Spectral and spatial filtering for enhanced thematic change analysis of remotely sensed data

Abstract

Post‐classification change comparison (PCC) is a widely applied method for extracting thematic change information. However, the main drawback associated with using classified image data for detecting change is the compounding of errors resulting from the overlay of classified images, each of which contains classification errors. A number of studies have examined this issue but several important questions remain. This paper examines two complementary approaches to change detection, one spectral‐based, the other spatial‐based, in an attempt to minimize spurious change identified through PCC. Two SPOT satellite images were acquired (1990 and 1999) over a human‐dominated landscape in North Sulawesi, Indonesia. Images were classified and validated using an eight‐class scheme and ‘no‐change’ reference data. Thematic change information was extracted by overlaying the two classified dates and validated using ‘change’ reference data. First, a new additive spectral filtering method for excluding areas incorrectly identified as changed was tested. Second, the efficacy of a spatial misregistration noise filtering method was evaluated. Finally, the two combined approaches were applied and validated. A standard PCC identified 44.70% of the land surface as having changed from 1990 to 1999. Tests showed that the spectral and spatial filtering methods reduced the amount of classified change from 44.70% to 12.89% and 31.80%, respectively. Combined spectral and spatial filtering further reduced the amount of classified change to 10.86%. The approach presented here is more effective than previous methods for correctly excluding spurious thematic change as part of the PCC process. These findings are important for all applications that aim to extract thematic change information using PCC.