DOI QR코드

DOI QR Code

Analysis on the Effect of Spectral Index Images on Improvement of Classification Accuracy of Landsat-8 OLI Image

  • Magpantay, Abraham T. (College of Computer Studies, Far Eastern University Institute of Technology) ;
  • Adao, Rossana T. (College of Computer Studies, Far Eastern University Institute of Technology) ;
  • Bombasi, Joferson L. (College of Computer Studies, Far Eastern University Institute of Technology) ;
  • Lagman, Ace C. (College of Computer Studies, Far Eastern University Institute of Technology) ;
  • Malasaga, Elisa V. (College of Computer Studies, Far Eastern University Institute of Technology) ;
  • Ye, Chul-Soo (Department of Aviation and IT Convergence, Far East University)
  • Received : 2019.08.05
  • Accepted : 2019.08.21
  • Published : 2019.08.31

Abstract

In this paper, we analyze the effect of the representative spectral indices, normalized difference vegetation index (NDVI), normalized difference water index (NDWI) and normalized difference built-up index (NDBI) on classification accuracies of Landsat-8 OLI image.After creating these spectral index images, we propose five methods to select the spectral index images as classification features together with Landsat-8 OLI bands from 1 to 7. From the experiments we observed that when the spectral index image of NDVI or NDWI is used as one of the classification features together with the Landsat-8 OLI bands from 1 to 7, we can obtain higher overall accuracy and kappa coefficient than the method using only Landsat-8 OLI 7 bands. In contrast, the classification method, which selected only NDBI as classification feature together with Landsat-8 OLI 7 bands did not show the improvement in classification accuracies.

Keywords

1. Introduction

Mapping of land use/land cover is one of the main issues in remote sensing using various sensor images. Accuracy of classification is still the important main issue in themapping ofland use/land cover.To improve the accuracy of classification, many researchers have proposed various classification methods especially focusing on advanced classifiers or multi-source data fusion. The classification methods using the spectral index images are very easy and useful to improve the classification accuracy.The classificationmethod based on spectral index image combination improves the classification accuracy using only band information without using any other special classifier or multisource information except band information. The spectral index imagesin image classification have been used as a new band containing a specific characteristic on the earth surface.

The normalized difference vegetation index (NDVI) (Rouse, 1974), which is calculated fromthe two spectral reflectance measurements of near-infrared and red bands, is widely used for estimating the area with strong vegetation on the earth surface.In addition, normalized difference water index (NDWI) (Mcfeeters, 1996) and normalized difference built-up index (NDBI) (Zha et al., 2003) are also used for extraction of water body and built-up area, respectively.

Each spectral index is useful not only for estimating pixelswith a specific property in land cover classification but also for improving accuracy of classification by combining it with other spectral index. For example, two spectral indices of the modified normalized difference water index (MNDWI) (Xu, 2006) and the automated water extraction index (AWEI) (Feyisa et al., 2014) were combined to extract water bodies from Landsat 8 Operational Land Imager (OLI) image (Japitana et al., 2019). The combination of the two waterindices producedhigheroverall accuracycompared to other methods using two water index combination or single water index methods. The combination of NDWI and near infrared (NIR) reflectance for water body extraction produced remarkable improvement in classification accuracy (Ye, 2015). The overall accuracy and kappa coefficient of the NDWI and NIR combination method were 97.3% and 94.1%, respectively, while the NDWI method produced the overall accuracy of 89.5% and kappa coefficient of 69.6%. The combination of multiple feature images such as mean and variance of pixels within a local area and derivatives of pixels was also proposed to find the corresponding feature points between stereo images(Ye, 2007). Gao (1996) proposed the NDWI as a vegetation index to measure vegetation liquid water from space, i.e., the NDWI is a complementary index formeasuring vegetation strength.Thakkar et al.(2015) added NDVI and NDWI to classification features together with green, red and NIR bands of Indian Remote Sensing (IRS)-R2 satellite image and obtained improved classification accuracy of 7.95% compared to the classification method using original four bands.

