Paul Doherty, Qinghua Guo, Wenkai Li, Otto Alvarez, and Jared Doke
“The application of GISystems to solve real-world problems continues to expand from reactive, where we simply document and visualize where and when a phenomenon happened, to proactive, where we are able to reliably forecast event locations based on what we have learned from previous events. During this process we often test GIScience theories and techniques, leading to new scientific discovery. This is especially true in the field of spatial epidemiology, which merges spatial analysis with studies from public health (Ostfeld et al. 2005, Robertson et al. 2010). The objectives of such studies are to collect information about spatially varying factors that may contribute to the occurrence of disease, illness or injury.
“Within spatial epidemiology, datasets often consist of incident coordinates or other locality descriptions that need to be georeferenced. Furthermore, most data describe locations where illness or injuries have previously occurred (presence) but not where they have not occurred (absence). Therefore, analyses have often been limited to descriptive analyses (density or “heat maps”) and spatial statistics (hot spot or Getis Ord G* maps; Getis and Ord 1992) because traditional modeling approaches requires presence and absence data to derive relationships from underlying factors (Hirzel et al. 2002). This limitation is known as the geographic one-class data issue (GOCD; Guo et al. 2011) and requires a specialized approach to generate probability maps.
“Here we study a real-world problem: wilderness or wildland search and rescue (WiSAR) incident prevention in Yosemite National Park using a novel GIScience technique. WiSAR is the process of locating, accessing, stabilizing, and transporting people in remote environments (Worsing 1993). Therefore, our objectives are to describe our methodology, present our results, and discuss the preliminary implications of our findings for WiSAR incident prevention, spatial epidemiology, and GIScience. To do so we used a novel machine learning approach based on GOCD (incident occurrence) to forecast areas of probable occurrence in the future.”
- Read the paper [PDF]