University of Southern California Dissertations and Theses

Assessment of land cover change in Southern California from 2003 to 2011 using Landsat Thematic Mapper

(USC Thesis Other)

Assessment of land cover change in Southern California from 2003 to 2011 using Landsat Thematic Mapper

PDF

Download a page range

Download transcript

Copy asset link

Request this asset

Transcript (if available)

Content

Assessment of Land Cover Change in Southern California
from 2003 to 2011 Using Landsat Thematic Mapper

A Thesis in Biostatistics
By
Yi Zhang
Department of Preventive Medicine
University of Southern California

Submitted in Partial Fulfillment
of the Requirements
for the Degree of
MASTER OF SCIENCE
May 2019

Contents
1 Abstract ........................................................................................................................................... 1
2 Introduction ..................................................................................................................................... 2
3 Methods........................................................................................................................................... 3
3.1 Landsat raster selection and preprocessing ...................................................................... 3
3.2 Greenness Calculations .................................................................................................... 6
Normalized Difference Vegetation Index (NDVI) .............................................. 6
Enhanced Vegetation Index (EVI) ...................................................................... 7
Normalized Burn Ration (NBR) ......................................................................... 7
3.3 Greenness Change ............................................................................................................ 8
3.4 Supervised Image Classification ...................................................................................... 9
Band Stacking and Image Enhancement............................................................. 9
Developing Training Sites................................................................................. 10
Random Forest Algorithm..................................................................................11
Post-Processing ................................................................................................. 12
4 Results and Conclusions ............................................................................................................... 13
4.1 Greenness Calculation and Changes .............................................................................. 13
4.2 Supervised Image Classification .................................................................................... 15
5 References ..................................................................................................................................... 19
6 Appendix ....................................................................................................................................... 21

1

1 Abstract
Land cover change plays an important role in socioeconomical and bioecological systems, and
recent research has shown its importance related to aspects of human health. In an urbanizing
world, built environments are replacing green spaces, and concurrently, environmental stressors
such as noise, air pollution, and heat are increasing. Southern California is home to 4 of the 10
most populous counties in the United States, and according to the last census, Los Angeles-Long
Beach-Anaheim has surpassed New York City as the most densely populated metro area. Focusing
Southern California, we examine the change in greenspace from 2003 to 2011 using 30 m
resolution satellite imagery from the Landsat Thematic Mapper (TM). Using open source software
QGIS, we calculated and visualized the Normalized Difference Vegetation Index (NDVI),
Enhanced Vegetation Index (EVI) and Normalized Burn Ratio (NBR) as indicators of greenspace
change. After visual analysis and training, land use change was classified into four forest and non-
forest categories using the random forests algorithm with error rates around 0.06% and out-of-bag
error rates around 6.34% respectively. Our findings show that over the 8-year period greenspace
in Southern California has significantly decreased. The biggest change from 2003 to 2011 is in the
NDVI greenness metric, where the mean decreased from 0.207 to 0.152 over this period. The mean
EVI changed from 0.006 to 0.003, and mean NBR changed from 0.192 to 0.162.

