A hybrid artificial intelligence approach using GIS-based neural-fuzzy inference system and particle swarm optimization for forest fire susceptibility modeling at a tropical area

Agricultural and Forest Meteorology 233 (2017) 32–44

Contents lists available at ScienceDirect

Agricultural and Forest Meteorology journal homepage: www.elsevier.com/locate/agrformet

A hybrid artificial intelligence approach using GIS-based neural-fuzzy inference system and particle swarm optimization for forest fire susceptibility modeling at a tropical area Dieu Tien Bui a,∗ , Quang-Thanh Bui b , Quoc-Phi Nguyen c , Biswajeet Pradhan d,f , Haleh Nampak d , Phan Trong Trinh e a Geographic Information System Group, Department of Business Administration and Computer Science, University College of Southeast Norway, Hallvard Eikas Plass, N-3800 Bø i Telemark, Norway b Faculty of Geography, VNU University of Science, 334 Nguyen Trai, Thanh Xuan, Hanoi, Vietnam c Department of Environmental Sciences, Hanoi University of Mining and Geology, Duc Thang, Bac Tu Liem, Hanoi, Vietnam d Department of Civil Engineering, Geospatial Information Science Research Center (GISRC), Faculty of Engineering, University Putra Malaysia, Serdang, Selangor Darul Ehsan 43400, Malaysia e Institute of Geological Sciences, Vietnam Academy of Sciences and Technology, 84 Chua Lang Street, Dong da, Hanoi, Vietnam f Department of Geoinformation Engineering, Choongmu-gwan, Sejong University, 209 Neungdong-Ro Gwangjin-Gu, Seoul 05006, Republic of Korea

a r t i c l e

i n f o

Article history: Received 20 June 2016 Received in revised form 25 September 2016 Accepted 5 November 2016 Keywords: Forest fire Particle swarm optimization Neural-fuzzy GIS Lam Dong Vietnam

a b s t r a c t This paper proposes and validates a novel hybrid artificial intelligent approach, named as Particle Swarm Optimized Neural Fuzzy (PSO-NF), for spatial modeling of tropical forest fire susceptibility. In the proposed approach, a Neural Fuzzy inference system (NF) was used to establish the forest fire model whereas Particle Swarm Optimization (PSO) was adopted to investigate the best values for the model parameters. Tropical forest at the province of Lam Dong (Central Highland of Vietnam) was used as a case study. For this purpose, historic forest fires and ten ignition factors (slope, aspect, elevation, land use, Normalized Difference Vegetation Index, distance to road, distance to residence area, temperature, wind speed, and rainfall) were collected from various sources to construct a GIS database, and then, the database was used to develop and validate the proposed model. The performance of the forest model was assessed using the Receiver Operating Characteristic curve, area under the curve (AUC), and several statistical measures. The results showed that the proposed model performs well, both on the training dataset (AUC = 0.932) and the validation dataset (AUC = 0.916). The usability of the proposed model was further assessed through comparisons with those derived from two benchmark state-of-the art machine learning methods, Random Forests (RF) and Support Vector Machine (SVM). Because the performance of the proposed model is better than the two benchmark models, we concluded that the PSO-NF model is a valid alternative tool that should be considered for tropical forest fire susceptibility modeling. The result in this study is useful for forest planning and management in forest fire prone areas. © 2016 Elsevier B.V. All rights reserved.

1. Introduction Forest fire is an environmental problem that poses threats to the safety of human life, infrastructures, and the environment (Podur et al., 2003), and is considered as an important agent of pattern formation of forests, such as succession and regeneration (Huebner et al., 2012). Due to changes in climate i.e. less rainfall and/or increment of day temperature, longer dry season, and interventions of

∗ Corresponding author. E-mail addresses: [email protected], [email protected] (D. Tien Bui). http://dx.doi.org/10.1016/j.agrformet.2016.11.002 0168-1923/© 2016 Elsevier B.V. All rights reserved.

