Extraction of Spatial Characteristics of Cadastral Land Category from RapidEye Satellite Images

Abstract

With rapid land development, land category should be updated on a regular basis. However, manual field surveys have certain limitations. In this study, attempts were made to extract a feature vector considering spectral signature by parcel, PIMP (Percent Imperviousness), texture, and VIs (Vegetation Indices) based on RapidEye satellite image and cadastral map. A total of nine land categories in which feature vectors were significantly extracted from the images were selected and classified using SVM (Support Vector Machine). According to accuracy assessment, by comparing the cadastral map and classification result, the overall accuracy was 0.74. In the paddy-field category, in particular, PO acc. (producer's accuracy) and US acc. (user's accuracy) were highest at 0.85 and 0.86, respectively.

1. Introduction

In a cadastre, land categories are given by parcel, and the classification of land category influences taxation, such as real estate taxes and environmental change. However, in reality, there might be several land uses for a single parcel, and land category could be wrongly classified. Therefore, land categories should be regularly investigated to provide accurate land category management information to the state for land management and administrative purposes. With rapid land development and changes, land category would also keep changing. Under these circumstances, the rapid acquisition of necessary information would be helpful for land development and accuracy of taxation. With current field survey techniques, which have labor constraints, changes in land category cannot be investigated regularly. Therefore, an aerial photo- or satellite image-based monitoring approach is required.

Thus far, it has been demonstrated that aerial photo- or satellite image-based cadastral land categories can determine current building management and land use on a parcel level by overlapping the images and cadastral maps. As resolution has sharply improved, the applicability of the cadastral resurvey of satellite images was previously proposed (Park et al., 1999; Song et al., 2006). Hong et al. (2004) suggested land-use or landcover classification by parcel using highresolution satellite images and digital maps, and then developed a cadastral non-coincidence analysis algorithm by parcel. Hong et al. (2004) also tested the possibility of land category surveys using hyperspectral aerial photos. Here, the applicability of vegetation and land-cover information, which can be converged with cadastral information, was analyzed, and a test on the current land category by parcel was not actually carried out (Lee and Hyun, 2014). Dixon and Candade (2008) studied the application of statistical learning theory called SVM and ANN (Artificial Neural Network) classifiers for landuse classification using multispectral image. Both SVM and ANN show comparable results. Lu and Weng (2006) developed an approach for urban landuse classification based on the combination of impervious surface and population density using Landsat 7 ETM+. The overall accuracy is quite good, however, only five classes mainly depends on population density were selected for classification.

In this study, many features were used to classify landuse category by using SVM. Four features (spectral features by parcel, imperviousness, texture, and vegetation index) were extracted from multispectral satellite images, their feature vectors were defined, and they were compared between the land categories by parcel. After deriving discriminant functions with which land category can be inferred, the type of land category, which can be classified with images, was analyzed.

 

2. Test Area and Experimental Data

For the test area, a region that had sufficient parcels to determine statistical significance along with percentage of seven land category and diversity was chosen. In this study, the city of Iksan, which meets these conditions, was selected for the target area as shown in Fig. 1. There were 26 land categories present in our study area. Agricultural fields occupy most of the study area, but there are a small number of high-rise residential areas and some public and commercial buildings. The area is partly covered by forest and has substantial tree cover.

Fig. 1.The study area. The red polyline indicates the administrative boundary of Iksan

Multispectral-band satellite images, which are available for monitoring over a broader swath than aerial photos, were used. The satellite images used in the test should exhibit a spatial resolution sufficient for interpretation of urban areas and a spectral resolution applicable for interpretation of vegetation, crop vitality, and crop status (Jensen, 2007). In addition, short-term and temporal characteristics should be considered in future technology-based monitoring. Among the images that met these requirements, five (5) bands of RapidEye were used. The RapidEye image was taken on September 4, 2013, and the cadastral map was established in 2007.

Ortho-rectification was performed on the RapidEye image based on a 5m DEM (Digital Elevation Model), then the image was resampled with a spatial resolution of 7m. However, it was not actually overlaid with the cadastral map. Therefore, a one dimension polynomial transformation processed with 30 GCPs (Ground Control Points) was selected on road intersections. The RMSE (Root Mean Square Error) was less than 0.1 pixels.

 

3. Methodology and Experimental Results

In order to evaluate the profile of several parameters, the cadastral map was rasterized by labeling each parcel with its parcel ID and land category. Then, statistics of feature vectors, including percentage of landcover class, PIMP, textures, and VIs were computed for each parcel. Parcel statistics were used to compute the mean and std. (standard deviation) of each land category, and the results were compared to find appropriate parameters for separating land categories. Finally, SVM was applied on the selected parameters to classify landuse categories. A brief flowchart is shown in Fig. 2.

