RELATIONSHIP OF THERMAL PROPERTIES AND VEGETATION INDICES IN REMOTE SENSING LAND COVER MAPPING

Tran Thi Van

Institute for Environment and Resources - Vietnam National University Ho Chi Minh City 142 To Hien Thanh St., Dist. 10, Ho Chi Minh City, Vietnam

Tel: (84)-8-38651132 Fax: (84)-8-38655670 E-mail: tranthivan@hcmier.edu.vn KEYWORDS: Emissivity, Land cover, Land surface temperature, Thermal remote sensing, Vegetation indices ABSTRACT: This study aims to investigate the thermal remote sensing for retrieving the land surface temperature with taking the surface object emissivity into consideration. Then the study applies the thermal properties combined with vegetation indices to determinate final classes from the unsupervised classification results for the land cover map. The experiments were carried out on Aster image with the spatial 90m-resolution of thermal bands suited the researches in region and city level. The results show that the NDVI method for retrieving emissivity to estimate land surface temperature gained the realizable accuracy with the bias less than 2oC. Classification results for land cover map gave a believable numbers with overall accuracy about 88.57% and Kappa index about 86.62%. This shows the potential of thermal remote sensing information in the earth surface detection. 1. INTRODUCTION Nowadays land is becoming scarce resources because of population and urbanization pressure. Land cover change is a core part in strategies of natural resources and environment management and monitoring. Application of remote sensing in mapping land cover is specially paid in environment and resources researches for change detection, erosion evaluation, agriculture and forestry potential investigation, regional planning… For large areas remote sensing technology is an effective method for land cover mapping through satellite images of various spatial, spectral and temporal resolutions. Different feature types have different spectral reflectance and emittance properties. Their recognition is carried out through the classification procedure. Two kinds of classification methods are known as supervised and unsupervised classification. These classifiers require the ground truth data for exactly identifying the surface features. But this is difficult for the historical images. So it is necessary to find some methods to supply information for those classifiers. Thermal remote sensing has ideal potential for monitoring the environment and natural resources. It captures the radiance from the sun and the earth, and then converts it into the useful information on the physical earth characteristics. Land surface temperature (LST) is one of its products. Besides, thermal infrared (TIR) ( 8 - 14µm) wave bands have been shown to provide important additional and supplementary information to that provided by the reflectance data measured in visible (0.4 - 0.7µm) and near-infrared (NIR) ( 0.7 – 1.3µm) bands for land cover mapping (Boyd et al., 1996). This study aims to understand the thermal remote sensing properties to retrieve the LST with taking the surface object emissivity into consideration. After that the study applies the thermal remote sensing information combined with vegetation indices to supplement the classified

results of the land cover map. The satellite Aster image will be used in this study because its sensor can take a wide range of spectrum from visible to thermal infrared wavelengths 2. STUDY AREA AND DATA SETS

Fig. 1. The study area in 2nd

smaller square

2.1. Study area Study area is the central part of Ho Chi Minh City which is located in the South of Vietnam in The Indochinese Peninsula (Fig. 1). This part as a nucleus of the urban areas. Surrounding agricultural land is changing into road and built-up land in process of rapid urbanization. The city sometimes has not controlled unplanned constructions making the change of land cover in this area. Therefore, the study area is focused on this part for orienting a reasonable urban management strategy.

2.2. Data sets Aster satellite image was used in this research. It acquired on 25 Jan, 2006 has 14 bands: three visible-near-infrared (VNIR) bands (15m pixel size), 6 shortwave infrared (SWIR) bands (30m pixel size) and 5 thermal infrared (TIR) bands (90m pixel size). For the image processing stage it was georeferenced in Universal Tranverse Mercator projection based on the topographical map with RMSE < 0.5 pixel. Radiometric calibration is needed for transforming the DN value to reflectance for VNIR and SWIR bands and radiance for TIR bands. 3. METHODS 3.1. Retrieval of LST Satellite thermal infrared sensors measure radiances at the top of the atmosphere (TOA), from which brightness temperatures TB (also known as blackbody temperatures) can be derived by using Plank's law (Markham, et al., 1986).

