When to Detect Changes in Object-Based Image Analysis: Before or after Classification?

This study compared two object-oriented land use change detection methods—detection after classification (DAC) and classification after detection (CAD) —based on a digital elevation model, slope data, and multi-temporal Landsat images (TM image for 2000 and ETM image for 2010). We noted that the overall accuracy of the DAC (86.42%) was much higher than that of the CAD (71.71%). However, a slight difference between the accuracies of the two methods exists for deciduous broadleaf forest, evergreen coniferous forest, mixed wood, upland, paddy, reserved land, and settlement. Owing to substantial spectrum differences, these land use types can be extracted using spectral indexes. The accuracy of DAC was much higher than that of CAD for industrial land, traffic land, green shrub, reservoir, lake, river, and channel, all of which share similar spectrums. The discrepancy was mainly because DAC can completely utilize various forms of information apart from spectrum information during a two-stage classification. In addition, the change-area boundary was not limited at first, but was adjustable in the process of classification. DAC can overcome smoothing effects to a great extent using multi-scale segmentations and multi-characters in detection. Although DAC yielded better results, it was more time-consuming (28 days) because it uses a two-stage classification approach. Conversely, CAD consumed less time (15 days). Thus, a hybrid of the two methods is recommended for application in land use change detection.

detection is essential to timely and accurately detect changes in Earth's surface features and deepen our understanding of the interactions between man and nature for real-time policy decisions. In addition, its products are important input data for hygrological models [12] and climate models [13 ]and a basis data for eco-environment management [14,15].
Pixel-and object-based image analyses are two methods for detecting changes in land cover/use using a satellite image. Pixel-based methods are generally simple and easy to understand, and numerous algorithms for monitoring the dynamics of land cover/use exist. However, pixel-based methods ignore many available information apart from the pixel spectral value in remote sensing images. Moreover, the results of pixel-based change detection have "salt and pepper" effects [16].
Therefore, traditional pixel-based change detection is not suitable for processing the full range of change information. In contrast, object is the basic unit in the object-oriented method. It consists of a group of homogeneous pixels representing a meaningful object in the real world, including more information regarding shape, texture, and context features.
The shift toward object-oriented image analysis is driven by pixel limitations. Object-based image analysis (OBIA) has been extensively used in numerous countries and regions around the world [17][18][19]. OBIA is less sensitive to registration errors [20,21] and can effectively solve the "salt and pepper" effect [11,22]. Furthermore, the monitoring results of OBIA is higher than that of pixel-based methods because it completely utilizes object units and pieces of image information 18,23,24 . Compared with land use classification, change detection using OBIA is still in its infancy stage despite several previous studies.
By summarizing previous literature [15,[25][26][27][28][29][30][31][32] and references to pixel-based land use change detection methods, we can conclude that two approaches to obtain change information through the object-oriented technique generally exist. One is classifying first and then comparing the classification results, which is a technique known as detection after classification (DAC). The other is detecting the area of change first via spectral similarity analysis and then classifying the results, which is a technique known as classification after detection (CAD). Unfortunately, only a few studies have been performed to compare the two approaches for detecting land cover changes using OBIA. We do not know whether detecting changes in OBIA should be conducted before or after classification. In addition, numerous conclusions from the pixel-based change detection cannot be applied to the object-based change detection of land cover/use because the basic units have essentially changed. Comparing the performance of the two methods further in relation to change detection analysis is necessary to determine which method is superior.
To address the concern above, the CAD and DAC approaches were compared in this study for object-oriented land cover change detection using digital elevation model (DEM) data, slope data, and multi-temporal ETM and TM images in combination with the eCognition® image processing software. First, we discussed the basic process and steps in detail. Subsequently, we conducted the comparison experiment in the eCognition® software. On the basis of the results, we compared the merits and disadvantages of both and further analyzed the reasons. Finally, we proposed a direction for future research.

