Land surface temperature with land cover dynamics: multi-resolution, spatio-temporal data analysis of Greater Bangalore

Results

Temporal LU details are displayed in figure 2 and class statistics are listed in table 1. The classified images of 1973, 1992, 1999, 2000, 2002, 2006 and 2007 showed an overall accuracy of 72%, 75%, 71%, 77%, 60%, 73% and 55%. Accuracy assessment was performed which showed higher accuracy for high resolution data (~ 70-75% for Landsat and IRS LISS-III) and decreasing accuracy with coarse spatial resolution (~ 55-60% for MODIS). Figure 3 (a) – (f) depicts the LU change based on differencing techniques of PCA and TC. The disappearance of water bodies from 1973 to 2006 is given in Figure 4. 55% decline (from 207 to 93) in the number of water bodies and 61% decline in the spatial extent (of water bodies) is noticed from the temporal analysis. Validation was done considering training data and Google Earth image, covering approximately 15% of the study area. Then, pixels corresponding to urban category were extracted for further analysis. Figure 5 shows the LST and NDVI of Greater Bangalore in 1992, 2000 and 2007. The minimum (min) and maximum (max) temperature was 12 °C and 21 °C with a mean of 16.5±2.5 from Landsat TM (1992, winter). Similarly MODIS data of 2000 (summer) show the min, max and mean temperature of 20.23, 28.29 and 23.71±1.26 °C respectively. Corresponding values for 2007 (summer) are 23.79, 34.29 with a mean of 28.86±1.60 °C. LC wise NDVI and LST are listed in table 2.


Figure 2. Greater Bangalore in 1973, 1992, 1999, 2000 and 2006.



Figure 3: PC1 of Landsat MSS-1973 (a), PC1 of IRS LISS-III-2006 (b) and the change in LU is highlighted in (c) using PCA differencing methods. The highlighted box in (a) and (b) are further enlarged in (d) to (f). Brightness values from TC transformation on Landsat TM-1992 (d) and ETM+2000 (e). The changes are highlighted in (f) by differencing the brightness values emphasizing new built up.

TABLE 1. Greater Bangalore land cover statistics.

Class
Year 		Built up Vegetation Water Bodies Others
1973 Ha 5448 46639 2324 13903
% 7.97 68.27 3.40 20.35
1992 Ha 18650 31579 1790 16303
% 27.30 46.22 2.60 23.86
1999 Ha 23532 31421 1574 11794
% 34.44 45.99 2.30 17.26
2000 Ha 24163 31272 1542 11346
% 35.37 45.77 2.26 16.61
2002 Ha 26992 28959 1218 11153
% 39.51 42.39 1.80 16.32
2006 Ha 29535 19696 1073 18017
% 43.23 28.83 1.57 26.37
2007 Ha 30876 17298 1005 19143
% 45.19 25.32 1.47 28.01



Figure 4. Temporal changes in water bodies from 1973 (using Landsat MSS) to 2006 (using IRS LISS-III) highlighted in rectangular boxes and circles.



Figure 5: LST from Landsat TM (1992), MODIS (2000 and 2007), NDVI from Landsat TM (1992), Landsat ETM+ (2000) and IRS LISS-III (2006).


TABLE 2. NDVI and LST (°C) for respective land uses.

Land cover 1992 (TM) 2000 (MODIS) 2007 (MODIS)
Mean LST
(SD) 		Mean NDVI
(SD)		 Mean LST
(SD) 		Mean NDVI
(SD)		 Mean LST
(SD) 		Mean NDVI
(SD)
Builtup 19.03
(1.47)		 -0.162
(0.096)		 26.57
(1.25)		 -0.614
(0.359)		 31.24
(2.21)		 -0.607
(0.261)
Vegetation 15.51
(1.05)		 0.467
(0.201)		 22.21
(1.49)		 0.626
(0.27)		 25.29
(0.44)		 0.348
(0.42)
Water bodies 12.82
(0.62)		 -0.954
(0.055)		 21.27
(1.03)		 -0.881
(0.045)		 24.00
(0.27)		 -0. 81
(0.27)
Open ground 17.66
(2.46)		 -0.106
(0.281)		 24.73
(1.56)		 -0.016
(0.283)		 28.85
(1.54)		 -0.097
(0.18)

