Assesing the spatial transferability of Random Forest models trained on the basis of the WSF3D dataset for cities

Abstract

The main objective of this research is to evaluate if the Random Forest models trained on the basis of spatial metrics derived solely from the WSF3D dataset can be transferred from one city to another. Though with the increased availability of remotely sensed data, new machine learning techniques are constantly emerging for land use mapping, the challenges of collecting validation data and spatial transferability are yet to be addressed. The WSF3D dataset and the technique of "Dissimilarity Index" are used to address these challenges. The main factors allowing model transferability; the association between prediction accuracies and transferability of cities; and the morphological similarities existing between transferable cities are analysed. The "Area of Applicability" is identified, to make assessments for successfully transferring a model to areas where validation data is not available.