˜ human activities (Arganaraz et al., 2015), the trend and frequency of forest fires are increased and reached an alarming rate in many ˜ areas of the world (Arganaraz et al., 2015; Arpaci et al., 2014; Littell et al., 2016; Mavsar et al., 2013; Moritz et al., 2012). Therefore, it is necessary to predict forest fires as accurately as possible. This will assist local authorities in forest management and planning, resource allocation, emergency services, and early warning systems (Eastaugh and Hasenauer, 2014). Various approaches have been proposed for modeling of forest fire behaviors and they can be classified into three groups including physics-based method, statistical method, and machine learning method. The physics-based method simulates and predicts

D. Tien Bui et al. / Agricultural and Forest Meteorology 233 (2017) 32–44

potential fire behaviors through a set mathematical equations of fluid mechanics, combustions of canopy biomass, and heat transfer mechanisms (Pastor et al., 2003). Therefore, they are capable to model fire behaviors in both space and time. The most common used physics-based models for forest fires are EMBYR (Hargrove, 1994), FARSITE- Fire Area Simulator (Keane et al., 1998), FDS (McGrattan et al., 2000), FIRETEC (Linn et al., 2002), FireStation (Lopes et al., 2002), and LANDIS-II (Sturtevant et al., 2009). The main disadvantage of these models is that it is difficult to quantify the magnitude of the errors (Massada et al., 2011). In addition, physicsbased models require detailed data, i.e. locations and dimensions of trees, fuel mass, soil moisture that are difficult to collect for large areas (Pimont et al., 2016). Statistical method is more suitable for forest fire modeling when study areas are large, especially in combinations with geographic information system (GIS) technology (Duarte et al., 2016; Teodoro et al., 2015; Teodoro and Duarte, 2013), due to its capacity to collect and process spatial data of large regions with different scales and resolutions (Bonham-Carter, 2014; Chuvieco et al., 2010; Tien Bui et al., 2016d; Verde and Zêzere, 2010; Wittenberg and Malkinson, 2009). Consequently, various statistical methods and techniques have been adopted for forest fire modeling, such as Poisson regression (Wotton et al., 2003), generalized Pareto distribution (Bermudez et al., 2009), favorability functions (Verde and Zêzere, 2010), generalized additive model (Vilar et al., 2010), Monte Carlo simulations (Conedera et al., 2011), multiple linear regression (Oliveira et al., 2012), logistic regression (Arndt et al., 2013; Chuvieco et al., 2010; Pourghasemi, 2015), weights of evidence (Amatulli et al., 2007), and geographically weighted regression (Oliveira et al., 2014). However, forest fire regime is a typical complex and non-linear process that is difficult to assess and predict, therefore prediction accuracy of these models is not always satisfied. Due to critical accuracy of forest fire models, machine learning methods have been explored and investigated such as Decision Tree learning (Camp et al., 1997), Kernel methods (Gonzalez-Olabarria et al., 2012), Random Forests (Arpaci et al., 2014; Oliveira et al., 2012), Kernel logistic regression (Tien Bui et al., 2016b), Maximum Entropy (Arpaci et al., 2014), artificial neural networks (ANN) (Cheng and Wang, 2008; Satir et al., 2015; Vasconcelos et al., 2001). In general, performance of machine leaning models is better when compared with the statistical models (Massada et al., 2013). Nevertheless, prediction of forest fire at regional scales is still difficult due to multiple and complex interactions of ignition factors. In addition, forest fire modeling requires collecting ignition factors with different resolutions i.e. climate and weather conditions (temperature, humidity, rainfall, and wind), topography (elevation, slope, aspect), and land use (Ganteaume et al., 2013). Therefore, it is difficult to eliminate uncertainties as well as imprecisions (Tien Bui et al., 2016c). As a result, these above models have difficulties in processing fire ignition maps with inaccuracies of GIS information, such as spatial errors due to age of data, different scales and resolutions (Benz et al., 2004; Short, 2014; Zald et al., 2014). To address this issue, fuzzy reasoning has been proposed (Loboda and Csiszar, 2007). However, the subjective determination of fuzzy membership values prevents this method from producing high accuracy results. Therefore, developments of new prediction models for forest fires that lead to deal with uncertainties and imprecisions and improve fire prediction capabilities are highly necessary. This paper attempts to partially fill this gap in literature by proposing a novel hybrid artificial intelligent model, named as Particle Swarm Optimized Neural Fuzzy (PSO-NF) model, for spatial prediction of tropical forest fire susceptibility with a case study at the province of Lam Dong in the Central Highland region (Vietnam), a typical province that has seriously affected by forest fires during the last ten years (Dan et al., 2014). In the proposed approach,

