This study was performed in order to suggest the possibility of tree species classification using high-resolution Quick-Bird-2 images spectral characteristics comparison(digital numbers [DNs]) of tree species, tree species classification, and accuracy verification. In October 2010, the tree species of three conifers and eight broad-leaved trees were examined in the areas studied. The spectral characteristics of each species were observed, and the study area was classified by image classification. The results were as follows: Panchromatic and multi-spectral band 4 was found to be useful for tree species classification. DNs values of conifers were lower than broad-leaved trees. Vegetation indices such as normalized differ-ence vegetation index (NDVI), soil brightness index (SBI), green vegetation index (GVI) and Biband showed similar pat-terns to band 4 and panchromatic (PAN); Tukey’s multiple comparison test was significant among tree species. However, tree species within the same genus, such as
, showed similar DNs patterns and, therefore, supervised classification results were difficult to distinguish within the same genus; Random selection of validation pixels showed an overall classification accuracy of 74.1% and Kappa coefficient was 70.6%. The classification accuracy of
, 89.5%, was found to be the highest. The classification accuracy of broad-leaved trees was lower than expected, ranging from 47.9% to 88.9%.
were classified as the same species because they did not show significant differences in terms of spectral patterns.
Forests account for over 60% of land area in South Ko-rea. However, city expansion plans and urban develop-ment projects have inevitably created more demands and increasing forest development. The Ministry of Environ-ment (ME)’s latest report on the time-series changes in the land cover map of the country shows significantly re-duced forest area: approximately 64% remained as of the end of 2000s from 66% at the end of the 1980s (Ministry of Environment 2010).
Forests, as resources for human activities and wild life habitat, play an indispensible role in the environment and in the entire ecosystem. They require an organized system of forest survey and management to prevent dam-ages by indiscriminate development.
The 1st National Natural Environment Survey kicked off in 1986, as part of ME’s nationwide natural environ-ment survey (as per Article 32 of the National Environ-ment Conservation Act) initiative to be carried out every 10 years. From the 2nd National Natural Environment Survey onwards, results and information collected were included in the GIS-DB building and then generated into an official report on the status of the eco-environment (published in April, 2007). Over 90% of the report con-cerns the results of the vegetation survey, in addition to covering various aspects of the ecosystem, including in-formation on the flora and the fauna (mammals, birds, amphibians and reptiles, insects, freshwater fish, and freshwater invertebrates) and topography (National Insti-tute of Environmental Research 2010).
With the 2013 4th National Natural Environment Sur-vey, the survey interval is to be reduced to 5 years, rather than the aforementioned 10. Thus, devising an efficient vegetation survey master plan has become a matter of ur-gency, in anticipation of the significant increase in vegeta-tion survey area to be covered annually.
Satellites have the merit of being able to obtain data for a large area. Since 1995, several countries have launched satellites that can obtain images with spatial resolution of equal to or less than 10 m. In addition, the QuickBird, launched October 2001, is a high-resolution satellite from Digital Globe. It offers high spatial resolution by simulta-neously acquiring 0.64-m panchromatic (PAN) and 2.50-m multispectral imageries and has four bands that simul-taneously record reflected or emitted radiation from the earth’s surface in the blue (band 1, 0.45 to 0.52 μm), green (band 2, 0.52 to 0.60 μm), red (band 3, 0.63 to 0.69 μm) and near-infrared (band 4, 0.76 to 0.90 μm) portions of the electromagnetic spectrum.
Forest classification is important for forest ecosystem management and scientific research. Field ground sur-veys of plant species composition and forest classification are time consuming and provide incomplete information of forest polygon delineation. The techniques of plant species classification, from satellites with high resolution images, could be the most effective way to classify and as-sess large areas of forest (Martin et al. 1998, Franklin et al. 2001).
Studies using IKONOS and QuickBird images have been used to identify trees (Ham et al. 2004, Katoh 2004, Kim 2008) and classify forest types (Chung et al. 2001, Lee 2001, Jayakumar et al. 2007) in forest stands with discern-ible crown diameter. However, it is difficult to distinguish among species in broadleaved forests, as they have the same tone in QuickBird images. The classification process of high resolution images, using spectral characteristics of bands and vegetative indices, needs to be developed.
Therefore, the objectives of this study are to use the high-resolution QuickBird-2 satellite images to compare and analyze the spectral characteristics of the tree spe-cies within the study area, to automatically classify and perform an accuracy assessment of the tree species, using the images, and to address the question of feasibility of QuickBird-2 imagery-based tree species classification by providing the relevant databases.
