Improved Land Cover Mapping Using Landsat 8 Thermal Imagery

1

• Paper ID78
• Title
• Name

Improved Land Cover Mapping Using Landsat 8
Thermal Imagery
By Jwan M. Aldoski
Prof. Dr. Shattri B. Mansor Prof. Dr. H'ng Paik
San Dr. Zailani Khuzaimah Address Universiti
Putra Malaysia, Faculty of Engineering,
Department of Civil Engineering, Jalan UPM
,43400 Serdang, Selangor, Malaysia
Emails Aldoski-Jwan_at_hotmail.com,
shattri_at_gmail.com,
ngpaiksan_at_upm.edu.my
zailanikhuzaimah_at_gmail.com
2
PRESENTATION OUTLINE 
3
INTRODUCTION

• Land Cover dynamics acts as a primary parameter
  in current strategies and policies for natural
  resource management and monitoring. Thus
  reasonable use of the available land is vital for
  the sustainable conservation of the
  bio-environment, which eventually enhances the
  socio-economic status of sustainable livelihoods.
  This includes the accurate estimation of current
  and past land cover dynamics.
• With the advance and development of advanced
  geospatial techniques that incorporate the use of
  Remote Sensing (RS), Geographic Information
  Systems (GIS) and Global Positioning System
  (GPS), the listing of spatial-temporal LC
  dynamics has become simple, fast, cost-effective
  and accurate.

4
INTRODUCTION

• Digital image processing on multi-temporal or
  multi-spectral satellite imagery has enormous
  potential for LC categorization, landscape
  dynamics, and change detection analysis.
• Digital classification methods involve
• Unsupervised classification approaches
• Supervised classification approaches the most
  widely utilized
• Object-based classification approaches showed
  better accuracy
• In addition, a Hybrid classification is
  frequently used to differentiate land features.
  Recently, the accuracy of classification can be
  enhanced by utilizing Multi-source Data .

5
INTRODUCTION

• Land surface temperature (LST) measured from the
  remotely sensed thermal band reveals a specific
  response to landscape dynamics including LC
  adjustment .Therefore, thermal infrared (TIR)
  sensors can assess the quantitative information
  of Ts throughout varying LC categories.
• There are intricate relationships among LST and
  several physico-chemical and biological processes
  of the Earth. As a consequence, LST serves as a
  key parameter in the physics of land surface
  processes, earth-atmosphere interactions, and
  energy fluxes between the ground and the
  atmosphere since it is part of the energy balance
  .