2

2 Introduction
Analysis of land cover (LC) and LC change is applicable to many areas of research and
decision making, and has been used for tracking deforestation and urbanization, monitoring the
impacts of disasters, and managing and planning land use (Hussain et al, 2013). Multispectral
satellite sensors detect and record the electromagnetic radiation reflected and emitted from Earth’s
surface at various wavelengths. With a suite of sensors, the Landsat program has collected a
continuous record of data about the Earth’s surface since 1972 (USGS, 2015). The longest running
sensor is Landsat 5, launched March 1, 1984 and decommissioned on June 5, 2013 after collecting
over 2.5 million images of Earth’s surface. After Landsat 6 failed to reach orbit in 1993, Landsat
7 was launched in April 1999 and continues to operate today along with Landsat 8, which, launched
in February 2013, is the most recent member of the Landsat program. Landsat 5 carried the
Thematic Mapper (TM) sensor and Landsat 7 and 8 carry the Enhanced Thematic Mapper Plus
(ETM+) sensor. Both TM and ETM+ collect data in spectral bands in the shortwave infrared
(SWIR), visible, and near infrared (NIR) bands and the delivered products are at 30 m resolution.
The ETM+ sensor is enhanced with a thermal and panchromatic band.
Analysis of satellite remote sensing data with geographic information systems (GIS) coupled
with machine learning methods has become a widely applied approach for LC classification (Xiao
et al, 2006, Nery et al 2017). For LC change, there are two routinely used methods; post-
classification change detection, and pre-classification change detection (Lunetta & Elvidge, 1998).
In pre-classification change detection images from two time periods are first stacked and then
differenced, whereas in post-classification comparison, LC is first classified in each image
separately before differencing (Lunetta & Elvidge, 1998). Since any misclassifications would
propagate in differencing, post-classification LC change is not recommended.
Interest in understanding the component of land use comprising of vegetation and natural
elements (Taylor and Hochuli, 2017) has increased in recent years due to evidence of its benefit to
human health and wellbeing (Frumkin, 2013). Vegetation-informative spectral bands include the
3

NIR and visible red bands, and when harnessed from satellite sensors such as Landsat are
combined into metrics representing greenness or greenspace. The most common metric of
greenness is the normalized vegetation index (NDVI), which has been used in several studies
examining local greenspace and health (De Vries et al, 2013). Two other metrics have been
developed for assessing greenness from multispectral satellite imagery: the enhanced vegetation
index (EVI) and the normalized burn ratio (NBR). The EVI was developed to address some
limitations in NDVI by reducing the influence of atmospheric conditions and background canopy
issues. The NBR harnesses the SWIR band and was designed to examine burn areas and estimate
fire severity.
The objective of this study is to examine the change in greenness as indicated by NDVI, EVI
and NBR over an 8-year period (2003-2011) in Southern California. We calculate these metrics
using Landsat 5 imagery from the same time period in 2003 and 2011, and classify the change
using the random forests algorithm. This study period coincides with a longitudinal study of over
5,000 children conducted by researchers at the University of Southern California as part of the
Children’s Health Study (CHS) (McConnell et al, 2006). Recent studies of the CHS have shown
the impact of multiple pollutants including air pollution and noise (Franklin & Fruin, 2017) on the
respiratory outcomes of these children, and understanding the impact of changes in greenspace on
these and other outcomes such as stress and sleep are of utmost importance to further study the
health effects associated with the urban landscape.

3 Methods
3.1 Landsat raster selection and preprocessing
For consistency in our comparison of imagery from 2003 to 2011, we use Landsat 5 TM
imagery acquired from the USGS website (https://earthexplorer.usgs.gov). Although Landsat 7
ETM+ was collecting data over this time period there are major artifacts in the product due to a
May 2003 failure of the scan line corrector. Since it is difficult to correct for these artifacts and the
4

failure occurred in the starting year of our study, we opted to use Landsat 5 TM imagery. Details
about the spectral bands in the Landsat 5 product are shown in Table 1 (Markham et al, 2004) and
the signature bands and wavelengths used in identifying greenness are illustrated in Figure 1.

Table 1. Characteristics of Landsat 5 Thematic Mapper
Bands Wavelengths ( m) Resolution (m)
1 – Blue 0.45-0.52 30
2 – Green 0.52-0.60 30
3 – Red 0.63-0.69 30
4 – Near Infrared (NIR) 0.76-0.90 30
5 – Shortwave Infrared
(SWIR) 1
1.55-1.75 30
6 – Thermal 10.41-12.5 120
7 – Shortwave Infrared
(SWIR) 2
2.08-2.35 30

Figure 1. Spectral signature of vegetation, soil and water. Image source: seos-project.eu

A flow diagram of image and data processing is shown in Figure 2 and described in detail
below. Our study region includes Santa Barbara southeast to Riverside, California. To cover this
area, three Level 1 terrain and precision corrected (L1TP) Landsat tiles were necessary (Figure 3).
5

In order to compare LC over different years, pairs of images in the same month were chosen to
control for the effect of sunlight angle and phenological differences. Excluding images with more
than 10% cloud contamination resulted in 3 images in August for 2003 and in 2011. Each image
consists of radiances in the separate bands outlined in Table 1.

