Multivariate Alteration Detection (MAD) and MAF Post-Processing in Multispectral, Bi-temporal Image Data: New Approaches to Change Detection Studies

Allan A. Nielsen¹, Knut Conradsen¹ and James J. Simpson²

¹Department of Mathematical Modelling, Technical University of Denmark,
²Digital Image Analysis Laboratory, Scripps Institution of Oceanography

Accurate and quantitative change detection in a temporal sequence of satellite scenes is an important requirement for many climate and global change studies. This paper introduces the Multivariate Alteration Detection (MAD) transformation and describes the Minimum/Maximum Autocorrelation Factor (MAF) transformation, and it combines these two multivariate statistical transformations to quantitatively and accurately detect changes in sequences of satellite data. As opposed to the popular Principal Component (PC) transformation both the MAD and the MAF transformations are invariant to affine transformations which means that they are both insensitive to for instance 1) changes in offset and gain in a measuring device such as an imaging spectrometer, and 2) simple multiplicative atmospheric corrections. Both Landsat Multispectral Scanner System (MSS) data and Advanced Very High Resolution Radiometer (AVHRR) data, together with the above cited analysis methods, are used to quantify case studies of urbanization and ENSO-related (El Niño-Southern Oscillation) alteration of ocean thermal structure. Change is also computed by applying the PC transformation to simple difference images. These latter analyses are compared with those derived using the MAF/MAD analyses. Results show that the combination of MAF/MAD analysis is superior to the PC analysis for accurate detection of patterns of spatial change in multivariate satellite scenes. Moreover, the MAF/MAD analysis is less effected by noise in the data than is the PC analysis and the MAF/MAD analysis also is more effective than PC analysis for the detection of outliers in data. MAF/MAD analysis, coupled with simple spectral differencing of properly intercalibrated images, is an effective combination to 1) compute accurate rates of change, and 2) place these rates in a statistically meaningful spatial context (i.e., a natural image context). This combination can be achieved by assigning significance to the simple differenced values only in regions where the cross-correlation between the simple differences and the MAF/MADs is high.

Remote Sensing of Environment, 64:1-19 (1998)