6
AIM OF STUDY
The current study analyzes the possibilities for
thermal data from Band 10 of Landsat 8 and
spectral data to improve the accuracy of the LC
classification over heterogeneous tropical forest
areas.
7
STUDY AREA
Kuala Krai District locate in the center of the
state of Kelantan in the northeast of
Malaysia. The study area in this research covers
an area of almost 2329 km2
Figure Study area LC map (Kuala Krai District,
Malaysia)
8
DATA SETS
Landsat 8 (L8) Operational Land Imager (OLI) and
the Thermal Infrared Scanner (TIRS) imagery used
2014 and obtained from (http/earthexplorer.usgs.
gov) freely . The Malaysian survey toposheets
were as well utilized to identify the LC along
with the satellite imagery on the UTM projection,
Level L1 T.
9
METHODOLOGY FLOW-DIAGRAM
10
METHODOLOGY
A preprocessing Landsat 8 OLI Data April 14, 2014
Analyzed by utilizing ENVI image processing
software (v5.1). 1- Radiometric and Atmospheric
Corrections through applying the latest
radiometric calibration coefficients published
(http//landsat.usgs.gov) with FLAASH model 2-
Geometric correction The positional precision in
x and y image directions as Root Mean Square
Error (RMSE) was 0.07, well below standard
requirements of less than 1 pixel using 15 GCPs
and (DEM) with (RMSE) 0.07 3- Band selection
2, 3, 4, 5, 6, and 7 bands 4- Sub-set and
Re-projection process re-projected to UTM
(Universal Transverse Mercator), DatumWSG-1984,
zone 48 N and resampled to a spatial resolution
of 30 m, then use it as a dataset to perform
classification methods and subsequent processes.
11
METHODOLOGY
B Landsat 8 Spectral VIs
Eq. Formula Comments
1 NDVI(NIR-R)/ (NIRR)/ Surface reflectance in near-infrared (NIR) and red (R) spectral bands
2 SAVI(NIR-R)(1-L)/(NIRRL) L is a constant whose value depends on the soil properties
3 LAI 2.3689SAVI 0.7877  
4 EVIG (NIR_R)/ NIRC1R-C2BL Surface reflectance in blue (B) spectral band, G2.5, C1 6, C2 7.5, L 1
5 EVI2 2.5 (NIR-R)/ (NIR2.4R1)  
6 e 0.047ln(NDVI) 1.009 e 0.003 (LAI) 0.97 for LAIlt3.0  
7 ashort 0.3a2 0.277 a3 0.233a40.143 a5 0.036 a60.12 a7 a short is shortwave broadband albedo, and a1, . . ., a7 are the reflectance of the respective band number of Landsat ETM
8 fractional vegetation cover Fc 1 _ (NDVIsc max-NDVIi / NDVIsc max-NDVIsc min)0625 NDVIscmax and NDVIscmin are the maximum and minimum NDVI values from the L8 scene and NDVIi is the NDVI value of ith pixel
9 Thermal Integrated Vegetation Index TLIVI (DN L8 (band10)-NDVI-LAI) / (DN L8 band10)NDVILAI)  In this study
10 Advanced Thermal Integrated Vegetation Index ATLIVI (DN L8 (band10)-NDVI-LAI-EVI2) / (DN L8 (band10) NDVI LAI EVI2)  In this study
12
METHODOLOGY
C Retrieval of Ts and Associated Parameters
Calibration for band 10 thermal band data is
conducted utilizing a two-step process 1- The
first phase includes converting the band to 10
digital number (DN) values into L? . 2-
Secondly, this L? is transformed to Tb in Kelvin.
Next, an emissivity adjustment is made utilizing
surface emissivities for the specified LCs
estimated from the NDVI and LAI .
13
METHODOLOGY
D SVM Implementation and post processing
The steps are 1) LC Image interpretation
and LULC Categories Collecting Training and 100
test data using Random Pixel Selection Strategy
Training points for LC classification have been
identified on the basis of knowledge acquired via
comprehensive ground survey and thorough field
analysis of the area the topographical sheets
and the Spot 5(2.5 m) image of the study area
have been taken into consideration. 11 LC
features, notably, Water Bodies (WB), Residential
Area (RA), Primary Forest (PF), Secondary Forest
(SF), Swamp Forest / Mangrove Swamp (SWF), Oil
Palm (OP), Rubber(R), Others Crops (OC),
Marshland (ML), Scrub Area (SA) Cleared Land
(CL). 2) Spectral separability
14
METHODOLOGY
D SVM Implementation and post processing
3) Apply SVM classifier
TLIVI and ATLIVI are combined with the NIR and
Red L8 OLI imagery bands as FCC set is utilized
as a reference for the SVM supervised
classification. SVM parameters Radial Base
Function RBF kernel function Penalty value
C120 Kernel parameter ? 0.15 probability
threshold value 1
.
4) Post Classification Processing using
Majority Filtering
Parameters windows size 33 pixels by pixels
with centre pixel weight (1 value)
5) Accuracy Assessment using Error Matrices
6) Correlation (R2) statistics
15
RESULTS AND DISCUSION
Figure Study area LC map (Kuala Krai District,
Malaysia)

