Landsat-based Land Cover Mapping in Île à Vache, Haiti
|
May 2013
|
ABSTRACT:
Haiti has experienced a myriad of environmental and political issues over the last century, and today the country’s environment is nearly entirely degraded as humans have harvested the forests in an unsustainable fashion. This study uses a combination of Landsat TM imagery and high resolution (30cm.) imagery to create a land cover map of a small island near Haiti. Results show that 25% of the island is devoted to agricultural production land cover type, and the other regions on the island are a combination of bare soil, wetlands, and natural shrubland. Overall map accuracy was 65% for this 2010 land-cover map.
Keywords: Remote sensing, imagery, tasseled cap, Landsat, Haiti, dark object subtraction.
Introduction
Land cover mapping in the Caribbean, and in Haiti in particular, is important in monitoring and understanding global patterns of biodiversity and change. Haiti has been one of the least examined locations due to remoteness and political turmoil, despite the high degree of interest and concern by the global community because of their environmental degradation. This degradation can be captures and quantified through the use of remotely sensed multispectral imagery.
Study Site
One area in particular is rich in land cover types and has not been studied to date, called Île à Vache. Île à Vache is a small island off of the southern coast of the Tiburon peninsula in western Haiti. The island has a rich history, for it served as a frequent base for notorious pirate Henry Morgan in the 17th century. The island is 20 sq. mi. (50km2) and the highest elevation on the island is 490 ft. (150 m). The island is composed of rolling hills primarily, with the highest point being located on the western part of the island. To the east, the island is swampy and houses one of the largest mangrove forests in Haiti. The island is administrated as a part of the Sud Department, and the island has between 10 and 15 thousand inhabitants. The island also is home to some of the most beautiful landscapes in Haiti, and prospects for tourism development are high. The development on the island could have complex implications on the regions wildlife and social wellbeing.
Data and Methods of Analysis
In this analysis I use two types of imagery: Landsat TM imagery and high resolution satellite imagery (unknown sensor). The Landsat imagery was collected on January 29, 2010, and the higher resolution imagery was collected a day earlier on January 28, 2010. This imagery was flown just two weeks after the notorious 7.0 magnitude earthquake whose epicenter struck just 10 miles away from the most densely populated region in Haiti, Port au Prince, and devastated the country. The Landsat imagery is used for the classification, and the higher resolution imagery is used as the ground truth, because the high resolution imagery allows for easy identification of land cover type for a point. The imagery was provided to me from Dr. S. B. Hedges of the biology department at the Pennsylvania State University.
Atmospheric Correction (Dark Object Subtraction)
The Landsat imagery was corrected for atmospheric scattering using the image-based dark object subtraction method (cite DOS). This method is one of the oldest methods developed for correcting remotely sensed imagery of atmospheric scattering and works to remove the additive scattering component of radiance which is caused by path radiance[2], [1].
Endmember Selection
Classes were determined in accordance with Anderson’s Land Cover Classification System (1976). I used the following classes:
1. Natural Shrubland
2. Agriculture
3. Bare Soil
4. Coastal Sand and Rock
5. Forested Wetland
6. Non-Forested Wetland
7. Forest
8. Lakes
9. Ocean
10. Commercial
ROIs were delineated by creating new vector layers and then drawing polygons over training areas. These areas were determined to be the classes of interest through a combination of ancillary imagery (high resolution and Google Earth) and my own memory; I spent several days on the island in June of 2012. One class was manually delineated which was the commercial and services class, that being the hotels on the island to accommodate tourists, and the lakes. Residential homes were not discernible with such a course resolution.
Maximum Likelihood Supervised Classification
For my supervised classification, I used the maximum likelihood method to classifying the image. This method classifies all of the pixels in the image based on probability; the class with the highest probability of being the pixel of interest is chosen.
Results
The classified land cover maps is shown in Figure 2 (top). The area in square meters and percent total land cover for each land cover type (bare soil, sand and rock, forestest wetland, etc.) are given in Table 1. A quarter (24.13%) of Île à Vache is agriculture and then just below that is forest (21.3%), coastal sand and rock (16.9%), and non-forested wetland (11.19%).
Accuracy Assessment
To perform an accuracy assessment using random points, I transferred my classified image into ArcMap. Once it was in ArcMap, I created as mask of just the land areas to constrain the random points that would be generated. Forty (40) random points were generated, and the points were used to assess the accuracy of the classified image. (Figure 3) shows the location of all 40 generated control points. Overall, the accuracy was 65%, with 25 of the 40 points being correct. Some error was expected, because the resolution of the Landsat image, 30 meters, is very coarse and so the mixing of objects within pixels can reflect incorrect classes. This simple one-to-one bimodal accuracy assessment is a simply way to quantify accuracy. Using the high resolution imagery, I am able to determine which land cover type the point falls into, and then compare it (by turning on) to the classified map image.
