Remote Sensing For Fire Risk Assessment Based On Semi-supervised Machine Learning Classification

Wildfire is one of the most common natural hazards in the world.It has tendency to destroy large fields of land and make them uncultivable effecting agriculture and hence economy of the region.China is one of the countries that have a serious wildfire problem and is among the top 20 countries most affected in terms of economic loss as a result of wildfire.Fire risk estimation for purposes of risk reduction is an important aspect in disaster studies around the world.Therefore,it is highly pertinent to estimate fire risk in the region as it will empower research to formulate plans and methodology for short and long terms risk reduction.Remote sensing has become an important tool for studying various atmospheric and earth surface processes.It provides many advantages such as synoptic view utilizing the spatial resolution of the sensors,frequent monitoring capability even in remote and inaccessible areas due to improved temporal resolutions in the past few decades and the ability to quantify physical quantities on the surface of the earth through spectral information in remote sensing data.Understandably remote sensing has also become an important technology for studying fire activity.Over the years,wild fire estimation and early warning has been extensively undertaken using remote sensing.Multi-platform remote sensing is not only able to detect rapid fire events in specific regions with high accuracy but also allow monitoring of their spread and consequences in different part of the world.Improvements in compatibility of different sensors and the global coverage of satellite remote sensing platforms have also enabled comparisons into the causes and consequences of fire activity.Various remote sensing sensors and data products have been utilized for research into fire activity,including Suomi-NPP VIIRS,MODIS,ATSR,SPOT-VEGETATION,ASTER,AVHRR,etc.Research has reported significant progress in fire aspects such as fire and burn area detection and classification.The aim of this research is to estimate fire risks.The study was divided into two parts,the first part dealt with the identification of climatic and environmental parameters that relate to fire variability in the South China region.Second part dealt with the modeling of these parameters to estimate fire risk.Literature has suggested various environmental and climatic parameters that explain fire activity such as precipitation(PPT),evapotranspiration(ET)and potential evapotranspiration(p ET)among others.Initial in the study these and other parameters that lead to fire prone conditions were selected from literature and parameters relevant to the fire variability in South China region were identified using a grid base correlation analysis between fire density and each individual parameter.The grid based correlation analysis was carried out for three time scales namely annual,winter and spring.The spring(March,April and May)and winter seasons(December,January and February)are identified as fire prone season in literature for the South China region.Our results revealed strong correlation on seasonal scale but with spatial and temporal context.The south east region exhibited correlations mostly in the winter season,whereas the south west region exhibited correlation mostly in the spring season.The best parameters correlated with fire activity in South China were potential evapotranspiration and ratio between evapotranspiration and potential evapotranspiration(ET/p ET).Other parameters such as moisture balance(PPT-p ET)and precipitation also correlated with fire activity but mostly in the South East region of the study area in winter season.These correlations suggest that these environmental parameters are related to fire activity and can be used as a proxy to identify fire risk in South China region.In the second part of the study the relationships identified between environmental and climatic parameters were used to model fire risk using a special case of semi-supervised machine learning method called Learning from Positive and Unlabeled Data employing the Support vector machine(SVM)algorithm.Study spatial resolution of the girds was improved at this stage of the study but the study area was reduced to South East China for amicable handling of huge volumes of data.Three natural land covers dominate the South East China region,namely Evergreen Broadleaf Forest,Mixed Forest and Woody Savannas.Different models were trained for each of these land covers.The model for Mixed Forest land cover performed the best compared to the Evergreen Broadleaf Forest model and Woody Savannas model.It was found that better representation of Mixed Forest in training samples made the model more reliable as compared to other.The methodology and the techniques used in this study are good foundation for understanding the fire risk problem.Improving the individual models constructed for different land covers and combining them can provide fire risk classification for a larger region.There is room to improve the spatial precision of fire risk classification.Introducing finer scale features that have higher correlation with fire activity and exhibit high spatial variability seems viable way forward.Overall,semi-supervised PU learning technique using remote sensing data was found to be effective way for fire risk modeling.This study has made an important distinction by applying semi-supervised machine learning technique in fire risk estimation.It has also used the transductive bagging SVM,a PU learning technique in fire risk estimation which has not been applied before.Another important distinction is that the model constructed in this study is completely based on climatic parameter extracted from remote sensing data sources.Most of the studies identified in the literature have used a mixture of data sources.Unlike the literature surveyed for this study the model was designed to estimate fire risk in large scale natural land covers.The methodology also shows potential for integration of different models for different natural land covers for comprehensive fire risk estimation for the region of China.