The objective of thisstudy isto analyze the effect of three representative indices, NDVI, NDWI and NDBI on improvement of classification accuracy. In this paper we evaluate the effectiveness of each index on improvement of classification accuracy by computing the classification accuracies of the methods of each index combination and three indices combination with Landsat-8 OLI seven bands, respectively. As a kind of feature selection problem, more spectral index images used in image classification may contribute to the improvement of classification accuracy. However, they do not necessarily contribute like that if one of the spectral index images has redundant property with other feature images. This study has the meaning of evaluating the redundancy of the three spectral index images in image classification and also analyze the effectiveness of each spectral index image on image classification.

In this paper,we first analyze the effect of combination of three spectral index images of NDVI, NDWI and NDBI and the Landsat-8 OLI band images based on ourinitial experiment(Park et al., 2018), in which three indices of NDVI, NDWI and NDBI were used together withLandsat-8OLIseven bandsforimage classification. After describing variousspectral index image combination with Landsat-8 OLIseven bands, we then evaluate the accuracies of classified images obtained from the various spectral index image combination.

2. Index-based Classification

The objective of this study is to analyze which of the spectral index images (NDVI, NDWI, NDBI) and Landsat- 8 OLI band (1 to 7) feature combination has a better accuracy for classifying the image into three classes, vegetation, built-up area, and water. In this study, four methods were proposed as feature image selection method as shown in Fig. 1.

OGCSBN_2019_v35n4_561_f0001.png 이미지

Fig. 1. Flow of the index-based classification.

In themethod 1, we selectLandsat-8 OLI bandsfrom 1 to 7 as feature images used for image classification. In the method 2, three spectral index images (NDVI, NDWI and NDBI) are used together with the Landsat8 OLIseven bands.The NDVI measuresthe vegetation area by using the red and NIR band images. We obtain the spectral index image using the following formula:

\(NDVI = {NIR - Red \over NIR + Red}\)        (1)

TheNDWI andNDBI are derived fromthe following formulas, respectively:

\(NDWI = {NIR - Green\over NIR + Green}\)        (2)

\(NDBI = {SWIRI -NIR \over SWIRI + NIR}\)        (3)

In the method 3, three spectral index images(NDVI, NDWI and NDBI) are used for image classification. In the method 4, each spectral index image is used together with the Landsat-8 OLIseven bands.We apply a simple histogram-based normalization to atmospheric correction to each band. We select minimum and maximum digital numbers (DNs) from the histogram of each band and rescale the DNs to eight-bit scale using the formula (4). To make sure that the pixels in the spectral index images and band images are in the range of 0 to 255, the Landsat-8 OLI bands and the obtained spectral index images are normalized using following formula:

\(I_o = {I_{IN}-I_{MIN}\over I_{MAX}+ I_{MIN}} \cdot 255\)        (4)

where IIN and IO are intensity values before and after normalization and IMIN and IMAX are minimum and maximum intensity valuesin the image. For each ofthe defined classes, vegetation, built-up area and water, a training site is selected. An area of N × N pixels is selected asthe training site for each of the classes with a predetermined starting point on the image. This training site is selected for all the band images and spectral index imagesincluded for each ofthe proposed method. After selecting the training site for each of the classes, the feature vector I is now ready to be calculated. For each of the band and spectral index images used in the proposed method, the mean value is calculated from the training site area of each class. The mean vector m of the training site for each classis computed using the formula:

\(\mathbf {m} = {1 \over N \cdot N} \sum {_{k=1}^{N\cdot N}} \mathbf {I}_k\)         (5)

where N × N isthe total number of pixels belonging to the training site for each class.