Figure 2. Flowchart of method for land cover change detection in 2003 and 2011
2003 Landsat 5 (3 tiles, <10% cloud,
each tile has bands 1-5 and 7)
2011 Landsat 5 images (3 tiles,
<10% cloud, each tile has bands 1-5)
2003 Land Cover after merging bands 2011 Land Cover after merging bands
Greenness Calculation Greenness Calculation
Greenness Change
Stack 2-date image
Develop training sets
Random Forest algorithm
Land cover change classification
6

Figure 3. Study area showing positioning of the three Landsat 5 tiles with symbology of the
merged raster changed to the combination of false color (over land).

In the sections below we describe the data processing and modeling steps shown in the Figure
2 flowchart in more detail. Briefly, images for all bands except for band 6 are merged, providing
the spectral information needed to calculate the three indexes of greenness. We then use the raw
pixel values in a combination of two images to identify change in the greenness over time. In order
to visualize the land cover use, image transformations are conducted with NDVI, EVI and NBR.
After judging and selecting areas with greenness change, the classification is conducted. By
comparing the highlighted desired change in known areas, we developed the training sets and
model the random forest algorithm to predict and classify other areas in the whole image.
3.2 Greenness Calculations
Normalized Difference Vegetation Index (NDVI)
We calculate three greenness metrics (NDVI, EVI, NBR). Normalized Difference Vegetation
Index (NDVI) is one of the most widely used method to detect pre-classification change (Lyon et
al., 1998). Healthy, green vegetation absorbs blue and red wavelengths lights for photosynthesis
process with chlorophyll, and they reflect green and infrared (invisible) wavelengths. Making use
of the difference between absorbance of red band and reflectance of near infrared band, NDVI
could show the distribution of vegetation canopy (Jensen, 2004). The formula of NDVI is listed
below:

7

𝑁𝐷𝑉𝐼 =
𝑁𝐼𝑅 − 𝑅𝐸𝐷 𝑁𝐼𝑅 + 𝑅𝐸𝐷

In this formula, RED and NIR represents the electromagnetic radiation of red (band 3 of
Landsat TM) and near-infrared (band 4 of Landsat TM) bands. The resultant NDVI ratio has a
possible range of -1.0 to 1.0, and the contribution of near-infrared wavelengths could generally
represent the density of forest. If the sensor captured more electromagnetic radiation in near-
infrared (invisible) wavelengths than red (visible) wavelength, then there might be some dense
vegetation in that pixel. If the difference between near-infrared and red wavelengths was small,
then the pixel might be barren and consisted of unhealthy vegetation, rock, or desert (Illustration
by Robert Simmon).
Enhanced Vegetation Index (EVI)
The Enhanced Vegetation Index (EVI) is a vegetation index that enhances the vegetation
signals determined with NDVI, but with adjustment of background noise, atmospheric noise and
saturation (USGS, 2017). It is computed as:

𝐸𝑉𝐼 = 𝐺 ∗
(𝑁𝐼𝑅 − 𝑅𝐸𝐷 )
(𝑁𝐼𝑅 + 𝐶 1 ∗ 𝑅𝐸𝐷 − 𝐶 2 ∗ 𝐵𝐿𝑈𝐸 + 𝐿 )

RED, NIR and BLUE are the reflectance of red (band 3), near-infrared (band 4) and blue (band
1) from the Landsat imagery, C1 and C2 are the coefficients of the aerosol resistance term using
the blue band to correct aerosol influences from the red band, L is the coefficient adjusting for
background canopy cover, and G is the gain factor. The coefficient values adopted for Landsat 5
are G = 2.5, C1 = 6, C2 = 7.5 and L = 1 (USGS, 2017). Comparing with chlorophyll sensitive
NDVI, EVI is more sensitive to canopy structural variations, such as canopy type, leaf area index
(LAI), plant physiognomy, and canopy architecture (Huete et al., 2002).
Normalized Burn Ration (NBR)
8

