we propose a new object-based method to detect contrails from satellite images. The advantage of the object-based method is that it offers new pos-sibilities to incorporate spectral values, texture, shapes, and contextual relationships in contrail detection. By taking into account both spectral characteristics of a single pixel and those of the surrounding contextual pixels and spatial characteristics of the surrounding pixels, this method may better recognize contrails in satellite images. The method may not only identify straight line–shaped contrails but also relatively curved contrails because it treats images as aggregated image objects instead of indi-vidual image pixels. Although object-based methods have been successfully applied to classify different spatial features from high-resolution satellite images such as IKONOS and QuickBird images (e.g. Herold et al., 2002; Mathieu and Aryal, 2005; Stow et al., 2007; Van Coillie et al., 2007; Zhou and Troy, 2008; Mladinich, 2010), and a few studies have used the method for feature detection in medium- or low- resolution satellite images (Geneletti and Gorte, 2003, Giada et al., 2003; Mitri and Gitas, 2004; Miliaresis and Kokkas, 2007), no research has reported on the method being used for contrail detection. The objective of this study is to test the suitability of an object-based method for contrail detection. We also compare the object-based method with the traditional pixel-based classification methods, such as the minimum distance method and the maximum likelihood method for recognizing contrails from low-resolution AVHRR satellite images.

METHODS

Figure 1 illustrates the methods used in this study. AVHRR images, which have a 1.1 km nadir point pixel resolution, are the primary input used for contrail detec-tion. We obtained the AVHRR data from the online Comprehensive Large Array-Data Stewardship System (CLASS) of the National Oceanic and Atmospheric Administration (NOAA) (www.nsof.class.noaa.gov). To ensure maximum resolution, only the High Resolution Picture Transmission (HRPT) data were obtained from within the broader AVHRR archive. AVHRR sensors have five bands or channels, which have different spectral characteristics.

Through visual inspection of many images, we determined that contrail features are most difficult to detect in the thermal infrared band (band three). Contrail features become vague in the visible red (band one) and near infrared (band two) bands and have similar radiance characteristics to the background in these two bands. Contrail features exhibit the most distinctive radiance characteristics relative to the background in thermal infrared bands four and five. Because bands four and five images are often duplicated, depending on the time of the images, we selected band four as the input data and used the Definiens Developer 8.0 software program to perform the object-based image analysis. The general process of the object-based image analysis consists of the following steps: the whole image was segmented first into meaningful pixel groups, or image objects, and then knowledge-based classification was performed to define these image objects into desired contrail or non-contrail classes. In the process, not only spectral information but also spatial information was considered. The two


main steps—segmentation and classification—in the object-based image analysis are outlined in the following sections.

Image Segmentation

As noted above, the first step in object-based image analysis is image segmen-tation. There are several segmentation algorithms available in Definiens Developer, ranging from simple algorithms (such as quad-tree based segmentation) to compli-cated procedures such as multi-resolution segmentation. Choosing an appropriate segmentation method and the method’s parameter settings are important for accurate contrail detection because the identification of contrail features is based on the image objects that are extracted through image segmentation, rather than individual pixels. The results of the image segmentation step will directly influence the overall perfor-mance of the object-oriented based classification. In this study, we used two segmenta-tion algorithms—quad-tree based segmentation and multi-resolution segmentation—to identify contrail features.

Quad-tree Based Segmentation. Quad-tree based segmentation algorithms divided the entire pixel-based region into a quad-tree grid, which is composed of square objects. Therefore, it is a top-down segmentation method and a typical segmentation algorithm using basically color differences within the image. The quad-tree structure

Fig. 1. Flowchart showing the methodology used in the study.

416 zhang et al.

was used to decompose continuously the image into blocks until a certain criterion was fulfilled. In the beginning, the quad-tree based segmentation method considered the entire image as a root node. In this process, an upper limit of color differences within each square, which is the segmentation criterion, was defined. In Definiens Developer, this parameter is referred to as the scale. It was used to check whether the root node satisfied the segmentation criterion, and if the segmentation criterion was not fulfilled in the region, the root node would be divided into four equal-sized square children regions. Otherwise, the region would be left unchanged. The four children regions were then checked successively to determine whether they fulfilled the crite-rion or not. This process was recursively applied to each region until the criterion for a node was accomplished, which means there were no child regions remaining. Thus, the quad-tree based segmentation method produced different-sized regions that might vary from a single pixel to the entire image.

