H11+V7+-+Haiti

Haiti faces many environmental problems. Deforestation, a well documented problem in the country, has caused the destruction of 90% of its trees, making Haiti the most deforested nation in the western hemisphere (1, 2). The stark visual contrast between vegetation in Haiti and its neighbor to the east, The Dominican Republic, suggests the involvement of political factors in the environmental condition of Haiti (3).



Reports of increased deforestation imply decreasing vegetation for the nation. But while one would expect this to be the case, NDVI data collected from MODIS sensors over the nine year period of 2000-2008 suggest an increase of vegetation in the country. My explanation as to why the NDVI shows increase over the period is not related to curbing of deforestation or reforestation efforts. I believe the NDVI increases are in part due to increased agricultural activity and error from the reflectance of exposed soil. My analysis of MODIS data from the period will allow me to test this theory in the coming weeks. The image below shows the MODIS Land Cover classifications for 2003:

The land cover classification done for Haiti could help substantiate a theory involving high agricultural activity. Agricultural lands can alter NDVI figures substantially, mainly owing to their abnormally higher red band reflectance. Agricultural lands can also alter NDVI figures because when the lands lay fallow, exposed soil offers varying reflectance.

In order to test the theory that exposed soil has caused distortion in the NDVI, I must look at another relevant index. The index that I think that will allow me to test this is SAVI (soil adjusted vegetation index). The SAVI equation is:

code NIR-red SAVI = --(1+L) NIR+red+L code

The L in the equation represents a correction factor for the amount of vegetation cover (4). I used an L value of .2 for my band math, this means I anticipate a higher than average level of vegetation cover. Because I don't know the actual vegetation cover at any given time, it will be important to experiment with different values periodically. The spring, summer, fall SAVI images for Haiti appear below:



The maps show that vegetation peaks during Fall. This data will become most illuminating once they are able to be compared to NDVI for the same period.


 * Assignment 4 **

Anthrome subset has 121 Lines, 164 Columns, 1 Band Time Series has 201 Lines, 274 Columns, and 184 Bands
 * Question 1:**

The time series has more lines, columns, and bands. The bands in the time series are different temporal periods. The number of columns and lines is more on the time series even though they are the same spatial subset. This means the time series has a smaller pixel size. There are 23 observations per year in the Time Series with a 16 day time step. In total there are 184 (23 * 8 years) time steps.
 * Question 2:**

The Anthrome pixels have values that correspond to ordinal classifications. A value in-between classifications doesn’t really make sense, so shrinking the anthrome pixels doesn’t make sense.
 * Question 3:**


 * Question 4:**


 * ~  ||~ # ||~ Class ||~ No. of pixels in area ||~ % of pixels in total area ||
 * 1 || #32: || Residential rainfed mosaic || 1,036 || 5.221% ||
 * 2 || #42: || Populated rangelands || 399 || 2.011% ||
 * 3 || #51: || Populated forests || 338 || 1.703% ||
 * 4 || #26: || Rainfed Mosaic villages || 295 || 1.487% ||
 * 5 || #34: || Populated rainfed cropland || 246 || 1.240% ||
 * 6 || #41: || Residential rangelands || 241 || 1.214% ||
 * 7 || #52: || Remote forests || 237 || 1.194% ||
 * 8 || #31: || Residential irrigated cropland || 157 || 0.791% ||
 * 9 || #12: || Dense settlements || 143 || 0.721% ||
 * 10 || #25: || Rainfed Villages || 103 || 0.519% ||
 * 11 || #24: || Pastoral villages || 85 || 0.428% ||
 * 12 || #11: || Urban || 80 || 0.403% ||
 * 13 || #43: || Remote rangelands || 64 || 0.323% ||
 * 14 || #22: || Irrigated Villages || 13 || 0.066% ||
 * 15 || #33: || Populated irrigated cropland || 12 || 0.060% ||
 * 16 || #23: || Cropped pastoral villages || 8 || 0.040% ||
 * 17 || #62: || Sparse trees || 6 || 0.030% ||
 * 18 || #35: || Remote croplands || 5 || 0.025% ||
 * 19 || #61: || Wild forests || 1 || 0.005% ||
 * 20 || #21: || Rice villages || 0 || 0.000% ||
 * 21 || #63: || Barren || 0 || 0.000% ||