33

the forest fire model is established based on a neural fuzzy inference system that combined fuzzy logic and neural networks (Jang, 1993). Development of the neural fuzzy model has a challenge of finding the best premise and consequent parameters that strongly influence the model performance, therefore, this study proposes Particle Swarm Optimization (PSO)(Eberhart and Kennedy, 1995) for optimizing the neural fuzzy model. PSO is considered a powerful algorithm that has been widely used in soft computing for model optimization, such as in rainfall–runoff modeling (Taormina and Chau, 2015), flood susceptibility modeling (Tien Bui et al., 2016e), dam behavior modeling (Bui et al., 2016), and slope instability analysis (Gordan et al., 2016). Consequently, the proposed model is a powerful inference system that has capabilities to process uncertainty data with high accuracy. Finally, the usability of the proposed model is confirmed through comparisons with benchmarks, Random Forests (RF) and Support Vector Machine (SVM), and concluding remarks are provided. 2. The study area and data used 2.1. Description of the study area Lam Dong province (Fig. 1) is located on the southern part of the Central Highland region of Vietnam, between latitudes 11◦ 12 00 N and 12◦ 15 00 N, and between longitudes 107◦ 15 00 E and 108◦ 45 00 E. The province occupies 9805.4 km2 with complex topographical variations. The altitude varies between 120 m and 2280 m above sea level, with the mean of 907.6 m and the standard deviation of 392.1 m. Climate in the province is influenced by tropical monsoon with moderate temperatures and high humidity. The local climate varies with altitude and can be divided into two distinct seasons, rainy and dry seasons. The rainy season extends from May to November, whereas the dry season is from December to April. The temperature varies significantly across the region, ranging from 18◦ C to 25◦ C with mild and cool weathers. The rainfall in the rainy season is up to 90% of the total annual rainfall, with values between 1600 and 2700 mm per year. The average relative humidity of the year ranges from 85% to 87%. The forest coverage of the province is approximately 60% of the total study area (Vu et al., 2013) whereas agricultural land and populated areas covers around 28% and 6%, respectively. The remaining is for other types of land covers. Tree species are dominated by Suzygium, Dipterocarpus, Anisopkera cochinchinensia, and Schima superba Gardner & Champ (between 1000 and 1300 m), Pinus merkusii (600 and 1000 m), Pinus khasya (>1000 m), Dipterocarpus obtusifolius and Shorea obtusa (1300 m). According to Vu et al. (2013), deforestation and forest degradation continue to be key threats to the forest cover of the province during the period 1990–2010. Although the plantation forest was increased 81.7% (5.1% of the total study area), however the total forest area was reduced by 118067.9 ha (12.04% of the total study area). Broadleaf forest, bamboo forest, and coniferous forest were reduced by 30.5%, 37.1%, and 28.2% respectively. These degradations are equivalent to 17% of the total study area and causes of forest loss can be named as fire, illegal logging, land conversion and inefficient management (Vu et al., 2013). In fact, there is a significant pressure of population growth on forest resources due to increased demands for residential lands and lands for productions. 2.2. Historical forest fires and ignition factors 2.2.1. Historical forest fires Because prediction models for forest fire susceptibility are developed based on relationship analysis of historical forest fires

34

D. Tien Bui et al. / Agricultural and Forest Meteorology 233 (2017) 32–44

Fig. 1. Location of the study area and historical forest fires.

Table 1 Temporal analysis of forest fire occurrence in this study. No.

Forest fires (%)

Month

1 2 3 4 5 6 7 8

12.8 19.8 39.1 20.7 5.0 1.5 1.1 0.0

January February March April May June November July, August, September, October, and December

and their ignition factors (Higuera et al., 2015; Tien Bui et al., 2016b), therefore preparation of a forest fire inventory map is a mandatory task. In this study, a forest fire inventory map with a total 540 historical fires locations was compiled. These fires that occurred in 2013 were provided by the Department of Forest Protection (Ministry of Agriculture and Rural Development of Vietnam, 2016) (now available at http://www.kiemlam.org.vn/firewatchvn/ ). This is the official national database on forest fire in Vietnam. In this analysis, only the forest fires in 2013 were selected, because in this year, the study area was suffered the worst drought during the last three decades (Dan et al., 2014). Our analysis of these fires locations shows that many forest fires occurred in March (39.1%). In contrast to March, no forest fire was reported for July, August, September, October, and December. Detail statistical analysis on the temporal occurrence of the forest fires is shown in Table 1.