Multi-resolution Segmentation. Multi-resolution segmentation extracted image objects at different hierarchical segmentation levels. Each subsequent level yielded image objects of a larger average size by combining objects from a lower level. Objects were grouped into a larger object based on a single parameter called heterogeneity, which was calculated based on three criteria including spectral similarity, contrast with the neighboring objects, and shape characteristics of the resulting object. There are four basic steps in the multi-resolution segmentation procedure. First, a single pixel in the selected image was chosen as the starting step of the segmentation pro-cedure. Second, the single pixel merged with other single pixels repeatedly, and they became larger units as long as they did not exceed the heterogeneity threshold. Third, a pixel group looked for its best suitable neighbor group for a potential merger. Because the multi-resolution segmentation algorithm is a mutual best-fitting approach, if the best suitable neighbor was not mutual, this neighbor would be chosen as a new root pixel to find its new best suitable neighbor. If the best fitting was mutual, then these two pixels would be merged together and became a new upper level image object as described in step two. Finally, all of the pixels in the image were tested until no further merging was possible.

Object-Based Classification

After segmentation, a nearest neighbor classification method was used to clas-sify the AVHRR image into two classes—contrail and non-contrail objects. This method used a set of training samples to assign contrail and non-contrail member-ship values. The classification results were tested based on the training image object samples. The entire image object domain was then classified based on the nearest sample neighbors.

Before objects can be distinguished as contrail and non-contrail classes, the spec-tral and spatial attributes of the objects need to be evaluated carefully for the purpose of determining their usefulness in the classification process. Shape, size, spectral char-acteristics, and the context of the contrail and non-contrail sample objects such as contrast and relative location were examined. Based on these results, contrail and non-contrail training samples from the segmented objects were selected. For example, if the contrail image objects are characterized by bright regions compared to non-contrail image objects in the whole image, 5–10% of the objects in the image histogram that


have high mean values as contrail samples were chosen, while 5–10% of the objects that have low mean values in the histograms were chosen as non-contrail background image objects.

In order to get good object-based classification results, several iterations of train-ing sample selection and refinement were conducted. Initially, only a small number of contrail and non-contrail samples were chosen. After the classification process, more contrail and non-contrail samples were added if many image objects were misclassi-fied. The process was repeated until the classification results were satisfactory.

Nearest neighbor classification returned a contrail or non-contrail fuzzy member-ship value to each of the segmented image objects based on its feature space distance to its nearest neighbors. The fuzzy logic values of the feature space distance were calculated based on the spectral and shape features. Brightness was used as a spectral feature and the shape index and width as shape features to compute the fuzzy logic values. The following formulae were used to calculate these three features:

Br = ci (1)

(2) S l4 A-----------=

W = #pv/γv (3)

where Br (brightness) is the spectral mean values ci of an image object; l is the border length of the image objects; A is the area of the image objects; S is the shape index, which is the border length l of the image object divided by four times the square root of its area A; #pv is the total number of pixels contained in pv; γv is the length/width ratio of an image object v; and W is the width, which is calculated by #pv/γv.

Brightness is the sum of mean values ci of the layers containing the spectral infor-mation divided by their quantity nb. The fuzzy logic values of each segmented image object were determined based on the calculated feature space distance using the fol-lowing equation:

(4) dvf

svfo

σf----------⎝ ⎠⎜ ⎟⎛ ⎞

f∑

2

=

where d is the distance value between sample objects and the image object o; vfs is the

feature value of the sample object; vfo is the feature value of the image object; and σf is

the standard deviation of the feature value.Because shape features and color features might not be represented under a com-

mon scale and might have different value ranges, the fuzzy logic values were stan-dardized by dividing the standard deviation of both the shape and color feature values. Thus the final obtained fuzzy logic values ranged from 0 to 1.0. A value of 0 means the object does not belong to the contrail class, whereas a value of 1.0 means the object most likely belongs to the contrail class.

