Hardware Implementation of Forest Fire Detection System using Deep Learning Architectures

Forests being called as lungs of earth play a very important role in maintaining a sustainable climate on the earth. They are instrumental in maintaining a quality eco-system by filtering the air, preventing soil erosion and help to maintain diverse life on the earth. Forest fires are a matter of concern in terms of economic growth and ecological damage and damage to animals and human life. Forest fires contribute to global warming and imbalances the climate on the earth making the lives harder. Early detection of forest fire can prevent the damage by a great extent. Sensor based and Image processing-based methods have been widely used followed by machine learning techniques to process the sensor data and detect the occurrence of forest fires. These methods are costly and difficult to install at different locations in the forest. As the dimensions of the forest area increases, the complexity of the system also increases. Deep Learning techniques such as variations of convolutional neural networks process image data and can provide an early warning about the occurrence of the fire. In the proposed system different pre trained deep neural network architectures such as Resnet 50, InceptionV3, GoogleNet, AlexNet, MobileNet have been employed using transfer learning approaches on two very important datasets namely Mendely dataset and Kaggle Datasets. The best performing architecture i.e Alexnet has been deployed on to Raspberry PI embedded hardware to work as a standalone module. The trained models have demonstrated a good accuracy of 99.45% on Mendely and 99.42 on Kaggle Datasets for Fire detection.

1. Introduction

Forests being our lifeline; mankind relies on them for survival, clean air, medicines, food, and a range of environmental resources. A forest is a heavily vegetated area or a complex ecosystem that is rich in biodiversity and supports a diverse range of living organisms. They sustain the ground by avoiding soil erosion. The trees assist to keep the environment clean, which supports the forest's vegetation and animals. They are an important component of the ecosystem because they clean the air, cool it during the day, and act as excellent sound absorbers. A forest fire is indeed an uncontrolled fire as a result of combustion and heat from surface and ground fires. A large forest fire quickly spread through the upper branches of the trees before involving the shrubbery or the forest floor. As a result, violent blow-ups in forest fires are common, and they might start taking on the characteristics of a fire storm. Plant and animal extinctions, wildlife habitat devastation, depletion and loss of natural regeneration and reduction of forest cover are only a few of the major losses caused by forest fires. Warmer temperatures and drier conditions can aid in the spread of fires and make them more difficult to extinguish. Warmer, drier weather can aid in the spread of the mountain pine beetle and other insects that can weaken or kill trees, contributing to the accumulation of fuels in a forest. Increased drought and a longer fire season are boosting these increases in wildfire risk. An average annual 1 degree C temperature increase would increase the median burned area per year as much as 600 percent [11]. Sensor based approach for the early detection of forest fires is very popular. These systems require very accurate sensors to avoid false positives. Fire can be detected by using the amount of smoke. The smoke sensors are used to measure the amount of smoke from the fire, and it could be compared with a threshold value and if it is beyond that value, it is considered as fire scenario. Using image processing, fire can also be detected. Fixing the CCTV camera in the vast forest field to detect the occurrence of fire by processing of these acquired images is not a cost effective solution. Machine learning and deep learning techniques are developed to aid signal processing tasks, with the added benefit of being generalizable to different problems. The signal processing way of solving a problem is optimum but it lacks generalizability. Machine learning techniques focus on the manual extraction of relevant features from the data, rank the features for better suitability and then applied to either regression or classification problem. Manual extraction of features is cumbersome and a tedious process. Deep learning architectures which are mainly modification of convolutional neural network are prepared with intermediate layers for automatic extraction of relevant features. Machine learning and deep learning techniques are used for variety of applications that involves prediction of events. Coupled with Computer vision, deep learning can be used for the early detection of forest fires. In this paper it is proposed to build an efficient deep neural network to detect the forest fires and deploy the high performance deep neural network on to Raspberry PI embedded hardware to make it act as a standalone module. Different pertained model architectures such as Resnet50, InceptionV3, GoogleNet, Alexnet etc., have been studied and best model has been deployed on the Raspberry PI hardware to act as a standalone module.