The Normalized Burn Ratio (NBR) is an index used to characterize burned areas and the
severity of burn. Instead of calculating with red and near-infrared band in NDVI, NBR included
information of shortwave infrared and near-infrared band:

𝑁𝐵𝑅 =
(𝑁𝐼𝑅 − 𝑆𝑊𝐼𝑅 )
(𝑁𝐼𝑅 + 𝑆𝑊𝐼𝑅 )

NIR and SWIR are the electromagnetic radiation of near-infrared (band 4) and shortwave
infrared (band 7) spectra. Health vegetation is represented more prominently in the NIR than SWIR
band, while burned land with scarred woody vegetation and earth are more strongly represented in
the SWIR than NIR band (Figure 4).

Figure 4. Different reflectance of SWIR and NIR between healthy vegetation and burned areas.
Image source: US Forest Service

3.3 Greenness Change
Since we created three single bands for three indexes in a certain year, we can detect the
difference of single-date transformations between 2003 and 2011 by subtracting one from the other.
Based on the definition of NDVI, the change of NDVI can be regarded as the change of the
green leaves in an area and then it could be used as the index of the greenness coverage change.
9

Similarly, EVI change would be more presentative to the vegetation variability because of its
sensitivity of canopy structural variation.
For NBR, calculating the difference between pre-fire and post-fire NBR is the best way to
detect the severity and effect of burning. Although we cannot collect this index immediately before
or after a fire, the slow plant regeneration in Southern California provided good conditions for us
to detect the land cover change with NBR. To interpret the meaning of dNBR (the change in NBR),
we create the table of burn severity categories as below (Table 2.):

Table 2. Burn Severity Categories
dNBR Burn Severity
< -0.5 High-severity burn
-0.5 to -0.25 High-moderate severity burn
-0.25 to -0.1 Moderate-low severity burn
-0.1 to +0.1 Unburned
0.1 to 0.25 Low vegetation regrowth
0.25 to 0.5 High vegetation regrowth

3.4 Supervised Image Classification
Supervised image classification is a method to classify the pixels in an image according to the
pixels in known cover types. We stack two-date images and use their raw pixel values to conduct
pre-classification method. After enhancing, stretching and comparing images, we visually inspect
the difference between the two dates and develop the training sites in the stacked image. The
random forest algorithm matches the pixels in the whole stacked image with the spectral signatures
in training sites and then classifies them into the closest classification.
Band Stacking and Image Enhancement
For each area, we stacked the bands from two images into one image and then created two
duplicates. Using false color composite and adjusting the images with standard deviations of pixels,
we enhanced three images to show the vegetation distribution in 2003, in 2001 and the difference
between two years respectively. As for the single-date images, the false color composite band
10

combination was SWIR1, NIR and Red bands. In our study, 2003 bands were stacked on the top
of 2011 bands, so the false color combination for the vegetation distribution in 2003 would be band
5, 4 and 3, and for the vegetation distribution in 2011 would be band 11, 10 and 9. To enhance the
multi-date image, the false color combination is SWIR1 in 2011 (band 11), SWIR1 in 2003 (band
5) and Blue in 2011 (band 7).
Developing Training Sites
In this study, we labeled the land cover change into four categories: Vegetation to Vegetation,
Vegetation to Non-Vegetation, Non-Vegetation to Vegetation, Non-Vegetation to Non-Vegetation.
Once we combined the bands in false color composite for the multi-date image, the land cover
change could be represented as colors. Areas in dark were generally areas that were vegetation
covered in both 2003 and 2011. Areas in magenta or purple were areas used to be covered by
greenness in 2003 while barren in 2011. Areas in light green were the areas used to be barren in
2003 while covered by vegetation in 2011. Areas in white or yellow represent areas that were non-
vegetation covered in neither 2003 nor 2011. We compare these colors with the vegetation
distribution in 2003 and 2011 to confirm the changes (Figures 5-7). With these composite images,
we selected more than ten polygons per category to develop the training sites.