418 zhang et al.

RESULTS AND DISCUSSION

Object-Based Classification Results

Six scenes of the AVHRR images with contrails (left column of Figure 2) were processed using the aforementioned methods. These scenes represent different sce-narios for evaluating the object-based classification method. Table 1 provides basic information about the six AVHRR images, including the date and time collected, their spatial extent, and a brief description of the contrails in each image. These images were chosen because they can help to test the capability of the object-based method in identifying contrails from different backgrounds in satellite images. The multi-resolu-tion segmentation method was used to divide the images into various objects for clas-sification. In the image segmentation step, the settings for the several parameters were established. The first parameter is scale. Through visual inspection, we found that with an increase in the scale value the total number of image objects decreased significantly. In all of these scenarios, the contrails were usually several pixels wide. Therefore, a larger scale value would decrease the precision of the contrail detection. After many trials, we found that the segmentation results depicted contrails reasonably well at a segmentation scale value of 5, and it was chosen in the segmentation process.

Both spectral and spatial heterogeneity criteria were applied in the segmenta-tion process to avoid generating branched and fractured objects. Three parameters are related to three heterogeneity criteria: color, smoothness, and compactness. The latter two are shape criteria. The three parameters have a total weight of 1.0. Due to the fact that a contrail pixel (or object) in the image has a much higher value than that of a non-contrail pixel (or object), we gave the color parameter an initial weight greater than 0.5. Different color parameter values (greater than 0.5) were tested to determine the best delineation of contrail objects. Finally, a color parameter of 0.9 was set for all of the scenarios to emphasize the brightness information. The shape parameter was set to 0.1 and both compactness and smoothness were 50% of the shape criterion,

Table 1. List of AVHRR Images

ID Date and time Spatial extent Scenario description

1 Oct 16, 2002 at 2:37 p.m.

–90.28 to –89.2837.86 to 38.17

One main contrail is in the image

2 Oct 14, 2002 at 1:43 p.m.

–86.35 to –85.2938.14 to 38.66

Contrails lie close to one another

3 Oct 16, 2002 at 2:37 p.m.

–88.47 to –87.7042.51 to 42.78

Contrails are densely distributed in the image with obvious noise data

4 Oct 3, 2002 at 7:33 p.m.

–74.64 to –74.1244.91 to 45.50

Contrail with curved shape and short-line- segment contrail appear in the image

5 Oct 16, 2008 at 10:44 p.m.

–82.71 to –82.2631.72 to 32.27

Contrails cross one another in the image

6 Oct 21, 2008 at 11:35 a.m.

–90.71 to –89.9835.79 to 36.63

Very short segment contrails are sparsely distributed in the image


Fig. 2. Results of the object-based classification under six different scenarios. The left column (A, C, E, G, I, K) shows the original AVHRR images. The right column (B, D, F, H, J, and L) shows the final classification results.

420 zhang et al.

because a certain degree of shape homogeneity often can improve the quality of object extraction.

Multiple contrail and non-contrail training samples, evenly distributed across the entire image, were selected. The purpose of feature selection is to reduce the compu-tational requirements while preserving the overall classification accuracy. Through visual inspection and checking of many object features of the segmented image objects, we found that brightness, shape index, and width could optimize the three-dimensional feature space.

The right column in Figure 2 shows the final results of the object-based classifica-tion for the six different scenarios. The first scenario represents the case of a simple contrail. The classification results show that the straight-line shape of the contrail was identified accurately in the image and the linear feature was retained well.

The second scenario shows contrails in close proximity to one another. It can be seen that in the middle of Figure 2C that two contrails are closely located relative to one other, yet do not overlap. The classification results (Fig. 2D) show that the object-based method confused the two contrails, joining them as a single segment, although preserving well the linear features of the contrails. This verifies that the object-based method captures the spatial structure and the linear nature of the contrails.

In the third scenario, the contrails were densely distributed in the image and there was obvious noise data in the upper part of the image. The classification results show that the object-based method distinguished the other image objects, which contain a very dark background, from the contrail image objects, and both the spatial continu-ity and patterns of the contrails are maintained. Basically, all of the contrails were detected using the object-based method.