After obtaining the feature vector I for each class, image classification is carried out using the minimum distance classifier for each method. We employ the minimum distance classifierin the image classification because it is a representative classifier, which does not need any assumption of spectral distribution of each class.We compute the vector distancesfrom each pixel to the mean vector m of each class and find the class with a minimum distance from the pixel. To assess the resulting classified image obtained by each of the proposed methods, confusion matrix is constructed. Using the confusionmatrix, overall accuracy and kappa coefficient are obtained tomeasure the accuracy of each proposed method. Reference points are selected from the image using the random sampling method. We generate two random numbers, which are used as the row and column coordinate values of each reference point, in the range 1 to height and width of image, respectively.For 200 reference points,we checkwhether each reference point belongs to a certain class.

3. Experimental Results

1) Study area and datasets

The image dataset used in thisstudy wastaken from Landsat-8 OLI bands 1 to 7 acquired on February 7, 2014. Fig. 2 shows two Landsat-8 OLI images of size 250 × 250 pixels over Metro Manila in Philippines and also three training sites of vegetation, built-up area and water. The training sitesin siteAand site B are 20 × 20 and 15 × 15 pixels in size, respectively. The site A contains large built-up area and vegetation area with relatively simple spectral characteristic, while the site B contains large vegetation area with more diverse spectral characteristic compared to the site A.

OGCSBN_2019_v35n4_561_f0002.png 이미지

Fig. 2. Two test Landsat-8 OLI images over Metro Manila in Philippines acquired on February 7, 2014 (a) site A (b) site B.

Three spectral index images of site A and siteB were created using the formulas(1),(2) and (3),respectively. Fig. 3 and Fig. 4 show the three spectral index images (NDVI, NDWI and NDBI) of site A and site B, respectively, together with some bands of Landsat-8 OLI. The relatively bright pixels in NDVI, NWI and NDBI, represent the areas belonging to vegetation, water body and built-up area,respectively.The intensity of pixels belonging to water is very noticeable in both NDVI and NDWI images.

OGCSBN_2019_v35n4_561_f0003.png 이미지

Fig. 3. Enhanced Landsat-8 OLI band images and spectral index images for site A (a) band 5 (b) band 6 (c) band 7 (d) NDVI (e) NDWI (f) NDBI.

OGCSBN_2019_v35n4_561_f0004.png 이미지

Fig. 4. Enhanced Landsat-8 OLI band images and spectral index images for site B (a) band 5 (b) band 6 (c) band 7 (d) NDVI (e) NDWI (f) NDBI.

2) Experimental results of site A

Fig. 5(a) shows the result of image classification by Method 1 forsiteA, which uses only Landsat-8 OLI bandsfrom 1 to 7 as classification feature. Many pixels in built-up area were classified into water. The user’s accuracy of water in the Method 1 produced very low value of 41.6%,while the overall accuracy of 87.8% and kappa coefficient of 77.2% were obtained, respectively, as shown in Table 1. One of the reasons is that the intensities of water and built-up area in Landsat-8 OLI bands from 1 to 3 are very similar in site A. The mean intensity of the training site for built-up area in band 1 is 40.4, while that of the training site for water in band 1 is 38.7. The difference of mean intensities between built-up area and water is 1.7, which is too small to discriminate clearly built-up area and water.

OGCSBN_2019_v35n4_561_f0005.png 이미지

Fig. 5. Results of image classification for site A (a) Method 1: Landsat-8 7 bands (b) Method 2: Landsat-8 7 bands +3 Index Images (c) Method 3: 3 spectral index images (d) Method 4.1: Landsat-8 7 bands+NDVI (e) Method 4.2: Landsat-8 7 bands+NDWI (f) Method 4.3:Landsat-8 7 bands+NDBI.

Table 1. Accuracy assessment of Method 1: Landsat-8 7 bands for site A​​​​​​​

OGCSBN_2019_v35n4_561_t0001.png 이미지

Fig. 5(b)showsthe result of image classification by the Method 2 for site A, which used Landsat-8 OLI bandsfrom 1 to 7 together with the three spectral index images,NDVI,NDWI andNDBI.Manypixels classified into water in Fig. 5(a) were classified into built-up area. The overall accuracy and kappa coefficient of the Method 2 produced higher values of 93.0% and 86.8%, respectively, than the Method 1 as shown in Table 2. The Method 2 produced a remarkable increase of 45.9% in user’s accuracy of water compared to the Method 1.