Discussion and Conclusions
The accuracy of this classification could be improved by changing the methodology of this analysis to incorporate a combination of using an automated method such as the supervised maximum likelihood classification with more heads-up or manual classification. In addition, if I were to create training sites using the 2D scatterplot, perhaps more accurate spectral signatures for each class would have been made. Before I provided this product to any client, I would need to revise my training sites, because the accuracy is not high enough at this point. If time allowed, I would certainly try to perform the classification again. Assuming the classified image provided in Figure 2 is correct, nearly 20% of the island is a wetland, both forested and nonforested. Due to the fragile nature of these types of ecosystems, the development planned for Île à Vache could have complex socio-ecological implications.
Further Research
A more precise accuracy assessment could involve classifying the high resolution imagery with a heads up approach, where I the analyst manually delineate classes by drawing polygons. Once this was done, I could compare the Landsat supervised image with the heads-up, higher resolution image to perform a statistical accuracy assessment. This may prove to be more accurate because it does not use a subset of sample points as the approach I took did. In addition, changes in land use for Île à Vache could be revealed in adding a temporal competent to this research.
Tables and Figures
Haiti has experienced a myriad of environmental and political issues over the last century, and today the country’s environment is nearly entirely degraded as humans have harvested the forests in an unsustainable fashion. This study uses a combination of Landsat TM imagery and high resolution (30cm.) imagery to create a land cover map of a small island near Haiti. Results show that 25% of the island is devoted to agricultural production land cover type, and the other regions on the island are a combination of bare soil, wetlands, and natural shrubland. Overall map accuracy was 65% for this 2010 land-cover map.
Keywords: Remote sensing, imagery, tasseled cap, Landsat, Haiti, dark object subtraction.
Introduction
Land cover mapping in the Caribbean, and in Haiti in particular, is important in monitoring and understanding global patterns of biodiversity and change. Haiti has been one of the least examined locations due to remoteness and political turmoil, despite the high degree of interest and concern by the global community because of their environmental degradation. This degradation can be captures and quantified through the use of remotely sensed multispectral imagery.
Study Site
One area in particular is rich in land cover types and has not been studied to date, called Île à Vache. Île à Vache is a small island off of the southern coast of the Tiburon peninsula in western Haiti. The island has a rich history, for it served as a frequent base for notorious pirate Henry Morgan in the 17th century. The island is 20 sq. mi. (50km2) and the highest elevation on the island is 490 ft. (150 m). The island is composed of rolling hills primarily, with the highest point being located on the western part of the island. To the east, the island is swampy and houses one of the largest mangrove forests in Haiti. The island is administrated as a part of the Sud Department, and the island has between 10 and 15 thousand inhabitants. The island also is home to some of the most beautiful landscapes in Haiti, and prospects for tourism development are high. The development on the island could have complex implications on the regions wildlife and social wellbeing.
Data and Methods of Analysis
In this analysis I use two types of imagery: Landsat TM imagery and high resolution satellite imagery (unknown sensor). The Landsat imagery was collected on January 29, 2010, and the higher resolution imagery was collected a day earlier on January 28, 2010. This imagery was flown just two weeks after the notorious 7.0 magnitude earthquake whose epicenter struck just 10 miles away from the most densely populated region in Haiti, Port au Prince, and devastated the country. The Landsat imagery is used for the classification, and the higher resolution imagery is used as the ground truth, because the high resolution imagery allows for easy identification of land cover type for a point. The imagery was provided to me from Dr. S. B. Hedges of the biology department at the Pennsylvania State University.
Atmospheric Correction (Dark Object Subtraction)
The Landsat imagery was corrected for atmospheric scattering using the image-based dark object subtraction method (cite DOS). This method is one of the oldest methods developed for correcting remotely sensed imagery of atmospheric scattering and works to remove the additive scattering component of radiance which is caused by path radiance[2], [1].
Endmember Selection
Classes were determined in accordance with Anderson’s Land Cover Classification System (1976). I used the following classes:
1. Natural Shrubland
2. Agriculture
3. Bare Soil
4. Coastal Sand and Rock
5. Forested Wetland
6. Non-Forested Wetland
7. Forest
8. Lakes
9. Ocean
10. Commercial
ROIs were delineated by creating new vector layers and then drawing polygons over training areas. These areas were determined to be the classes of interest through a combination of ancillary imagery (high resolution and Google Earth) and my own memory; I spent several days on the island in June of 2012. One class was manually delineated which was the commercial and services class, that being the hotels on the island to accommodate tourists, and the lakes. Residential homes were not discernible with such a course resolution.
