Slope unit-based genetic landform mapping on Tibetan plateau- a terrain unit-based framework for large spatial scale landform classification

Abstract

Genetic landform classification and mapping provide crucial information concerning the formation and evolution of the earth surface, aiding in understanding geological processes, climate variation, and driving mechanisms of earth dynamics. Conventional pix and object unit-based research paradigms poss challenges in classifying and mapping the genetic landform on a large spatial scale, while the terrain unit-based method offers an inspiring opportunity for this typical landform task. In this context, we herein developed a slope unit (SU)-based landform classification framework, involving improving the mean curvature method for SU segmentation, developing the spectra method to determine the minimal area size of SU, introducing a specific quantitative indicator system for describing SU, and a universal classification model based on XGBoost. Using the entire Tibetan Plateau as the study area, we first implemented the genetic landform map and divided it into nine genetic landform types including aeolian, arid, loess, karst, periglacial, fluvial, glacial, volcanic-lava, and lacustrine. Classification results achieved an encouraging performance with overall accuracy, kappa coefficient, and mean accuracy of up to 90.2%, 84.71%, and 88.65%, respectively. Experimental results suggested that delineated genetic landform regions and their boundaries are generally closely correlated with regional climatic conditions, geological factors, or surface processes. Compared to pix or object-unit-based methods, the SU delineation results possess internal homogeneity and clear terrain boundaries with geo-meaning. Compared to the watershed-unit-based method, our method is better suited for large-scale landform classification. In the end, we confirmed that the proposed framework can be extended to large-scale planetary genetic landform mapping via two case studies on Mars and Moon, which is transferable yet straightforward. In brief, we argued that the proposed framework offered a powerful and effective method in large-scale genetic landform classification, somewhat effectively circumventing the limitations of existing landform classification methods.