2.2.2. Forest fire ignition factor Selection of the appropriate ignition factors for forest fire modeling is an important issue (Verde and Zêzere, 2010) that influences the quality of resulting prediction models. It is wellknown that forest fires are influenced not only on non-climatic factors i.e. topography, vegetation characteristics, and human factors (Eastaugh and Hasenauer, 2014; Guo et al., 2016; Pourtaghi et al., 2016; Tien Bui et al., 2016b), but also on climatic factor i.e. temperature, rainfall, and wind, therefore these factors should be considered. In this research, ten ignition factors were considered: slope (o ), aspect, elevation (m), land use, NDVI (Normalized Difference Vegetation Index), distance to road (m), distance to residence area (m), temperature (◦ C), wind speed (m/s) and rainfall (mm). Three topographical factors including slope, aspect, and elevation were extracted from a Digital Elevation Model (DEM) of the study area. The DEM was generated from national topographic maps at scale of 1:50,000 using ArcGIS 10.2 software. These topographic maps used 5 m, 10 m, and 20 m contour intervals for slopes areas of 0–2◦ , 2–15◦ , and >15◦ , respectively (Tien Bui et al., 2016a). Slope is selected for this analysis because it is a well-known factor for fire progression, and the steeper slopes, the faster fire travelling is (Lentile et al., 2006). Regarding elevation, this factor may influence precipitation, temperature, humidity, and evapotranspiration (Camp et al., 1997; Verde and Zêzere, 2010), therefore it was selected for this analysis. In this study, the slope map (Fig. 2a) was generated with seven classes, whereas for the elevation map (Fig. 2b), six classes were considered. These classes of the two factors were determined based on the natural breaks algorithm (North,

D. Tien Bui et al. / Agricultural and Forest Meteorology 233 (2017) 32–44

2009) that is available in ArcGIS 10.2. This algorithm that is based on Jenks’s optimization (Jenks and Caspall, 1971) is capable to naturally determine breakpoints for homogenous classes. Accordingly, the sum of variance of each class is minimized but the variance

35

between these classes is maximized. For the case of aspect, this factor has been widely used for forest fire modeling (Camp et al., 1997), because slope directions influence soil moistures and wind speeds that affect fire behaviors (Schmidt et al., 2008). In this study,

Fig. 2. Forest fire ignition factors used in this study: (a) Slope map; (b) Elevation map; (c) Aspect map; (d) Land use map; (e) NDVI map; and (f) Distance to road map. (CCC: Public land; CDG: Specially used land; CHN: Annual crop land; CLN: Perennial crop land; CSD: Unused land; CSK: Land for non-agriculture production and business; LNP: Other forest land; NKH: Other annual crop land; NTS: Water surface for aquaculture; OTC: Residential land; RDD: Specially used forest land; RPH: Protective forest land; RSX: Productive forest land; SMN: Water surface; TTN: Religious land). (g) Distance to residential area map; (h) Temperature map; (i) Wind speed map; and (j) Rainfall.

36

D. Tien Bui et al. / Agricultural and Forest Meteorology 233 (2017) 32–44

Fig. 2. (Continued)

the aspect map (Fig. 2c) with nine classes was considered, including one plain area class and eight slope direction classes. Human land use activities could be ignition sources that induce forest fire susceptibility (Huesca et al., 2009; Nepstad et al., 2008), therefore, land use was selected for this analysis. The land use map (Fig. 2d) with fifteen classes at scale of 1:50,000 for the study area was provided by the local authority of the Lam Dong province. This map was a result from the national land use inventory project carried out in 2010 and was updated in 2013. NDVI is also considered as an important factor in forest fire modeling because NDVI reflexes vegetation health status (Bajocco et al., 2015) that is a proxy for load fuel distribution (Yi et al., 2013). In this study, Landsat-8 Operational Land Imagery with 30 m resolution obtained from the USGS archive (available at http://earthexplorer.usgs.gov) was used to compute NDVI using the equation as follows: NDVI = (NIR − RED)/(NIR + RED)