Table 2. Accuracy assessment of Method 2: Landsat-8 7 Bands + 3 Index Images for site A​​​​​​​

OGCSBN_2019_v35n4_561_t0004.png 이미지

Fig. 5(c) shows the result of image classification by the Method 3, which used only the three spectral index images, NDVI, NDWI and NDBI in image classification. The Method 3 produced similar result compared to the Method 2, except that some pixels belonging to the boundary of water were classified into built-up area.The overall accuracy of 90.5% and kappa coefficient of 81.9% for the Method 3 as shown in Table 3 were lower than the Method 2 and higher than the Method 1.

Table 3. Accuracy assessment of Method 1: 3 spectral index images (NDVI, NDWI, NDBI) for site A​​​​​​​

OGCSBN_2019_v35n4_561_t0002.png 이미지

In the Methods 4.1, 4.2 and 4.3, Landsat-8 OLI bands from 1 to 7 were used together with one of the three spectral index images. The Method 4.1, which used NDVI as the spectral index image, produced the overall accuracy of 93.0% and kappa coefficient of 86.9% asshown inTable 4 (Fig. 5(d)).The Method 4.2, which used NDWI as the spectral index image, also produced a little higher overall accuracy of 93.5% and kappa coefficient of 87.8% as shown in Table 5 than the Method 2 and Method 4.1 (Fig. 5(e)). The result of the Method 4.3, however, showed lower overall accuracy of 86.0% and kappa coefficient of 75.4% than all the other methods(Table 6).The classified image of the Method 4.3 is similar to the Method 1 as shown in Fig. 5(f) and many pixels in built-up area were classified into water like Fig. 5(a). The distribution of misclassified pixels with blue color in built-up area in Fig. 5(f)shows visually a similar pattern asin Fig. 5(a), while other classified images shows visually similar classification result except that some pixels in nonwater area in Fig. 5(d) were misclassified into water by the Method 4.1.

Table 4. Accuracy assessment of Method 4.1: Landsat-8 7 bands+NDVI for site A

OGCSBN_2019_v35n4_561_t0005.png 이미지

Table 5. Accuracy assessment of Method 4.2: Landsat-8 7 bands+NDWI for site A

OGCSBN_2019_v35n4_561_t0003.png 이미지

Table 6. Accuracy assessment of Method 4.3: Landsat-8 7 bands+NDBI for site A​​​​​​​

OGCSBN_2019_v35n4_561_t0006.png 이미지

One of the reasons is that the some pixel intensities in water and built-up area in the NDBI image are similar in site A. The difference of mean intensities between training sites of built-up area and wateris 21.1 and some pixels in built-up area have intensity values similar to those of water. This seems to cause the class separability between two classes of built-up area and water decrease in image classification.The twomethods of Method 4.1 and Method 4.2, which did not use NDBI image,show a little higher overall accuracy and kappa coefficient than the Method 2, which used the three spectral index images including NDBI.

3) Experimental results of site B

Fig. 6 and Table 7~Table 12 show the classified images and accuracy assessment of each method for site B. The classified images for site B shows similar tendency to site A. Some pixels in built-up area and vegetation area were classified into water in the Method 1 (Fig. 6(a)). The user’s accuracy of water for the Method 1 produced very low value of 27.5%, while the overall accuracy of 72.5% and kappa coefficient of 55.5% were obtained, respectively. The three methods of Method 2, Method 4.1 and Method 4.2, which contained NDVI and NDWI images together with Landsat-8 OLI bands from 1 to 7 as feature images show relatively high overall accuracy of about 84.0% and kappa coefficient of about 70%, while the Method 4.3 shows low overall accuracy of 72.5% and kappa coefficient of 55.9%.The visual pattern ofthe classified images ofsiteBissimilarto that ofsiteA, where we can see some pixelsin non-water areaweremisclassified into water by Method 1 and Method 4.3.

