Study On Data Assimilation Of Large Scale Wildland Fire Spread Prediction Based On Ensemble Transform Kalman Fiter

In recent years,wildland fire events occur frequently in the world,and the occurrence frequency and burned area are on the rise.Wildland fires are extremely destructive.Large-scale wildland fires not only lead to large-scale displacement of people,but also may cause casualties of residents and firefighters,houses and buildings damaged,and even cause severe social impacts and ecological disasters.To predict the spread of large-scale wildland fire in real-time and effectively is beneficial to formulate fire fighting tactics and evacuation strategy reasonably and reduce the loss caused by fire.For the traditional fire spread prediction method,given the location of the initial fire perimeter,terrain and vegetation conditions of the fire site,wind field and other environmental and climatic conditions,the result of fire spread prediction is certain.However,the uncertainty of the location of the initial fire perimeter,the intervention of the fire fighting,the error of the terrain and vegetation data,the change of wind speed and direction,and so on,will lead to a large deviation between the prediction results of the traditional methods and the real situation of fire spread,bringing adverse consequences to the emergency rescue.With the development and maturity of remote sensing technology,we can obtain the real-time information of fire perimeter positions through satellites or drones.Through the data assimilation method,the state and parameters of the model are updated in real time to reduce the state and parameter errors of the model,so as to greatly improve the accuracy of wildland fire spread prediction.The purpose of this paper is to establish a framework and system for data assimilation prediction of large-scale wildland fire spread,propose a more realistic state and parameter estimation method,and improve the data assimilation algorithm under bad observation data.The specific work of this paper is as follows:A data assimilation prediction framework to integrate Kalman filter based on ensemble predict and fire spread model was proposed,and a data assimilation prediction system for wildland fire spread was established.The system mainly includes observation module,data assimilation module,fire spread simulation module and control layer.Observation module,which is used to input observation data;data assimilation module,which is written by MATLAB language and compiled into executable file;fire spread simulation module,which is the fourth version of FARSITE command line;control layer,which is written by Python language,controls the data flow process of the whole system.In this system,the reliability of the algorithm is tested when the error of the observation data changes.An ensemble transform Kalman filter(ETKF)algorithm for state estimation of wildland fire spread is proposed.ETKF overcomes the limitation that the ensemble Kalman filter needs artificial disturbance of observation data;Based on an observing system simulation experiment(OSSE)approach,the performance of ETKF and the ensemble Kalman filter in state estimation is compared under the condition that the wind field conditions are known and unknown.The results of fire propagation prediction including burned area,fire perimeter length and fire perimeter position are studied,and the performance of data assimilation algorithm is evaluated by using Hausdorff distance(HD),a more conservative index compared with root mean square error(RMSE).The results show that,when the wind field conditions are known,ETKF and ensemble Kalman filter perform well in dealing with the errors of initial conditions and boundary conditions.However,when the wind field conditions are unknown,that is,when there are model parameter errors,both the predicted and analyzed values of ETKF are significantly closer to the real values than the results of ensemble Kalman filter.It highlights the application advantage of ETKF when the error source is more consistent with the actual situation.A strategy for estimating the position of fire perimeter and fuel adjustment factors is constructed.The estimation of fire perimeter position is the key to reduce model state error and the estimation of fuel adjustment factors is an important part to reduce model parameter uncertainty.For this reason,we still use ETKF-based algorithm to modify the fire perimeter position and use Monte Carlo based radial basis function neural network(RBFNN)to estimate the fuel adjustment factors.It is verified by the data of fireflux I experiment and camp fire event.The results show that the proposed state parameter estimation strategy can improve the prediction accuracy and highlight the advantages of the proposed hybrid estimation strategy.The estimation of fire perimeter position can detect fire spotting and merging and update fire perimeter position.The different stages of wildland fire development(such as "explosive" stage)are fully modelled.In addition,the Monte Carlo based RBFNN has a low computational cost,which is helpful for emergency rescue decision-making.The improved ETKF algorithm is systematically studied under adverse observation data.In the prediction of wildland fire spread data assimilation,it is usually assumed that the observation data of fire perimeter position is complete.In addition,the traditional observation error also assumes only one variance.Due to instrument failure or the presence of hot plumes caused by fire or clouds,observations of the location of wildland fires obtained by aircraft or satellite-mounted sensors,may be incomplete and/or spatial variations in the error variance of observations.In order to overcome these problems,we introduce a vertex weight to modify the "nudge" term to make up for the shortcomings of ETKF algorithm in processing such observation data.We can adjust the extent to which the predicted fire perimeter position is pushed to the observed fire perimeter position at the predicted fire vertex without observation or with lower fidelity of observation.Through a series of OSSEs,the performance of VWETKF is compared with that of ETKF,and the superiority of the state estimation of VWETKF's flexible spatial distribution is verified.We consider the practical situations that the observed fire perimeter is a part of the real fire perimeter and/or part of the observed data is subject to a larger noise.The results showed that VWETKF with the new "nudge"term improved the accuracy of analyzed and predicted fire perimeter positions.