MATERIALS AND METHODS
- Study site
The study area was defined as the 122,500 m
(350 m × 350 m) land located at 127°00′00″ E-127°00′15″ E longitude and 37°40′35″ N-37°40′50″ N latitude from the Bukhansan National Park office at Ui-dong, Gangbuk-gu, Seoul, Ko-rea (
). The study forest is composed of broadleaved trees such as
Quercus mongolica, Q. serrata
, the conifer-ous tree
, and the
plantation. The altitude ranges from 193 to 304 m above mean sea level. The study area includes steep slopes (63%) and gentle ones. The mean annual precipitation is 1,450.5 mm, and the mean annual temperature is 12.5℃.
was showed tree species condition in this study.
Satellite images obtained from Digital Globe’s Quick-Bird-2 were used as primary data for this study. These of-fer spatial resolution by simultaneously acquiring 0.64-m PAN and 2.50-m multispectral imageries. There are four bands that simultaneously record reflected or emitted ra-diation from the earth’s surface in the blue (band 1, 0.45 to 0.52 μm), green (band 2, 0.52 to 0.60 μm), red (band 3,
Tree species conditions in the study and their abbreviationsDBH, diameter at breast height; CD, crown diameter.
Tree species conditions in the study and their abbreviations DBH, diameter at breast height; CD, crown diameter.
Map of the study area in Ui-dong valley in Seoul and training data of tree crowns derived from field survey on QuickBird PAN image. Ll Larix leptolepis; Pd Pinus densiflora; Pr Pinus rigida; Alh Alnus hirsute; Prs Prunus sargentii; Pts Pterocarya stenoptera; Qa Quercus acutissima; Qal Quercus aliena; Qm Quercus mongolica; Qs Quercus serrata; Rop Robinia pseudoacacia.
0.63 to 0.69 μm) and near-infrared (band 4, 0.76 to 0.90 μm) portions of the electromagnetic spectrum. The digi-tal map, detail vegetation map, and boundary data of the Bukhan Mountain National Park were obtained at the courtesy of relevant institutions and authorities, and the digital elevation model (DEM) was produced hands-on, using the digital map for this particular study (
The study was carried out in 3 major parts: field survey, image processing, and tree species analysis. Refer to
for a study process flow chart.
The site map, produced from QuickBird-2 images, was used to carry out the field survey. The diameter at breast
List of materials used in the study
List of materials used in the study
height (DBH) and the crown diameter of the tree species were measured, and the crown boundary was marked on the map, scanned, and geometrically corrected to build the GIS-DB with the crown boundary, DBH, and crown diameter information for each tree species.
For image processing, rational polynomial coefficient data was provided at the purchase of images and was re-flected to remove internal errors. Ground control point and DEM were each marked and built on a digital map with a ratio scale of 1:5,000 for orthography correction of PAN (resolution 0.64M) and multispectral (resolution 2.50M) QuickBird images.
Field survey results were used to compare and analyze the band patterns of the orthography-corrected images and the spectral characteristics of each vegetation index.
Flow chart of the study. DBH diameter at breast height; RPC rational polynomial coefficient GCP ground control point DEM digital elevation model.
Tukey’s multiple comparison tests were performed for a statistical significance assessment, and corresponding band combinations for tree species classification were set. Supervised classification, using the maximum likeli-hood classifier, was applied to the field survey area desig-nated as the training region.
The final stage of the study consisted of comparing the results of supervised classification against the field survey findings for the final accuracy assessment necessary for
Transformation of vegetation indices*Ratio vegetation index (Jordan, 1969).†Normalized difference vegetation index (Kriegler et al. 1969).‡Transformation normalized difference vegetation index.§Soil brightness index.¶Green vegetation index.**Yellow stuff vegetation index.††Water band index.
Transformation of vegetation indices *Ratio vegetation index (Jordan, 1969). †Normalized difference vegetation index (Kriegler et al. 1969). ‡Transformation normalized difference vegetation index. §Soil brightness index. ¶Green vegetation index. **Yellow stuff vegetation index. ††Water band index.
determining the feasibility of image-based tree species classification and its proposal.
Programs used for this study included: Arcview 3.2 and ArcGIS 9.2 for DEM construction; ERDAS 9.1 for image processing (ERDAS 1997). Also, QuickBird-2 images were used to produce and analyze a total of 7 vegetation indi-ces (
- Statistical analysis
Analysis of variance (ANOVA) was used to test for statis-tically significant differences among groups, using SPSS ver. 17.0 (SPSS Inc., Chicago, IL, USA), and Tukey’s mul-tiple comparison test was then used (
< 0.001), as shown in
. The classification accuracy was examined from a random selection of 800 pixels.
