USING CONVOLUTIONAL NEURAL NETWORK-BASED DEEP LEARNING AND DIGITAL TERRAIN MODEL-DERIVED LAND SURFACE PARAMETERS FOR GEOMORPHIC MAPPING (Invited Presentation)

Geomorphic and surficial features, both natural and anthropogenic, are primarily differentiated based on terrain textures and patterns, topographic position, and spatial relationships with other features and landforms. Land surface parameters (LSPs), such as slope, topographic position index, topographic roughness index, and surface curvatures, derived from high spatial resolution digital terrain models (DTMs) can capture such characteristics and serve as key input variables for mapping processes based on manual interpretation, supervised learning, or a combination of these processes. In contrast to more traditional supervised learning and machine learning methods, convolutional neural network (CNN)-based deep learning semantic segmentation methods can capture and model spatial relationships and textures, making them particularly applicable to geomorphic and surficial mapping tasks. This talk explores key issues in applying CNN-based semantic segmentation methods to geomorphic and surficial mapping including selecting input variables, collecting training data, selecting and parameterizing algorithms, and customizing training processes. Frontiers and research needs are discussed including transfer learning, semi-supervised learning, and transformer-based architectures.