MASTER data processing and derivation of spectral reflectances and land surface temperature (LST)

The MODIS/ASTER Simulator, or MASTER, sensor acquires data over the visible through mid-infrared wavelengths (0.46 – 12.817 μm) in 50 spectral bands, which were used to derive land surface temperature (LST), Normalized Difference Vegetation Index (NDVI), and albedo at high spatial resolution (Hook et al. 2001). The 2011 airborne data collection campaign was completed over the Phoenix metropolitan area during July 12-13 (daytime flights collected from 10am – 1pm local time) and July 15-16 (nighttime flights collected from 12:30am – 3am local time) using a Beechcraft B-200 aircraft to obtain high spatial resolution data, approximately 7 m/pixel. While the original intent was to collect a true diurnal dataset, this was not possible due to the large area of data collection, and pilot flight rules that constrain daily flying time and mandate stand-down periods for flight safety. Fortuitously, a stable weather pattern prevailed during the week of July 12 that enabled collection of an approximate diurnal dataset.

As the MASTER data were also acquired over several days, it was necessary to perform atmospheric correction to obtain apparent surface reflectance and surface temperature/emissivity and also to allow for comparison of the flight line data collected on different days. The ENVI/IDL image processing environment was used to perform all processing of the MASTER data. Atmospheric correction of the mid-infrared wavelength data was accomplished using an in-scene atmospheric compensation technique (Johnson and Young 1998), followed by separation of surface temperature and emissivity by the emissivity normalization approach (Kealy and Hook 1993). Atmospheric correction of the visible through shortwave infrared wavelength data was accomplished using the Quick Atmospheric Correction (QUAC) algorithm (Bernstein et al. 2005) within ENVI/IDL. Reasonableness of both surface temperature and apparent reflectance results were assessed qualitatively by random examination of returned pixel values (surface temperature) and image spectra (vegetation and soil pixel spectra were compared to laboratory spectra of similar materials).

While the MASTER data has been processed to at-sensor radiance, it did not include any geolocation data. As georeferencing the entire dataset was not possible, the processed data were extracted for the 41 PASS neighborhoods within the MASTER flight boundary. We manually geo-referenced each neighborhood to its corresponding National Agriculture Imagery Program (NAIP) data with geolocation information using an image warping approach incorporating polynomial nearest neighbor resampling to a target registration error of no more than 0.5 pixel. In cases where severe distortions of the original MASTER data did not allow registration to this level of accuracy, registration to the lowest possible error was performed. In four neighborhoods the MASTER images were too distorted for acceptable georeferencing, and so these highly distorted images were not included in further analyses. The final processed MASTER dataset included high spatial resolution LST data for neighborhoods that included direct estimate of emissivity, atmospheric corrections, and georeferencing. In total we were able to obtain data for 37 neighborhoods during the day and 41 at night. We linked the MASTER and PASS data with the previously developed high resolution (1m) land cover classification for the Phoenix metropolitan region, which had a classification accuracy of 92% and a kappa statistic of 0.91 (Li et al. 2014). This classification was conducted using NAIP aerial photography and following an object-oriented expert rule algorithm (Baatz et al. 2008), which is becoming increasingly common for urban regions (Myint et al. 2011). We then overlaid the classification and MASTER data with parcel boundaries and extracted all residential plots. This corresponded with 21,589 parcels during the day, 25,011 parcels at night, and a subset of 721 PASS survey respondents.

As image acquisition occurred over approximately a 3-hour flight period we expected during the day LST would increase during sampling while at night we expect LST would decrease throughout sampling. We investigated the potential for these differences by comparing means across all parcels in maximum, minimum, and mean temperature. During the day we only observed increasing temperatures in the minimum parcel temperatures but not the mean or maximums, therefore no correction was performed. At night, we observed a significant cooling in the maximum, minimum, and mean parcel temperatures. We used linear regression with time analysis to identify the rate of cooling and then corrected all parcel temperatures to the median sampling time, approximately 1am. Subsequent analyses conducted using both corrected and uncorrected results had similar qualitative results in regard to significance tests.

Normalized Difference Vegetation Index (NDVI) calculation

Spectral vegetation indices (SVI) are based on the reflectance properties of green leaves that strongly absorb in red wavelengths and strongly reflect in near-infrared wavelengths, which is a function of the chlorophyll content. Most aerial and satellite-based studies of vegetation dynamics have relied on the Normalized Difference Vegetation Index (NDVI) (Tucker 1979, Huete et al 2002), a proxy for many vegetation characteristics including percent ground cover, biomass, leaf area index (LAI), and net primary production (NPP). However, in its original form, NDVI served as a proxy for the amount of photosynthetically active radiation intercepted by green leaves (FPAR, or fraction of photosynthetically active radiation). NDVI time-series are now increasingly used for monitoring vegetation dynamics. When the index is related to climatic variables important details of climate-ecosystem interactions can be effectively inferred. Therefore, estimates of actively photosynthesizing vegetation abundance in our study area were obtained by deriving NDVI from processed MASTER spectral bands. NDVI was calculated as follows:

NDVI = (Band9 – Band5) / (Band9 + Band5)

where Band9 and Band5 correspond to MASTER spectral channels 9 (0.87 µm) and 5 (0.66 µm). The index has a range of values from -1 to +1, with values below 0.2 generally signifying no photosynthetically active vegetation.

Albedo estimation

Albedo is the fraction of solar energy (shortwave radiation) reflected from the Earth’s surface. It is quantified as the proportion, or percentage, of solar radiation of all wavelengths reflected by a surface to the amount of downwelling radiative flux. An ideal white surface has an albedo of 1, or 100%, and an ideal black surface, 0%. Ice, especially with snow on top of it, has a high albedo, whereas water (albedo as little as 0.03) absorbs most of the incoming radiation and reflects very little. Forest can have an albedo anywhere between 0.05 and 0.20. Light toned dry soil and sand have the albedo of 0.4 – 0.45. Albedo is a very important variable for understanding climatic processes. In our study albedo was estimated by utilizing an empirical relationship found between spectral bands of the MODIS (Moderate Resolution Imaging Spectroradiometer) sensor on-board of Terra and Aqua satellites operated by NASA and the so-called broadband albedo. Specifically this was done by converting narrowband to broadband albedo using empirical formulae (Liang 2001). We derived the total visible broadband albedo (α_vis) as follows:

α_vis = 0.331α₅ + 0.424α₁ + 0.246α₃

where α₁, α₃, and α₅ are reflectances in MASTER spectral bands 1 (0.46 µm), 3 (0.54 µm), and 5 (0.66 µm), respectively. The chosen MASTER bands match closely those of MODIS.