Modeling Bare Ground With Classification Trees in Northern Spain

Abstract

Bare ground abundance is an important rangeland health indicator and its detection is a fundamental part of range management. Remote sensing of bare ground might offer solutions for land managers but also presents challenges as modeling in semiarid environments usually involves a high frequency of spectral mixing within pixels. Classification tree analysis (CTA) and maximum likelihood classifiers were used to model bare ground in the semiarid steppes of the middle Ebro valley, Aragon, Spain using Satellite Pour l’Observation de la Terre 4 (SPOT 4) imagery and topographic data such as elevation, slope, aspect, and a morphometric characterization model. A total of 374 sample points of bare-ground fraction from sixteen 500-m transects were used in the classification and validation process. Overall accuracies were 85% (Kappa statistic = 0.70) and 57% (Kappa statistic = 0.13) from the CTA and maximum likelihood classifiers, respectively. Although spectral attributes were essential in bare-ground classification, the topographic and morphometric properties of the landscape were equally critical in this modeling effort. Although the specific layers best suited for each specific model will vary from region to region, this study provided an important insight on both bare-ground modeling and the potential advantages of CTA.