Top 5 classifications are:


 * 1 || #32: || Residential rainfed mosaic || 1,036 || 5.221% ||
 * 2 || #42: || Populated rangelands || 399 || 2.011% ||
 * 3 || #51: || Populated forests || 338 || 1.703% ||
 * 4 || #26: || Rainfed Mosaic villages || 295 || 1.487% ||
 * 5 || #34: || Populated rainfed cropland || 246 || 1.240% ||

These represent 11.662% of the total area (because most of the MODIS tile is water) or 191,231,440,211.84 Meters2






 * Question 5: Discuss the differences between NDVI and the coefficient of variation for the 5 classes.**

All 5 of my top land classes are increasing in both mean NDVI and Variability for the 2000-2008 period. Populated Rangelands have noticeably lower NDVI values and Coefficient of Variation values comparatively to the other 4 classes for the period.


 * Question 6: Can you determine when the growing season starts and ends for each class? Does it change by year?**

When the growing season starts can be determined each year by finding the first strings of increasing NDVI values for the year. The growing season pretty consistently begins late March.


 * Question 7: How do you think the phenology in this biome would have looked if there were no humans?**

It seems like the green up periods as well as the brown down period are becoming shorter for the 2000-2008 period. Thus vegetation is simply staying at a higher level all year round, with softer peaks and bases. The smoothing of the peaks could be the result of either human intervention (more people growing crops during the non-standard growing season) or simply changes in precipitation and temperature for the period.


 * Question 8: Do you find anything remarkable in the time series? Are there trends?**

The most remarkable things are the overall increases in NDVI throughout the period and the shortening of the green up and brown down periods (vegetation is becoming less seasonal).


 * Presentation**




 * Lab 7 + 8**

I now take a look at some LandSat 5 TM images for my region. I ended up selecting two images with the least cloud cover. The two images I selected were from September 1996 (left) and September 2007 (right). A quick look of each image:



The raw images received from LandSat have distortions from atmosphere (far too high blue reflectance and far too low red and green reflectance) below is a pixel's spectral profile before and after atmospheric correction. The stages correspond to (1) raw image reflectance (2) raw image radiance, and (3) atmospherically corrected image for the 1996 image:

(1.) (2.) (3.)

The table containing all possibly appropriate LandSat 5 images:


 * ~ LandSat 5 Table ||
 * Cloud Cover || Date ||
 * 10 || 9/2/1984 ||
 * 10 || 2/9/1985 ||
 * 0 || 2/25/1985 ||
 * 10 || 3/13/1985 ||
 * 0 || 12/10/1985 ||
 * 10 || 1/11/1986 ||
 * 0 || 2/28/1986 ||
 * 10 || 3/16/1986 ||
 * 10 || 5/3/1986 ||
 * 10 || 7/6/1986 ||
 * 0 || 7/22/1986 ||
 * 0 || 8/7/1986 ||
 * 0 || 9/8/1986 ||
 * 10 || 10/10/1986 ||
 * 0 || 10/26/1986 ||
 * 0 || 12/13/1986 ||
 * 10 || 12/29/1986 ||
 * 0 || 1/30/1987 ||
 * 10 || 3/3/1987 ||
 * 0 || 3/29/1988 ||
 * 10 || 10/23/1988 ||
 * 0 || 11/8/1988 ||
 * 0 || 12/10/1988 ||
 * 0 || 12/26/1988 ||
 * 0 || 1/11/1989 ||
 * 0 || 10/10/1989 ||
 * 0 || 1/22/1990 ||
 * 10 || 8/2/1996 ||
 * 10 || 9/3/1996 ||
 * 0 || 9/19/1996 ||
 * 10 || 10/5/1996 ||
 * 10 || 10/21/1996 ||
 * 0 || 7/20/1997 ||
 * 0 || 9/6/1997 ||
 * 0 || 10/24/1997 ||
 * 10 || 5/20/1998 ||
 * 0 || 6/21/1998 ||
 * 0 || 7/7/1998 ||
 * 0 || 9/9/1998 ||
 * 0 || 8/11/1999 ||
 * 10 || 8/27/1999 ||
 * 0 || 9/12/1999 ||
 * 0 || 9/28/1999 ||
 * 0 || 10/30/1999 ||
 * 10 || 5/9/2000 ||
 * 10 || 5/25/2000 ||
 * 0 || 6/10/2000 ||
 * 10 || 7/12/2000 ||
 * 5 || 8/29/2000 ||
 * 5 || 9/30/2000 ||
 * 10 || 10/16/2000 ||
 * 10 || 2/5/2001 ||
 * 10 || 2/21/2001 ||
 * 4 || 9/18/2007 ||
 * 10 || 11/26/2009 ||
 * 5 || 1/29/2010 ||

