DEVELOPING TRANSFERABLE SPATIAL PREDICTION MODELS: A CASE STUDY OF SATELLITE BASED LANDCOVER MAPPING

Abstract

The mapping of environmental information based on remote sensing requires a workflow that involves image processing, model training usually based on machine learning, as well as model application and validation. Remote sensing data processing capabilities are nowadays simplified by cloud computing platforms. State of the art machine learning methods for spatial data which involve a reduction of spatial overfitting, handling of extrapolation situations and a spatially explicit error assessment, however, are currently mostly implemented in local computation frameworks. Here we present a workflow that combines the improved processing capabilities of the cloud computation platform openEO with state-of-the-art machine learning model development in R. OpenEO is used for standardized imagery acquisition and preprocessing to provide predictors for model training. To reduce overfitting, predictors which are meaningful for the mapping are identified via spatial variable selection as implemented in R packages. The mapping accuracy is assessed via spatial cross-validation and predictions are limited to the 'Area of Applicability' of the model. The workflow is designed to enhance and assess the spatial transferability of machine learning models which is demonstrated by a case study of a landcover classification based on Sentinel-2 imagery.