1. Interpretation of the LC feature

 The accuracy evaluation carried out for this
classification demonstrated an overall accuracy
of 93.57 and a Kappa accuracy of 0.92 
16
RESULTS AND DISCUSION
2. Retrieval of Ts parameters
  Numerous spatial parameters such as NDVI, SAVI,
LAI, EVI2, Surface Emissivity, LST and LC maps
connected to heterogeneous tropical forests of
the Kuala Krai district have been calculated
. The spatial variance of NDVI varied from
values less than 0 (0 to-0.7) in areas with no
vegetation cover and water bodies to 0.8 in areas
with a high vegetation cover density.
17
RESULTS AND DISCUSION
2. Retrieval of Ts parameters
  The spatial variance of LAI revealed that the
values varied from negative values of-2.25 in the
water bodies to positive values as high as 1.9 in
areas defined by vegetation.
18
RESULTS AND DISCUSION
2. Retrieval of Ts parameters
  Emissivity is directly correlated to NDVI and
LAI values and has similar trends in spatial
distribution as NDVI and LAI .
19
RESULTS AND DISCUSION
2. Retrieval of Ts parameters
  In the spatial sense, EVI2 revealed an average
higher value of almost 0.8 for heavily vegetated
areas, whereas an average value of-0.18 for
cleared land was received.
20
RESULTS AND DISCUSION
2. Retrieval of Ts parameters
The spatial pattern of Ts in the Kuala Krai
district and surrounding areas ranged between
almost 18Co in the water bodies and 19Co in the
primary forest at a minimum to a maximum of
30.5-31.9Co in the scrub and residential areas.
Hence, the temperature difference among different
LC classes reached nearly 16Co.
21
RESULTS AND DISCUSION
3. Heat flux radiation univariate statistics and
related parameters for LC categories
LC categories LC categories Minimum Maximum Mean Standard Error  
NDVI Water Bodies -0.3551 0.7847 0.4508 0.0381 0.0381
NDVI Residential Area 0.1279 0.7831 0.5934 0.0147 0.0147
NDVI Primary Forest 0.2365 0.8169 0.7001 0.0039 0.0039
NDVI Secondary Forest -0.2009 0.8352 0.7286 0.0029 0.0029
NDVI Swamp Forest / Mangrove Swamp 0.4932 0.7091 0.6546 0.0022 0.0022
NDVI Oil Palm -0.1294 0.8117 0.6770 0.0038 0.0038
NDVI Rubber -0.5136 0.8108 0.6843 0.0033 0.0033
NDVI Others Crops -0.3475 0.7985 0.6340 0.0092 0.0092
NDVI Marshland -0.3441 0.7904 0.6337 0.0105 0.0105
NDVI Scrub Area -0.3567 0.8010 0.6089 0.0057 0.0057
NDVI Cleared Land 0.0638 0.8218 0.6738 0.0067 0.0067
           
SAVI Water Bodies -0.1467 0.5573 0.2873 0.0223 0.0223
SAVI Residential Area 0.0941 0.5813 0.3763 0.0100 0.0100
SAVI Primary Forest 0.0389 0.6482 0.4130 0.0039 0.0039
SAVI Secondary Forest -0.0879 0.6678 0.4658 0.0025 0.0025
SAVI Swamp Forest / Mangrove Swamp 0.3214 0.4714 0.3950 0.0016 0.0016
SAVI Oil Palm -0.0490 0.5996 0.4452 0.0026 0.0026
SAVI Rubber -0.1910 0.6134 0.4323 0.0023 0.0023
SAVI Others Crops -0.1386 0.5975 0.3992 0.0058 0.0058
SAVI Marshland -0.1418 0.6051 0.4027 0.0068 0.0068
SAVI Scrub Area -0.1478 0.6182 0.3933 0.0038 0.0038
SAVI Cleared Land 0.0391 0.6551 0.4580 0.0057 0.0057
All LC categories have confirmed the variation in
the univariate statistical values of the
radiation thermal flux parameters, because these
are continuous spatial parameters characterized
by a gradational change in the values of each
parameter.
22
RESULTS AND DISCUSION
3. Heat flux radiation univariate statistics and
related parameters for LC categories
 LC categories  Maximum  Minimum  Mean  Standard Error
EVI2 Water Bodies -0.1231 0.5823 0.2884 0.0220
EVI2 Residential Area 0.0867 0.6131 0.3733 0.0107
EVI2 Primary Forest 0.0333 0.7021 0.4099 0.0044
EVI2 Secondary Forest -0.0750 0.7282 0.4713 0.0029
EVI2 Swamp Forest / Mangrove Swamp 0.3113 0.4770 0.3886 0.0018
EVI2 Oil Palm -0.0419 0.6392 0.4489 0.0029
EVI2 Rubber -0.1580 0.6547 0.4330 0.0024
EVI2 Others Crops -0.1165 0.6337 0.3975 0.0060
EVI2 Marshland -0.1192 0.6446 0.4019 0.0071
EVI2 Scrub Area -0.1241 0.6617 0.3923 0.0039
EVI2 Cleared Land 0.0350 0.7117 0.4664 0.0065
         
