五、報告內容
4. RESULTS 1 Land cover classification results
4.2 Results of landscape analysis
Based on the classification results, from 2003 to 2009, the percentage of built-up area increased by 3.95%, and the percentage of sandy area and cropland decreased by 1.51% and 3.99%, respectively. To analyze the landscape changes of the study area during these years, FRAGSTATS software was used to derive landscape indices from the classified images. Table 3 shows the NP, LSI, AWMSI, and SHDI indices of these three images. Table 4 lists the t-test results of Shannon diversity index, indicating significant landscape changes between 2006 and 2009. The results indicate that the number of patches (NP) increased from 5,362 to 7,501 during 2003 and 2006, and then dropped rapidly to 4,210 in 2009. Similar trend is also found for LSI and AWMSI. This is an evidence of landscape fragmentation followed by aggregation of landscape elements.
The Hsuehshan Tunnel, also called Snow Mountain Tunnel, is the longest tunnel (12.942 km) in Taiwan. The tunnel is bored through the Hsuehshan Range, connecting Taipei to Yilan, reducing the travel time from two hours to just half an hour.
Construction of the tunnel spurred rapid development of tourism and real estate market in Yilan for the past few years.
Consequently, a lot of farm lands and bare lands were converted to buildings and roads. Nowadays we can see many large-scale hotels surrounded by farm lands and country roads in Yilan. The result of landscape analysis support the findings in landscape changes.
LandscapeIndex 2003 2006 2009
NP 5362 7501 4210
LSI 47.758 54.385 44.687
AWMSI 11.670 12.683 12.605
SHDI 1.519 1.531 1.510
Table 3. Landscape indices derived from images
Time period 2003 v.s. 2006 2006 v.s. 2009 Computed t
value 1.69 2.82
Degree of
freedom 10822 7841
Theoretical t
value (5%) 0.67 0.67
Table 4. t-test of Shannon diversity index 4.3 Results of Logistic analysis
The logit model derived from the sample data is as follows:
8
The predictor variables included elevation, distance to embankment, distance to stream, distance to protection forest, wind speed in fall, wind speed in winter, wind direction in winter, and slope. The logit model of sandy areas distribution between 2003 and 006 is shown in Table 5. The Waldχ2 and
p-value show that these factors have significant effects on the spatial distribution of sandy areas in the coastal zone.
Predictor variables Regression coefficients
Wald χ2 Pr.＞χ2
Iintercept 11.1725 45.1913 0.0001
Elevation -0.4305 97.2758 0.0001
Distance to embankment
0.0009 52.8008 0.0001 Distance to stream -0.0049 70.0173 0.0001 Distance to
protection forest
0.0018 6.1519 0.0001 Wind speed in fall 55.0056 79.8379 0.0001 Wind speed in
winter -25.2362 65.9785 0.0001
Wind direction in winter
-9.0171 88.2118 0.0001
Slope 0.2379 52.5402 0.0001
Table 5. The logit model of sandy areas distribution between 2003 and 2006
The Spatial Analyst of ArcGIS was used to derive the probability distribution map of the coastal zone, which is shown in Figure 8. The cell values are between 0 and 1, which represent the probability of being sandy area. The map shows that the areas with high probability (higher than 0.5) coincide with the sandy areas in the coastal zone. Furthermore, by comparing the classification results of 2006 and 2009 images, we can evaluate the prediction accuracy of the logit model.
Figure 9 shows the predicted probability distribution map using data of 2006. Table 6 shows the validation results. By examining the 2006 and 2009 maps, the prediction accuracy is above 80% if we choose 0.5 as the probability threshold of changing to sandy area. In the distribution map of 2009, there was more than 75% of sandy area fell in the area with predicted probability higher than 0.5. Overall, the estimation accuracy was acceptable.
Figure 8. Probability distribution of sandy area
±
0 2,000 4,000 6,000
Meters between 0.0-0 .5 between 0.5-0 .7 between 0.7-0 .9 9
Figure 9. Predicted distribution using 2006 data
Table 6. The validation results
5. CONCLUSIONS
The results of this study show that the image classification process yielded quite acceptable classification accuracy. The DMC images provide much more detailed information than the SPOT images, which can benefit the classification process. In this particular study, integration of GIS with quantitative landscape analysis techniques has been proven a valid method for analyzing landscape changes. Moreover, logit model was able to predict land cover changes in costal zone with accuracy over 80%. In this study we selected 15 predictor variables, which represent topographic, environmental, and climatic factors. However, land cover changes may also be affected by some other factors such as population distribution, the amount of tourists, and development of industrial areas. Further study will incorporate more predictor variables in order to improve the accuracy of prediction.
