Tomàs Artés

Real-time genetic spatial optimization to improve forest fire spread forecasting in high-performance computing environments

Abstract Forest fires are a kind of natural hazard with a high number of occurrences in southern European countries. To avoid major damages and to improve forest fire management, one can use forest fire spread simulators to predict fire behavior. When providing forest fire predictions, there are two main considerations: accuracy and computation time. In the context of natural hazards simulation, it is well known that part of the final forecast error comes from uncertainty in the input data. These data typically consist of a set of GIS files, which should be appropriately conflated. For this reason, several input data calibration methods have been developed by the scientific community. In this work, the Two-Stage calibration methodology, which has been shown to provide good results, is used. This calibration strategy is computationally intensive and time-consuming because it uses a Genetic Algorithm as a solution. Taking into account the aspect of urgency in forest fire spread prediction, it is necessary to maintain a balance between accuracy and the time needed to calibrate the input parameters. In order to take advantage of this technique, one must deal with the problem that some of the obtained solutions are impractical, since they involve simulation times that are too long, preventing the prediction system from being deployed at an operational level. A new method which finds the minimum resolution reduction for such long simulations, keeping accuracy loss to a known interval, is proposed. The proposed improvement is based on a time-aware core allocation policy that enables real-time forest fire spread forecasting. The final prediction system is a cyberinfrastructure, which enables forest fire spread prediction at real time.

Enhancing computational efficiency on forest fire forecasting by time-aware Genetic Algorithms

A way to overcome data input uncertainty when simulating forest fire propagation, consists of calibrating inaccurate input data by applying computational-intensive methods. Genetic Algorithms (GA) are powerful and robust optimization techniques. However, their main drawback is their overall run time, which can easily become unacceptable, especially when dealing with natural disasters forecast. The prediction system has been parallelized using a hybrid MPI-OpenMP approach where the number of cores allocated to each GA individual is based on a priori time-aware population classification, which allows to keep bounding the optimization process bound to a predetermined deadline. In this work, an efficient time-aware GA is introduced that estimates the required number of cores to keep the calibration process under imposed time limits and also takes into account an efficient use of the computational resources.

Time aware genetic algorithm for forest fire propagation prediction: exploiting multi‐core platforms

Forest fire propagation prediction is a crucial issue when fighting these hazards as efficiently as possible. Several propagation models have been developed and integrated in computer simulators. Such models require a set of input parameters that, in some cases, are difficult to know or even estimate precisely beforehand. Therefore, a calibration technique based on genetic algorithm (GA) was introduced to reduce the uncertainty in input parameters values and improve the accuracy of the predictions. Such a technique requires the execution of a set of simulations and several iterations of the process to calibrate the values of the input parameters. To reduce the execution time of this calibration stage, an Message Passing Interface master/worker scheme was developed to distribute the simulations of one iteration among the worker processes. However, the execution time of each simulation varies drastically depending on the particular input parameters used, provoking a significant load imbalance. To overcome this imbalance and reduce execution time to operational requirements, core allocation policies have been developed. These policies are based on execution time estimation and classification of simulations according to the estimated execution time. Then, multicore capabilities of the current systems are applied to devote more resources (cores) to the longest simulations reducing the load imbalance. These simulations that are estimated as taking too long, even when many resources are devoted to them, require especial consideration. So, a generation time limit has been introduced, and three different strategies have been designed considering individuals that exceed the generation execution time limit. In the first one, the longest individuals are replaced before starting the execution with shorter individuals (Time Aware Core allocation with replacement). In the second one, these individuals are executed, but when the generation limit is reached, the individuals still executing are killed (Time Aware Core allocation without replacement). In the third one, all the individuals are executed normally, and when the generation time limit is reached, the GA is applied considering the individuals that have finished their executions, while the individuals still executing are allowed to continue running and are considered by the GA when they finish. The three strategies have been tested in real scenarios, and the results show these policies significantly improve the calibration accuracy within the superimposed deadlines.