⎟⎟⎠

⎞⎜⎜⎝

⎛+

=

λ

1BKln

KT1

2B (1)

In order to determine an actual surface temperature it is necessary to do atmospheric correction and know the emissivity of the surface land cover. Due to lack of atmospheric measures during image acquisition, the atmospheric correction was ignored. However, these images were acquired in dry season in the study area, so they appeared fairly clear. In this context, the atmospheric effects on these images were not significant. The emissivity (ε) was calculated by using the formula of Valos and Caselles (1996):

ε = εv Pv + εs (1 – Pv) (2) where εv, εs are the emissivity of the full vegetation and bare soil, Pv is the vegetation cover fraction. They can be calculated by NDVI. With the known of land surface emissivity, the LST (TS) can be calculated by Stefan Boltzmann law (Gupta, 1991): (3) 4

B4

S TTB σ=εσ=

Therefore, BS T1T4

1ε

= (4)

where σ is the Stefan Boltzmann constants (5.67x10-8Wm-2K-4). Aster has 5 thermal bands from 10 to 14 in the window 8.125-11.25µm, but 2 bands 13 and 14 in the atmosphere window of 10.25-11.65µm will be used for calculating temperature, because approximately 80% of the energy thermal sensors receive in 10.4-12.5µm wavelength region is emitted by the land surface (Czajkowski et al., 2004) and maximum LST is usually obtained in this region (French et al., 2007). The maximum of those 2 temperature values from Aster bands 13 and 14 is considered to be the final LST value. The results gave the spatial distribution of LST in the whole study area. 3.2. Land cover mapping Vegetation index is a number that is generated by some combination of remote sensing bands and may have some relationship to the amount of vegetation in a given image pixel. Rouse et al. (1974) had detected that the ratio of the difference to the sum of the reflectance values of NIR and Red (named Normalized Difference Vegetation Index - NDVI) useful for measuring the amount of greenness in the vegetation cover.

NIRred

NIRredNDVIρ+ρρ−ρ

= (5)

In vegetated areas, the NDVI typically ranges from 0.2 to 0.8, in proportion to the density and greenness of the plant canopy. Clouds, water and snow, which have larger visible reflectance than NIR reflectance, will yield negative NDVI values. Rock and bare soil areas have similar reflectance in the two bands and result in NDVI near zero. Besides, Boyd et al. (1996) had shown that the potential of radiance data in mid-infrared and thermal infrared bands combined with green band for detecting forest stages of regeneration and mapping the thermal characteristics of land features.

MIR

TIRgreenIR

*VI

ρ

ρρ= (6)

For Aster image, the bands green, red, near-infrared, mid-infrared and thermal infrared match bands 2, 3, 9 and 14. Land cover was determined using ISODATA classification and scatter diagram of vegetation indices (VI) versus LST by relationship scheme of Lambin and Ehrlich (1996). According to them variations in surface temperature are highly correlated with variations in surface water content over bare soil, i.e. low VI and high LST happen to dry bare soil or low VI, low LST for moist bare soil). As the fractional vegetation cover increases, surface temperature decreases as a result of several biophysical mechanisms: (1) high VI and relatively high LST happen for continuous vegetation canopies with a high resistance to evapotranspiration and low soil water availability; (2) high VI and low LST happen for continuous vegetation canopies with low resistance to evapotranspiration on well-water surfaces. The spectral signatures of all classes from unsupervised classification were used to determine the mean values of correlative LST and vegetation indices. They were plotted on scatter diagrams to find instances of strong relationship for deciding the land cover class group. The flow diagram of land cover mapping from thermal properties and vegetation indices is shown on figure 2.