6. ACKNOWLEDGEMENT
The authors thank the National Science Council of Taiwan for the sponsorship of this research (NSC 99-2621-M-004-003).
7. REFERENCES
Backer, W. L., 1989. A Review of Models of Landscape Change.
Landscape Ecology, 2(2), pp. 111-133.
Chang, C.Y., Yang, Y.L., Liu, Y.Y., 1991. The Report of Surveying the Environmental Resources in Coastal zone in Taiwan. The Graduate Institute of Geography, National Taiwan University Press.
Cheng, C.C., Wu C.D., Wang, S.F., 2005. Application of Markov and Logit Models on Monitoring Landscape Changes.
Taiwan Journal of Forest Science, 20(1), pp. 29-36.
Dunn, C.P., Sharpe, D.M., Guntenspergen, G.R., Stearns, F., Yang, Z., 1991. Methods for analyzing temporal changes in landscape pattern, in: Turner, M.G., Gardner, R.H. (Eds.), Quantitative methods in landscape ecolog, Springer-Verlag, New York, pp. 173-198.
Forman, R.T.T., Godorn, M., 1986. Landscape ecology. John Wiley and Sons, New York.
Hansom, J.D., 2001. Coastal Sensitivity to environmental Change: A View from the Beach. Catena, 42, pp. 291-305.
Hulshoff, R.M., 1995. Landscape indices describing a Dutch landscape. Landscape Ecology, 2, 45-61.
Jensen, J.R., 2004. Introductory Digital Image Processing, 3rd ed., Pearson Education, Inc.
Lang, R., Shao, G., Pijanowski, B.C., Farnsworth, R.L., 2008.
Optimizing unsupervised classifications of remotely sensed imagery with a data-assisted labeling approach. Computers and Geosciences, 34, pp. 1877-1885.
Lo, C.P., Choi, J., 2004. A hybrid approach to urban land use/cover mapping using Landsat 7 Enhanced Thematic Mapper Plus (ETM+) images. International Journal of Remote Sensing, 25(14), pp. 2687-2700.
McGarigal, K., Marks, B.J., 1995. FRAGSTATS: Spatial Pattern Analysis Program for Quantifying Landscape Structure.
Gen. Tech. Report PNW-GTR-351, USDA Forest Service, Pacific Northwest Research Station, Portland, OR.
Turner, M.G., 1993. Landscape changes in nine rural counties in Georgia. Photogrammetric Engineering and Remote Sensing, 56(3), pp. 379-386.
Williams, A.T., Alveirinho-Dias, J., Novo, F.G., García-Mora, M.R., Curr, R., Pereira, A., 1995. Integrated Coastal Dune Management: Checklist. Continental Shelf Research, 21, pp.
1937-1960.
Predicted probability 0.0-0.5 0.5-0.994 Percentage of sandy area in
2006
19.9% 80.1%
Percentage of sandy area in 2009
25.1% 74.9%
1
INTEGRATION OF GIS AND MULTI-SOURCE IMAGES FOR COASTAL CHANGE DETECTION
Jihn-Fa Jan, Chih-Da Wu Young-Chung Chuang, Pin Liang
DMGIS 2011
1
OUTLINES
INTRODUCTION
STUDY AREA AND MATERIALS
METHODS AND RESULTS
CONCLUSION
DMGIS 2011
2
INTRODUCTION
The soft coastal landscapes of beaches, sand dunes, and mudflats represent fast-responding and mobile geomorphic systems that are highly sensitive to environmental changes
In the last few decades, urbanization and industrialization drastically change the sand dune landform of Yilan coastal zone, located in northeastern Taiwan.
DMGIS 2011
3
INTRODUCTION
Development of remote sensing offers a better alternative for large scale resource monitoring.
GIS technology provides a variety of analytical tools for quantifying landscape changes.
DMGIS 2011
4
2012/8/2
2 INTRODUCTION
The objective of this study was
to integrate remote sensing and GIS techniques to derive landscape indices from land cover maps obtained using multitemporal imageries
to utilize spatial statistical method to analyze the land-use change patterns and possible factors in Yilan coastal landscape both temporally and spatially.