Figure 5. Example of the Vegetation Difference in the Multi-Date Image
Non-Vegetation to Non-
Vegetation
Vegetation to Non-
Vegetation
Non-Vegetation to
Vegetation
11

Figure 6. Example of the Vegetation Distribution in 2003

Figure 7. Example of the Vegetation Distribution in 2011

Random Forest Algorithm
Random forest is one of the most widely used machine learning algorithms; it combines
bootstrap aggregation (bagging) and decision trees to classify data into groups (James et al 2013).
Non-Vegetation
Vegetation
Non-Vegetation
Non-Vegetation
Non-Vegetation
Vegetation
12

The algorithm samples a random set with replacement from the training data (in this case, pixels
from the imagery) as one bootstrap sample and builds the classification tree on each random set
(Figure 8).

Figure 8. Random Subsets in Random Forest Algorithm. Source: NASA’s Applied Remote
Sensing Training Program

The unsampled data serve as the testing set and are used to compute the error rate. If the data
in the testing set is not in any bootstrap samples, it is called out-of-bag (OOB) testing (Breinman,
2001). We used OOB testing since there were about 36% of the times on average for each
observation to not be in the bootstrap samples, and we calculated the average OOB as the OOB
error rate.
Compared to standard decision trees, random forests are preferred because by using random
samples the trees are not overfit and aggregate many decision trees to find the optimal groupings.
Predicting data from a fitted tree thus depends on the occurrences of different classifications from
different trees (Breiman, 2001). For example, if many trees have a few common features, then
predictions will be improved as classification errors will be reduced.
Post-Processing
The Random Forest classification output was filtered, removing noise from the image with
majority filter tool in QGIS, and we changed the symbology to visualize the classification.

13

4 Results and Conclusions
4.1 Greenness Calculation and Changes
The images for all three indices for 2003 and 2011 are shown in the Appendix (Figure a-1 to
Figure a-3). Images of the greenness changes obtained by subtracting the metrics in 2003 from
those in 2011 are shown in the Appendix (Figure a-4 to Figure a-6).
Visually, we see many more red than green points over the study area (Figure a-4 to Figure a-
6), indicating a decrease in greenness. All three indexes showed severe decreases in Angeles
National Forest from 2003 to 2011. Specifically, decreasing NDVI indicates the greenness was
greatly reduced over this time period, and as evidenced by EVI, the complexity of canopy structure
was also reduced given smaller EVI in 2011 than in 2003. The decrease of NBR showed the
evidence of burning. According to the record, Station Fire, the largest wildlife of the 2009
California wildfire season and the largest wildfire in the history of Los Angeles County, burned in
the Angeles National Forest on August 26, 2009 (Pasadena Star-News). This is consistent with our
results.
To make the metrics of greenness change more interpretable and to connect them with the
locations of the study subjects from the Children’s Health Study, we summarize the distribution of
the three indexes within 1,000 random spatial points selected from eastern Santa Barbara to
Victorville and to Northern Temecula (Table 3), which was used to represent the changes in the
whole study region.

Table 3. Summary of statistics of greenness metrics in 2003 and 2011
Min 1
st
Quantile Median 3
rd
Quantile Max Mean SD
NDVI in 2003 -0.226 0.088 0.183 0.308 0.659 0.207 0.165
NDVI in 2011 -0.188 0.051 0.122 0.253 0.608 0.152 0.141
EVI in 2003 -0.006 0.003 0.006 0.009 0.032 0.006 0.005
EVI in 2011 -0.010 0.001 0.002 0.005 0.018 0.003 0.004
NBR in 2003 -0.272 0.063 0.171 0.301 0.697 0.192 0.179
NBR in 2011 -0.265 0.052 0.153 0.256 0.607 0.162 0.153
14

From the table above, we notice that both means and standard deviations of all three metrics
decreased over 8 years. The distribution of the difference between metrics in 2003 and 2011 are
shown in Figures 9-11.

