VII. DISCUSSION |
Sustainable management of natural resources entails spatio-temporal information related to various surface features such as vegetation, water, and soil. In this regard, LC mapping using temporal RS data helps in capturing the dynamics associated with these LC components [33]. In principle, LC mapping can be done with any multispectral dataset, but currently available data have limitations such as poor spatial and temporal resolutions, or high spatial but low spectral resolution (IKONOS PAN, CARTOSAT-II) or moderate spatial and spectral resolutions (IRS LISS-III/IV, SPOT, Landsat TM/ETM+) as well as limited coverage, reduce the suitability of the available data types. Most of these RS data are commercial and many scenes would be required if mapping is to be carried out for larger regions. Even though some of the data (such as ASTER) are free, they are not available for all parts of the world and at variable temporal resolution. In this context, high temporal resolution MODIS data are freely available and are useful in providing insights into LC dynamics.
The present study evaluated the scope of CLSU technique for LC mapping using low spatial resolution MODIS data. Endmembers and abundance maps for the six LC categories were generated using MODIS data through linear unmixing. The proportion of the classes in each pixel of MODIS was computed and compared with the classified image of better spatial resolution and also through field data.
The results show that five of the six identified classes exhibit high interclass variability, allowing linear spectral unmixing. The variability of the pixels that represent the local pure classes is also responsible for the uncertainity of the mixture proportions. Probability density functions and parametric representation of the constituent pure classes did represent local conditions evident from the post classification validation. Assigning the categories to endmembers in consultation with the field data proved to be critical as it helped to discern the adjacency effect that exists between contrasting features such as forest and plantation, agriculture and plantation, etc. [34], [35]. This highlights that LMM, with improved estimates of proportional abundance values, has greater scope compared to other hard classification techniques in handling coarse spatial resolution data. The spectra of the different endmembers were modelled by linear equations. This introduces another form of quality assurance, as these equations can be inverted, taking the endmember spectra to reconstruct the spectra of each pixel from the original image. Divergences between the original and reconstructed image then give idea of the goodness of fit of the model, and provide insights to which bands add more to the errors. When no more recognizable patterns are found and the error is overall small, it can be deduced that a near perfect model has been reached.
Nevertheless, even though the number of spectral bands in MODIS is higher compared to LISS-III, with the CLSM [36], only six categories can be discerned (number of bands available minus one). The condition of identifiability (the number of classes should be one less than the number of bands) can be solved by a two step process provided that many spectral endmembers are available. A subset with a prefixed number of endmembers that optimally decompose the candidate pixel is first selected by a Gram-Schmidt orthogonalisation process [37]. This restricted subset is then used in conventional LMM. The final result is the decomposition of the scence into all end-members considered while reducing the residual errors. Also, this approach fails in heavily fragmented landscapes or small isolated areas of high contrasting nature. This can be addressed through image fusion techniques [38] of low spatial resolution (MODIS) with high resolution (LISS-III/Landsat-TM/ETM+). To account for contributions from the neighbourhood pixels at the same time, non-linear mixture models are suitable for unmixing of coarse spatial resolution data.