FC Water Bodies 0.3225 0.8924 0.7254 0.0191
FC Residential Area 0.5639 0.8915 0.7967 0.0073
FC Primary Forest 0.6182 0.9084 0.8500 0.0020
FC Secondary Forest 0.3996 0.9176 0.8643 0.0014
FC Swamp Forest / Mangrove Swamp 0.7466 0.8546 0.8273 0.0011
FC Oil Palm 0.4353 0.9058 0.8385 0.0019
FC Rubber 0.2432 0.9054 0.8421 0.0016
FC Others Crops 0.3262 0.8993 0.8170 0.0046
FC Marshland 0.3280 0.8952 0.8168 0.0052
FC Scrub Area 0.3217 0.9005 0.8045 0.0029
FC Cleared Land 0.5319 0.9109 0.8369 0.0033
23
RESULTS AND DISCUSION
3. Heat flux radiation univariate statistics and
related parameters for LC categories
TLIVI LC categories Minimum Maximum Mean Standard Error
TLIVI Water Bodies 8.3272 10.0926 9.1070 0.0368
TLIVI Residential Area 8.8023 10.5773 9.5591 0.0366
TLIVI Primary Forest 8.1937 10.1878 9.1265 0.0131
TLIVI Secondary Forest 8.3923 10.0759 9.3230 0.0077
TLIVI Swamp Forest / Mangrove Swamp 8.8461 9.0163 8.9347 0.0018
TLIVI Oil Palm 8.6029 10.2863 9.3942 0.0071
TLIVI Rubber 8.3122 10.2410 9.3029 0.0065
TLIVI Others Crops 8.2899 10.3146 9.3822 0.0155
TLIVI Marshland 8.4841 10.0524 9.3375 0.0146
TLIVI Scrub Area 8.3946 10.3790 9.5077 0.0097
Cleared Land 8.7310 10.0462 9.4056 0.0160
ATLIVI Water Bodies 8.5939 10.1316 9.3798 0.0349
ATLIVI Residential Area 9.0753 10.7666 9.7983 0.0309
ATLIVI Primary Forest 8.3198 10.1836 9.4198 0.0130
ATLIVI Secondary Forest 8.6473 10.2711 9.6521 0.0076
ATLIVI Swamp Forest / Mangrove Swamp 9.1108 9.2680 9.2021 0.0020
ATLIVI Oil Palm 8.8262 10.3515 9.6892 0.0058
ATLIVI Rubber 8.6267 10.2824 9.6048 0.0064
ATLIVI Others Crops 8.5919 10.3921 9.6525 0.0147
ATLIVI Marshland 8.9122 10.0936 9.6110 0.0130
ATLIVI Scrub Area 8.6974 10.5008 9.7628 0.0088
ATLIVI Cleared Land 8.9937 10.1887 9.6873 0.0156
24
RESULTS AND DISCUSION
4. Thermal Vis Maps
(b) ATLIVI map
(a) TLIVI map
25
RESULTS AND DISCUSION
5. Accuracy of LC classification Imagery
LC Categories Data Used Data Used Data Used Data Used Data Used Data Used
LC Categories Standard FCC Standard FCC TLIVI TLIVI ATLIVI ATLIVI
LC Categories UA PA UA PA UA PA
Water Bodies 74.11 90.62 83.67 88.76 82.12 90.20
Residential Area 86.03 92.16 92.03 96.30 92.87 95.42
Primary Forest 91.06 92.32 90.13 90.32 91.35 91.47
Secondary Forest 76.12 84.06 99.92 92.74 92.88 92.74
Swamp Forest / Mangrove Swamp 98.10 98.81 69.17 98.29 98.63 98.12
Oil Palm 64.11 60.93 76.93 61.82 69.25 72.19
Rubber 76.77 71.52 56.27 87.33 78.32 93.21
Others Crops 46.92 51.58 93.62 82.36 68.02 78.49
Marshland 92.87 76.97 92.18 86.35 93.61 87.91
Scrub Area 90.19 75.23 89.74 87.36 86.26 89.71
Cleared Land 82.33 76.28 87.68 79.92 87.95 88.65
OA () 84.15   91.32   93.57  
K 0.8103   0.8727   0.9203  
Note UA User Accuracy, PAProducer Accuracy,
OAOverall Accuracy, k Kappa coefficient
26
RESULTS AND DISCUSION
6. Statistical Assessment using R Correlation
2
The Table below indicate a fairly good
correlation between LST and EVI2 (R2 0.68)
compared to NDVI, SAVI, and LAI (R2 0.55-0.65)
  LST (in K) LAI SAVI NDVI FC TLIVI EVI2 ATLIVI
LST (in K) 1  
LAI 0.554 1  
SAVI 0.592 0.626 1  
NDVI 0.653 0.574 0.943 1  
FC 0.359 0.574 0.543 0.887 1  
TLIVI 0.825 0.508 0.614 0.758 0.652 1  
EVI2 0.681 0.601 0.997 0.918 0.918 0.577 1  
ATLIVI 0.887 0.461 0.796 0.681 0.581 0.759 0.530 1
27
RESULTS AND DISCUSION
7. Differences in the mean value for TLIVI and
ATLIVI per each LC classes
Note (Water Bodies (WB), Residential Area
(RA), Primary Forest (PF), Secondary Forest
(SF), Swamp Forest / Mangrove Swamp (SWF), Oil
Palm (OP), Rubber(R), Others Crops (OC),
Marshland (ML), Scrub Area (SA) and Cleared
Land (CL)).
28
CONCLUSIONS 

