A multi-resolution forest stand segmentation algorithm integrating Landsat imagery and forest structural, age, and species attributes

Strategic forest inventories are required to support management, enable monitoring, and inform policy development. These inventories are typically developed using photointerpretation of aerial imagery, followed by manual stand delineation and attribution. The process of delineating and attributing these segments can be costly, time consuming, and subjective. Many spectrally driven segmentation algorithms have been developed, but few have incorporated pixel-based, wall-to-wall, forest attributes such as stand height or species. As such, we propose a two-phase segmentation algorithm to automatically delineate forest stands from remotely sensed Landsat spectral reflectance data and forest attributes, including height and height variation, stem volume, age, and tree species. Initially at a micro-segmentation phase, small groups of pixels from the Landsat imagery are grouped based on spectral similarity. Next, in the segment-development phase, these groups of pixels are further merged into forest segments based on structure, age, and species information. We applied this approach to five study sites totalling 45 Mha selected to represent distinct forest landscapes across Canada. Our results found that grouping micro-segments by stand height variation minimized variance within forest segments in western Canada, while stem volume minimized variance within forest segments in boreal forest sites. Stand height variation also consistently showed the lowest Global Moran’s I (ranging from 0.1 to 0.2), indicating that variance between segmented forest stands was greatest and most heterogeneous from their neighbours. Interestingly, while using species produced relatively homogeneous forest stands, it was less successful in discriminating between stands. The analysis showed that structural attributes, particularly those related to height and stem volume, were most effective in delineating homogeneous forest stand polygons that were distinct from their neighbours. This work expands upon existing spectrally-driven segmentation algorithms to integrate forest structure, age, and species information, mimicking manual forest stand delineation and attribution practices in a transparent and automated manner.