Can training data counteract topographic effects in supervised image classification? A sensitivity analysis in the Cantabrian Mountains (Spain)

Supervised classification converts digital data of satellite imagery into categorical land-cover classes suitable for end users. Far away from being an easy process, many factors such as landscape heterogeneity and topography cause radiometric exchange among classes that needs to be filtered out prior quantitative applications. To address this problem, specific processing methods of image correction allow normalizing surface reflectance of different land-cover types across topographic gradients. However, differences in spectral responses of sunny and shady slopes cannot be related solely to topography, since they could represent a mixture of plant species or functional types on the ground. In these cases, topographic correction may represent an unfounded aggressive modification of original reflectance values that contributes to the ongoing controversy about the adequacy of its application before image classification. An easy alternative to overcome this problem may consist in carrying out an extensive sampling of training data across topographic gradients. This will provide a full description of the spatial (and spectral) variability of the informational classes sought, which is a prior requirement of image classifiers. To evaluate whether a comprehensive sampling of training data can counteract topographic effects in the classification of medium resolution satellite imagery, we evaluated four methodological options: (1) the collection of reference points across all landscape heterogeneity, (2) different sampling schemes of training data, (3) the application, or not, of topographic correction, and (4) the use of different classification algorithms. Results demonstrated that normalizing spectral responses should be applied as a processing step for image classification in rugged terrains, mainly when reference data are not representative of all environmental variability and we do not explore different decision-making for selecting only optimal results.