1. A suitable and reliable methodology for LC
   classification and mapping is still required for
   time.
2. Landsat 8 satellite OLI and TIRS data were
   utilized for the LC classification using SVM
   classifier algorithm.
3. This study indicated improvement in the accuracy
   of classification using combined thermal and
   spectral satellite imagery data almost 6 in the
   overall accuracy of the LC classification
   compared to the SVM algorithm using the Landsat 8
   standard False Colour Composite (FCC) satellite
   image as a reference.
4. Two thermal indices TLIVI and ATLIVI revealed
   fairly good correlations (R20.65 and 0.7,
   respectively) with the Ts.
5. Many factors also affect the retrieval of LST
   from satellite thermal infrared data,
   transmission, atmospheric moisture, radiance,
   etc. that are difficult to assess from remote
   satellite observations.
6. Retrieval of Ts of heterogeneous tropical forests
   by L8 demonstrates comparatively low ranges of Ts
   in forested vegetative areas, which was quite
   apparent.
7. Water bodies showed lower Ts values as opposed to
   barren areas.
8. Typically, there was an inverse relationship
   among LAI and surface albedo caused by increased
   absorption of the canopy and reduced reflection
   from the relatively lighter soil below the
   vegetation. Nevertheless, this has not been
   strongly noticed in this study

29
Thank You