Satellite images

Image Preprocessing

VIS NIR MIR

Unsupervised classification

Fig. 2: Flow diagram of land cover mapping from thermal properties and vegetation indices

4. RESULTS AND DISCUSSION 4.1. The LST distribution map By the NDVI-based method for emissivity retrieval, this study found that the emissivity of vegetation is dependent on the vegetation density, where denser vegetation is where higher emissivity is. Bare land always has lower emissivity value. As shown in formula 4 the LST is in inverse proportion to the emissivity. Consequently, the resulting map in figure 4 were produced to show completely fitted spatial distribution of emissivity-corrected LST. The retrieved LST map shows the picture of LST distribution in an area. The statistics of LST indicates that the highest temperature was 49.4oC. The highest LSTs (greater than 45oC) were found in the industrial zones, where the temperature was created from the production activities plus the solar radiance receiving. The urban areas with prevalent impervious surfaces has suffered the temperature within 36-40oC. In the suburban and rural areas where the crop presents with the full vegetation cover the LST usually is lower, because of higher emissivity. Cloud was detected with the lowest temperature below 22oC.

0.9 0.91 0.92 0.93 0.95 0.98

Fig. 3. Map of NDVI-based emissivity

Fig. 4. Map of LST distribution

4.2. Land cover mapping Land cover of the study area could be visually identified from the image with different type such as: vegetation, soil and water. ISODATA unsupervised classification result generated 16 classes. The scatter diagrams of vegetation indices and LST (Fig. 5 and 6) helped to determine and group these classes into 7 classes (in red ellipse shapes): urban and built-up land, rangeland, water, bare land and agricultural land with different growth stages of crop (1) crop in mature stage, (2) crop in growth, (3) nurseling crop in wetland. Cloud was identified as one class, but it was used for only masking (Fig. 7). The figure 5 shows the negative T-NDVI slope. The increase in green biomass is often associated with a reduction in surface resistance to evapotranspiration, greater transpiration, and a larger latent heat transfer resulting in lower surface temperature. The T-NDVI scatter diagram shows a clear discrimination of land cover classes and aggregation of classes with similar spectral signatures. Water body has lower LST and lowest NDVI and is shown on the lower left corner of the diagram. Crop in wetlands have warmer surface temperature than do open water bodies, and due to their vegetative cover they have higher NDVI than water. Rangeland (grassland and brush) has higher surface temperature than wetlands because the soil is unsaturated. The infrared reflectance of the rangeland is not much different from that of the non-forested vegetated wetland. Agriculture land has higher NDVI than rangeland, and depending on recent irrigation and growth stage, the temperature varies. Bare land has the highest surface temperature and higher NDVI than the urban and built-up areas. Urban and built-up areas have also higher surface temperature and lower NDVI than all other classes but soil and water. LST variation in vegetated surfaces results from variations in the proportion of surrounding bare soil visible to the thermal sensor of Aster and the thermal inertia of the surface. Thermal inertia is the measure of thermal response of surfaces to temperature changes (Boyd et al. 1996). It is a function of thermal conductivity and heat capacity, and is affected by surface characteristics such as soil moisture and albedo. Surfaces with higher thermal inertia possess a strong inertial resistance to temperature fluctuations at a surface boundary hence they show less temperature variation per heating/cooling cycle than those with lower thermal inertia. The thermal inertia of vegetation canopies is lower than that of soils and water has higher thermal inertia than soils.

Figure 6 shows that scatter diagram of VIIR-NDVI has an approximate positive slope unlike the T-NDVI diagram. This scatter diagram contains radiance data from visible, near-infrared, thermal and mid-infrared spectra (as formulas 5 and 6). Green vegetation has higher green and near-infrared band reflectance and lower red and mid-infrared reflectance. This results in higher values of NDVI and VIIR.

Fig. 5. Scatter diagram of T-NDVI for unsupervised-classified classes

Fig. 6. Scatter diagram of VIIR-NDVI for unsupervised-classified classes

Fig. 7. Land cover map extracted from relationship of LST and vegetation indices

4.3. Accuracy evaluation The accuracy evaluation was carried out for retrieved LST and land cover classification. (1) The study had set up 10 surface temperature observed points on 10:30 a.m. at the acquisition time of Aster image on 25 December 2006. This measured value was used to compare with the estimated LST from Aster image. In addition, the other methods such as Emissivity Normalization Method (NOR) (Gillespie, 1985) and method of Artis and Carnahan (1982) were used for retrieving LST. The results were compared with each other. It showed that the bias of measured LST and estimated LST with NDVI-based emissivity correction in this study has smallest value less than 2oC in comparison with others (Table 1).