OGCSBN_2019_v35n4_561_f0006.png 이미지

Fig. 6. Results of image classification for site B (a) Method 1: Landsat-8 7 bands (b) Method 2: Landsat-8 7 bands +3 spectral index images (c) Method 3: 3 spectral index images (d) Method 4.1: Landsat-8 7 bands+NDVI (e) Method 4.2: Landsat-8 7 bands+NDWI (f) Method 4.3:Landsat-8 7 bands+NDBI.

Table 7. Accuracy assessment of Method 1: Landsat-8 7 bands for site B

OGCSBN_2019_v35n4_561_t0007.png 이미지

Table 8. Accuracy assessment of Method 2: Landsat-8 7 Bands + 3 spectral index images for site B

OGCSBN_2019_v35n4_561_t0010.png 이미지

Table 9. Accuracy assessment of Method 1: 3 spectral index images (NDVI, NDWI, NDBI) for site B

OGCSBN_2019_v35n4_561_t0008.png 이미지

Table 10. Accuracy assessment of Method 4.1: Landsat-8 7 bands+NDVI for site B

OGCSBN_2019_v35n4_561_t0011.png 이미지

Table 11. Accuracy assessment of M

OGCSBN_2019_v35n4_561_t0009.png 이미지

Table 12. Accuracy assessment of Method 4.3: Landsat-8 7 bands+NDBI for site B​​​​​​​

OGCSBN_2019_v35n4_561_t0012.png 이미지

4) Analysis on the effect of spectral index image on classification accuracy

From the experimentresults obtained fromsiteAand siteB, we observed that image classification containing Landsat-8 OLI bands from 1 to 7 and two spectral index images of NDVI and NDWI produced higher overall accuracy and kappa coefficient than the Method 1 using only Landsat-8 OLI 7 bands.When the spectral index image of NDBI isinserted to the Landsat-8 OLI 7 bands, the accuracies remain similar or decreased a little compared to those of the Method 1, which uses only Landsat-8 OLI 7 bands. This seems to be due to the low class separability of NDBI image between the two classes of built-up area and water. When many feature images are used in image classification like the Method 2, the effect of NDBI on classification accuracies decreases relatively, i.e., the accuracies of the Method 2 are a little lowerthan the Method 4.1 and the Method 4.2.The effect of NDBIincreasesrelatively when the number of feature images is small like the Method 3, which uses only three index images. Fig. 7 shows the comparison of classification accuracies of each classification method for site A and site B. We can see easily that Method 2 (M2), Method 4.1 (M4.1) and Method 4.2 (M4.2) are the three methods with relatively high classification accuracies.

OGCSBN_2019_v35n4_561_f0007.png 이미지

Fig. 7. Comparison of overall accuracy (O.A.) and kappa coefficient (K.C.) for site A and site B according to classification methods. (A) and (B) in the legend represent site A and site B, respectively.

4. Conclusions

In this paper, we analyzed the effect ofspectral index images on improvement of classification accuracies of Landsat-8 OLI image. We created the images of three representative indices of NDVI, NDWI and NDBI using Landsat-8 OLI band images and then used these spectral index images as classification features in various methods together with Landsat-8 OLI bands from 1 to 7.

When the spectral index image of NDVI or NDWI is used as one ofthe classification featurestogether with the Landsat-8 OLI bands from 1 to 7, we can obtain higher overall accuracy and kappa coefficient than the method using only Landsat-8 OLI 7 bands. When the NDBI is inserted to the Landsat-8 OLI 7 bands, however, the accuracies remain similar or decrease a little compared to those ofthe method, which uses only Landsat-8 OLI 7 bands.Thisseemsto be due to the low class separability of NDBI image between the two classes of built-up area andwater.Fromthe experiments, we can conclude that NDVI and NDWI are useful index imagesfor image classification and when NDVI and NDWI are used as classification features together with Landsat-8 OLI 7 bands, then the classification accuracies are not heavily influenced by the addition of NDBI to classification features. This means that as a kind of redundant feature, the NDBI with low class separability between the two classes of built-up area and water, does not contribute to the improvement of classification accuracies.

