Near Infrared Mapping

In the AVALMAPPER project, we aim to use NIR imagery for mapping alpine areas. Previous work has
shown that the NIR spectrum is the best choice in the alpine environment due to the challenging lighting and reflection conditions (Yves Bühler et al. 2017) . Thus, a key requirement for this project will be to create methods for identifying features in NIR imagery.

Robust camera-based mapping methods often rely on the identification and description of features in
an image. Visual features are distinct locations that can be found across multiple images. By triangulating these features we can generate a set of 3D positions to create a map. The same features can then be used to localise the camera (and hence the vehicle) within the map. A ‘good’ feature can be identified, located, and used to generate a unique descriptor (a lower-dimensional coded representation of the feature), that can be matched between images (with different viewpoints and potentially different lighting conditions). There is a long history of varying approaches for detecting and describing visual features. Key requirements are often a subset of computational speed, repeatability and descriptor invariance under changes in viewpoint, rotation, lighting conditions, and scale in the image. Corners, edges and lines often form the basis for traditional visual-spectrum computer vision features in structured, man-made environments.

However, alpine areas present very different visual conditions compared to structured environments.
Snow reflects nearly 100% of the incoming solar radiation in the visible part of the electromagnetic spectrum ( λ =0.35-0.7um) while it absorbs up to 75% in the NIR range ( λ =0.7-1um) (Y. Bühler et al. 2009; Yves Bühler et al. 2017) . Due to this observation, most approaches to detect avalanches are based onNIR imagery. As illustrated in Fig. 5, Y. Bühler et al. (2009) use the imagery from two NIR cameras mounted rigidly on a manned aircraft to compute the NormalisedDifference Angle Index (NDAI). The NDAI values in combination with geometric constraints derived from the slope and the numerical simulation tool RAMMS (Rapid Mass Movements) are used as input for the entropy-based texture analysis and a subsequent object-based classification method by comparing the homogeneity values of adjacent pixels. As depicted in Fig. 5 (right), the method proposed by Korzeniowska et al. (2017) use NIR imagery to compute the brightness, normalised difference vegetation index (NDVI), the normalised difference difference water index (NDWI), and its standard deviation to segment avalanches from other land surface elements. The major drawback of these methods is that they are based on heuristics and hand-craftedmetrics, limiting the generality of their application.