Automated extraction and discrimination of open land areas from IRS-1C LISS Ⅲ imagery

摘要

Availability of open land area is a key indicator in assessing and planning urban environments. However, accurate uninhabited surface extraction is still a challenge. In this paper an attempt is made to combine the advantages of partial differential equation (PDE) and random forest (RF) method for segmentation of IRS-1C LISS Ⅲ satellite image for mapping and discrimination of open land areas. Spatial variations of the pixels in the image are analyzed to identify open land pixels. PDE method denoise and conserve finer details simultaneously by using correlation between the spectral bands and directions of the edges. Low-resolution resultant image patch is mapped to high-resolution image patch by linear regression model. Variance of classification is reduced by training many classifiers using interpolated RF method. This methods elevates the accuracy of the direct RF method and achieves 3.35 dB improvement in PSNR and 6.47 reduction in MSE. Further, discrimination of open land areas is done into distinct classes, using inherent spatial information. Accuracy assessment indicates an overall accuracy of 87% over direct RF method.