A novel strategy for wetland area extraction using multispectral MODIS data

MODIS (Moderate Resolution Imaging Spectro-radiometer) is an imaging sensor onboard Terra/Aqua platform which provides information in 36 spectral bands. Spectral mixing of features due to high data dimensionality and high inter-band correlations constitutes the biggest challenge in the analysis of MODIS data. In view of this, present study attempts to develop a novel strategy involving Principal Component Analysis (PCA), Band to Band Correlation (BTBC) analysis, Stepwise Discriminant Analysis (SDA), and separability analysis to reduce the data dimensionality for extracting reliable and maximum information for wetland areal extent using coarse resolution MODIS (1km) data. The PCA explains variability in data and removes data redundant information; BTBC analysis eradicates the high correlated bands providing best bands suitable for wetland mapping; SDA evaluates the discriminatory power of different MODIS bands to discriminate the wetlands from other class types. Further, Normalized Difference Vegetation Index (NDVI) and Wetland Model Index (WMI) were also incorporated into the study to improve the classification accuracy. Finally, separability analysis was conducted to optimize the selected MODIS bands and indices. Results of rigorous data mining reveal that out of 24 input layers of MODIS (1km) data (22 optical MODIS bands, NDVI and WMI), only 4 input layers (WMI, NDVI, MODIS bands – NIR band 2 and SWIR band 6) are best suited for delineation and mapping of wetlands. This study also corroborates the usage of WMI, a newly developed index with combination of visible and short wavelength infra red (SWIR) MODIS bands, as the most optimal input layer to separate wetlands from the other land use class types (barren land, agricultural land, rivers/canals/streams, forest, wasteland/gullied or riverine land). It was observed that the magnitude of mapped wetland areal extent using MODIS (1km) data varied from 105,053ha in 2010–2011 to 111,479ha in 2011–2012 accounting for 0.44% (2010–2011)–0.46% (2011–2012) of the total geographical area of Uttar Pradesh, India. Seasonally, monsoon season displayed maximum wetland area with overall accuracy of 92% whereas summer season exhibited minimum wetland area with overall accuracies varying from 85 to 87% during both the sampling years. The output of the present research work will not only facilitate to improve the wetland area estimates but also provide an important input for climate change predictions in wetlands over large