Characterizing vegetation patch proximity and size across Australia with skimage and dask map overlap

How a patch of vegetation burns partly depends on its size, proximity to neighbours, and the size of the neighbourhood patches. Using dask map overlap we scaled an image processing pipeline using functions from scikit-image and SciPy to identify, label, and categorize vegetation patches that are isolated and small, and so unable to sustain full intensity fires.

Many image processing tasks are pixel independent and are embarrassingly parallel. In our pipeline, labelling continuous vegetation patches requires calculations across neighbourhoods of pixels and means parallelization is more complex. Buffered tiles solves parallelization by allowing non-parallel pixel operations within tiles to be run simultaneously across many tiles without edge effects in parallel. Dask provides map overlap as a general purpose function for buffered tile operations. Dask map overlap readily integrated with scikit-image and SciPy and allowed us to scale our approach across Australia at high resolution.

This talk will introduce some basics of landscape morphology relevant to fire behaviour, it will show how image processing algorithms can be applied to characterise vegetation in a landscape and demonstrate how Dask can be practically used to run this process fast and at scale using cloud based clusters.