On Drivers of Subpixel Classification Accuracy-An Example from Glacier Facies

Abstract

Subpixel classification (SPC) extracts meaningful information on land-cover classes from the mixed pixels. However, the major challenges for SPC are to obtain reliable soft reference data (RD), use apt input data, and achieve maximum accuracy. This article addresses these issues and applies the support vector machine (SVM) to retrieve the subpixel estimates of glacier facies (GF) using high radiometric-resolution Advanced Wide Field Sensor (AWiFS) data. Precise quantification of GF has fundamental importance in the glaciological research. Efficacy of the approach was first tested on the synthetic data followed by the input AWiFS and reference MultiSpectral Instrument data, including ancillary data. SPC of synthetic data resulted in overall accuracy (OA) of 95%, proving the merit of SVM. Classification accuracy is inversely related to the glacier's surface heterogeneity. Reducing the number of classes enhanced the OA by ∼18%. Source and timing of RD invariably controls the SPC accuracy. OA improved by ∼5% on addressing the issue of temporal gap between input and RD. ∼11% increase in OA with the inclusion of ancillary data confirmed their positive effect on the accuracy. Input and reference fractional area of GF were strongly correlated (r > 0.9) with each other substantiating the results.