Study area and data
Wuhan is located at the intersection of the middle reaches of the Yangtze River and Hanshui River (the biggest tributary of the Yangtze River). Wuhan has an area of 8,494 km 2 , covering an expanse of 29°58′-31°22′N and 113°41′-115°05′E [33] over an average elevation of 27 m. Wuhan is in a typical subtropical monsoon climate region where the annual mean precipitation is 1,205 mm. The maximum temperature in the hot and rainy summer is 30 °C and the minimum temperature in the cold and dry winter is 3 °C. Surface water (lakes, rivers, and wetlands) is richly distributed in the region, with wetlands accounting for 39% of the 8,494 km 2 study area. The steps used to pre-process the images included: i) radiometric correction to compensate for the effects of atmospheric decay and ii) geometric correction to place the different images in the same reference system. In this study, the entire image was re-projected into the Gauss-Kruger coordinate system on the Krasovsky datum.

Detection After Classification (DAC) and Classification After Detection (CAD)
Preprints (www.preprints.org) | NOT PEER-REVIEWED | Posted: 29 September 2017 doi:10.20944/preprints201709.0155.v1 DAC often involves two main steps: i) independently classifying multi-temporal images and ii) comparing the object pairs in the same place, which are identified as change regions for different classes. The first step requires extracting land use information for two respective dates and then determining how objects change over time. The DAC process is graphically depicted in Figure 2.

Image segmentation and classification
Image segmentation is the process of delineating an image into homogeneous polygons, laying the basis for classification. Herein, we adopted the bottom-up region-merging [24] algorithm to segment images. In the process, the segmentation process was stopped when the smallest growth of an object exceeded a specific threshold, which is the maximum value or the so-called scale parameter. Trial-and-error and experience are still the best ways because no efficient method for choosing scale parameters exists [34 17]. The multi-resolution segmentation module of eCognition® software was used for segmentation and classification.
A few land use/land cover classification systems exist, including the USGS classification system and the  Table 1. Object-oriented classification in the eCognition® software provides two classifiers: decision tree (DT) and nearest neighbor (NN). The classification accuracy of DT is higher than that of the NN approach [34,35]. The DT model (also known as the rule-based model) first builds a classification tree and then completes the classification based on developmental rules [23,24]. In the DT approach, human interaction with the computer exists to allow the selection of suitable features. The DT classifier approach was applied in the DAC and CAD approaches, wherein similar features (Table 2) ensured comparability. For the DAC method, the two outputs of classification before determining the area of change included one class for the 2000 image data and another for the 2010 image data ( Figure 5). For the CAD method, two classes (one for each of the 2000 and 2010 image data) also existed within the detected change area.

Change detection function
The vector similarity function was adopted to detect the area of change based on the images acquired in winter. The images for both the benchmark and change periods were used as the multi-dimension vectors as conducted for band dimensions [37] and the measurement standard of vectors as the X, Y vector angles. The greater the angle and vector module, the larger the vector difference [34,38]. Thus, number of objects. For a more robust accuracy assessment, kappa statistic was incorporated into the error matrix [39].

Change direction
Being aware of changes in land cover is not only necessary but also helps in further understanding the direction of change (what changes into what) [21]. The accuracy of change direction is the overall and final estimation of land use change. The results are based on polygonal objects and not on pixels. A sample cannot completely reflect the real situation of an object consisting of a group of pixels. Therefore, the percent area method is used to assess the accuracy of change direction as follows: where i δ is the accuracy of the change direction of i class, o S is the extracted area of the change area of i class, and a S is the actual area of change of i class.

Image segmentation and classification
Multi-scale segmentation at different levels was used in this study to extract different land use classes. As shown in Figure 4, we failed to extract upland, paddy, and grassland at scales 80, 60, and the land use information was finally extracted at six levels (with segmentation scales of 80, 60, 30, 20, 15, and 10) after repeated efforts (   The results of the classification were smooth and compact, which were attributed to the insufficiently scattered and isolated pixels ( Figure 5). The object-oriented method provided several objects of homogeneous pixels with low fragmentation at reasonable levels using multi-scale.
Therefore, this method significantly reduced the so-called salt and pepper noise. However, two observations were prominent. 1) A difference between the accuracies of the CAD and DAC methods was slight, especially for deciduous broadleaf forest, evergreen coniferous forest, mixed wood, upland, paddy, reserved land, and settlement. The spectrum difference of these land cover types was relatively large and extractable by spectral indexes.
2) The accuracy of CAD was much lower than that of DAC for industrial land, road, green shrub, river, channel, reservoir, and lake. Extracting these land cover types using only spectrum information was difficult because of the easy interference and confusion with other land cover types at the spectrum scale. For example, confusion was easy between lake and reservoir, channel and river, and shrub and green shrub.
Using the indexes of others rather than those tied with the spectrum, including shape, texture, adjoin, context, and structure, was necessary for the extraction of these land cover types. However, the use of indexes other than the spectral information was limited to change area because of the sole existence of fragmented change objects and the insufficient corresponding background and context.

