@article {bnh-5279, title = {A lightning-caused wildfire ignition forecasting model for operational use}, journal = {Agricultural and Forest Meteorology}, volume = {253-254}, year = {2018}, month = {05/2018}, pages = {16}, chapter = {233}, abstract = {

Lightning-caused wildfires are responsible for substantial losses of lives and property worldwide. Convective storms can create large numbers of ignitions that can overwhelm suppression efforts. Both long- and short-term risk planning could benefit from daily, spatially-explicit forecasts of lightning ignitions. We fitted a logistic regression generalised additive model to lightning-caused ignitions in the state of Victoria, Australia. We proposed a new method for model selection that complemented existing methods and further reduced the number of variables in the model with minimal change to predictive power. We introduced an approach for deconstructing ignition forecasts into contributions from the individual covariates, which could allow model output to be more readily integrated with existing intuitive understandings of ignition likelihood. Our method of model selection reduced the number of variables in the model by 37.5\% with little change to the predictive power. The final model showed good predictive ability (AUC 0.859) and we demonstrated the utility of the model for short term forecasting by comparing model predictions with observed lightning-caused fires over three time periods, two of which had extreme fire conditions, while the third was randomly chosen from our validation dataset. The model presented in this paper shows good predictive power and advancements in model output could allow fire managers to more easily interpret model forecasts.

}, doi = {10.1016/j.agrformet.2018.01.037}, url = {https://www.sciencedirect.com/science/article/pii/S0168192318300376}, author = {Nicholas Read and Thomas Duff and Peter Taylor} } @mastersthesis {bnh-5260, title = {Statistical models for the location of lightning caused wildfire ignitions}, year = {2018}, month = {05/2018}, pages = {258}, school = {The University of Melbourne}, address = {Melbourne}, abstract = {

Lightning-caused wildfire is a significant concern for fire management agen\ cies worldwide. Unlike other ignition sources, lightning fires often occur in\ remote and inaccessible locations making detection and suppression particularly challenging. Furthermore, individual lightning storms result in a large\ number of fires clustered in space and time which can overwhelm suppression\ efforts. Victoria, Australia, is one of the most fire prone environments in the\ world and the increased frequency of large-scale landscape fires over the last\ decade is of particular concern to local wildfire management authorities.\ This thesis is concerned with modeling lightning-caused wildfire ignition
locations in Victoria. Such models could be used for predicting daily lightning\ caused ignition likelihood as well as simulating realistic point patterns for use\ in fire spread models for risk analyses.

The first half of this thesis looks at regression models. We review methods\ for the model selection, validation, approximation and interpretation of generalised additive models. A review of performance metrics, such as the AUC,\ shows the difficulties and subtleties involved in evaluating the predictive performance of models.\ We apply this theory to construct a non-linear logistic regression model\ for lightning-caused wildfires in Victoria. The model operates on a daily time\ scale, with a spatial resolution of 20 km and uses covariate data including fuel\ moisture indices, vegetation type, a lightning potential index and weather.\ We develop a simple method to deconstruct model output into contributions\ from each of the individual covariates, allowing predictions to be explained in\ terms of the weather conditions driving them. Using these ideas, we discuss\ ranking the relative {\textquoteleft}importance{\textquoteright} of covariates in the model, leading to an\ approximating model with similar performance to the full model.\ The second half of this thesis looks at point process models for lightning\ caused ignitions. We introduce general theory for point processes, focusing
on the inhomogeneous Poisson process, cluster processes and replicated point\ patterns. The K-function is a useful summary function for describing the\ spatial correlation point patterns and for fitting models. We present a method
for pooling multiple estimates of the K-function, such as those that arise when\ using replicated point patterns, intended to reduce bias.

We fit an inhomogeneous Poisson process model as well as a Thomas and\ Cauchy cluster process model to the Victorian lightning-caused ignition dataset. The cluster process models prove to have significantly better fit than the\ Poisson process model, but still struggle to reproduce the complex behaviour\ of the physical process.

}, url = {https://minerva-access.unimelb.edu.au/bitstream/handle/11343/214157/Statistical\%20models\%20for\%20the\%20location\%20of\%20lightning-caused\%20wildfire\%20ignitions\%2C\%20Nicholas\%20Read.pdf?sequence=1\&isAllowed=y}, author = {Nicholas Read} }