RESULTS AND DISCUSSION
- Comparison of the tree species derived from QuickBird-2 bands
QuickBird-2 band analysis show lower digital num-bers (DNs) for conifers,
, than broad-leaved trees in Band 4. This is consistent with the previous image-based study findings of Landsat TM (Katoh 1988, 1995) and IKONOS (Katoh, 2004) conifers showed lower DNs than wide-leaved trees. On the con-trast,
, though a conifer, showed similar or higher DNs to wide-leaved trees. Such observations may be attributable to
which, being a deciduous conifer, was already showing a similar range of DNs at the time the images were taken in September (
- Comparison of the tree species derived from vegetation indices
A comparison of vegetation indices-based tree spe-cies spectral characteristics showed clear differences in the average DNs between tree species in the normalized
Comparison of digital numbers values of tree species derived from QuickBird bands. PAN panchromatic; LI Larix leptolepis; Pd Pinus densiflora; Pr Pinus rigida; Alh Alnus hirsute; Prs Prunus sargentii; Pts Pterocarya stenoptera; Qa Quercus acutissima; Qal Quercus aliena; Qm Quercus mongolica; Qs Quercus serrata; Rop Robinia pseudoacacia.
Tukey's multiple comparison test of the QuickBird-2 bands and vegetation indices with ANOVA analysis (N = 1450)***P< 0.001PAN, panchromatic; NDVI, normalized difference vegetation index (Kriegler et al. 1969); TNDVI, transformation NDVI; SBI, soil brightness index; GVI, green vegetation index; YVI, yellow stuff vegetation index; WBI, water band index.
Tukey's multiple comparison test of the QuickBird-2 bands and vegetation indices with ANOVA analysis (N = 1450) *** P < 0.001 PAN, panchromatic; NDVI, normalized difference vegetation index (Kriegler et al. 1969); TNDVI, transformation NDVI; SBI, soil brightness index; GVI, green vegetation index; YVI, yellow stuff vegetation index; WBI, water band index.
difference vegetation index (NDVI), green vegetation in-dex (GVI), soil brightness index (SBI) and Biband. Such a distribution pattern was similar to the one obtained from image band spectral characteristics comparison (
- Statistical significance assessment
Tree species were classified into
groups, according to their spectral characteristics (
). The group of trees with high DNs was designated as
. Oth-erwise, the group with low DNs was depicted as
Comparison of digital numbers values of tree species derived from vegetation indices. NDVI normalized difference vegetation index; TNDVI transformation NDVI; SBI soil brightness index; GVI green vegetation index; YVI yellow stuff vegetation index; WBI water band index; LI Larix leptolepis; Pd Pinus densiflora; Pr Pinus rigida; Alh Alnus hirsute; Prs Prunus sargentii; Pts Pterocarya stenoptera; Qa Quercus acutissima; Qal Quercus aliena; Qm Quercus mongolica; Qs Quercus serrata; Rop Robinia pseudoacacia.
Using the significant differences in the QuickBird bands (
), the group with high DNs included
; the next group included
Al-nus hirsute, Prunus sargentii, Q. acutissima, Q. aliena, Q. mongolica, Q. serrata, R. pseudoacacia
; and the group with the lowest DNs contained the conifers
. For the vegetation indices (
) that pro-duced significant differences,
belonged to the group with high DNs; while
belonged to the group with low DNs. The QuickBird-2 bands and vegetation indices produced similar groups.
A statistical significance assessment revealed varia-tions in the spectral characteristics among tree species in the PAN, Band 4, Biband, NDVI, SBI, and GVI. However, species’ signatures between
were indistinguishable; both showed indiscernible differ-ences in QuickBird images and in DNs values for spectral characteristics.
Based on the results of the statistical significance test, images of 5 QuickBird-2 bands and vegetation indices were collected and the maximum likelihood classifier was applied for automatic classification. Biband, however,
Areas of each tree species in supervised classificationLl,Larix leptolepis; Pd,Pinus densiflora; Pr,Pinus rigida; Alh,Alnus hirsute; Prs,Prunus sargentii; Pts,Pterocarya stenoptera; Qa,Quercus acutissima; Qal,Quercus aliena; Qm,Quercus mongolica; Qs,Quercus serrata; Rop,Robinia pseudoacacia.
Areas of each tree species in supervised classification Ll, Larix leptolepis; Pd, Pinus densiflora; Pr, Pinus rigida; Alh, Alnus hirsute; Prs, Prunus sargentii; Pts, Pterocarya stenoptera; Qa, Quercus acutissima; Qal, Quercus aliena; Qm, Quercus mongolica; Qs, Quercus serrata; Rop, Robinia pseudoacacia.
was excluded from automatic classification in order to prevent potential errors because it failed to show normal distribution, despite optimal results in the statistical sig-nificance assessment.
