OBJECT-BASED CHANGE DETECTION IN HIGH RESOLUTION IMAGE
Gang Hong Yun Zhang Department of Geodesy and Geomatics Engineering University of New Brunswick 15 Dineen Dr., P.O. Box 4400 Fredericton, New Brunswick, CANADA E3B 5A3 Telephone: 1-506-4473397
[email protected] Daniel A. Lavigne Defence R&D Canada - Valcartier 2459, Pie-XI Blvd North, Val-Belair (QC), CANADA G3J 1X5
ABSTRACT Change detection is the process of identifying differences in the state of an object or phenomenon by observing it at different times and is one of the most important image processing and analysis tools for a variety of military and civilian operations. Numerous change detection techniques have been developed since the successful launch of Landsat 1 in 1972. Because of the constant improvements of spatial resolutions and other image characteristics of available airborne/spaceborne electro-optic imagery, object-based change detection is becoming available and is often required in civilian and military application area. However, many challenges have been brought for the traditional change detection to implement the object-based change detection among high resolution images. For example, the post-classification comparison method is very popular and widely used in change detection, but when it is used in object-based change detection, the problem is that it is difficult to classify the interested objects correctly. This paper has introduced a change detection method which is based on the machine learning algorithm, which can overcome the traditional change detection method limitation and find the interested changed objects. High resolution airborne imagery is used to test the method. Keywords: airborne, spaceborne, change detection, high resolution.
INTRODUCTION Change detection is the process of identifying differences in the state of an object or phenomenon by observing it at different times. With the launch of the first Earth Resources Technology Satellite in 1972, the remote sensing method for change detection became available. The basic premise in using remote sensing data for change detection is that changes in the object of interest will result in changes in radiance values or local texture that are separable from changes caused by other factors, such as differences in atmospheric conditions, illumination and viewing angle, soil moisture etc (Singh, 1989). With the technology development, more and more Earth observation satellites have been launched, and different resolution images have been available for meeting different application requirements. For example, AVHRR (1km resolution) provides ideal data for global area change detection; while the high resolution images such as IKONOS (1m resolution) and QuickBird (0.7m resolution) make urban change detection possible and more economical. Because of the constant improvements of spatial resolutions and other image characteristics of available airborne/spaceborne electro-optic imagery, object-based change detection is becoming available and is often required in civilian and military application areas. While high resolution image also brings challenges to the traditional image classification methods: with the spatial resolution refinement, the internal variability within homogenous land cover units increases. The increased variability decreases the statistical separability of land-cover classes in the spectral data space, which leads to reduce per pixel classification accuracies by using traditional classification algorithm, because these methods have predominately used basic statistical methods, such as Maximum Likelihood. Also these methods require a priori statistical assumptions and often fail on small-feature capture because of their inability to (a) classify disjunctive concepts, (b) take into account spatial context, and (c) ASPRS 2006 Annual Conference Reno, Nevada May 1-5, 2006
remove clutter (Opitz, 2003). The increased variability was attributed to the imaging of diverse class components by higher resolution sensors, whereas at coarser resolutions, sensors integrated the reflected spectral radiance of the various components, and classes appeared more homogeneous than high resolution image (Carleer, et al., 2005). Thus, the classification algorithm to process high resolution image must be more detailed than traditional classification algorithm because the spectral response of each pixel is associated with very specific earth and manmade materials. In this research, the objective of change detection is object-specific, the first step in this research should be how to extract the object-specific existing different images from other non-interested objects respectively, and then compare the extracted object in two images for change detection analysis. In this paper, one method based on machine learning algorithm which includes spatial context of objects is used to classify the specific object. High resolution airborne image is used to test the method.
METHODOLOGY The change detection method used in this research is implemented by using software called Feature Analyst, which employs machine learning algorithms, such as Nearest Neighbor, Neural Networks, Decision Trees, Genetic Ensemble Feature Selection (a patent pending technology) and others, to efficiently extract user-defined features. Hierarchical learning methods are involved to iteratively improve classification results. Feature Analyst uses a technique called Foveal Vision (patent pending) that allows the learning algorithms to take into account the spatial context of a target feature. It integrates advanced image processing transformations, such as Spectral Mixture Analysis, into the feature extraction process. It is more accurate than hand digitizing of many features and more accurate than image processing of almost all features (Visual Learning Systems, 2002).