(1)

where NIR is the near-infrared band (0.76–0.90 ␮m, Band 4) and RED is the red band (0.63–0.69 ␮m, Band 3). The NDVI map (Fig. 2e) for the study area was computed with nine classes and these classes were determined using the aforementioned Natural Breaks algorithm. Many forest fires in Vietnam are related anthropogenic activities such as grass burning, hunting with fire, and forest exploitation (Le

et al., 2014), therefore distance to road and distance to residential area were included in this analysis. The road network was extracted from the national topographic maps at scale of 1:50,000, and then, the distance to road map (Fig. 2f) with five classes was constructed by calculating the Euclidean distance from the road lines using the buffer tool in ArcGIS 10.2. Residential areas for the Lam Dong province were extracted from the land use map, and then were used to construct the residential area map using the aforementioned buffer tool. This map was generated with six classes (Fig. 2g). Regarding climatic factors, variations of air temperature and wind speed have found influencing severity and frequency of forest fires (Gillett et al., 2004; Heimann and Reichstein, 2008), whereas variation of rainfall influences soil moisture and drought (Guttman, 1998; Zaitchik et al., 2013). Therefore, temperature, wind, and rainfall were selected for this analysis. In this study, the temperature, wind, and rainfall data for the Lam Dong province (in 2013) were derived from the Climate Forecast System Reanalysis, available at https://www.ncdc.noaa.gov/. The temperature map (Fig. 2h) with six classes was considered, whereas in the wind speed map (Fig. 2i), five classes were used, and for the rainfall map (Fig. 2j), seven classes were used. These classes were determined using the aforementioned Natural breaks algorithm in ArcGIS 10.2.

D. Tien Bui et al. / Agricultural and Forest Meteorology 233 (2017) 32–44

37

Fig. 3. Structure of the neural-fuzzy inference system.

3. Theoretical background of neural-Fuzzy inference system and particle swarm optimization

Layer 4 (Consequence node): This is the adaptive layer where the output for each note is estimated using Eq. 5.

3.1. Neural-Fuzzy inference system

¯ j (p0j + p1j x1 + p2j x2 + ... + p10j x10 ) Opj = ω

Neuro-fuzzy inference system is a hybrid framework based on fuzzy logic and artificial neural networks (ANN) to infer relationships between inputs and outputs. In this study, the neural-fuzzy structure with Takagi-Sugeno inference engine proposed by Jang (1993) was used for forest fire susceptibility modeling. This fuzzy inference engine was selected because it has capacity to model complex nonlinear problems more accurate and fewer rules than others, such as Mamdani engine (Abraham, 2001; Kumar and Verma, 2015; Lilly, 2010). The neural fuzzy structure (Fig. 3) consists of five layers with two types of nodes, adaptive nodes and fixed nodes. The first and forth layers are designed with adaptive nodes whereas the other layers contain fixed nodes only. In the adaptive nodes, connection weighs are adjusted in training phases to fit training data whereas fixed nodes simply do summation or normalization of all incoming signals. n Let (xi , y)N i=1 is the training dataset where xi ∈ R is the vector of input variables with n dimension, N is the number of training data samples. y denotess the class labels with “1” is for these forest point points and “0” is for these non-forest points. In the current research, 10 input variables were used (slope, aspect, elevation, land use, NDVI, distance to road, distance to residence area, temperature, wind speed, and rainfall). The process flow of the neural fuzzy inference system can be expressed as follows: Layer 1 (Input node): The ten forest fire ignition factors (described in Section 2.2.2) were assigned fuzzy membership values through conducting a fuzzification process. In this analysis, Gaussian function showed in Eq.2 was used Ai,k (x)

−(x − c)2 = exp 2ı2

(2)

where Ai,k (x) are the membership function associated with each of ten input variables; ıand c are the parameters of the function and called premise parameters; i = 1,., 10 and k is the number of classes in each input factor. Because this is the adaptive layer, so the premise parameters will be adjusted and optimized in the training phase. Layer 2 (Rule node): This is the fixed node layer where the weights for these nodes are calculated using Eq. 3. ␻j = A1,k (x1 ) .A2,k (x2 ) ., ..., .A10,k (x10 )

