Balancing signature variance between local and global minima/maxima: Restricted maximum likelihood (REML) classification and the search for plagioclimax

Due to long-term anthropogenic disturbance, plagioclimax results in vegetation compositions that cannot develop to a climax state. The overarching goal of the study was to develop a method to identify plagioclimax sub-regionally in the eastern Mediterranean (Israel) through signature extension. A case study was established in the Carob-Mastic (typicum judaicum) vegetation sub-association, demonstrating plagioclimax in both a southern, moderately affected stand and a northern, heavily impacted one. Training sets for supervised classification were constructed in the southern stand from an existing 1 m land cover classification with global and local class sets. Signature extension was employed to identify the plagioclimax in the northern stand using 30 m Landsat-9 data. The specific objectives were twofold: (1) to construct a mixture modeling technique that enabled fusing the 1 m and 30 m data sets; (2) to devise a classification method by which the complex segmentation of the plagioclimax, as an interstitial garrigue between phanerophyte shrub matrices, could be identified. An experimental method was devised in which five levels of density-restricted training sets, based on minimum pixels per patch, were built from the patch spatial structures of shrub community-related classes. Patch ecology metrics were derived directly from the restriction levels to develop an understanding of the landscape mosaic. Entropy (a measure of disorder) and emptiness (a proxy for fragmentation) measures were designed into bin tables and examined relative to the variance in the spectral signatures. A restricted maximum likelihood (REML) classifier that relies on limiting variance was chosen to identify local maximum clusters (the unique plagioclimax classes), and the five classification results were compared. The results were successful in identifying the plagioclimax at the local maximum. This strategy is appropriate for cases where disturbance has caused continuous ensembles of vegetation compositions, which result in unstructured remotely sensed data.