Independent two-step thresholding of binary images in inter-annual land cover change/no-change identification

Abstract

Binary images from one or more spectral bands have been used in many studies for land-cover change/ no-change identification in diverse climatic conditions. Determination of appropriate threshold levels for change/no-change identification is a critical factor that influences change detection result accuracy. The most used method to determine the threshold values is based on the standard deviation (SD) from the mean, assuming the amount of change (due to increase or decrease in brightness values) to be symmetrically distributed on a standard normal curve, which is not always true. Considering the asymmetrical nature of distribution histogram for the two sides, this study proposes a relatively simple and easy 'Independent Two-Step' thresholding approach for optimal threshold value determination for spectrally increased and decreased part using Normalized Difference Vegetation Index (NDVI) difference image. Six NDVI differencing images from 2007 to 2009 of different seasons were tested for inter-annual or seasonal land cover change/no-change identification. The relative performances of the proposed and two other methods towards the sensitivity of distributions were tested and an improvement of ~ 3% in overall accuracy and of ~0.04 in Kappa was attained with the Proposed Method. This study demonstrated the importance of consideration of normality of data distributions in land-cover change/no-change analysis.