Lab 9

A subset of my image (area around Port-Au-Prince) was enhanced so that features in the image would be more apparent. This is a subset of my 1996 image, with a high pass filter, 3 X 3 kernel size and 60% image add-back shown in color infrared:



Since my image subset is of an urban area of Haiti, smoothing really doesn't add much or make the image more visually pleasing. I apply two low pass filters and compare the affect of neighborhood size on the smoothing:

7X7 neighborhood, 20% image add-back:

11X11 neighborhood, 40% image add-back:



Principle Component Analysis:

Original true color version of image subset:



Images were taken from the first 6 principle component bands (Band 1 left, Band 2 center, Band 3 right):



(Band 4 left, Band 5 center, Band 6 right):



Statistics for the principle component analysis:


 * ~ Basic Stats ||~ Stats ||~ Min ||~ Max ||~ Mean ||~ Stdev ||~ Num ||~ Eigenvalue ||
 * || Band 1 || -273 || 4873 || 403.3377 || 350.4131 || 1 || 3186649.153 ||
 * || Band 2 || -145 || 8293 || 659.6845 || 440.7392 || 2 || 611358.9362 ||
 * || Band 3 || -131 || 8691 || 621.1157 || 527.2184 || 3 || 92296.99603 ||
 * || Band 4 || -208 || 9221 || 2149.159 || 1408.134 || 4 || 8331.847343 ||
 * || Band 5 || -97 || 6955 || 1440.532 || 961.7381 || 5 || 7344.488296 ||
 * || Band 6 || -73 || 8849 || 851.4871 || 636.8067 || 6 || 2323.340673 ||

