Introduction

Land cover (LC) changes induced by human and natural processes are linked to climate and weather in many complex ways. These linkages between LC dynamics and climate include the exchange of greenhouse gases (water vapor, carbon dioxide, methane, etc.) between the land surface and the atmosphere, the radiation balance of the land surface, the exchange of sensible heat in the atmosphere, and the roughness of the land surface. Because of these linkages between LC and climate, changes in LC are important for climate studies and its variability. This has fuelled the research in LC mapping, with more recent technical developments in object oriented analysis and ontology (Benz et al., 2004; Camara et al., 2000; Sun et al., 2005). Recently, there have been attempts of LC mapping in many parts of the world including China, the European Union, and India (EEA & ETC/LC, Corine LC Technical Guide, 1999; Natural Resources Census, 2005; Torma and Harma, 2004), etc. based on monotemporal remote sensing (RS) data with the analysis being done on an annual basis. However, monitoring LC dynamics with time series data would not be economical for regional or national level mapping with commercial data. Also, RS data such as ASTER are inexpensive and have a better spatial resolution, but are not regularly available for all geographical regions. Relatively, temporal MODIS data with more spectral bands (7 bands composite-data every 8 days availability with Level 3 processing and MODIS 36 bands product every 1-2 day availability with Level 1B processing) with spatial resolution ranging from 250 m to 1 km can be downloaded freely and are suitable for many applications, especially for countries with large area ground coverage. Their frequent availability is 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. LC mapping can be performed using various algorithms by processing the RS data into different themes or classes. The principle and the purpose behind each of these techniques may be different and each of these algorithms may also result in different output maps.

Many methods have been proposed to obtain the classified image. This study comparatively analyses the performance of neural network (NN) based multilayer perceptron (MLP) classifier and decision tree (DT) that have been proposed for classification of superspectral MODIS data. The motivation for using MLP and DT in this study came from the poor result obtained by experimenting with the conventional classification techniques such as Maximum Likelihood Classifier (MLC) and Spectral Angle Mapper (SAM) that gave lesser overall accuracies (76%, 30%, 42% with MLC and 69%, 35%, 49% with SAM for three different types of MODIS based inputs (as explained in the “Results - MODIS data classification” in section 4). On the other hand, the results obtained from neural based classifier and DT (which is a non-parametric classifier) gave quite motivating results for regional LC mapping using the coarse spatial resolution data.