AUTHOR=Kale Manish P. , Mishra Asima , Pardeshi Satish , Ghosh Suddhasheel , Pai D. S. , Roy Parth Sarathi TITLE=Forecasting wildfires in major forest types of India JOURNAL=Frontiers in Forests and Global Change VOLUME=5 YEAR=2022 URL=https://www.frontiersin.org/journals/forests-and-global-change/articles/10.3389/ffgc.2022.882685 DOI=10.3389/ffgc.2022.882685 ISSN=2624-893X ABSTRACT=

Severity of wildfires witnessed in different parts of the world in the recent times has posed a significant challenge to fire control authorities. Even when the different fire early warning systems have been developed to provide the quickest warnings about the possible wildfire location, severity, and danger, often it is difficult to deploy the resources quickly to contain the wildfire at a short notice. Response time is further delayed when the terrain is complex. Early warning systems based on physics-based models, such as WRF-FIRE/SFIRE, are computationally intensive and require high performance computing resources and significant data related to fuel properties and climate to generate forecasts at short intervals of time (i.e., hourly basis). It is therefore that when the objective is to develop monthly and yearly forecasts, time series models seem to be useful as they require lesser computation power and limited data (as compared to physics-based models). Long duration forecasts are useful in preparing an efficient fire management plan for optimal deployment of resources in the event of forest fire. The present research is aimed at forecasting the number of fires in different forest types of India on a monthly basis using “Autoregressive Integrated Moving Average” time series models (both univariate and with regressors) at 25 km × 25 km spatial resolution (grid) and developing the fire susceptibility maps using Geographical Information System. The performance of models was validated based on the autocorrelation function (ACF), partial ACF, cumulative periodogram, and Portmanteau (L-Jung Box) test. Both the univariate- and regressor-based models performed equally well; however, the univariate model was preferred due to parsimony. The R software package was used to run and test the model. The forecasted active fire counts were tested against the original 3 years monthly forecasts from 2015 to 2017. The variation in coefficient of determination from 0.94 (for year 1 forecast) to 0.64 (when all the 3-year forecasts were considered together) was observed for tropical dry deciduous forests. These values varied from 0.98 to 0.89 for tropical moist deciduous forest and from 0.97 to 0.88 for the tropical evergreen forests. The forecasted active fire counts were used to estimate the future forest fire frequency ratio, which has been used as an indicator of forest fire susceptibility.