Enhanced detection of burned area using cross- and autocorrelation

New and effective methods for burned area detection from remote sensing data are required by diverse applications. Burned area detection from a large heterogeneous landscape is a challenging task because of the low-interclass dissimilarity of burned and unburned pixels. Burned and unburned pixels may reflect similar spectral signatures but retain distinct temporal and spatial dependence in its vicinity. This change in temporal and spatial domains can be used to supplement spectral information. The existing methods use only spectral information of the images for detecting burned area and ignore the local temporal and spatial associations. We propose a new method to improve burned area detection accuracy by combining spectral information with local temporal and spatial associations. A method for simultaneous incorporation of these associations by means of cross-and autocorrelation is put forward and tested here under the logistic regression framework. The proposed method was assessed over two different study areas and showed significant increment of 5% to 9% in Kappa accuracy. We demonstrated that the use of cross-and autocorrelation leads to substantial increase in burned area detection accuracy. (C) 2015 Society of Photo-Optical Instrumentation Engineers (SPIE)Beijing Natural Science Foundation (clustering and forecasting of large scale temporal data based on knowledge-guidance and optimal granulation of information) [4112031]; Fundamental Research Funds for the Central Universities; Talent Introduction Research Project of Yunnan Minzu UniversitySCI(E)EIARTICLEtauqir_mughal@yahoo.com1