In the fourth scenario (Fig. 2G), a curve-shaped contrail is located in the upper part of the image. Existing methods have problems detecting such curved contrails as they focus mainly on detecting straight-line segments. The object-based classification results (Fig. 2H) demonstrate that curve-shaped contrails can be detected, because the object-based method uses not only spectral features in the analysis, but also other object features such as the extent and shape index.

In the fifth scenario, the detection of two overlapping contrails in some parts of the image was tested. The detection of spatially overlapping contrails has not been previously reported in the literature. The classification results show that most of the overlapping parts of the contrails can be detected and the two contrails, though they intersect one another, are still represented as two separate contrails after classifica-tion. Non-contrail background cells were basically excluded from the contrail cells, although the black circle in Figure 2J suggests that image objects adjacent to the overlapped part of the contrails can still pose a problem. Overall, contrails having a straight-line shape were detected well and not confused with each other.

The sixth scenario tested contrails with very short lengths within the image. Existing methods usually cannot detect these contrail pixels because of their small size. However, these short segments can be effectively detected as contrails using the object-based classification method because the method not only uses color or bright-ness in the classification criteria, but it also uses other object-related spatial features such as a shape index and width to constitute the feature space. Using more compre-hensive feature criteria instead of only one spectral criterion improves the classifica-tion accuracy.


Comparison between Quad-tree Based Segmentation and Multi-resolution Segmentation

Because the first step in object-based image analysis is image segmentation, two segmentation algorithms—quad-tree based segmentation and multi-resolution segmentation—were compared to find out the best segmentation algorithm for this study. In order to compare the segmentation results, the same scale of 5 that indicates the upper limit of brightness differences within the image was used in the segmenta-tion process. The middle column of Figure 3 shows the results of the quad-tree based segmentation and the right column of Figure 3 shows the results of multi-resolution segmentation. From these results, it can be seen that the quad-tree based method can-not provide a clear view of the contrails included in the original images. The segmen-tation method split the whole image into many tiny squares and the contrails cannot be readily identified. Another drawback of this segmentation method is that there is not much difference between it and the simple pixel-based segmentation methods, since quad-tree based segmentation also only considers pixel values in the segmentation process. Therefore, quad-tree based segmentation is not as suitable as multi-resolution segmentation for contrail detection. Compared to the quad-tree based segmentation, the multi-resolution segmentation depicts the rough borders of the contrails well, as illustrated in the right column of Figure 3. This method can handle objects with differ-ent spectral properties and geometric shapes, and allows for more precise and accurate extraction of contrail object features. Thus, multi-resolution segmentation is a better method for the preparation step in the object-based image analysis.

Comparison between Object-Based Classification and Pixel-Based Classification

For the pixel-based classification, we used maximum likelihood (ML) and mini-mum distance (MD) classification methods. Image 1 in Figure 4A and Image 4 in Figure 4E were randomly selected for comparison tests with the object-based method. The images were classified into two classes: contrail and non-contrail. Figure 4B shows the results of the classification using the ML classification method in the first image. The white area was classified as the contrail class and the dark area was classi-fied as the non-contrail class. From the figure, it can be seen that the contrails cannot be detected well and contrail class pixels and non-contrail class pixels are confused with each other. Contrail pixels and non-contrail pixels are scattered throughout the image, and many contrail pixels are incorrectly separated. Therefore, it is difficult to distinguish the two classes using the ML classification method.

Figure 4C shows the results of the classification using the MD method for the first image. Compared with the maximum likelihood method, the results are improved to some extent. The stripe-like contrails can be roughly detected and most of the contrail pixels have become continuously detected in the image. However, the contrails still have been broken into several parts, hence appearing fractured. Figure 4D shows the results of the object-based classification for the first scenario. There are no discon-nected parts in the contrails. Contrail and non-contrail objects are clearly separate.

The pixel-based classification and object-based classification results for the fourth image can be seen in Figures 4F–4H. Figure 4H illustrates that there are no discon-nected parts of the contrails in the object-based classification results, and the contrail

422 zhang et al.

