Back | 4. Results and Discussion |
4.1 Land Cover Analysis using LISS-3 MSS: NDVI was generated using LISS-3 data for land cover analysis ranging from 0.71 to -0.50. NDVI gave land cover (vegetation/green versus non-vegetation/non-green) information showing that 46.03% of the area has vegetation (agriculture, forest and plantations /orchards) and the remaining 53.98 % has non-vegetation (built up land, waste/barren rock/stony and water bodies).
4.2 Land Cover Analysis using MODIS data: Red (Band 1) and near-infrared band (Band 2) of the MODIS sensor at 250 m spatial resolution were used to compute NDVI, ranging from 0.35 to -0.54, indicating that 47.35% of the area under vegetation and the remaining 52.65 % under non-vegetation.
4.3 Classification of high resolution LISS-3 MSS data: The class spectral characteristics for the six land cover classes for LISS-3 MSS bands 2, 3 and 4 were generated to see the inter class seprability. The Transformed Divergence matrix also helped in distinguishing different classes indicating that the ROI pairs have a very good separability. Ground truth obtained from field and other ancillary data were used for the LISS-3 MSS classification. This was done in two steps: unsupervised classification and supervised classification.
False colour composite (FCC) was generated from the LISS-3 MSS data. The heterogeneous patches (training polygons) were chosen for the field data collection. Supervised classification using GMLC was performed with the ground truth data. Care was taken to see that these training sets are uniformly distributed representing/covering the study area. The supervised classified image shown in figure 2(A) was validated by field visit and by overlaying the training sets used for classification. The land cover statistics are listed in Table 1.
Back |
4.4 Classification of MODIS data: The class spectral characteristics for the six classes defined in this study across the first seven bands, PCs and MNF components of the 36 bands of the MODIS sensor were determined showing their good separablity. The Transformed Divergence Matrices were also computed which showed a similar pattern and helped in determining the separability among the various classes.
The MODIS data (bands 1 to 7), the first five PCs and the first five MNF components were classified using Neural Network as shown in figure 2(B), (C), and (D). The process of training the neurons was time consuming. Although NN is considered to be one of the most robust techniques for classification of remotely sensed data, yet, controlling the training process in NN was difficult. The training process for training the neurons converged at 1000 iterations. The number of hidden layer was kept at 1 and the output activation function was kept at 0.001. The output activation function was increased in steps to see the variations in the classification. The training momentum was initially 0 and was increased gradually. The RMS error at the completion of the process was 0.09, 0.39 and 0.29 for the three different inputs. Table 1 shows the land cover statistics for the classification results.
Table 1: Percentage wise distribution of classes obtained from LISS-3 MSS (using MLC) and MODIS classification on Surface Reflectance Bands (1 to 7), Principal Components and MNF components of MODIS bands 1 to 36 using NN.
Classes |
LISS-3 MSS |
MODIS Surface Reflectance bands (Bands 1 to 7) |
MODIS derived PCs (36 bands) |
MODIS derived MNF Components (36 bands) |
Agriculture (%) |
19.03 |
21.88 |
21.49 |
19.38 |
Built up (Urban/Rural) (%) |
17.13 |
26.44 |
15.78 |
17.55 |
Evergreen/ Semi-Evergreen Forest (%) |
11.41 |
7.68 |
12.04 |
11.32 |
Plantation/ orchards (%) |
10.96 |
19.31 |
09.45 |
10.84 |
Waste land/Barren Rock / Stony waste (%) |
40.39 |
24.38 |
40.33 |
39.97 |
Water bodies (%) |
1.08 |
00.31 |
0.91 |
00.94 |
Total (%) |
100.00 |
100.00 |
100.00 |
100.00 |
Figure 2: Supervised Classification using (a) MLC on LISS-3 MSS (b) NN on MODIS bands 1 to 7 (c) NN on PCs and (d) NN on MNF components.