I. INTRODUCTION |
Land cover (LC) relates to the discernible Earth surface expressions, such as vegetation, geology, and hydrologic or anthropogenic features, and thus describes the Earth’s physical state in terms of the natural environment and the man-made structures. Essentially, LC can have only one class or category at a given time and location, and can be mapped using suitable image data with spectral signatures. Land use is an expression of human uses of the landscape, e.g. for residential, commercial, or agricultural purposes, and has no spectral basis for its unique identification. Thus it can not be explicitly derived from image data, but only inferred by visual interpretation or assessed in the framework of object-based contextual analysis. It involves both the manner in which the biophysical attributes of the land are altered and the intent underlying that alteration, and the purpose for which the land is used. LC changes induced by human and natural processes play a major role in global as well as at regional scale patterns of the climate and biogeochemistry of the Earth system. Variations in topography, vegetation cover, and other physical characteristics of the land surface influence surface-atmosphere fluxes of sensible heat, latent heat, and momentum of heated air particulates caused by conduction, convection and radiation, which in turn influence weather and climate [1].
Many remote areas of the world are now being opened to exploration and development, generating a growing demand for up-to-date knowledge of topography, LC and other geo-spatial information. By using as many data sources as possible, a more complete and accurate knowledge of a landscape can be obtained. Therefore, users are seeking to integrate a multitude of spatially referenced information into their management and decision-making systems, a step that is facilitated by the standardisation of digital formats and the rapidly expanding market of GIS (Geographic Information System). There is a growing need for a global standardised LC and land use mapping system, similar to the CORINE approach. Many classification systems and innumerable map legends exist, but in most cases even from the same country, they are incompatible with each other. This system would enable a variety of end-users to use the results for their specific application (e.g. from rural planning to energy planning) and would enable intercomparison of existing data and a harmonised approach of data collection in areas where this information is not available or obsolete, whilst minimising the data and processing cost. Therefore, LC mapping remains an important research field, one what has grown more sophisticated with more recent technical developments in object oriented analysis or ontology [2], [3], [4].
LC features such as vegetation, water, and soil are important components in regional planning and management for monitoring the dynamics associated with the Earth. LC changes induced by human and natural processes play a major role in the climate, hydrology and biogeochemistry at global as well as regional scales. LC features can be classified using remotely sensed satellite imagery of different spatial, spectral and temporal resolutions. Recently, LC mapping of the entire territory of China was done at a 1:1 million scale to understand the LC change, based on existing LC maps, field surveys, NOAA’s (National Oceanic and Atmospheric Administration) AVHRR (Advanced Very High Resolution Radiometer) imagery and aerial photos. The CORINE LC map of the European Union includes 44 LC classes divided into 5 main categories (agricultural areas, artificial surfaces, forests & semi natural areas, wetlands and water bodies). It is based primarily on Landsat TM (Thematic Mapper bands 4, 5 and 7) data of different vegetation periods with additional information in the form of topographic maps and orthogonal photos [5], [6].
In India, spatial accounting and monitoring of LC have been carried out at a national level at 1:250,000 scale, using multi-temporal IRS AWiFS (Indian Remote Sensing Satellite Advanced Wide Field Sensor) with 4 bands (Green, Red, NIR and SWIR) at 56 m resolution to address the spatial and temporal variability in cropping patterns and other LC classes. A decision tree classifier method was adopted to account for the variability of temporal datasets [7].
The above attempts are based on monotemporal remote sensing (RS) data with the analysis being done on an annual basis. Monitoring LC dynamics with time series satellite data would not be economical for regional or national level mapping with commercial data such as IRS LISS (Linear Imaging Self Scanner)-III, LISS-IV, SPOT or Landsat TM/ETM+ (Enhanced Thematic Mapper plus). This imposes a major limitation on the use of such data despite their high spatial resolution. RS data such as ASTER (Advanced Spaceborne Thermal Emission and Reflection Radiometer) are inexpensive and have a better spatial resolution, but are not regularly available for whole regions. MODIS (Moderate Resolution Imaging Spectroradiometer) data, with a spatial resolution of 250m to 1 km, have better spectro-temporal resolution (7 bands, and composite-data with Level 3 processing and 36 bands every 8 days, or every 1-2 day availability with Level 1B processing) can be downloaded freely and may be suitable for regional mapping and planning activities in many developing countries. Their frequent availability is especially useful to account for seasonal variations and changes in LC pattern.
In order to obtain these LC types, remotely sensed data are classified by identifying the pixels according to user-specified categories, by allocating a pixel to the spectrally maximally “similar” class, which is expected to be the class of maximum occupancy within the pixel [8]. MODIS based LC mapping addresses large area coverage but is limited to classification of whole pixels [9], [10]. A variety of other classification methods exist, including spectral matched filter [11], mixture tune matched filtering and spectral angle mapper [12], which are appropriate when pixels do not contain mixtures of materials with correlated spectra, especially in higher spatial resolution data. However, MODIS pixels generally have the problem of spectral mixing. The mixed pixel problem is normally found at boundaries between two or more mapping units, or along gradients, etc. when the occurrence of any linear or small subpixel object takes place, and is usually dealt with by subpixel classifiers that assume either linear or non-linear mixtures [8] and [13]-[16]. It is often the case in RS that one wants to deal with identification, detection and quantification of fractions of the target materials for each pixel for diverse coverages in a region using unmixing approaches to discern the proportion of heterogeneity. A suitable way of extracting information is to estimate the composition of each pixel (proportion of the category contents) by spectral unmixing, i.e. soft classification techniques. The concepts of spectral unmixing emerged in the early 1970’s [17], [18] and gained more prominence in 80’s and 90’s [8], [13]-[16] and [19]. During the last two decades, methods have been proposed ranging from modelling the component mixtures to solving the linear combinations to obtain abundances. Later, a number of techniques were developed to estimate and extract endmembers from the scenes or use spectral libraries.
In the present study, constrained linear spectral unmixing (CLSU) technique is applied on MODIS data to assess the suitability of the method for regional LC mapping with six LC classes. Linear Mixture Model (LMM), also known as a macro spectral mixture model, assumes no interaction between materials, and a pixel is treated as a linear combination of signatures resident in the pixel with relative concentrations [14]. This approach is different from the Spectral Angle Mapper (SAM) technique, where for the identification of pixel signature spectra only the angular information is used, which is based on the idea that an observed reflectance spectrum can be considered as a vector in a multidimensional space, and where the number of dimensions equals the number of spectral bands.