Fig. 3. The results of the quad-tree based segmentation and multi-resolution segmentation under the six different scenarios. The left column (A, D, G, J, M, and P) shows the original AVHRR images with contrails. The middle column (B, E, H, K, N, and Q) shows the results of the quad-tree based segmentation. The right column (C, F, I, L, O, and R) shows the results of the multi-resolution segmentation.


pixels remain continuous. The results of the ML and MD classification methods both show disconnected parts of contrails.

To assess the accuracy of the pixel-based and object-based approaches, 30 samples were randomly selected and assessed for their correctness. Table 2 shows the accuracy

Fig. 4. Contrail detection comparison between pixel-based classification and object-based clas-sification. A. First original AVHRR image. B. Result of the ML classification method for the first image. C. Result of MD classification for the first image. D. Result of object-based method for the first image. E. Fourth original AVHRR image. F. Result of the ML classification method for the fourth image. G. Result of the MD classification for the fourth image. H. Result of object-based method for the fourth image.

424 zhang et al.

assessment results. Three assessment indices—producer’s accuracy, user’s accuracy, and the kappa statistic—are included. From Table 2, one can see that the object-based classification produced more accurate results. For the first scenario, the overall accu-racy for the object-based method is 90%, whereas the overall accuracy for the ML classification method is 73% and the overall accuracy for the MD method is 86%. Among the three accuracy assessment results, the object-based method also produced the highest overall kappa statistic, 0.80. For the fourth image, the overall accuracy for object-based classification method is 90%. The overall accuracy for ML classi-fication method is 73%, and the overall accuracy for MD method is 83%. Moreover, the overall kappa statistics increase markedly for the object-based method, increasing from 0.47 to 0.80 and from 0.67 to 0.80, respectively. The results demonstrate that object-based classification can produce more accurate results because the object-based classification method includes important object feature information for interpreting the AVHRR contrail images.

DISCUSSION AND CONCLUSION

The results of the experimental study demonstrate that it is practical to use the object-based classification method for contrail detection. The object-based method overcomes the limitations of pixel-based methods by combining both spatial and spectral information into the classification process. The method takes advantage of

Table 2. Accuracy Assessment Results from Pixel-Based and Object-Based Classification

Class name

Maximum likelihood method

Minimum distance method

Object-based method

Producer’s accuracy

%

User’s accuracy

%


%

User’s accuracy

%


%

User’s accuracy

%

First scenario

Contrail 100.00 46.67 92.31 80.00 100.00 80.00

Non- Contrail

65.22 100.00 82.35 93.33 83.33 100.00

Fourth scenario

Contrail 81.82 60.00 91.67 73.33 87.50 93.33Non-

contrail 68.42 86.67 77.78 93.33 92.86 86.67

First scenario

Overall accuracy

73.33 86.67 90.00

Overall Kappa

0.4667 0.7333 0.8000

Fourth scenario

Overall accuracy

73.33 83.33 90.00

Overall Kappa

0.4667 0.6667 0.8000


using other supplemental information besides spectral brightness to differentiate con-trail pixels from non-contrail pixels. The incorporation of spatial information makes it possible to achieve more accurate results. The results also show that the object-based classification method has advantages in processing curve-shaped contrails and short-segment contrails, likely due to the method segmenting the whole image into contrail objects instead of treating pixels individually; the formation of the contrail objects under homogeneity criteria improved the accuracy for identifying the curved contrails by considering the shape information.

A comparison of quad-tree based segmentation versus multi-resolution segmen-tation for contrail pixel classification shows that the multi-resolution segmentation performed better. The quad-tree based segmentation considered only the pixel values, so it produced many tiny squares and could not segment the entire image into clear contrail view. However, the multi-resolution method could segment the entire image into well-defined contrail objects, which are readily analyzed and classified.