(3)

Layer 3 (Average node): This is also the fixed node layer but the weights calculated in the previous layer are normalized using Eq. 4 ω ¯ j = ωj /



ωj

(4)

(5)

where p0j , p1j , ...p10j are the consequent parameters of the affine function. Layer 5 (Output node): This is the aggregation layer that summaries all incoming signals to calculate forest fire susceptibility index in Eq. 6 Forest fire susceptibility index =



Opj

(6)

3.2. Particle swarm optimization Proposed by Eberhart and Kennedy (1995), Particle Swarm Optimization (PSO) is a powerful optimization technique based on population. PSO is proposed for this study because it has been applied successfully in a wide range of engineering fields and proven outstanding performance in many real-world problems, such as hydrological modeling (Taormina and Chau, 2015), flood mapping (Li et al., 2015), and landslide modeling (Zhou et al., 2016). This section describes the PSO algorithm that is proposed in this paper for optimizing the premise and consequent parameters of the neural-fuzzy model for forest fire modeling in this study. The PSO algorithm is based on a pattern matrix called a flock of particles or swarm, with the goal to find the best position for the swarm. In this position, the difference (RMSE, Eq. 7) of the outcome and the target values is smallest. OPTi)2nRMSE = SQRT[(1/n) ∗



(TAGi -OPTi )2 ]

(7)

where n is total samples in the training dataset or the validation dataset; TAGi is the target values of the training dataset or the validation dataset; values from forest fire  it itOPTi is output   it the  model. it it , . . ., vit , xi2 , . . ., xiS and vit = v , v are the Let xiit = xi1 i i1 i2 iS position and velocity of particle i at iteration it in the constrained dimensional space (S). During the training process, the position and velocity of the particle i will be changed and updated to find the best position using equations as follows: it+1 it xis = xis + vit is

(8)



 it



 it

it vitid = W vit−1 + C1 r1 Pisit − xis + C2 r2 PGs − xis , with s = 1, 2, · · ·, S is

(9)

where W is the inertia weight; C1 denotes the personal learning factor, whereas C2 is the social learning factor; r1 and r2 are random it is the best position of the particle i at values in the range [0,1]; Pis

it is the best position of the swarm at iteration it. iteration it; PGs The operating mechanism of PSO for optimizing the forest fire model in this study is shown in Fig. 4.

38

D. Tien Bui et al. / Agricultural and Forest Meteorology 233 (2017) 32–44

MathWorks, 2014). Meanwhile the proposed PSO-NF model was developed by the authors in Matlab environment. In addition, a C++ application was also programmed to convert the forest fires susceptibility values to a raster format for further works in a GIS environment. 4.1. Stage 1: Construction of the forest fire database and preparation of the training and validation datasets

Fig. 4. The operating mechanism of PSO for optimizing the forest fire model in this study.

3.3. Model evaluation The goodness-of-fit and prediction power of the forest fire susceptibility models were evaluated based on statistical measures (Tien Bui et al., 2016b) such as overall success rate, positive predictive value, negative predictive value, specificity, sensitivity (Eqs. 10 and 11). In addition, Receiver Operating Characteristic (ROC) curve, and Area Under the Curve (AUC) were also used. Overallsuccessrate =

Specificity =

TP + TN TP ; Sensitivity = TP + TN + FP + FN TP + FN (10)

TN TP TN ; PPV = ; PPv = FP + TN FP + TP FN + TN

(11)

