THE USE OF KNOWLEDGE-BASED RECLASSIFIERS IN MAPPING LAND COVER CLASSES

This research discusses the similarities and differences between visual and digital classification cues and strategies for mapping land cover using remotely sensed imagery. A preliminary analysis of mapping land cover classes for an inland wetland area demonstrated the fundamental problems using the traditional single-stage per-pixel classification strategy: mixed pixels, spectrally similar classes, shadows, heterogeneous classes, and the minimum mapping unit.;Two strategies for mapping wetland land cover classes were compared to the results of a visual classification of a wetland area. A rule-based reclassification logic was designed for mapping land cover classes from the biophysical components of a study region. The results of a rule-based reclassification strategy and traditional digital classification were qualitatively and statistically compared. The rule-based reclassification strategy was ninety-one percent accurate as compared to the traditional classification of only seventy-two percent.;Five knowledge-based reclassification strategies used to remove shadows from digital imagery were developed and tested. Individual class and overall accuracy assessments were performed for each reclassification strategy. The best shadow removal logic, using contextual information and neighborhood constraints, correctly predicted the land cover class shadowed for eighty percent of the observations.