Figure 9. Distribution of NDVI difference

The vertical line at zero indicates no change, showing that over the region the density of
positive NDVI change was much smaller than that of negative NDVI change. This indicates that
there was a substantial decrease in NDVI from 2003 to 2011.

Figure 10. Distribution of EVI difference
15

Like the NDVI change, the decrease in EVI between 2003 and 2011 was far greater than the
increase of EVI over the time period. Thus, in addition to overall greenness decreasing, the
complexity of canopy structure and vegetation variety were reduced over the study period.

Figure 11. Distribution of NBR difference

As mentioned in the methods, -0.1 to 0.1 of the dNBR would be considered as unburned, so
we focused on the tails rather than the center of the distribution. As the densities of NBR difference
in -0.2 and -0.4 were larger than those in 0.2 and 0.4, there were more moderate-high severity of
burnings in this area than vegetation regeneration.
Since all three distributions located slightly left to the zero, we can conclude that the negative
NDVI, EVI and NBR were more than positive ones. Therefore, although there were some
vegetation regenerations between 2003 and 2011, the reduction in vegetation coverage was very
prominent in Southern California.
4.2 Supervised Image Classification
To make quantitative analysis of land cover change, the classification was defined by four
categories. Results were shown below (Figure 12):

16

Figure 12. Land Cover Change Classification by Random Forest Algorithm

From the map shown in Figure 19, it is evident by the red shading that Alamo Mountain, Sespe
Condor Sanctuary, Rocky Peak Park and Angeles National Forest encountered severe vegetation
loss between 2003 and 2011. There was also some reduction in greenness among forest and parks
in South, such as Whiting Ranch Wilderness Park and Capitan Grande Reservation. Changes from
non-forest to forest are seen in areas around the Alamo Mountain, Sespe Condor Sanctuary, Rocky
Peak Park and Angeles National Forest. These changes can also be seen on many mountains in the
east of the study region.
As before, to connect the classification with the locations of the study subjects from the
Children’s Health Study, we randomly selected 1,000 points from eastern Santa Barbara to
Victorville and to Northern Temecula and extracted the land cover change classification of these
points, which was used to represent the changes in the whole study region (Table 4). The average
17

errors and out-of-bag error rates of the random forests for the entire study area were listed below
(Table 5 and 6).

Table 4. Summary of land cover change classification
Forest to Forest Forest to Non-
Forest
Non-Forest to
Forest
Non-Forest to
Non-Forest
Frequency 143 127 45 87
Percentage (%) 35.57 31.59 11.19 21.16

Table 5. Confusion Matrix
Confusion Matrix
1 2 3 4 class.error
1 292 10 15 0 0.079
2 17 294 3 3 0.073
3 15 0 293 9 0.076
4 1 2 8 309 0.034

Table 6. Summary of OOB error rate
area number OOB error rate
1 10.35%
2 2.86%
3 6.34%

Among locations of the study subjects from the Children’s Health Study, 35.57% of the area
remained vegetation covered between 2003 and 2011. 31.59 % of the area were barren in both
2011 and 2003. It is worth mentioning that 31.59% of the area used to be covered by greenness in
2003 changed to be non-forest in 2011. In contrast, only 11.19% of the area changed from non-
forest coverage to forest coverage. Therefore, the change from a green metric (forest) to a non-
green metric (non-forest) in the interested area between 2003 and 2011 is clear and substantial.
18

For all three areas, the classification errors were less than 0.2, which were pretty low. Besides,
the out of bag error rates were also pretty low in these three areas. The OOB error rates in area 1
is a little bit high because it’s the largest area among our study region. In general, it showed a great
performance of this classification.
Our results highlight that over the study period there was measurable changes in land cover
related to greenness. Some of this change may be attributed to natural phenomenon such as forest
fires and drought, but in more urban areas the observed decreases in greenness suggest
urbanization has increased over the 8-year period. Future work to determine if there are any health
effects associated with these land use changes is warranted.

