Statistical Analysis of Topographic and Climatic Controls and Multispectral Signatures of Rock Glaciers in the Dry Andes, Chile (278–338S)

• Published in: Permafrost and Periglacial Processes
• Main Authors: Brenning, A., Azócar, Gerardo
• Format: Article in Journal/Newspaper
• Language: English
• Published: John Wiley 2010
• Subjects: generalised additive model; logistic regression; digital elevation model; terrain analysis; rock glacier; Andes; Gam; Random Point; Ice; permafrost
• Online Access: https://doi.org/10.1002/ppp.670; http://www.captura.uchile.cl/handle/2250/11070

The dual nature of rock glaciers as ice-rich mountain permafrost and sediment storage systems results in a combination of geomorphic processes and energy balance components controlling their distribution. We use the generalised additive model (GAM), a semi-parametric nonlinear method, to empirically analyse environmental controls and spectral characteristics of rock glaciers in the dry Andes of Chile based on presence/absence data at random point locations and predictor variables derived from digital elevation models and Landsat data. A combination of nonlinearly transformed local and catchment-related terrain attributes (especially local and catchment slope and potential incoming solar radiation, PISR) characterises the geomorphic and climatic niche of rock glaciers. The influence of (latitude adjusted) relative PISR varies with mean annual air temperature (MAAT): high-PISR sites are favourable for rock glacier development at lower MAATs and low-PISR sites at higher MAATs. TM/ETMþ band 6 (thermal infrared) is an additional nonlinear predictor. The combination of topographic, climatic and multispectral data in a GAM achieves an excellent general discrimination (area under the ROC curve 0.87 on the model domain and 0.94 overall). In automatic rock glacier detection at a sensitivity of 70 per cent, this model achieves a false-positive rate (FPR) of 6.0 per cent overall and 12.8 per cent on the model domain (bootstrap estimates: 7.9% and 16.8%). Dropping the multispectral data significantly increases the bootstrapped FPR by 36 per cent. Thus, the fusion of multisource data using modern nonlinear classification techniques is a promising step towards automatic rock glacier detection.