Band 1 correlates most with bands 2 and 3. Band 2 correlates most with bands 1 and 3. Band 3 correlates most with bands 2 and 1 Band 4 correlates most with bands 5 and 6. Band 5 correlates most with bands 4 and 6 Band 6 correlates most with bands 5 and 3. The least correlated bands are bands 1 and 4, bands 1 and 5, bands 2 and 4, bands 3 and 4.
 * ~ Covariance ||~ Band ||~ Band 1 ||~ Band 2 ||~ Band 3 ||~ Band 4 ||~ Band 5 ||~ Band 6 ||
 * || Band 1 || 122789.4 || 142053.9 || 165478.5 || 52511.45 || 138500.8 || 144747.8482 ||
 * || Band 2 || 142053.9 || 194251.1 || 227187.4 || 264605.9 || 286739.7 || 231392.6642 ||
 * || Band 3 || 165478.5 || 227187.4 || 277959.3 || 311981 || 356267.6 || 288985.8365 ||
 * || Band 4 || 52511.45 || 264605.9 || 311981 || 1982842 || 1175475 || 604339.5878 ||
 * || Band 5 || 138500.8 || 286739.7 || 356267.6 || 1175475 || 924940.3 || 570205.89 ||
 * || Band 6 || 144747.8 || 231392.7 || 288985.8 || 604339.6 || 570205.9 || 405522.7616 ||
 * ~ Correlation ||~ Band ||~ Band 1 ||~ Band 2 ||~ Band 3 ||~ Band 4 ||~ Band 5 ||~ Band 6 ||
 * || Band 1 || 1 || 0.919795 || 0.895717 || 0.106422 || 0.410975 || 0.64867 ||
 * || Band 2 || 0.919795 || 1 || 0.977714 || 0.426357 || 0.676471 || 0.824442 ||
 * || Band 3 || 0.895717 || 0.977714 || 1 || 0.420236 || 0.702634 || 0.860753 ||
 * || Band 4 || 0.106422 || 0.426357 || 0.420236 || 1 || 0.867986 || 0.673953 ||
 * || Band 5 || 0.410975 || 0.676471 || 0.702634 || 0.867986 || 1 || 0.931038 ||
 * || Band 6 || 0.64867 || 0.824442 || 0.860753 || 0.673953 || 0.931038 || 1 ||
 * || Band 6 || 0.64867 || 0.824442 || 0.860753 || 0.673953 || 0.931038 || 1 ||

Majority of Eigenvalues are in the first three bands:



81.5 % of the variability is explained by Band 1 15.6 % of the variability is explained by Band 2 2.4 % of the variability is explained by Band 3

A true color image was then produced using the top three bands of the Principle Component Analysis. The result looks substantially better than the original true color image:




 * Lab 10**

For this lab we did three types of change detection on our atmospherically corrected Landsat 5 TM images: Two-Color Multiview, Image Transform (principal component analysis), and Subtractive.

Two Color Multi-View Image (below) takes a difference between the near infrared reflectance for both images; if the second image (2007) was brighter, the area is shown in cyan, if the first image was brighter, the area is shown in red. There are some problems with the image: areas which are clouded in the first image but not in the second show brightest red while images while images clouded in the second but not the first show brightest blue. Disregarding the distortions caused by cloud cover, most areas appear a grayish color, with slightly more blue. This image shows a gradual increase in the amount of vegetation.



A principal component analysis was then done on the images to determine areas of greatest change (Below). The goal of a principal component analysis is to find a combination of an image’s spectral bands such that the image produced by performing the analysis shows the most unique image or the one combination of bands which explains the highest amount of variability in the data. Interestingly, in this case the principal component bands which had the highest concentration of Eigen values did not best show change between the two images. Band 3 seemed to show the areas most changed (shown in black). Areas of greatest change appear around the cities and clouds with some lighter changes occurring in vegetated areas.

Lastly, a subtractive analysis was done on the NDVI values for both images. The Image (below) is produced by taking NDVI values from the first image and subtracting the NDVI values of the second. The image shows lowest vales (areas of greatest vegetation increase) in dark green and highest values (areas of vegetation decrease) in light green. A linear 2% stretch was applied to the image in order to more clearly show the range of values. Again, there are areas of distortion—areas with clouds in the second but not in the first show darkest green on the map while areas with clouds in the first but not in the second appear white. Ideally the image would be re-compared with a mask covering both the water and clouded areas to reduce distortion. Even so, some non-clouded areas show some respectable increases in vegetation values between the two time periods.