Fig. 2.Experimental flowchart

3.1 Percentage of landcover class

To obtain the percentage of landcover classes for each land category, the RapidEye image was classified into seven classes: (1) Man-made str. (Man-made structure), (2) Forestry, (3) Agr. + Gra. Land (Agricultural and grass land), (4) Water, (5) Shadow, (6) Bare land, and (7) Sand. This was achieved by using MLC (Maximum Likelihood Classification) available in ENVI. The percentage of pixels for each landcover class in each parcel was computed using MATLAB. The mean of each by land category is shown in Fig. 3. Most categories contain a main landcover class, which occupies over 50% of the land area. The main landcover class was either man-made structures or Agr. + Gra. land. Land categories 4, 6–13, 20, 23, 24, and 26 mostly consists of most man-made structures, while other categories mostly consisted of vegetation. Several categories were supposed to contain only vegetation such as paddy-field and forestry, but they also contained some artificial structures in Fig. 3. For instance, in paddy-field areas there are many greenhouses, which are usually classified as Man-made str. There were 26 land categories in our study area, and their names are shown in Table 1.

Table 1.Land categories in Korea. Mineral spring and salt marsh were excluded (FAOLEX, 2001)

Fig. 3.Percentage of landcover class by land category (FAOLEX, 2001)

3.2 Percent imperviousness

Generally, an impervious surface is any Man-made str. that does not allow natural infiltration of water, such as buildings, residential areas, industrial areas, roads, and parking lots made from asphalt, concrete or bricks (Kim et al., 2008; Schuler and Kastdalen, 2005). Various approaches have been developed to estimate PIMP based on remote sensing. In this research, impervious surface was obtained from landcover classification image. The PIMP of each parcel was simply calculated by computing the percentage of a catchment area that was made up of impervious surfaces, such as roads, roofs, and other paved surfaces. The mean PIMP by land category was then computed from the PIMP of parcels that belonged to that category. The results with std. are shown in Fig. 4.

Fig. 4.PIMP by land category

3.3 Texture

To evaluate textural information, eight texture features were calculated from each band of RapidEye image based on the co-occurrence matrix by performing co-occurrence measures available in ENVI with a moving window size of 3×3 pixels. The eight textural features were HOM (Homogeneity) in Eq. (1), CON (Contrast) in Eq. (2), DISSIM (Dissimilarity) in Eq. (3), MEAN (Mean) in Eq. (4), VAR (Variance) in Eq. (5), ENT (Entropy) in Eq. (6), SM (Second moment) in Eq. (7), and COR (Correlation) in Eq. (8).

where p(i, j) : (i, j)th entry in a normalized gray-tone spatial dependence matrix, Ng : Number of distinct gray levels in the quantized image, and μx, μy, αx, αy : Mean and standard deviations of px and py (Haralick et al., 1973)

The averages of textural features were computed for each parcel, and then the means for each land category were obtained. Fig. 5 shows the means with std. of the eight textural features by land category for band #1. It is clear that there is no significant difference in SM and ENT by land category. Therefore, they are not good candidates for separating land category. On the other hand, the other features, such as COR, DISSIM, and MEAN were quite good for distinguishing several land categories. For instance, COR, MEAN, CON, and DISSIM were good for classification of building site and factory site, while COR, DISSIM, HOM, and MEAN could be used to classify dry paddies and rice paddies. The same patterns were found for the other bands.

Fig. 5.Texture distribution for band #1

3.4 Vegetation indices

VIs represent features such as vegetation vigor, and many types of VIs have been proposed. In this study, VIs were calculated according to land category and reviewed on whether they could be specialized. Because RapidEye has red-edge bands, their indices were estimated as well (Kim et al., 2011; Lee and Hyun, 2014).

In this section, five VIs, which are defined in Table 2, were computed from RapidEye images. Fig. 6 shows mean VIs and std. by land category, which were calculated from a parcel’s mean VIs. The patterns of VIs are almost the same, except for a slight difference in the link between categories #7 and #8 and categories #11 and #12, which were almost level in RVI_RE (Red Edge) and NDRE.