The relationship between LST and NDVI were investigated for each LC type through the Pearson’s correlation coefficient (CC) at a pixel level, which are listed in table 3. It is apparent that values tend to negatively correlate with NDVI for all LC types. NDVI values ranges from -0.05 to -0.6 (built up) and 0.15 to 0.6 (vegetation). Temporal increase in temperature with the increase in the number of urban pixels is noticed during 1992 to 2007 (63%) and ‘r’ confirms this relationship for the respective years. The decrease in vegetation is reflected by the respective increase in temperature. Further analysis is done by considering vegetation abundance.

Landsat ETM+ (band 1, 2, 3, 4, 5 and 7) were unmixed to get the abundance maps of 5 classes (1) dense urban (commercial/industrial/residential), (2) mixed urban (urban with vegetation and open ground), (3) vegetation, (4) open ground and (5) water bodies. We considered only dense urban, mixed urban and vegetation abundance for further analysis as shown in figure 6. The min and max temperature from ETM+ data was 13.49 and 26.32 °C with a mean of 21.75±2.3. These abundance images were further analysed to see their contribution to the UHI by separating the pixels that contains 0-20%, 20-40%, 40-60%, 60-80% and 80-100% of the commercial/industrial/residential (dense urban), mixed urban and vegetation. Table 4 gives the mean and standard deviation (SD) of the LST for various LU. Application of decision based unmixing approach, systematically exploited the information from both the sources (sub-pixel class proportions and classified image based on training data collected from the ground) for achieving more reliable classification, shown in figure 7. Table 5 lists LC wise LST, NDVI and correlation coefficient. Relationship of population density with LST (Landsat ETM+) is evident in figure 8, which corroborate that the increase in LST is due to urbanisation and consequent increase in population.


Figure 6. Abundance maps and LST obtained from Landsat ETM+ data (2000).


TABLE 3. Correlation coefficients between LST and NDVI by LC type (significant at 0.05 level).

LU 1992 2000 2007
Builtup -0.7188 -0.7745 -0.7900
Vegetation -0.8720 -0.6211 -0.6071
Open ground -0.6817 -0.5837 -0.6004
Water bodies -0.4152 -0.4182 -0.4999

TABLE 4. Mean LST for different LC classes for various abundances.

Class
Abundance 		Mean Temperature ± SD
of dense urban 		Mean Temperature ± SD 
of mixed urban		 Mean Temperature ± SD
of vegetation
0-20% 21.99±2.37 21.57±2.36 17.91±2.19
20-40% 22.06±2.15 21.58±2.36 17.39±1.37
40-60% 22.27±2.00 21.67±2.41 17.22±0.89
60-80% 22.33±2.22 22.28±2.02 17.13±0.85
80-100% 22.47±1.96 22.37±2.17 17.12±0.91


Figure 7: Classified image obtained from combining unmixed images and classified image using spectral signatures from ground as input to Baye’s classifier from 6 MSS bands of Landsat ETM+.



Figure 8: LST with ward wise population density.


TABLE 5. LST, NDVI and correlation coefficient for different LC classes.

Landuse LST
Mean ± SD		 NDVI
Mean ±SD		 Correlation coefficient between LST and NDVI
Dense builtup 23.09± 1.16 -0.2904± 0.395 -0.7771
Mixed builtup 22.14± 1.06 -0.138± 0.539 -0.6834
Vegetation 19.27± 1.59 0.3969± 0.404 -0.8500
Open Ground 22.40± 1.97 -0.0193± 0.164 -0.6319
Water Bodies 19.57± 1.72 -0.301± 0.47 0.2319

TABLE 6. MMU sizes for different RS data sources used.

Data source MODIS Landsat MSS Landsat TM/ETM+ IRS LISS-III
MMU (ha) 6.25 0.62 0.09 0.055