Some Advanced Techniques For Spot 4 XI Data Handling
III. Automated Classification Of Land Cover Using Spot 4 Hrvir Data
The conventional classification methods (supervised or unsupervised) are based on statistical models which use mean vectors, standard deviation and distances such as Euclidean or Mahalannobis as the major classifiers. Different land cover objects have different spectral reflectance properties that can be visualized as a spectral reflectance curve, so it is possible to use this curve as one of the principal measures for classification (Nguyen Dinh Duong, 1997). The automated classification method developed by the authors uses this spectral reflectance curve along with other quantitative values such as band ratio and band differences for classification. The classification algorithm which is based on graphical analysis of the spectral reflectance curve (GASC) works well with LANDSAT TM data that has 6 spectral bands in visible region. SPOT 4 is equipped with a new short wave infrared band at 1.5 mm that provides higher spectral resolution and enhanced sensitivity for leaf moisture content and canopy structure. These improvement is essential for successful application of the GASC algorithm to SPOT 4 XI data in automated classification of land cover. SPOT 4 XI data of scene 277/329 acquired on April 24, 2000 near to Hochiminh City, Vietnam has been chosen as a study area.
The study area is located in south of Vietnam near to Hochiminh City. Its landscape is dominated by features of coastal zone: mangrove forest, wetland agricultural activities. The scene covers also a part of Mekong river's delta which is well known as area of highly productive rice cultivation. On the upper right quarter of the scene are the famous rubber plantation farms. Hochiminh City is located on the upper left part of the image. Land cover categories are enough diverse for land cover classification. The scene is partly cloudy from the middle towards the top. Standard false color composite of the study area is shown on Figure 1.
Figure 1. False color composite of the study area
For automated classification a module named as GASC_G07.F90 has been used. This program was developed based on GASC algorithm (Nguyen Dinh Duong 1997, 1998). For this study area a digital legend of 23 land cover categories was developed. In this legend each land cover is described by a set of image invariants (Nguyen Dinh Duong, 2000) composed of: Spectral curve modulation, total reflected radiance index TRRI, band ratios and difference of normalized spectral values. Major land cover categories such as forest, mangrove of different coverage density, rice crop, water body etc. has been automatically extracted using GASC_G07 module. On the Figure 2 is classification result of the study area.
Figure 2. Result of automated land cover classification
By visual comparison of classification result on the Figure 2 and standard color composite on Figure 1 we can recognize advantages of the proposed approach. Water body is extracted very precisely. Different vegetation types and its distribution has been correctly classified. Mangrove forest, forest plantation (rubber), shrub and grass land including rice crop are possible to be automatically extracted using information derived only from the image data. Bare soil of different level of moisture content is also well identified. Built up area such as urban and housing area is extracted reliably, however, some thin cloud is misclassified into this class. Thick cloud is subject of classification without any doubt, but cloud shadow remains as one of weak point of the GASC algorithm. One of disadvantages of application of different gain modes during observation is needed absolute calibration and working with image data in real number instead integer values which will slow down obviously overall computation performance of the program.
