Comparison of different vegetation indices for very high-resolution images , specific case UltraCam-D imagery

Today digital aerial images acquired with UltraCam sensor are known to be a valuable resource for producing high resolution information of land covers. In this research, different methods for extracting vegetation from semi-urban and agricultural regions were studied and their results were compared in terms of overall accuracy and Kappa statistic. To do this, several vegetation indices were first tested on three image datasets with different object-based classifications in terms of presence or absence of sample data, defining other features and also more classes. The effects of all these cases were evaluated on final results. After it, pixel-based classification was performed on each dataset and their accuracies were compared to optimum object-based classification. The importance of this research is to test different indices in several cases (about 75 cases) and to find the quantitative and qualitative effects of increasing or decreasing auxiliary data. This way, researchers who intent to work with such high resolution data are given an insight on the whole procedure of detecting vegetation species as one of the outstanding and common features from such images. Results showed that DVI index can better detect vegetation regions in test images. Also, the object-based classification with average 93.6% overall accuracy and 86.5% Kappa was more suitable for extracting vegetation rather than the pixel-based classification with average 81.2% overall accuracy and 59.7% Kappa.


INTRODUCTION
Vegetation is counted as one of the most important components of each ecosystem.Determining vegetation regions is crucial to understand the interactions between earth and atmosphere, its effects on climate, soil erosion and drought, and for natural resource management.Now, photogrammetry and remote sensing are one of the most important technologies to acquire and/or produce such information.Nowadays, digital aerial images taken with advanced UltraCam sensor are used extensively in our country and are known to be a valuable resource to produce high resolution land cover information.
Image processing techniques are categorized in two broad classes.The first ones, are those in which their processing unit is a single pixel.These techniques are called pixel-based and are highly used for extracting descriptive information.In cases where the region under study has the same size with image pixel, pixel-based methods can achieve acceptable results (Blaschke, 2010).Therefore, these methods can perform properly on low or mid resolution images (Wei, 2009).In recent years, with the advent of high resolution images, limitations of pixel-based methods are gradually revealed.Despite spectral features, additional information such as texture, geometry and shape indices can be extracted from high resolution images with low and mid-level processes.Because of low variation in spectral information of high resolution images, using only spectral features can make some problems in high level processes.As an example, for urban applications of such images, objects like buildings and roads might contain a number of pixels.So, using a single pixel instead of a group of pixels may lead to errors in final results.Some researchers have reported the pepper-salt noise in results of implementing pixel-based methods on high resolution images (Liu, 2008;Pu, 2011).Also, results of such methods are not matched well with reality.
The second class of image processing techniques are objectbased.The application of such methods holds back to 70s (Yan, 2006).Here, processing unit is a homogenous group of pixels (as image object) and the image is processed in object space instead of pixel space.In such methods, the image is first segmented and then classified.Each segment is known as a meaningful object.These methods have some advantages over pixel-based methods including: 1-Because of using objects instead of single pixels, all aspects of remote sensing field such as spectral and spatial information, morphological and structural information, context information and also temporal information can be involved.2-These methods can use geospatial information systems as an auxiliary data, distance criteria etc. 3-Objects can be formed in different sizes and in multi-scale levels of the image.So, one can first remove the unwanted additional objects in a bigger scale and then focus on extracting desirable objects.Moreover, to create image objects in different levels, parent-child relationships can be used to improve the process of detecting and extracting objects for change detection applications (Navulur, 2006).
There are many researches in literature for extracting vegetation in urban areas.Li and Shao (2014) produced land cover maps of west America using object-oriented classification and multi-scale segmentation of high resolution aerial images.They used NDVI to extract vegetation.Several other studies have been dedicated to compare different vegetation indices.Barati et al. (2011) evaluated different vegetation indices on IRS-LISIII image and found that DVI was the best index for the region.Also, many researches have compared pixel-based and objectoriented classifications.Myint et al. (2011) extracted vegetation from Quickbird image using both classifiers.Results showed that object-oriented classification could achieve 93.03% overall accuracy and 91.81% Kappa while pixel-based classifier got 80% and 76.31% respectively.They reported the lowest accuracy of extracting vegetation and soil classes due to their mix with other classes.
In this research, after studying recent methods with their results and their pros and cones, several cases (about 75 ones) were designed to compare quantitative and qualitative effects of increasing or decreasing auxiliary data.These cases evaluate the presence or absence of sample data, different vegetation indices (10 indices), using other features (redness, coloration, hue, saturation, brightness and NDWI), defining other classes (soil, artificial objects, shadow etc.), imaging conditions, whether to use near infrared bands or not (i.e.only RGB bands), all to find the optimum object-based classification and vegetation index for Ultracam images.Also, object-based and pixel-based classifications were compared to each other.

Data
Test images used in this study cover two different climates.The first study region is a part of Alborz area in Mazandaran, Bandar-Anzali and the second region is in Tehran, Damavand and near Absard region.
Both datasets are images form UltraCam D with 8 cm ground pixel size.One image from the first region (image 1) and two images from the second region acquired in two flight lines (images 2 & 3) are used to study different imaging situations.Ground truth map was provided visually based on archived information.

Segmentation
Segmentation refers to dividing image into uniform nonoverlapped conceptual regions which are named image objects or image segments.In other words, a group of neighborhood pixels with a similar property such as gray level, texture, shape and scale is called an image object.
In this research, multi-resolution segmentation is used.This method can use texture, shape, size, spectral and spatial characteristics, and previous information.It is able to extract objects with different sizes and weighting parameters like color, size, shape etc.In this segmentation, image objects are given weights according to their pixel size and the average heterogeneity of two adjacent objects is quantified based on "degree of fitting" criteria (Eq.1).
Where c= number of channels  1 &  2 = size of adjacent objects   = weight of each channel If the "degree of fitting" for adjacent objects is lower than the defined "least degree of fitting" the objects are merged together (Baatz and Schape, 2000).