Core Allocation Policies on Multicore Platforms to Accelerate Forest Fire Spread Predictions

Software simulators are developed to predict forest fire spread. Such simulators require several input parameters which usually are difficult to know accurately. The input data uncertainty can provoke a mismatch between the predicted forest fire spread and the actual evolution. To overcome this uncertainty a two stage prediction methodology is used. In the first stage a genetic algorithm is applied to find the input parameter set that best reproduces actual fire evolution. Afterwards, the prediction is carried out using the calibrated input parameter set. This method improves the prediction error, but increments the execution time in a context with hard time constraints. A new approach to speed up the two stage prediction methodology by exploiting multicore architectures is proposed. A hybrid MPI-OpenMP application has been developed and different allocation policies have been tested to accelerate the forest fire prediction with an efficient use of the available resources.

A High Performance Computing Framework for Continental-Scale Forest Fire Spread Prediction

Abstract Many scientific works have focused on developing propagation models that predict forest fire behavior. These models require a precise knowledge of the environment where the fire is taking place. Geographical Information Systems allow us determining and building the different information layers that define the terrain and the fire. These data, along with meteorological information from weather services, enables the simulation based on real conditions. However, fire spread prediction models require a set of input parameters that, in some cases, are difficult to know or even estimate precisely. Therefore, a framework, based on a genetic algorithm calibration stage, was introduced to reduce the uncertainty in the input parameters and improve the accuracy of the predictions. This stage is implemented using a MPI master/worker scheme and an OpenMP parallel version of the fire spread simulator. Additionally, the whole system is run using suitable automatic worker-assignment and core-allocation policies to respect the existing time restrictions, inherent to this real-world problem. This paper details the process of obtaining the necessary input data as well as the parallel evolutionary framework that delivers the final prediction. A real case study is presented to illustrate the way this framework works.

Non-supervised method for early forest fire detection and rapid mapping

Natural hazards are a challenge for the society. Scientific community efforts have been severely increased assessing tasks about prevention and damage mitigation. The most important points to minimize natural hazard damages are monitoring and prevention. This work focuses particularly on forest fires. This phenomenon depends on small-scale factors and fire behavior is strongly related to the local weather. Forest fire spread forecast is a complex task because of the scale of the phenomena, the input data uncertainty and time constraints in forest fire monitoring. Forest fire simulators have been improved, including some calibration techniques avoiding data uncertainty and taking into account complex factors as the atmosphere. Such techniques increase dramatically the computational cost in a context where the available time to provide a forecast is a hard constraint. Furthermore, an early mapping of the fire becomes crucial to assess it. In this work, a non-supervised method for forest fire early detection and mapping is proposed. As main sources, the method uses daily thermal anomalies from MODIS and VIIRS combined with land cover map to identify and monitor forest fires with very few resources. This method relies on a clustering technique (DBSCAN algorithm) and on filtering thermal anomalies to detect the forest fires. In addition, a concave hull (alpha shape algorithm) is applied to obtain rapid mapping of the fire area (very coarse accuracy mapping). Therefore, the method leads to a potential use for high-resolution forest fire rapid mapping based on satellite imagery using the extent of each early fire detection. It shows the way to an automatic rapid mapping of the fire at high resolution processing as few data as possible.

Error Function Impact in Dynamic Data-driven Framework Applied to Forest Fire Spread Prediction☆

Abstract In order to use environmental models effectively for management and decision-making, it is vital to establish an appropriate measure of confidence in their performance. There are different ways and different methodologies to establish and measure the confidence of the models. In this paper, we focus on the forest fire spread prediction. Simulators implementing forest fire spread models require diverse input parameters to deliver predictions about fire propagation. However, the data describing the actual scenario where the fire is taking place are usually subject to high levels of uncertainty. In order to minimize the impact of the input-data uncertainty a Two-Stage methodology was developed to calibrate the input parameters in (1) an adjustment stage so that the calibrated parameters are used, and (2) the prediction stage to improve the quality of the predictions. Is in the adjustment stage where the error formula plays a crucial role, because different formulas implies different adjustments and, in consequence, different wild fire spread predictions. In this paper, different error functions are compared to show the impact in terms of prediction quality in DDDAS for forest fire spread prediction. These formulas have been tested using a real forest fire that took place in Arkadia (Greece) in 2011.