Table 1. Comparison of bias of measured and estimated LST obtained with different methods

LST with ε correction methods Bias (oC) LST with NDVI-based ε correction 1.95 LST from Artis and Carnahan method 2.01 LST from NOR method (ε=constant) 6.83

(2) Error matrix and Kappa coefficient are a common and typical method for evaluating remote sensing classification accuracy. The total 140 ground truth points were compared with the classified pixels shows the overall accuracy of 88.57% and Kappa coefficient of 86.62%. The classification results shows that vegetation areas have the better estimation (with producer’s accuracy more than 90%) than of bare land and water (about 80% – 85%). But in real this result can be on trust for mapping the land cover map on the large area, especially with historical images without ground truth data. It can be based on the physical thermal properties of the earth surface materials. 5. CONCLUSION This paper has presented the thermal remote sensing potential of Aster sensor in environment and natural resources monitoring and management. It can help to determine the LST on the entire area, it is pre-eminent the in-situ measure method from meteorological stations, giving the only one measurement on site. This study has showed that the application of thermal information based on thermal properties of materials and combined vegetation indices for supplying the land cover classification results can be realizable (with accuracy greater than 80%). This study contributes a technique to enrich classification methods of land surface objects in land cover mapping. Particularly, it is useful in case of missing ground truth data. It will base on the material thermal properties to decide the class. This study can wide-open to others for discriminating land cover types with the historical images from sources with long continuity such as Landsat program. REFERENCES Artis, D.A. and Carnahan, W.H., 1982. Survey of emissivity variability in thermography of

urban areas, Remote Sensing of Environment, Vol. 12, pp. 313-329 Boyd, D.S., Foody, G.M., Curran, P.J., Lucas, R.M., and Honzaks, M., 1996. An Assessment of

Radiance in Landsat TM Middle and Thermal Infrared Wave Bands for the Detection of Tropical Regeneration, Int. J. of Remote Sensing, 17, 249-261.

Czajkowski, K.P., Goward, S.N., Mulhern, T, Goetz, S.J., Walz, A., Shirey, D., Stadler, S., Prince, S.D. and Dubayah, R.O., 2004. Estimating environmental variables using thermal remote sensing, in Thermal Remote Sensing in Land Surface Processes, CRC Press.

French, A.N., Schmugge, T.J., Ritchie, J.C., Hsu, A., Jacob, F. and Ogawa, K., 2007. Detecting land cover change at the Jornada Experimental Range, New Mexico with ASTER emissivities, Remote Sensing of Environment, doi: 10.1016/j.rse.2007.08.020 (in press).

Gillespie, A.R., 1985. Lithologic Mapping of Silicate Rocks Using TIMS. The TIMS Data User’s Workshop, JPL Publication 86-38, pp. 29-44.

Gupta, R.P., 1991. Remote Sensing Geology, Springer-Verlag Berlin Heidelberg, Germany Lambin, E.F. and Ehrlich, D., 1996. The surface temperature-vegetation index space for land

cover and land-cover change analysis, International Journal of Remote Sensing, 17, pp. 463-487

Markham, B.L. and Barkewr, J.L., 1986. Landsat MSS and TM Post Calibration Dynamic Ranges, Exoatmospheric Reflectance and at-satellite Temperatures, EOSAT landsat Technical Notes, 1, 3-8

Rouse, J. W., Haas, R. H., Schell, J. A., and Deering, D. W., 1974. Monitoring Vegetation Systems in the Great Plains with ERTS. Proceedings of Third Earth Resources Technology Satellite-1 Symposium, Greenbelt: NASA SP - 351, 3010 – 3017.

Valor, E. and Caselles, V., 1996. Mapping Land Surface Emissivity from NDVI: Application to European, African, and South American Areas, Remote Sensing of Environment, 57, 167-184.