Perform image to image registration
Create training samples for two images
Set up learning parameters
Classify the images respectively
Identify the correct and incorrect samples respectively
iteration
Add missed objects and reclassify the image respectively
Obtain the change image
Interpret the change
Figure 1. Workflow of feature extraction based on machine learning algorithm.
ASPRS 2006 Annual Conference Reno, Nevada May 1-5, 2006
The details of the change detection workflow are listed in Figure 1: perform image to image registration; create a training set for these two images respectively; set up learning parameters for both images; classify two images respectively and get two initial classified images; remove clutter by identifying the correct and incorrect samples then reclassify the images-this step may need to be performed several times before a good result may be obtained; begin to add missed by manually adding missed objects then reclassify the image again, and get two classified images; from two classified image, the change image can be obtained; and interpret the change image.
EXPERIMENT Two data sets are used to test the change detection method in this research, one is airborne image and the other is satellite image. The spatial resolution of the airborne image is 0.056m. These two images are airborne stereo pair and time difference is very small. The change detection activity happened in these images is aircraft motion. In this change detection case, the interested object is aircraft; the task of change detection is to find the change of aircraft during time difference between these two images. The change detection result has been listed in Figure 2, in this figure, (a) is airborne image of date 1, (b) is airborne image of date 2, (c) is aircraft extraction result of date 1, (d) is the aircraft extraction result of date 2 and (e) is the change image. In Figure 2(e), red means objects existing only in date 1 image, green means objects existing only in date 2 image, black means objects existing in both images (no change). The result of aircrafts extraction is very good in general. First, all aircrafts can be extracted although the final result has included some non-aircrafts. Second, the shadow of the aircraft was not extracted into the extraction result. From the change image (Figure 2 (e)); it can be easy to find that there are two aircrafts were moving during that time period. However, the image registration error affects the final change; this is also a common problem for change detection based on high resolution image. The completeness rate of extraction could also affect the final change result, for example, a static plane in image A, has been extracted mostly, but one wing was missed, while this plane has been extracted completely in B image, thus, these two extractions are compared together to get final change result, the missed wing in image A will be counted as false change in the final change image.
ASPRS 2006 Annual Conference Reno, Nevada May 1-5, 2006
(a) airborne image of date 1
(b) airborne image of date 2
(c) Aircraft extraction result of date 1 image
(d) Aircraft extraction result of date 2 image
(e) change detection image Figure 2. Change detection based on airborne images.
ASPRS 2006 Annual Conference Reno, Nevada May 1-5, 2006
CONCLUSION There are a lot of change detection methods, while object-based change detection is different from common change detection purpose, which requires finding the interested object first, then compares the extracted interested for change detection analysis. High resolution image has brought availability and convenience to detect the objectspecific change detection, while it also brought challenges to the traditional image classification methods. This paper has introduced a change detection method which is based on machine learning algorithm for object-based change detection purpose. The method could overcome the difficulty met in processing the high resolution image by using the traditional image classification method. A stereo pair of high resolution airborne data has been used to find the aircraft change in the airport. The final change detection result could reflect the aircraft changed during this time period; however it still included the false changes caused by wrongly extracted non-aircraft. Thus, the final result has to be refined manually because it could not be refined by the software automatically. The results shown in Figure 2 is the original results by using Feature Analyst in order to show the original capability of the software to find the object-based change, no manual intervention has been involved into it.
REFERENCE Carleer, A.P., O. Debeir, and E. Wolff (2005). Assessment of Very High Spatial Resolution Satellite Image Segmentations. Photogrammetric Engineering and Remote Sensing, 71(11): 1285-1294. Opitz, D. (2003). An Automated Change Detection Method for Specific Features. Proceedings of the International ESRI User Conference Proceedings. Singh, A. (1989). Digital Change Detection Techniques Using Remotely Sensed Data. International Journal of Remote Sensing. Vol. 10, No. 6, pp. 989-1003. Visual Learning Systems. (2002). User Manual, Feature Analyst Extension for ArcView and ArcGIS. Visual Learning Systems, Inc., Missoula, MT.
ASPRS 2006 Annual Conference Reno, Nevada May 1-5, 2006