Table 2.Equations to compute several vegetation indices

Fig. 6.Vegetation index distribution

 

4. SVM classification

SVM are particularly appealing for use in the remote sensing field because of their ability to handle small training data sets successfully, often producing higher classification accuracy than traditional methods (Burges, 1998; Kim et al., 2013; Warner and Nerry, 2009). The complete mathematical formulation of an SVM can be found in Vapnik (1995) and Burges (1998). The underlying principle benefitting SVM is the learning process, which follows what is known as “structural risk minimization”. Under this scheme, SVM minimize classification errors in unseen data without prior assumptions on the probability distribution of the data (Fauvel et al., 2008; Kim et al., 2013; Mountrakis et al., 2011). SVM are non-parametric, which makes them particularly useful for complex data sets. Even though SVM were primarily defined as binary classification, they can be used for multiclass classification (Foody and Mathur, 2004; Hsu and Lin, 2002).

The data used in the study was obtained from a RapidEye image with a spatial resolution of 7m. Therefore, there are many small parcels containing only one pixel, which can introduce noise during classification. To avoid noise, a size filter of eight pixels was applied for categories with small parcels, such as building site, and for categories containing large parcels. The minimum size was 20 pixels.

Among the attributes of the land categories mentioned in previous sections, 43 attributes were selected to classify land categories using SVM. These attributes were the percentage of seven landcover classes, PIMP, five VIs, and six texture parameters of each band. Among the 28 landuse categories, several sets of categories were selected to test possibility of classification. An analysis showed that nine categories, including paddy-field, forestry, building site, gas station, warehouse, river, marsh, water supply site, and religious site could be separated well based on the 43 parameters. SVM was originally designed for binary classification (two-class classification). However, it can function as a multiclass classifier by combining several binary SVM classifier. Two common methods to enable this adaptation are “one against all” (1AA) and “one against one” (1A1). The main disadvantage of 1AA over 1A1 is that its performance can be compromised due to unbalanced training datasets (Gualtieri and Cromp, 1998). Besides, the 1A1 strategy is substantially faster to train and seems preferable for problems with a very large number of classes (Jia, 2005). In our study, SVM with “pair wise” classification strategy also known as “one against one” (Knerr et al., 1990; KreBel, 1999) for multiclass classification was performed to classify these nine categories by using MATLAB. Firstly, 36 pairs of classes were generated from 9 selected categories. Then SVM was constructed for each pair of classes, resulting in 36 machines. These machines were used to classify test data, each classification gives one vote to the winning class. Finally, pixel was labeled with the class having most votes. A brief summary of SVM classification can be seen in Fig. 7. RBF (Radial Basic Function) kernel was used. Training and reference samples were randomly selected for each land category based on the cadastral map. The number of samples was determined based on the number of parcels in each category. If a category contained over 10,000 parcels, approximately 2,000 samples were selected as training samples, and approximately 200 parcels were selected as reference samples for accuracy assessment. For categories containing less than 1,000 parcels, 50% of the parcels were selected as training samples, and the others were set as reference samples. The accuracy of nine land categories classification is shown in Table 3. The overall accuracy is quite high at approximately 74%. Among the nine categories, the accuracy of the paddy-field category was highest with a PO acc. and US acc. of 85% and 86%, respectively. The accuracy of the religious category was lowest with a PO acc. and US acc. of 55% and 54%, respectively. Some building site parcels were classified into the gas station, ware building site, and Reli. (religious categories), which also contained Man-made str. Meanwhile, some forest parcels were classified into the paddy-field category and vice versa. This is because these classes are mostly covered by vegetation. Moreover, reference data is a cadastral map established in 2007, meanwhile RapidEye image taken in 2013. Therefore, some changes could have taken place during this period. This also might reduce the overall accuracy in accuracy assessment.

Fig. 7.Flowchart of SVM classification based on “one against one” strategy (c is number of categories for classification)

Fig. 8.A sub-classified image with cadatral map overlaid. (a) RapidEye image, (b) classified image, (c) reference cadastral map

Table 3.Confusion matrix of landuse category classification using SVM

 

5. Conclusions

The accurate classification of land category plays an important role in taxation. For this reason, it should be regularly investigated. However, it is difficult to regularly update land category with only manual field surveys in a rapidly changing land. Therefore, this study considered satellite images as a way to monitor land category and tested its applicability. In the test, a multispectral RapidEye image and a cadastral map were overlapped, and 3D feature vectors were extracted from spectral features by parcel, PIMP, texture, and VIs. Then, they were classified using SVM. According to an analysis of accuracy, in comparison with the cadastral map, the overall accuracy was 0.74. In paddy-field fields, in particular, PO acc. and US acc. were highest at 0.85 and 0.86, respectively.

In future studies, it is necessary to analyze differences in the accuracy of land category by securing seasonal and periodic ortho-rectification multispectral images and collecting additional variables to analyze the compatibility of 28 land categories. Moreover, considering spatial and spectral resolutions, there should be further studies using diverse satellite images.