A Clustering Algorithm to Organize Satellite Hotspot Data for the Purpose of Tracking Bushfires Remotely

TL;DR

A spatiotemporal clustering algorithm that makes enhancements to cluster points spatially in conjunction with their movement across consecutive time periods and allows for the adjustment of key parameters, if required, for different locations and satellite data sources.

Abstract

This paper proposes a spatiotemporal clustering algorithm and its implementation in the R package spotoroo. This work is motivated by the catastrophic bushfires in Australia throughout the summer of 2019-2020 and made possible by the availability of satellite hotspot data. The algorithm is inspired by two existing spatiotemporal clustering algorithms but makes enhancements to cluster points spatially in conjunction with their movement across consecutive time periods. It also allows for the adjustment of key parameters, if required, for different locations and satellite data sources. Bushfire data from Victoria, Australia, is used to illustrate the algorithm and its use within the package.

Figures (9)

• Figure 1: Illustration showing Step 2 of the clustering algorithm on a sample of 20 hotspots in one time window $\boldsymbol{S}_t$. Initially (a), a hotspot is selected randomly ($\boldsymbol{P}$) in order to seed a cluster. The circle indicates the maximum neighborhood distance ($adjDist$). Nearby hotspots as shown in red are clustered with $\boldsymbol{P}$ (b) to initialize list $\boldsymbol{L}$. The neighborhood is moved following every point in that collected list $\boldsymbol{L}$ and new observations are added (c), until there no more points that can be grouped (d). Then a new hotspot is selected external to the existing cluster, and the process is repeated (e). At the end, all the hotspots will be clustered (f).
• Figure 2: Illustration of clustering Step 3, which involves combining results from one time window to the next. There are 33 hotspots at $\boldsymbol{S}_t$, where 20 (green) of them have been previously clustered at $\boldsymbol{S}_{t-1}$ (Figure 1 f) and 13 (orange) of them are new hotspots. The connected graph show the clustering in this time window. Hotspots previously clustered at $\boldsymbol{S}_{t-1}$ keep their cluster labels. The 13 new hotspots are assigned labels of the nearest hotspot's cluster label. This might mean that a big cluster $\boldsymbol{S}_t$ (indicated by the graph) would be split back into two, if it corresponded to two clusters at $\boldsymbol{S}_{t-1}$ (e.g. clusters 2, 4). New clusters of hotspots are assigned a new label (e.g. cluster 5).
• Figure 3: This is the default plot for visualizing the spatial distribution of clusters. In the results shown there are six clusters, which correspond to six fires, shown using different colors. The black dots indicate the ignition site for each fire.
• Figure 4: This is the fire movement plot for visualizing the fire dynamics. Here there are six clusters, corresponding to six different fires. The path between the ignition point and the end point is drawn with black line, where the triangle is the ignition point and the circle is the end point. (Note that the aspect ratio of the plot reflects the relative spatial ratio of latitude and longitude.)
• Figure 5: This is the timeline plot for providing an overview of the bushfire season. The x-axis is the date and the y-axis is the cluster membership. The observed time of hotspots are shown as dot plots (green). The density plot at the top display the temporal frequency of fire occurrence over the timeframe. The dot plot at the bottom (orange) shows the observed time of hotspots that are considered to be noise.