Maximum Likelihood Supervised Classification
For my supervised classification, I used the maximum likelihood method to classifying the image. This method classifies all of the pixels in the image based on probability; the class with the highest probability of being the pixel of interest is chosen.
Results
The classified land cover maps is shown in Figure 2 (top). The area in square meters and percent total land cover for each land cover type (bare soil, sand and rock, forestest wetland, etc.) are given in Table 1. A quarter (24.13%) of Île à Vache is agriculture and then just below that is forest (21.3%), coastal sand and rock (16.9%), and non-forested wetland (11.19%).
Accuracy Assessment
To perform an accuracy assessment using random points, I transferred my classified image into ArcMap. Once it was in ArcMap, I created as mask of just the land areas to constrain the random points that would be generated. Forty (40) random points were generated, and the points were used to assess the accuracy of the classified image. (Figure 3) shows the location of all 40 generated control points. Overall, the accuracy was 65%, with 25 of the 40 points being correct. Some error was expected, because the resolution of the Landsat image, 30 meters, is very coarse and so the mixing of objects within pixels can reflect incorrect classes. This simple one-to-one bimodal accuracy assessment is a simply way to quantify accuracy. Using the high resolution imagery, I am able to determine which land cover type the point falls into, and then compare it (by turning on) to the classified map image.
Discussion and Conclusions
The accuracy of this classification could be improved by changing the methodology of this analysis to incorporate a combination of using an automated method such as the supervised maximum likelihood classification with more heads-up or manual classification. In addition, if I were to create training sites using the 2D scatterplot, perhaps more accurate spectral signatures for each class would have been made. Before I provided this product to any client, I would need to revise my training sites, because the accuracy is not high enough at this point. If time allowed, I would certainly try to perform the classification again. Assuming the classified image provided in Figure 2 is correct, nearly 20% of the island is a wetland, both forested and nonforested. Due to the fragile nature of these types of ecosystems, the development planned for Île à Vache could have complex socio-ecological implications.
Further Research
A more precise accuracy assessment could involve classifying the high resolution imagery with a heads up approach, where I the analyst manually delineate classes by drawing polygons. Once this was done, I could compare the Landsat supervised image with the heads-up, higher resolution image to perform a statistical accuracy assessment. This may prove to be more accurate because it does not use a subset of sample points as the approach I took did. In addition, changes in land use for Île à Vache could be revealed in adding a temporal competent to this research.
Tables and Figures
Table 1: Class statistics. Developed land was less than one percent. I mistakenly left it out when constructing the above table and no longer have the data.
Figure 1: Workflow Diagram.
Figure 2: Classified Image of Île à Vache, Haiti (top) and randomly generated points used for ground truth (bottom).
AWKNOWLEDGEMENTS:
I would like to thank Dr. D. Miller for his help in knowledge construction and methodology for this research project. I would also like to thank Dr. S. B. Hedges for providing me with the high resolution imagery and for instilling a profound interest in Haiti in me. Without both of these advisors this research would not have been possible.
REFERENCES:
[1] Chavez, Pat S., Jr. "Image-Based Atmospheric Corrections - Revisited and Improved." PE & RS (1996): n. pag.
[2] "Evaluation of the dark-object subtraction technique for adjustment of multispectral remote-sensing data", Proc. SPIE 1819, Digital Image Processing and Visual Communications Technologies in the Earth and Atmospheric Sciences II, 176 (March 26, 1993); doi:10.1117/12.142198 [3] 1976, Anderson, James R.; Hardy, Ernest E.; Roach, John T.; Witmer, Richard E. USGS Professional Paper: 964.
I would like to thank Dr. D. Miller for his help in knowledge construction and methodology for this research project. I would also like to thank Dr. S. B. Hedges for providing me with the high resolution imagery and for instilling a profound interest in Haiti in me. Without both of these advisors this research would not have been possible.
REFERENCES:
[1] Chavez, Pat S., Jr. "Image-Based Atmospheric Corrections - Revisited and Improved." PE & RS (1996): n. pag.
[2] "Evaluation of the dark-object subtraction technique for adjustment of multispectral remote-sensing data", Proc. SPIE 1819, Digital Image Processing and Visual Communications Technologies in the Earth and Atmospheric Sciences II, 176 (March 26, 1993); doi:10.1117/12.142198 [3] 1976, Anderson, James R.; Hardy, Ernest E.; Roach, John T.; Witmer, Richard E. USGS Professional Paper: 964.