Forest Fire Smoke Detection Base On Multispectral Remote Sensing Imagery

As a result of global warming,forest fire is occurring frequently around the world.It is a sudden,destructive and difficult to control natural disaster that causes huge losses to ecosystems and human life and property.Early detection of forest fire plays a crucial role in controlling fire spread and fire suppresion.Remote sensing satellites have been widely used in forest fire monitoring due to their advantages such as low cost,strong real-time performance,unrestricted terrain conditions,and wide monitoring area.In recent years,researchers have developed many algorithms and fire products for forest fire detection that have been well-ver ified in different parts of the world for different remote sensing satellite sensors.However,the current fire detection algor ithms have weak ability to recognize small-magnitude fire and smoldering combustion.Smoke detection algorithms have poor adaptability and cannot meet the requirements for early detection of forest fires.The objective of this study is to overcome the shortcomings of current algorithms.Based on the results of current computer vision development,an improved method is proposed,and redundant data in remote sensing bands are analyzed to extract sensitive band data.The article is organized as follows:First,in order to address the current problems of fire detetion algorithms which cannot detect small-area wildf ire and smolder ing combustion,we selected a high-resolution terrestrial observation satellite-data to establish a forest fire smoke dataset which contains variety regions and different seasons.Second,due to powerful feature extraction capabilities,deep learning algorithms are used to mine the sementic information of multispectral remote sensing images.Inspired by the U-net network,we present a smoke segmentation algor ithm based on self-encoding and SEBlock which fuses multispectral data and multi-remote sensing indexs.The results show that the accuracy of this method is more than72.5%,which can effectively detect forest fire smoke of different scales,and the method has good scalability and adaptability to various types and different surface backgrounds.Finally,we analyze the sensitivity of the bands data and select the most highly sensitive bands as the data source by comparing the impact of different bands combination on the remote sensing image segmentation results.