19

5 References
Hussain, M., Chen, D., Cheng, A., Wei, H., & Stanley, D. (2013). Change detection from
remotely sensed images: From pixel-based to object-based approaches. ISPRS Journal of
Photogrammetry and Remote Sensing, 80, 91–106.
https://doi.org/10.1016/j.isprsjprs.2013.03.006
Xiao, J., Shen, Y ., Ge, J., Tateishi, R., Tang, C., Liang, Y ., & Huang, Z. (2006). Evaluating urban
expansion and land use change in Shijiazhuang, China, by using GIS and remote sensing.
Landscape and Urban Planning, 75(1–2), 69–80.
https://doi.org/10.1016/j.landurbplan.2004.12.005
Lunetta, R.S., and C. D. Elvidge. (1998). “Remote Sensing Change Detection.” Michigan: Ann
Arbor Press
Jensen, J. R. (2004). Digital change detection. Introductory digital image processing: A remote
sensing perspective (pp. 467–494). New Jersey: Prentice-Hall.
USGS. (2017). Landsat Surace Reflactance-Derived Spectral Indices, Product Guide,
(December), 1–31. https://doi.org/10.1016/0042-207X(74)93024-3
Lyon, J. G., Yuan, D., Lunetta, R. S., & Elvidge, C. D. (1998). A change detection experiment
using vegetation indices. Photogrammetric Engineering and Remote Sensing, 64, 143–150.
A.Huete, K. Didan, T. Miura, E. P. Rodriguez, X. Gao, L. G. Ferreira. Overview of the
radiometric and biophysical performance of the MODIS vegetation indices. Remote Sensing of
Environment 83(2002) 195-213 doi:10.1016/S0034-4257(02)00096-2
Breiman. Random forests. Machine Learning, 45(1): 5–32, 2001. 18
Nery T, Sadler R, Solis-Aulestia M, White B, Polyakov M, Chalak M. 2016. Comparing
supervised algorithms in Land Use and Land Cover classification of a Landsat time-series. Int
Geosci Remote Sens Symp 2016–November:5165–5168; doi:10.1109/IGARSS.2016.7730346.
United States Geological Survey (USGS), Earth Resources Observation and Science Center.
Landsat – Earth Observation Satellites https://pubs.usgs.gov/fs/2015/3081/fs20153081.pdf, 2015.
Markham BL, Storey JC, Williams DL, Irons JR. 2004. Landsat Sensor Performance: History
and Current Status. IEEE Trans Geosci Remote Sens 42: 2691.
20

Taylor L, Hochuli DF. 2017. Defining greenspace: Multiple uses across multiple disciplines.
Landsc Urban Plan 158:25–38; doi:10.1016/j.landurbplan.2016.09.024.
De Vries S, van Dillen SME, Groenewegen PP, Spreeuwenberg P. 2013. Streetscape greenery
and health: Stress, social cohesion and physical activity as mediators. Soc Sci Med 94:26–33;
doi:10.1016/j.socscimed.2013.06.030.
James, G., Witten, D., Hastie, T, Tibshirani, R. An Introduction to Statistical Learning with
Applications in R, Springer, 2017.
"New fire breaks out near Angeles Crest Highway; forces road closure. Vetter mountain fire
lookout tower was also lost in this fire". Pasadena Star-News

21

6 Appendix

Figure a-1. NDVI

Figure a-2. EVI
22

Figure a-3. NBR

Figure a-4. Change of NDVI

23

Figure a-5. Change of EVI

Figure a-6. Change of NBR

Abstract (if available)

