Forest Guard: An Integrated Sensor cum AI-based Fire-prone Area Mapping and Early Forest Fire Detection System

Abstract

Abstract Ever-increasing global warming and human disruptions have resulted in unprecedented frequency, spread and intensity of destructive wildfires across continents. Thus, effective systems to tackle this are the need of the hour. “Forest Guard” is an integrated dispersed multi-variable sensor network cum AI-based system that promises pre-emptive prediction of fire-prone areas and proactive detection of wildfires. A deep learning AI model classifies the forest area on the basis of a numerical % probability of forest fire occurrence, calculated according to value intensity and degree of correlation between moving averages of parameters like Atmospheric. Temperature, Humidity and Soil Moisture gathered from sensors and other static variables. It then plots a color encoded percentage risk assessment map after extrapolating individual sensor node results to an area of 0.031km2. As this is a novel approach, the model was trained on reinforcement learning and tested on 3 simulations of 10000 data points each. It received accuracy of 98% ± 2%. Forest Guard detects distinctive wildfire and bird (distress-callings) sounds through analyzing and differentiating sounds recorded every 2 minutes by running them through a short Fast-Fourier-Transformation Deep Learning model and then raises a preliminary alarm. A confirmatory fire alarm is raised after the algorithm detects simultaneous sudden deviations with respect to the calibrated baseline in continuously monitored parameters (Atmospheric. Temperature, Humidity, Smoke, Soil Temperature and Moisture). As tested in 3 different control burns of 2500m2 each(avg.), the fire was detected within 2-3minutes (Minimum- time recorded- Preliminary-Alarm- 2:23 minutes, Confirmatory-Alarm- 3:58 minutes from ignition). All simulations/ experiments proved the hypothesis of using localized/ on-ground parameters for detection/prediction.