Think objects, not pixels! Semi-automated object-based analysis forgeomorphic identification and mapping from digital elevation data: the case of planation surfaces

International audience; Many geomorphological maps have been produced thanks to advances in GIS software (e.g. processing performance) and digital elevation, satellite imagery or airborne datasets breakthrough (resolution, accuracy, coverage, availability). These maps are mainly built either by manual digitalization or through semi-automated pixel classification, which do not consider relationships between neighboring pixels, can introduce scale bias, and can be highly time-consuming and subjective. However, GEOgraphic Object Based Image Analysis (GEOBIA) avoid these biases by using objects rather than pixels to map geomorphological features. This semi-automated method relies on i) a segmentation step that automatically divides the image into features through pixel grouping algorithms followed by ii) feature classification (e.g. landforms or land-cover types) by their characteristics (e.g. shape, size) and statistics (e.g. elevation, slope, curvature, aspect).Here, we performed GEOBIA to identify and map planation surfaces, which are widespread and useful to understand earth-surface dynamics and denudation chronology in source-to-sink studies. We developed two different protocols using SRTM 30arcsecond data and its derivatives (slope, curvature, ruggedness, etc.) on a study area where planation surfaces were recently mapped (Orange river, Namibia-South Africa). Both protocols differ in their classification way of thinking. The first uses an unsupervised classification based on a clustering algorithm and then a fuzzy logic chart to define feature classes. The second uses a supervised classification based on a machine-learning algorithm from user-defined landform samples. Our first results and benchmarking analysis of both protocols show i) that slope and curvature parameters should be preferred to elevation and other derivatives and ii) that they identify planation surfaces with accuracy around 80 to 90% despite their different philosophy. We will test the reproducibility and universality of both protocols from different control areas such as Argentina, Armorican Massif or even the French Massif Central before discussing the further steps required to discriminate the different generations of planation surfaces.