A comparison with pixel-based classification methods also demonstrates that the classification accuracy is significantly improved with the object-based method. Although the maximum likelihood classification performed better among the pixel-based classification methods, it still produced poorer results by only using spectral information without considering texture and contextual information. The pixel-based classification is more likely to mistakenly identify those neighboring pixels, which have a similar pixel value to contrail pixels but are not real contrail objects, as contrail pixels. The object-based classification method is more robust because all pixels of the contrail objects are guaranteed to be assigned to the same class and hence the classifi-cation results are closer to those based on human vision. It is necessary to create image objects that can represent the real contrail objects using more information than simply pixel values. In this study, although there are undesirable conditions in the satellite images such as noise in the data, heterogeneous atmospheric conditions, and some interference of the background data, the performance of the object-based classification method is still good.

Although the object-based classification method can detect contrails successfully in the AVHRR satellite images, there are still limitations of the method. For example, the classifier used in this paper is fuzzy nearest neighbor classifier, which looks like a “black box.” Given the number of dimensions in the feature space, although the feature space can be optimized, users do not know the mechanism that differentiates an object into a certain class and thus cannot control the classification process. In addi-tion, the choices of the training samples usually have to be repeated many times to be properly evaluated. This restricts the automation of contrail detection using the object-based method. Finally, the object-based method cannot effectively detect atypical con-trails such as spread ones, which are hard to distinguish from natural cirrus clouds. When contrails diffuse over a large area, they can form cirrus clouds. Distinguishing contrails from cirrus clouds can be difficult because the cirrus clouds usually have a homogeneous appearance and similar brightness values to contrails. These clouds and contrails usually are so extensive that they are virtually indistinguishable from one another and the individual line shapes of contrails disappear. Thus it is difficult to distinguish contrails from cirrus clouds using the normal object features—brightness, shape index, and width. Some possible object characteristics that may prove useful in distinguishing contrails from cirrus clouds should be studied in future work.

426 zhang et al.

ACKNOWLEDGMENTS

We thank the anonymous reviewers and the editor for their comments and sugges-tions on the manuscript. We especially appreciate the helpful revision from Dr. Jason A. Tullis. This research is partially supported by USA NSF grant No-0819396. The authors have sole responsibility for all of the viewpoints presented in the paper.

REFERENCES

Bakan, S., Betancor, M., Gayler, V., and H. Grassl, 1994,“Contrail Frequency over Europe from NOAA-Satellite Images,” Annales Geophysicae, 12(10):962–968.

Carleton, A. M., and P. J. Lamb, 1986, “Jet Contrails and Cirrus Clouds: A Feasibility Study Employing High-Resolution Satellite Imagery,” Bulletin–American Meteo-rological Society, 67(3):301–309.

Degrand, J. Q., Carleton, A. M., and P. J. Lamb, 1991, “A Mid-season Climatology of Jet Condensation Trails from High-Resolution Satellite Data,” in Proceedings of the 7th Conference on Atomspheric Radiation, San Francisco, California, 309–311.

Engelstad, M., Sengupta, S. K., Lee, T., and R. M. Welch, 1992, “Automated Detec-tion of Jet Contrails Using the AVHRR Split Window,” International Journal of Remote Sensing, 13(8):1391–1412.

Forkert, T., Strauss, B., and P. Wendling, 1993, “A New Algorithm for the Automated Detection of Jet Contrails from NOAA AVHRR Satellite Images,” in Proceedings of the 6th AVHRR Data User’s Meeting, Belgirate, Italy, 29 June–2 July 1993, 513–519.

Geneletti, D., and B. G. H. Gorte., 2003,“A method for object-oriented land cover classification combining landsat TM data and aerial photographs,”International Journal of Remote Sensing, 24(6):1273–1286.

Giada, S., De Groeve, T., Ehrlich, D., and P. Soille, 2003,“Information Extraction from Very High Resolution Satellite Imagery over Lukole Refugee Camp, Tanzania,” International Journal of Remote Sensing, 24(22):4251–4266.

Herold, M., Scepan, J., Müller, A., and S. Günther, 2002, “Object-Oriented Mapping and Analysis of Urban Land Use/Cover Using IKONOS Data,” in Proceedings of the 22nd Earsel Symposium Geoinformation for European-Wide Integration, 4–6.

Hetzheim, H., 2007, “Automated Contrail and Cirrus Detection using Stochastic Prop-erties,” International Journal of Remote Sensing, 28(9):2033–2048.

