Use of TanDEM-X and Sentinel products to derive gully activity maps in Kunene Region (Namibia) based on automatic iterative random forest approach

• Verfasst von: Vallejo Orti, Miguel [VerfasserIn]
• Verfasst von: Winiwarter, Lukas [VerfasserIn]
• Verfasst von: Corral-Pazos-de-Provens, Eva [VerfasserIn]
• Verfasst von: Williams, Jack G. [VerfasserIn]
• Verfasst von: Bubenzer, Olaf [VerfasserIn]
• Verfasst von: Höfle, Bernhard [VerfasserIn]
• Titel: Use of TanDEM-X and Sentinel products to derive gully activity maps in Kunene Region (Namibia) based on automatic iterative random forest approach
• Verf.angabe: Miguel Vallejo Orti, Lukas Winiwarter, Eva Corral-Pazos-de-Provens, Jack G. Williams, Olaf Bubenzer, Bernhard Höfle
• E-Jahr: 2021
• Jahr: 24 November 2020
• Jahr des Originals: 2020
• Umfang: 17 S.
• Fussnoten: Gesehen am 27.04.2021
• Titel Quelle: Enthalten in: Institute of Electrical and Electronics EngineersIEEE journal of selected topics in applied earth observations and remote sensing
• Ort Quelle: New York, NY : IEEE, 2008
• Jahr Quelle: 2021
• Band/Heft Quelle: 14(2021), Seite 607-623
• ISSN Quelle: 2151-1535
• DOI: doi:10.1109/JSTARS.2020.3040284
• Datenträger: Online-Ressource
• Sprache: eng
• Sach-SW: Agriculture
• Sach-SW: Arid regions
• Sach-SW: automatic classification
• Sach-SW: Degradation
• Sach-SW: gully erosion
• Sach-SW: iterative learning
• Sach-SW: land degradation
• Sach-SW: Namibia
• Sach-SW: random forest (RF)
• Sach-SW: Random forests
• Sach-SW: Soil
• Sach-SW: soil erosion mapping
• Sach-SW: Three-dimensional displays
• Sach-SW: Training data
• Sach-SW: Vegetation mapping
• K10plus-PPN: 175601499X

Abstract

Gullies are landforms with specific patterns of shape, topography, hydrology, vegetation, and soil characteristics. Remote sensing products (TanDEM-X, Sentinel-1, and Sentinel-2) serve as inputs into an iterative algorithm, initialized using a micro-mapping simulation as training data, to map gullies in the northwestern of Namibia. A Random Forest Classifier examines pixels with similar characteristics in a pool of unlabeled data, and gully objects are detected where high densities of gully pixels are enclosed by an alpha shape. Gully objects are used in subsequent iterations following a mechanism where the algorithm uses the most reliable pixels as gully training samples. The gully class continuously grows until an optimal scenario in terms of accuracy is achieved. Results are benchmarked with manually tagged gullies (initial gully labeled area <; 0.3% of the total study area) in two different watersheds (408 and 302 km2, respectively) yielding total accuracies of >98%, with 60% in the gully class, Cohen Kappa >0.5, Matthews Correlation Coefficient >0.5, and receiver operating characteristic Area Under the Curve >0.89. Hence, our method outlines gullies keeping low false-positive rates while the classification quality has a good balance for the two classes (gully/no gully). Results show the most significant gully descriptors as the high temporal radar signal coherence (22.4%) and the low temporal variability in Normalized Difference Vegetation Index (21.8%). This research builds on previous studies to face the challenge of identifying and outlining gully-affected areas with a shortage of training data using global datasets, which are then transferable to other large (semi-) arid regions.