DMGIS 2011
5
STUDY AREA AND MATERIALS
Study Area: Yilan plain and costal zone located in northeastern Taiwan, about 7,571 ha
DMGIS 2011
6
Date 2003/6/1 2006/4/11 2009/5/29 Spatial
resolution
10 m Spectral
resolution
green (0.5 to 0.59µm) red (0.61 to 0.68µm) near infrared (0.78 to 0.89 µm)
mid infrared (1.58 to 1.75µm) Processing Level 3 (ortho)
DMGIS 2011
7
Materials: SPOT 5 IMAGES SPOT 5 IMAGES
DMGIS 2011
8
3
Acquired on 5/22/2009.
The spatial resolution is 0.15 m.
Includes four multispectral bands (red, green, blue and near-infrared) and high-resolution panchromatic band.
Image size: 7,680 x 13,824 pixels.
This study used a total of 73 DMC images to compose a mosaic image of the study area.
DMGIS 2011
9
Materials: DMC (Digital Mapping Camera) Aerial Images
DMC AERIAL IMAGES
DMGIS 2011
10
ANCILLARY DATA
Digital Elevation Model (DEM)
Weather data
Vector data:
National Land use inventory
Road network maps
Administration boundary maps
DMGIS 2011
11
METHODS AND RESULTS
Land cover classification using supervised classification method
Quantitative landscape analysis
Evaluation of the change on landscape patterns by Shannon t-test
Logistic regression analysis on spatial distribution of sandy areas
DMGIS 2011
12
2012/8/2
4 IMAGE CLASSIFICATION
Supervised classification approach with maximum likelihood algorithm was used to classify the 2006 image data into five land use categories, which included water body, built-up area, cropland, sandy area, and forest land.
The training areas and validation areas were selected by comparing the images with the national land inventory.
13 DMGIS 2011
IMAGE CLASSIFICATION RESULT
the overall classification accuracy was 93.6%, with a kappa statistics of 91.75%
14 DMGIS 2011
SPOT IMAGE CLASSIFICATION RESULT
15 DMGIS 2011
DMC IMAGE CLASSIFICATION RESULT
16 DMGIS 2011
5 QUANTITATIVE LANDSCAPE ANALYSIS
Four indices of landscape structure were computed by using FRAGSTATS program.
NP (number of patches)- total number of patches in a landscape
LSI (landscape shape index)- aggregation or disaggregation of patches
AWMSI (area-weighted mean shape index)- irregular or uniform of patch shape
SHDI (Shannon’s diversity index)- proportions of patch types in a landscape
17 DMGIS 2011
SHANNON T-TEST
18
Shannon’s diversity index t-test was used to test whether the land-cover patterns change significantly
DMGIS 2011
LANDSCAPE ANALYSIS RESULT
19 DMGIS 2011
Landscape
Index 2003 2006 2009
NP 5362 7501 4210
LSI 47.758 54.385 44.687
AWMSI 11.670 12.683 12.605
SHDI 1.519 1.531 1.510
LANDSCAPE ANALYSIS RESULT
20 DMGIS 2011
2012/8/2
6
LOGISTIC REGRESSION OF SANDY AREA CHANGE
21
15 factors were considered:
elevation, slope, distance to roads, distance to stream, distance to embankment, distance to protection forest, precipitation, wind speeds of all four seasons, and wind directions of all four seasons.
DMGIS 2011
RESULTS OF LOGISTIC ANALYSIS
Predictor variables Regression coefficients
Wald χ2 Pr.＞χ2
Intercept 11.1725 45.1913 0.0001 Elevation -0.4305 97.2758 0.0001 Distance to embankment 0.0009 52.8008 0.0001 Distance to stream -0.0049 70.0173 0.0001 Distance to protection
forest
0.0018 6.1519 0.0001
Wind speed in fall 55.0056 79.8379 0.0001 Wind speed in winter -25.2362 65.9785 0.0001 Wind direction in winter -9.0171 88.2118 0.0001 Slope 0.2379 52.5402 0.0001
22 DMGIS 2011
SPATIAL VALIDATION OF LOGIT MODEL
23 DMGIS 2011
CONCLUSION
The DMC images provide much more detailed information than the SPOT images, which can benefit the classification process.
Integration image classification with quantitative landscape analysis techniques has been proven a valid method for analyzing landscape changes
Further study will incorporate more social-economic predictor variables to improve the accuracy of logit prediction.
24 DMGIS 2011
7