Land Cover Classification by Multisource Remote Sensing: Comparing Classifiers for Spatial Data

Land cover classification is a standard remote-sensing task in which typically multispectral satellite data is used to identify features such as land use. The detection of rock glaciers is a particularly challenging task that requires the combination of satellite data with terrain analysis data because their spectral signature alone is not distinct enough for their classification based on satellite imagery alone. The performance improvements that can be achieved by selecting an optimal classifier in this particular land cover classification problem are investigated. In the case study, eleven statistical and machine-learning techniques are compared in a benchmarking exercise, including logistic regression, generalized additive models (GAM), linear discriminant techniques, the support vector machine, and bootstrap-aggregated tree-based classifiers such as random forests. Penalized linear discriminant analysis (PLDA) achieves a median false-positive rate (mFPR, estimated by cross-validation) of 8.2% in early detection of rock glaciers at a sensitivity of 70%, which is significantly better than all other classifiers. The GAM and linear discriminant analysis are second best (mFPR: 8.8%). The mFPR of the worst three classifiers is about one-quarter higher compared to the best three classifiers. The land cover classification problem is further analyzed in general terms from a methodological perspective, highlighting potentials and pitfalls related to phenomena including error estimation in the presence of spatial dependence, high dimensional problems in hyperspectral remote sensing, and indirect models.