Exploring knowledge-based and data-driven approaches to map earthflow and gully erosion features in New Zealand

In New Zealand, earthflows and gullies are - next to shallow landslides - important erosion processes and sediment sources in hill country areas. They can cause damage to infrastructure, affect the productivity of farmland, and impact water quality due to fine sediment input to streams. Implementing effective erosion mitigation measures requires detailed information on the location, extent, and spatial distribution of these features over large areas. Remote sensing provides an excellent opportunity to gain such knowledge, whereby different approaches can be applied. In this study, we present two approaches for detecting earthflow and gully erosion features on the North Island of New Zealand.

Earthflows are complex mass movement features that can occur on gentle to moderate slopes in plastic, mixed, and disturbed earth with significant internal deformation, whereby vegetation cover usually remains on the earthflow bodies during movement. High-resolution aerial photography and a LiDAR digital elevation model (DEM), including a range of derived products such as slope, surface roughness, terrain wetness index, were used within a knowledge-based object-based image analysis (OBIA) workflow to semi-automatically map potential earthflows. Specific earthflow characteristics discernible from the optical imagery, such as the presence of bare ground at the toe and rushes, were identified on different hierarchical segmentation levels and subsequently aggregated. Additionally, morphological and contextual properties (e.g. connection to streams) were integrated into the mapping workflow. Gully erosion is an indicator of land degradation, which occurs due to the removal of soil along drainage channels through surface water runoff. We tested a region-based convolutional neural network (Mask-RCNN) deep learning approach for object detection to map gully features. The deep learning was performed on three LiDAR DEM terrain derivatives, namely, slope length and steepness (LS) factor, hillshade and terrain ruggedness index. Labelled chips for training data were generated with reference gully features mapped manually on historical aerial photography.

Semi-automated earthflow detection appeared to be very challenging due to their complexity and the lack of distinct characteristics to differentiate them from other features. The initial results suggest the knowledge-based OBIA workflow has potential, but a major challenge is the creation of objects that represent one earthflow. Hence, the current mapping results may better indicate terrain susceptible to potential earthflow occurrence rather than correctly detecting single earthflows. As for gully mapping, the data-driven deep learning framework shows promising results regarding gully presence and absence. Validation resulted in detected gullies overlapping 60% of the reference gully area. However, a limiting factor related to the available reference data that was mapped on historical aerial photography and does not align with the LiDAR DEM. Given the significant impact of earthflows and gullies, it is essential to develop reliable and targeted analysis methods to better understand their spatial occurrence and enable improved representation of these erosion processes in catchment sediment budget models.