Classification
To perform the object-oriented classification, nearest neighbor method is used.In this approach, a group of n-pairs {(x1, Ɵ1), …, (xn, Ɵn)} is defined where X stands for image object and Ɵ stands for class.The nearest neighborhood rule studies the following equation (Eq.2): If the above equation is satisfied, then x is related to class Ɵ ' n which contains x ' n objects (Cover and Hart, 1967).

THE PROPOSED METHOD
Different methods for vegetation extraction (pixel-based and object-based) where studied in semi-urban and agricultural regions.Fig. 1 describes the flowchart of object-based classification.
To find the optimum vegetation index, 10 different vegetation indices (Table 1) were studied according to the flowchart illustrated at Fig. 1.Then using this index (DVI) other cases of object-oriented classification were evaluated.These cases include the following 4 ones:

RESULTS AND SUGGESTIONS
In this research, 10 different vegetation indices were studied using one cases of object-based classifications to find the optimum index (Table 3).In addition, three other classification cases were evaluated (Table 4).After it, pixelbased classification was carried out for the three test images and its results were compared to optimum object-based classification (Table 5).Where L = 0.5 (Huete, 1988    Because Kappa of RDVI is not proper on the image 2, DVI is selected as the optimum vegetation index with 92.6% overall accuracy and 83.8% Kappa.DVI is less affected by background spectra (Roujean and Breon, 1995).Furthermore, although NDVI is a proper index for extracting vegetation, its relation with vegetation amount is non-linear (Fig. 2) and doesn't perform well in shrub and grassland classes (Montandon and Small, 2008).These prove that DVI's performance is better for vegetation extraction compared to NDVI.In all test images, object-based approach gives better results than pixel-based approach and its outputs are visually more interpretable and better understood.Results showed that the object-based classification with average 93.6% overall accuracy and 86.5% Kappa was better for vegetation extraction than the pixel-based classification with average 81.2% overall accuracy and 89.7% Kappa.The high difference between overall accuracy and Kappa in pixelbased method illustrates its low accuracy and the effects of omission and commission errors which have decreased Kappa considerably .
The order of applying features also affects classification results.It was found that in image 1, if the defined indices for soil and shadow are first applied and then the merging feature is used for the building class, there will be more classification errors compared to the case where the merging feature is first applied to the building class.Therefore, extracted features form an image may lead to good results only if a specific order of actions is taken.When this order changes, features seem to be unsuitable for the purpose.
All tested cases with DVI showed that if samples are used, results will improve.The amount of this difference in all images is approximately 2% for overall accuracy and 6% for Kappa which is negligible.In practice, sampling is done only when a high accuracy is required.
Definition of other classes decreased the overall accuracy in image 1 & 2 but increased it 4% in the image 3. So, to use this case or not depends on the image itself.The image 1 & 2 had denser vegetation and higher reflectance in NIR band in comparison to the image 3 These differences makes the definition of other classes to be effect less in these images because vegetation segments are well segmented due to the higher weight of NIR band and less combination with the covers like soil which degrades the accuracy in most images.
Definition of other features like NDWI increased the accuracy in image 1 & 2 (2% and 4% overall accuracy increase, and Kappa increase of 4% and 7% respectively when using samples) while decreased the accuracy in the image 3.So to use this case or not depends also to the image itself.Image 1 has numerous water basins and because NDWI improves results of images with highly water contained vegetation and water decreases reflectance in NIR band, using this index improves the results of this image.

Fig. 2
Fig. 2 illustrates result of optimum OBC, ground truth image and pixel based classification for image 1, fig. 3 for image 2 and fig. 4 for image 3.

Fig 4 .
Fig 2. Results for parts of image 1(NIR band instead of Red band), a) optimum OBC (green edges are boundaries of vegetation (merged)), b) ground truth image, c) pixel based classification

Fig 2 .
Fig 2. Calculated variation for vegetation indices as a function of fractional vegetation cover(Jones and Vaughan, 2010)

Table 1 .
Vegetation indexes investigated in test images

Table 3 .
Results of different cases of object oriented classification for test images

Table 5 .
Comparison the result of object oriented classification with pixel based classification

Table 4 .
Comparison of different cases of object oriented classifiction Vegetation is counted as one of the most important component in each ecosystem.Performance of indices on images of different sensors are not same and there is no optimum index for all images.Also different extraction methods leads to different results.To study this matter on high resolution images different vegetation indices were tested on UltraCam images and their overall accuracy and Kappa were calculated and compared to each other.In addition, various vegetation extraction methods including pixel-based and object-oriented were studied in semi-urban and agricultural regions.Next, different cases of objectoriented classification (presence or absence of sample data, using other features such as NDWI, Brightness etc. and definition of other classes like water, soil and others.)and pixel-based classification were evaluated and compared.After all these, it is concluded that among different indices, DVI leads to better results for extracting vegetation in test images.The reason for this is that DVI is less sensitive to background spectra, has linear relation with vegetation amount and better preforms in areas with low vegetation.Moreover, results showed that object-oriented classification with 93.6% overall accuracy and 86.5% Kappa could better extract vegetation in test images and got more real results in comparison to pixel-based classification with 81.2% overall accuracy and 59.7% Kappa.