Lee, T. F., 1989, “Jet Contrail Identification Using the AVHRR Split Window,” Jour-nal of Applied Meteorology,28:993–995.

Liou, K. N., 1986, “Influence of Cirrus Clouds on Weather and Climate Processes: A Global Perspective,” Monthly Weather Review, 114:1167–1199.

Mannstein, H., Meyer, R., and P. Wendling, 1999,“Operational Detection of Con-trails from NOAA-AVHRR Data,” International Journal of Remote Sensing, 20(8):1641–1660.

Mathieu, R. and J. Aryal, 2005, “Object-Oriented Classification and IKONOS Multi-spectral Imagery for Mapping Vegetation Communities in Urban Areas,” in Proceedings of the 17th Annual Colloquium of the Spatial Information Research


Centre, University of Otago, Dunedin, New Zealand [http://otago.ourarchive.ac.nz/handle/10523/740], last accessed March 22, 2012.

Meinert, D., Stuhlmann, R., and E. Raschke, 1997, “Impact of Jet Condensation Trails upon the Radiation Budget over Northwest Europe and the North Atlantic,” in Pol-lutants from Air Traffic—Results of Atmospheric Research 1992–1997, Cologne, Germany: DeutschesZentrumFuerLuft-und Raumfahrt, 161–169.

Meyer, R., Buell, R., Leiter, C., and H. Mannstein, 2007, “Contrail Observations over Southern and Eastern Asia in NOAA/AVHRR Data and Comparisons to Contrail Simulations in a GCM,” International Journal of Remote Sensing, 28(9):2049–2069.

Miliaresis, G. and N. Kokkas, 2007, “Segmentation and Object-Based Classification for the Extraction of the Building Class from LIDAR DEMs,” Computers & Geo-sciences, 33(8):1076–1087.

Mitri, G. H. and I. Z. Gitas, 2004, “A Performance Evaluation of a Burned Area Object-Based Classification Model When Applied to Topographically and Non-topographically Corrected TM Imagery,” International Journal of Remote Sens-ing, 25(14):2863–2870.

Mladinich, C., 2010, “An Evaluation of Object-Oriented Image Analysis Techniques to Identify Motorized Vehicle Effects in Semi-arid to Arid Ecosystems of the American West,” GIScience & Remote Sensing, 47(1):53–77.

Palikonda, R., Minnis, P., Heck, P. W., Sun-Mack, S., Q. Trepte, Z., and H. Mannstein, 2001, “Contrail Cover and Radiative Properties from High-Resolution Satellite Data,” in Proceedings of the AMS 11th Conference on Satellite Meteorology and Oceanography, 508–511.

Seaver, W. L. and J. E. Lee, 1987, “A Statistical Examination of Sky Cover Changes in the Contiguous United States,” Journal of Climate and Applied Meteorology, 26:88–95.

Stow, D., Lopez, A., Lippitt, C., Hinton, S., and J. Weeks, 2007, “Object-Based Clas-sification of Residential Land Use within Accra, Ghana Based on QuickBird Satellite Data,” International Journal of Remote Sensing, 28(22):5167–5173.

Travis, D. J., 1996, “Variations in Contrail Morphology and Relationships to Atmo-spheric Conditions,” Journal of Weather Modification, 28(4):50–58.

Van Coillie, F., Verbeke, L. P. C., and R. R. De Wulf, 2007, “Feature Selection by Genetic Algorithms in Object-Based Classification of IKONOS Imagery for Forest Map-ping in Flanders, Belgium,” Remote Sensing of Environment, 110(4):476–487.

Weiss, J. M., 1998, “Automatic Contrail Detection and Segmentation,” IEEE Transac-tions on Geoscience and Remote Sensing, 36(5):1609–1619.

Zhou, W. and A. Troy, 2008, “An Object-Oriented Approach for Analysing and Char-acterizing Urban Landscape at the Parcel Level,” International Journal of Remote Sensing, 29(11):3119–3135.

an object-based method for contrail detection in avhrr ...gis.geog.uconn.edu/personal/paper1/journal...

Documents