Abstract Land cover change plays an important role in socioeconomical and bioecological systems, and recent research has shown its importance related to aspects of human health. In an urbanizing world, built environments are replacing green spaces, and concurrently, environmental stressors such as noise, air pollution, and heat are increasing. Southern California is home to 4 of the 10 most populous counties in the United States, and according to the last census, Los Angeles-Long Beach-Anaheim has surpassed New York City as the most densely populated metro area. Focusing Southern California, we examine the change in greenspace from 2003 to 2011 using 30 m resolution satellite imagery from the Landsat Thematic Mapper (TM). Using open source software QGIS, we calculated and visualized the Normalized Difference Vegetation Index (NDVI), Enhanced Vegetation Index (EVI) and Normalized Burn Ratio (NBR) as indicators of greenspace change. After visual analysis and training, land use change was classified into four forest and non-forest categories using the random forests algorithm with error rates around 0.06% and out-of-bag error rates around 6.34% respectively. Our findings show that over the 8-year period greenspace in Southern California has significantly decreased. The biggest change from 2003 to 2011 is in the NDVI greenness metric, where the mean decreased from 0.207 to 0.152 over this period. The mean EVI changed from 0.006 to 0.003, and mean NBR changed from 0.192 to 0.162.

Linked assets

University of Southern California Dissertations and Theses

Conceptually similar

PDF

Using multi-angle imaging spectroradiometer aerosol mixture properties and meteorology for PM₂.₅ assessment in Iran

PDF

Machine learning approaches for downscaling satellite observations of dust

PDF

Downscaling satellite observations of dust with deep learning

PDF

Inference correction in measurement error models with a complex dosimetry system

PDF

Identifying prognostic gene mutations in colorectal cancer with random forest survival analysis

PDF

Uncertainty quantification in extreme gradient boosting with application to environmental epidemiology

PDF

Comparison of models for predicting PM2.5 concentration in Wuhan, China

PDF

Hierarchical regularized regression for incorporation of external data in high-dimensional models

PDF

Age-period-cohort analysis of hepatocellular carcinoma (HCC) incidence trends among Asians and non-Hispanic Whites in California, 1988-2017

PDF

Instability of heart rate and rating of perceived exertion during high-intensity interval training in breast cancer patients undergoing anthracycline chemotherapy

PDF

Essays on bioinformatics and social network analysis: statistical and computational methods for complex systems

PDF

Relationship of blood pressure and antihypertensive medications to cognitive change in the BVAIT, WISH, and ELITE clinical trials

Asset Metadata

Creator Zhang, Yi (author)

Core Title Assessment of land cover change in Southern California from 2003 to 2011 using Landsat Thematic Mapper

School Keck School of Medicine

Degree Master of Science

Degree Program Biostatistics

Publication Date 04/28/2019

Defense Date 04/24/2019

Publisher University of Southern California (original), University of Southern California. Libraries (digital)

Tag classification,EVI,land cover,landsat,NBR,NDVI,OAI-PMH Harvest,QGIS,random forest,Southern California

Format application/pdf (imt)

Language English

Contributor Electronically uploaded by the author (provenance)

Advisor Franklin, Meredith (committee chair), Lewinger, Juan Pablo (committee member), Marjoram, Paul (committee member)

Creator Email yizh2018@gmail.com,zhan039@usc.edu

Permanent Link (DOI) https://doi.org/10.25549/usctheses-c89-158256

Unique identifier UC11662924

Identifier etd-ZhangYi-7256.pdf (filename),usctheses-c89-158256 (legacy record id)

Legacy Identifier etd-ZhangYi-7256.pdf

Dmrecord 158256

Document Type Thesis

Format application/pdf (imt)

Rights Zhang, Yi

Type texts

Source University of Southern California (contributing entity), University of Southern California Dissertations and Theses (collection)

Access Conditions The author retains rights to his/her dissertation, thesis or other graduate work according to U.S. copyright law. Electronic access is being provided by the USC Libraries in agreement with the a...

Repository Name University of Southern California Digital Library

Repository Location USC Digital Library, University of Southern California, University Park Campus MC 2810, 3434 South Grand Avenue, 2nd Floor, Los Angeles, California 90089-2810, USA