Land cover is important for many planning and management activities and is considered an essential element for modelling and understanding the earth as a system. Land cover analysis relates to identifying the type of feature present on the surface of the earth. It deals with the identification of land cover features on the ground, whether vegetation, geologic, urban infrastructure, water, bare soil or others. Variations in land cover and its other physical characteristics influence weather and climate of our earth. Therefore the study of land cover plays an important role at the local/regional as well as global level for monitoring the dynamics associated with the earth. Land cover analysis has been done most effectively through satellite images of various spatial, spectral and temporal resolutions. Due to the spectral resolution limitations of conventional multispectral imageries, sensors that could collect numerous bands in precisely defined spectral regions were developed, leading to Hyperspectral Remote Sensing.

Hyperspectral images have ample spectral information to identify and distinguish spectrally unique materials that allow more accurate and detailed information extraction. These imageries are classified into different land cover categories using various algorithms. The genesis and the underlying principle behind each of these algorithms are different and essentially produce land cover maps with varied accuracies. This report evaluates the utility of various classification algorithms for land cover mapping using MODIS data having high spectral resolution and coarse spatial resolution.

NDVI (Normalised Difference Vegetation Index) was used to segregate the green (vegetation) and the non-green (non-vegetation) in the study area. Principal Component Analysis (PCA) and Minimum Noise Fraction (MNF) were used to reduce the data dimensionality and noise from the data. The hard classification algorithms used in this research include K- Means Clustering, Maximum Likelihood classification (MLC), Spectral Angle Mapper (SAM), Neural Network (NN) based classification, and Decision Tree Approach (DTA) to classify the image into 6 land cover classes (Agriculture, Built up (Urban / Rural), Evergreen / Semi-Evergreen forest, Plantations / orchards, Waste lands / Barren rock / Stony waste / Sheet rock and Waterbodies / Lakes / Ponds / Tanks / Wetland).

The utility of a soft classification technique, Spectral unmixing using a Linear Mixture Model (LMM), was evaluated to calculate the proportions of the different land cover categories in a pixel. The endmembers, pure pixels or pixels with only one class within them, were identified using the Pixel Purity Index (PPI) and Scatter plot. These endmembers were used in spectral unmixing to yield abundance maps; a map indicating the proportion of each category within each pixel (fraction images).

The accuracy of the MODIS classified maps were assessed by generating error matrix at the administrative boundary level (for Chikballapur Taluk) for which ground truth was collected from field. A comparative study of the percentage area for each land cover class for various classification algorithms was also done across all the taluks with the high spatial resolution LISS-3 MSS classified map as reference. At the pixel level, pixel to pixel analysis was done with respect to LISS-3 classified map. These results were also validated on the field. Results of this study show that NN on MNF components with an overall accuracy of 86.11%, and MLC on MODIS bands 1 to 7 with accuracy of 75.99% gave the best output among the various techniques. MLC on MODIS PC gave the lowest accuracy of 30.44%. The results are quite motivating for the application of land cover mapping using coarse spatial resolution data at a regional scale.

Keywords: Land cover, Remote sensing, GIS, Hyperspectral data, MODIS, Multispectral data, algorithms, hard classification and soft classification