The proposed method wastested using the two sites where the spectral characteristic of water is very clear to be differentiated from other land cover types while the discrimination between built-up area and vegetation is not easy due to diverse spectral characteristic of builtup area. Although the two tested sites are sufficient to demonstrate the relative improvement of classification accuracy by spectral index image combination, further studies on testsites withmore diverse and complex land cover types are needed.​​​​​​​

References

  1. Feyisa, G. L., H. Meilby, R. Fensholt, and S. R. Proud, 2014. Automated water extraction index: a new technique for surface water mapping using Landsat imagery, Remote Sensing of Environment, 140: 23-35. https://doi.org/10.1016/j.rse.2013.08.029
  2. Gao, B.C., 1996. NDWI - normalized difference water index for remote sensing of vegetation liquid water from space, Remote Sensing of Environment, 58(3): 257-266. https://doi.org/10.1016/S0034-4257(96)00067-3
  3. Japitana, M.V., C. S. Ye, and M. E. C. Burce, 2019. Combining water indices to detect water bodies using Landsat 8 OLI, Journal of Institute of Control, Robotics and Systems, 25(5): 470-475. https://doi.org/10.5302/J.ICROS.2019.18.0220
  4. Mcfeeters, S.K., 1996. The use of normalized difference water index (NDWI) in the delineation of open water features, International Journal of Remote Sensing, 17(7): 1425-1432. https://doi.org/10.1080/01431169608948714
  5. Park, J.H., R. T. Adao, J. L. Bombasi, A.C. Lagman, and C. S. Ye, 2018. Improvement of classification accuracy using index image combination, Proc. of 2018 International Symposium on Remote Sensing, Pyeongchang, Korea, May 9-11, pp. 619-620.
  6. Rouse, J.W., R.H. Haas, J.A. Schell, and D.W. Deering, 1974. Monitoring vegetation systems in the Great Plains with ERTS, Proc. of Third Earth Resources Technology Satellite-1 Symposium, vol. 1, pp. 309-317.
  7. Thakkar, A., V. Desai, A. Patel, and M. Potdar, 2015. Land use/land cover classification using remote sensing data and derived indices in a heterogeneous landscape of a khan-kali watershed, Gujarat, Asian Journal of Geoinformatics, 14(4): 1-12.
  8. Xu, H., 2006. Modification of normalised difference water index (NDWI) to enhance open water features in remotely sensed imagery, International Journal of Remote Sensing, 27(14): 3025-3033. https://doi.org/10.1080/01431160600589179
  9. Ye, C. S., C. G. Moon, and J. H. Jeon, 2007. Stereo matching method using directional feature vector, Journal of Control, Automation, and Systems Engineering, 13(1): 52-57. https://doi.org/10.5302/J.ICROS.2007.13.1.052
  10. Ye, C.S., 2015. Extraction of water body in before and after images of flood using Mahalanobis distance-based spectral analysis, Korean Journal of Remote Sensing, 31(4): 293-302. https://doi.org/10.7780/kjrs.2015.31.4.2
  11. Zha, Y., J. Gao, and S. Ni, 2003. Use of normalized difference built-up index in automatically mapping urban areas from TM imagery, International Journal of Remote Sensing, 24(3): 583-594. https://doi.org/10.1080/01431160304987

Cited by

  1. Evaluating the Contribution of Spectral Features to Image Classification Using Class Separability vol.36, pp.1, 2019, https://doi.org/10.7780/kjrs.2020.36.1.5
  2. Hue 채널 영상의 다중 클래스 결합을 이용한 객체 기반 영상 분류 vol.37, pp.6, 2021, https://doi.org/10.7780/kjrs.2021.37.6.3.9