Change area
For DAC, two classified maps for 2000 and 2010 were superimposed in ArcMap to calculate the change fields by subtraction after the land cover output from the eCognition® software. The detailed process is illustrated in Figure 6 and the change area is depicted in Figure 7.
For the CAD, using a vector similarity function to detect change area is essentially dividing the entire image into two categories: changed and unchanged classes. The detection function threshold is a critical step. If the function value is smaller than the threshold, the difference between the objects is large and is classified as a change area, and the opposite phenomenon is the unchanged area. A sample approximation method integrated with visual interpretation was used to determine the threshold. First, 30 completely changed samples were selected from the land cover maps, which were subsets of the change area. The value of the detection function was t when the change area contained only 30 samples. The threshold must belong to [0, t], which was set at t = 0.74 through repeated trials at an increment of 0.2 per trial ( Figure 6). The change area is plotted in Figure 7.

Accuracy assessment
The missing and pseudo rates were 11.84% and 16.45% for the CAD analysis (Table 5) and 8.13% and 1.97% for the DAC analysis (Table 6), respectively. Also, the overall accuracy were 71.71% for the CAD analysis (Table 5) and 86.42% for the DAC analysis (Table 6). Both the pseudo and missing rates of the DAC analysis were lower than those of the CAD analysis. Furthermore, the DAC analysis also had a much higher overall accuracy than the CAD analysis. This suggested that the DAC analysis is better than the CAD analysis in extracting land use change variables. This can be attributed to the DAC analysis on the feature scale, whose accuracy depended on the classification results. The object-oriented classification technique completely utilizes several variables (e.g., spatial, textural, spectral, and contextual), which obviously improve classification results [5,41] and is an important merit of this method. Furthermore, the boundary of change area is initially unlimited, but adjustable in the process of classification and is later decided by differences in categories. This is where the difference from the earlier land use to the latter land use is considered as the land use change.  (Table 7). Despite the object-oriented technique, "the same object with different spectrums" or "different objects with the same spectrum" easily occurred in the CAD because the detection function only uses the spectral information of images. Thus, the results further suggested the existence of several pseudo change objects. In the spectrum, pixels that are grouped into the same object are classified together, whereas the spectral information is the mean value of the grouped pixels [21,23]. Thus, the "smoothing effect" easily occurred [2], interfering with the detection and producing numerous pseudo changes. Moreover, the boundary of the area of change was initially limited for CAD, thereby preventing the modification or removal of pseudo change objects. By contrast, the comprehensive utilization of various pieces of information in the DAC analysis considerably reduced the "smoothing effect." Furthermore, the change-area boundary was unlimited at first, but was adjustable in the classification process. Compared with the pixel-based method, the advantages of object-oriented technology lie in the use of spatial, textual, and contextual information rather than spectral information. A smoothing effect attributed to spectral information smoothing in pixel grouping occasionally occurred 2,27 . The accuracy of the CAD approach based on medium resolution and high-resolution imagery were possibly low because the change detection function was determined by spectral information.
Smoothing effect confuses the increase in spectral variability, which further increases errors of change area. OBIA first groups pixels into objects [9]. Therefore, the advantage of high spatial spectral variability for high-resolution imagery was not presented. The accuracy of the DAC approach was possibly higher because the high-resolution imagery provided more information for classification comparison with medium-resolution imagery.

5.Conclusion and summary
Both change detection methods had their own advantages and disadvantages. DAC had higher accuracy but was more time-consuming, taking 28 days to complete the classification for the two periods and requiring suitable decision trees and regulations to determine the overall change detection. By contrast, CAD only involved the classification of change area. Therefore, change detection using CAD was fast, taking only 15 days to complete the same classification.
Focusing further studies on developing a hybrid land cover change detection approach that combines CAD and DAC based on the merits of the approaches is necessary. However, this requires a significant improvement in the object-oriented technology and the field of image segmentation for the optimal scale. This also requires more experiments to determine better features that can be used in classification.