Lab 11

ROI separability table # 1 Pair Separation (least to most);

Beach and Sand [Maroon] 777 points and Urban [Green] 1567 points - 1.97555327 Forest [Magenta] 787 points and Croplands [Red] 2593 points - 1.97570472 Ocean [Blue] 3255 points and Beach and Sand [Maroon] 777 points - 1.99998103 Forest [Magenta] 787 points and Urban [Green] 1567 points - 1.99998848 Forest [Magenta] 787 points and Beach and Sand [Maroon] 777 points - 1.99999099 Urban [Green] 1567 points and Croplands [Red] 2593 points - 1.99999725 Ocean [Blue] 3255 points and Forest [Magenta] 787 points - 1.99999903 Beach and Sand [Maroon] 777 points and Croplands [Red] 2593 points - 1.99999969 Ocean [Blue] 3255 points and Urban [Green] 1567 points - 1.99999993 Ocean [Blue] 3255 points and Croplands [Red] 2593 points - 2.00000000

Confusion Matrix Table # 1:



The confusion matrix shows that my image has pretty good accuracy in the classifications. The lowest user accuracy is natural vegetation, the lowest production accuracy is croplands. The reason the natural vegetation has low user accuracy is probably due to my selecting both forested and grassland areas for the classification.

ROI Separability Table # 2

Pair Separation (least to most);

Beach and Sand [White] 1266 points and Urban [Cyan] 1694 points - 1.91196509 Croplands [Yellow] 1129 points and Natural Vegetation [Green] 1093 points - 1.91805304 Urban [Cyan] 1694 points and Croplands [Yellow] 1129 points - 1.99436762 Urban [Cyan] 1694 points and Natural Vegetation [Green] 1093 points - 1.99672340 Beach and Sand [White] 1266 points and Croplands [Yellow] 1129 points - 1.99926523 Beach and Sand [White] 1266 points and Natural Vegetation [Green] 1093 points - 1.99992669 Water [Blue] 3392 points and Natural Vegetation [Green] 1093 points - 1.99994414 Urban [Cyan] 1694 points and Water [Blue] 3392 points - 1.99996784 Beach and Sand [White] 1266 points and Water [Blue] 3392 points - 1.99999221 Croplands [Yellow] 1129 points and Water [Blue] 3392 points - 2.00000000

Confusion Matrix Table # 2



A classified image was then produced for each of the two observations:



The classified image from 2007 seems quite different from the image from 1996, despite both images being taken only 1 day from each other in their respective years. Changes occurred mostly between coastline, croplands, and natural vegetation. 2007 has greater amounts of natural vegetation and sand, less croplands. The 2007 scene was much drier by comparison than the 1996 scene. The much drier conditons allowed farmland to be more distinguishable from natural vegetation, thus I think the 2007 image is actually more accurate. 1996 was much wetter, so it is difficult to distinguish farmland from natural vegetation.

Change Statistics:

Area:

Percentage:

These statistics reveal that the biggest changes occurring between the images are in the classifications of Natural Vegetation (77% increase), Croplands (56% decrease), and Sand (52% increase).

Compared to the earlier change detection images, these statistics don't really mesh. The previous change detection analysis suggested two conflicting things for the Port Au Prince Area: (1) NDVI was increasing and (2) near infrared was decreasing. These seemingly conflicting results go back I think to the comparative dryness of 2007 to 1996. Alot of the natural vegetation in the area is withering away, being replaced by sand and bare earth. Bare soil (increasing during the two periods) has a decent near-infrared reflectance but it is still distinguishably less near-infrared than vegetated areas (so the decrease in NIR is expected). Appearance of an increase in NDVI really does not make sense for this region around Port Au Prince. The only thing that might explain the discrepancy between the NDVI image and the classification statistics is that something wrong may have been done in the process of producing the NDVI image (The type of stretching applied to help produce a comprehensible image?).

Out of interest to how the January 12th, 2010 earthquake in Haiti affected classification I produced a supervised classification from a January29th, 2010 Landsat 5 scene. The image shows increases in bare earth with decreases in urban area. I think that this change is partially the result of the earthquake and partially the result of the image being taken in January (during the least vegetationally active season).



1. [|1998 Reuters article explaining problems of deforestation] 2. [|2003 New York Times suggesting 90% deforestation] 3. [|2002 Google Earth image link] 4. [|T.Rays FAQ on vegetation indices]

Page maintained by: Grant DeLozier