Keywords: Nighttime NOAA-AVHRR, Land Cover Type Classification
Abstract
Nowadays, land-cover type classification using only daytime AVHRR data is limited by accuracy and number of land cover classes. The new method developed in this study is to investigate nighttime AVHRR data as an additional band for classifying land cover types with respect to improve accuracy and specify more land cover classes. The classification accuracies of various band combinations, NDVI, ordinary brightness temperatures and/or land surface temperature (LST), were comparatively assessed. Both in cool and hot seasons, overall accuracies using combination between day and night time data show better than using only daytime data. The combination of three bands of NDVI, daytime LST, and nighttime LST shows the highest accuracy. Three-band combination using only daytime shows lower accuracy than two bands using day and nighttime. Adding nighttime data obviously increases the accuracies of forest and built up classes. The nighttime data can well discriminate forest from active agriculture (or mature crops), deciduous forest in hot season from inactive agriculture (or non-mature crops), and built up from harvested or fellow agriculture. The results indicate that this approach using nighttime data behaved well as a new method for classifying land cover types at the landscape scale using AVHRR low-resolution data.
1. Introduction
Recently, many practical satellites have been increased to monitor land cover information around the globe. The use of high-resolution satellite data, for example, derived from Landsat-TM, and SPOT have suffered from the infrequent coverage, high data volume and high costs. Thus, recent efforts have been directed toward the use of low-resolution data (1.1 km at nadir) obtained by NOAA AVHRR sensor that provides much better temporal resolution with daily coverage (Cihlar, Ly and Xiao, 1996). For this coarser AVHRR data, a main problem is pixels containing a mixture of land cover types. The supplementation of high-resolution image is generally used to overcome this problem but this is cost consumption. Thus, the methodological challenge is to extract meaningful land cover information by using only AVHRR data. For land cover type classification, vegetation index, computed as the pixel difference between AVHRR channels 1 and 2 reflectance divided by the sum of the two channels, is firstly used. The thermal infrared data is subsequently used. More recent research use the combination among NDVI, thermal, and ratio between these two variables. However, most of these works use only thermal data in daytime while the AVHRR sensor provides daily useful data both in day and night times. The land cover types and land characteristics can be specified by temperature difference over a time or can be described by thermal inertia, a physical variable describing the impedance to variations of temperature. High thermal inertia values lead to small changes in temperature, for a given transfer of heat (Xue and Cracknell, 1995). Water resist to the change of temperature through a time compared with other surface types. The fractional vegetation cover increases, surface temperature in daytime and temperature difference between daytime and nighttime decreases (Lambin and Ehrlich, 1997).
This study does not intend to produce a definitive land cover map, but rather attempt to investigate various band combinations of day and night time AVHRR data that can be the greatest possibility for classifying land cover types. The innovation of this research is to use nighttime AVHRR data as an additional band with respect to improve classification accuracy of land cover types derived from the AVHRR coarse resolution data.
2. Data and Method
The AVHRR images derived from the NOAA-14 passes at around 2 AM and 2 PM were selected from least cloud 3-continuous days in December 1997 (cool season) and March 1998 (hot season). Prior to image classification, the images were subjected firstly to geo-reference correction. Secondly, the images of NDVI and land surface temperature (LST) were generated. The adjusted LST equation using the split window technique (Kremer and Running, (1993), Deschamps and Phulpin (1980)) and NDVI equation in this study show as follows:
NDVI = (band 2-band 1)/ (band 2+band 1)
LST = band 4 +1.11 (band 4-band 5) -273, (Chada, 2000)
Thirdly, The composite images of AVHRR bands 3, 4, 5, LST and NDVI were generated using maximum values. Only the NDVI images were derived from daytime data. Others were retrieved from both daytime and nighttime data. Forth, cloud remaining in MVC LST images was masked on pixels with less than 13 degree Celsius and with temperature difference between day and night less than 0 and greater than 20 degree Celsius. Next, all corrected composite images were re-scaled from 0 to 100 by relative equation:
Finally, the Maximum Likelihood classification method was used. The transected training sites or ROI (region of interest) were randomly sampled based on the overlaid digitized land use map and supplemented by recent land cover types interpreted from NDVI and temperature profiles taken from the images. At least two trainning sites and more than 150 pixels per site were labelled for seven main classes: Water bodies (Wt): The dam and reservoir covered by water. Built up areas (Bu): The urban lands have least vegetation covered. Field crops (Cr): The vegetated areas grown for annual crops such as cassava, sugar cane, maize, and mixed field crops. Paddy fields (Pa, Ph, and Pu): The vegetated areas cultivated for rice crops. This class was labeled into three subclasses: lowland active paddy (Pa) and lowland harvested paddy (Ph) and upland paddy (Pu). Orchards (Or): The vegetated lands grown for fruit trees. Para rubber (Rb): The lands planted for para rubber tree. Forests (Fe, Fd): This class was labeled into two subclasses: (1) deciduous forest (Fd), the vegetative areas mostly covered by deciduous forest that somewhere mixed with the sparse evergreen forest, dry dipterocarp forest, plantation forest and disturbed forest, and (2) evergreen forest (Fe), the vegetated areas permanently covered by high density of evergreen forest.
The classified images were overlaid with the existing land use map in 1995 of Thailand classified from Landsat-TM to compute their overall accuracy (OAE, personal contact).
