Automating delineation and classification of alluvial fans by the use of object-based geomorphometry and machine learning

  • Automatische Identifikation und Klassifizierung von Schwemmfächern durch den Einsatz von objektbasierter Geomorphometrie und maschinellem Lernen

Pipaud, Isabel; Lehmkuhl, Frank (Thesis advisor); Wellmann, Jan Florian (Thesis advisor); Stauch, Georg (Thesis advisor)

Aachen : RWTH Aachen University (2021)
Dissertation / PhD Thesis

Dissertation, Rheinisch-Westfälische Technische Hochschule Aachen, 2020

Abstract

Automating the mapping of landforms via digital elevation data is a key research topic in the field of geomorphometry. The acquisition of extensive thematic datasets has the potential to foster scientific progress in geomorphology and related disciplines by enhancing our understanding of how formative processes are reflected in geomorphometric variables and by facilitating the transfer of in-field results to larger study areas. In the wake of continuously improving spatial resolutions of DEMs, traditional pixel-based classification approaches fail in accounting for the irregular boundaries of landforms. To address this fundamental issue, object-based frameworks developed in the field of remote sensing have been adopted for geomorphometric research. With respect to alluvial fans, all workflows presented hitherto are, however, deficient in delineating laterally adjoining fans as individual landforms. Since coalescing fans represent a common scenario for (semi-) arid mountain ranges, further research is warranted. Although object-based image analysis (OBIA) gained considerable emphasis during the last two decades, scientific progress is complicated by substantial challenges. (1.) The sensitivity of segmentation algorithms to local discontinuities requires a proper choice of parameters. The appropriate selection of scaling factors or bandwidths is, however, an issue which is not comprehensively resolved to this date, because segmentation is an ill-structured problem which is not conducive to an a-prori assessment. (2.) Existing adaptions of object-based image analysis (OBIA) in the field of geomorphometry neglect that by formalizing the local altitude field, morphometric variables act (i) as an interface to the land surface characteristics of interest and (ii) are strongly influenced by the terrain rendition characteristics of the DEM in use. This doctoral dissertation first sets the stage by evaluating freely available near-global DEM data— ASTER GDEM, AW3D30, SRTM1 (30 m posting each), and a self-processed TanDEM-X DEM (15 m posting). Both geomorphological mapping and geomorphometric processing attest TanDEM-X a superior terrain rendition up to moderate relief. In settings of strong topographic contrasts, however, challenges toward interferometric processing result in a deterioration of terrain representation. SRTM1 shows an overall consistent depiction of terrain, but is characterized by an autocorrelated grainy texture of moderate magnitude. The most recent 30 m-digital elevation model (DEM) product, ALOS World-3D DEM with 1'' (~30 m) posting (AW3D30), shows improvements over Shuttle Radar Topography Mission DEM with 1'' posting (SRTM1) by displaying only uncorrelated systematic noise of relatively low magnitude, which suitable filtering can remove. In contrast to the aforementioned DEM products, ASTER GDEM displays a striking bumpy texture, and it is strongly discouraged to use the DEM in geomorphometric research. Since DEM characteristics exert a strong influence on significant ranges of morphometric variables, this doctoral dissertation challenges the de facto standard of utilizing rule-based classification schemes for object-based analysis in the field of geomorphometry. As an alternative, a framework is proposed where both segmentation and classification are controlled by supervised machine learning in the following way: (1.) The approach recognizes that segmentation and classification need to be addressed with individual compilations of morphometric variables. In the case of alluvial fans, a robust segmentation of laterally adjoining fans can be accomplished by incorporating the sine and cosine of slope aspect in addition to slope and curvature values. For classification, tailored algorithms have been developed to formalize the morphometric signature of alluvial fans. (2.) A high-performing object-based workflow is thus compulsorily feature-specific. To formalize the morphometric signature of a landform, one-class estimators are for the first time integrated into an object-based workflow. (3.) To account for the ill-structured nature of segmentation, the issue of selecting appropriate bandwidths for segmentation is tackled from an a-posteriori perspective. Specifically, segmentation is first performed for a range of parametrizations, and all raw datasets are classified. Subsequently, the final thematic dataset is compiled by applying an overlap resolve algorithm on the raw segmentation data, selecting final objects according to their estimator scores and topological relationships. The framework was first tested for the delineation and discrimination of alluvial fans situated in a single study area within the Mongolian Altai, using a spatial variant of the mean-shift algorithm for segmentation and a one-class support vector machine (OCSVM) for classification. The framework is able to accurately delineate laterally coalescing fans with the exception of fans grading into a bajada, i.e., fans displaying diffuse and ambiguous lateral delimitations. By successively adapting the feature space representation yielded by the one-class support vector machine (OCSVM) estimator, a promising diagnostic ability could be achieved by using just eleven training samples. The sensitivity of the OCSVM algorithm to the choice of its parameters hampers, however, the transferability of the workflow to different landforms and DEM datasets. To address this issue, this thesis introduces a new one-class random forest (OCRF) and trials the algorithm with a training dataset of 306 fans compiled from seven different study areas. Owing to a high diagnostic ability and a robustness towards its parametrization and morphometric variables with little discriminatory potential, the OCRF improves the delineation and classification of alluvial fans at the expense of requiring a larger training dataset. Furthermore, the non-parametric nature of OCRF can be capitalized on by performing a fully data-driven optimization of parameters and DEM preprocessing. Eventually, the geomorphometric variables developed in this thesis are utilized to assess whether controlling factors can be inferred from the geomorphometry of alluvial fan surfaces. While the findings demonstrate that scaling relationships exert a dominant influence on the expected range of morphometric parameter values, it could be shown that the differentiation between tectonic activity and the capability of the fan–catchment system in evacuating sediment can be improved by quantifying the degree of fan progradation. In essence, this doctoral dissertation showcases that object-based analysis in the field of geomorphometry poses different demands on methodological frameworks as is the case in the field of remote sensing, and the term object-based morphometric analysis (OBMA) is suggested to emphasize this difference. The methodological framework developed in this thesis offers excellent prospects for an extensive use of object-based techniques in the fields of geomorphology and geomorphometry. Libraries of feature perimeters and feature signatures can be set up in a staggered, semi-automated manner, by first performing an OBMA analysis using an OCSVM or a comparable algorithm, and switching to a one-class random forest (OCRF) once suffcient training data has been acquired. For the thematic mapping of larger portions of the geomorphological inventory, it is envisaged to pursue a stratified approach where landform types are delineated and identified in a serial manner, utilizing object-based morphometric analysis (OBMA) workflows tailored towards each geomorphic feature of interest.

Identifier

Downloads