Results of the classification, as shown in
, revealed that the
covered the largest area-20,700 m
- Accuracy assessment
An accuracy assessment showed an overall accuracy of 74.1%. The three highest accuracy results were shown in
(88.9%), and the
(74.4%). However, in 2 cases, spectral patterns did not show significant differences even within the same genus, resulting in them being clas-sified as the same species: between
Using the results from a detailed vegetation survey, an area rich in tree species within the Bukhansan National Park was selected for tree species survey, QuickBird-2 image-based supervised classification, and accuracy as-sessment of the classification outcomes.
Results of the DNs value analysis of tree species showed that Band 4 and PAN images were useful for tree species classification using image bands. Conifers showed lower DNs values than broad-leaved trees in Band 4, with the exception of the
, which showed other-wise.
, a deciduous conifer, may have already begun shading at the time the images were taken (Sep-tember), showing similar DNs values as broad-leaved trees. The distribution of DNs values for each vegetation index was consistent with the differences/similarities in spectral characteristics obtained from image band obser-vation.
A statistical significance assessment of tree species spectral characteristics, based on Tukey’s test, showed similar spectral patterns among trees of the same genus, namely between
, and be-tween
. Hence, species differentiation was indiscrete.
showed the largest distribution in super-vised classification, followed by:
Alnus hirsuta, Q. serrata, Robinia pseudoacacia, P. rigida
, and the
The study area for this research was limited to the Ui-
Tree species classification map derived from QuickBird-2 data with maximum likelihood classifier. Alh Alnus hirsute; LI Larix leptolepis; Pd Pinus densiflora; Pr Pinus rigida; Prs Prunus sargentii; Pts Pterocarya stenoptera; Qa Quercus acutissima; Qal Quercus aliena; Qm Quercus mongolica; Qs Quercus serrata; Rop Robinia pseudoacacia.
Accuracy assessment of randomly selected pixelsEtc.1, residential area; Etc.2, barren land; Etc.3, shadow.Alh,Alnus hirsute; LI,Larix leptolepis; Pd,Pinus densiflora; Pr,Pinus rigida; Prs,Prunus sargentii; Pts,Pterocarya stenoptera; Qa,Quercus acutissima; Qal,Quercus aliena; Qm,Quercus mongolica; Qs,Quercus serrata; Rop,Robinia pseudoacacia.
Accuracy assessment of randomly selected pixels Etc.1, residential area; Etc.2, barren land; Etc.3, shadow. Alh, Alnus hirsute; LI, Larix leptolepis; Pd, Pinus densiflora; Pr, Pinus rigida; Prs, Prunus sargentii; Pts, Pterocarya stenoptera; Qa, Quercus acutissima; Qal, Quercus aliena; Qm, Quercus mongolica; Qs, Quercus serrata; Rop, Robinia pseudoacacia.
dong valley region and, thus, inevitable restrictions were imposed; a limited number of tree species, geographical signatures of the study area, and the season images were obtained, as variables in the study, and these were prone to affect the spectral characteristics of the tree species. Monthly or “all-season” image analysis in multiple areas is recommended for more reliable results.
Classifica-tion of forest type using resolution imagery of satellite IKONOS.
Korean J Remote Sens
(in Korean with English summary)
ERDAS Field Guide.
Using spatial co-occurrence texture to increase forest structure and spe-cies composition classification accuracy.
Photogramm Eng Remote Sens
Estimation of spa-tial distribution and occurrence probability of oak spe-cies using IKONOS satellite imagery and GIS.
Korean J For Meas
(in Korean with English summary)
Object-oriented classification and QuickBird multi-spectral im-agery in forest density mapping.
Korean J Remote Sens
Derivation of leaf-area index from quality of light on the forest floor.
Estimation of rates of the crown area in yezo spruce plantations with Landsat thematic mapper data.
Proceedings of the 16th ISPRS/IUFRO
The estimation of crown densities in todo-matsu plantations using Landsat thematic mapper data.
Proceedings and IUFRO on sustainable forest manage-ment
1995 Oct 17-21
Classifying tree species in a northern mixed forest using high-resolution IKONOS data.
J For Res
Use of crown feature analysis to separate the two pine species in QuickBird imagery.
Korean J Remote Sens
Preprocessing transformations and their effects on mul-tispectral recognition.
University of Michigan
Ann Arbor MI
Proceedings of the 6th Interna-tional Symposium on Remote Sensing of Environment
1969 Oct 13-16
A study on classification and spatial distribu-tion of forest types using IKONOS satellite imagery and GIS.
De-termining forest species composition using high spec-tral resolution remote sensing data.
Remote Sens Envi-ron
Ministry of Environment.
Mapping of the Detailed-class Land Cover Map Using Satellite Images. Report.
Ministry of Environment
National Institute of Environmental Research.
Con-struction of GIS-DB for National Ecosystem Informa-tion. Report.
National Institute of Environmental Re-search