where TP (True Positive) and TN (True Negative) are samples in the training or validation datasets that area correctly classified. FP (False Positive) and FN (False Negative) samples in the training or validation datasets that are misclassified. Overall success rate is the number of forest fires and non-fire points that are correctly classified divided to the total points. Positive Predictive Value (PPV) is the probability of points that is classified to forest fires, whereas Negative Predictive Value (NPV) is the probability of points that is classified to non-fire. Sensitivity is the percentage of correct forest fires points whereas specificity is the percentage of correct non-fire points in the training or validation datasets. The global performance of the forest fire models was evaluated using the ROC curve that was constructed based on sensitivity (true positivity rate) and specificity (false negative rate) (Tien Bui et al., 2016a). To quantify the global performance, AUC that varies between 0.5 and 1 was also used. AUC values of 0.5–0.6 indicate insufficient whereas values of 0.6–0.7 indicate poor performance. Models with AUC values of 0.7–0.8 denote moderate performance while models with AUC values 0.8–0.9 denote good performance. Models with AUC values of 0.7–0.8 will indicate very good performance (Peterson et al., 2008). 4. Proposed hybrid artificial intelligence approach for tropical forest fires susceptibility modeling using GIS This section describes the proposed hybrid artificial intelligence model for forest fire susceptibility modeling. The model was constructed through a combination of neural fuzzy system and the PSO algorithm (Fig. 5). In this research, the data were processed and visualized using IDRISI Selva 17.01 and ArcGIS10.2. The neural fuzzy system was implemented via Fuzzy Logic Toolbox (The

In the first step, national topographic maps at scale 1:50,000, Landsat-8 Operational Land Imagery (30 m resolution), land use at scale 1:50,000, climatic data (temperature, wind, and rainfall), and 540 historical forest fires were collected, preprocessed, and compiled to construct a forest fire database for the study area for the year 2013 (Fig. 5). All the factors were converted to a raster format with resolution of 30 m. Since the neural fuzzy model requires input factors are vectors of values, therefore the conversion method proposed by Tien Bui et al. (2012b) was used to assign attribute values to the classes of the ten ignition factors, and then, the max-min normalization was adopted to rescale these attribute values to a range 0.01 to 0.99. The normalization will help to avoid bias when magnitude differences of attribute values are big (Cheng and Hoang, 2015). In forest fire modeling, it is necessity to validate predictive models using independent data, and without validation, the models will have no scientific significance. For this purpose, the forest fire points were also randomly divided into two subsets with the first one (70%, 378 forest fires) were used for training forest file models, whereas the second one (30%, 162 forest fires) were used for model validation (Tien Bui et al., 2016d). Since the forest fire modeling using soft computing methods can be considered as a binary classification, therefore it is necessary to use non-forest points. For this purpose, the same amounts of non-forest points were randomly sampled from no-tree areas with NDVI less than 0 such as water bodies and bare lands. The forest fire points were assigned value of “1”, whereas the non-forest points were assigned value of “0”. Lastly, values of the ten ignition factors were extracted for these points using the sampling tool in ArcGIS 10.2 software to build the training and validation datasets. 4.2. Stage 2: configuration of the neural fuzzy model It is noted that a neural fuzzy with smaller numbers of rules is considered to be better for interpretability (Paiva and Dourado, 2004), therefore the fuzzy c-means clustering method was used to generate the training data into clusters (Tien Bui et al., 2012c). For this purpose, a test was conducted to check numbers of cluster versus RMSE, and 15 clusters were found a good choice for this study. Based on these clusters, an initial neural fuzzy model was generated, and in next step, the premise and consequent parameters for these rules were optimized in the next step using the PSO algorithm. 4.3. Stages 3 and 4: Optimization of neural fuzzy model, fitness evaluation, and stopping criteria The aim of this stage was to find a forest fire model that has the lowest RMSE by searching the best values for the premise and consequent parameters using the PSO algorithm. Since the sample size in the PSO algorithm should have good population diversity, a trial-and-error test (Tien Bui et al., 2016e) was carried out and 45 individuals were the best for the study area. The inertia weight (w) of 0.9 is employed to guarantee a balance between the global exploration and local search (Poli et al., 2007). To measure the fitness of particle positions, RMSE (Eq.7) was employed whereas the stopping criterion of 500 iterations

D. Tien Bui et al. / Agricultural and Forest Meteorology 233 (2017) 32–44

39

Fig. 5. The flow chart of the proposed PSO-NF model with Particle Swarm Optimization for spatial modeling of tropical forest fires.

Table 2 Predictive ability of 10 fire ignition factors using Pearson correlation with a 10-fold cross validation.

Table 3 Performance of the proposed PSO-NF model, the RF model, and the SVM model using the training dataset.

No.

Forest fire related factor

Predictive ability

Standard deviation

No

Evaluation metrics

PSO-NF

RF

SVM

1 2 3 4 5 6 7 8 9 10

NDVI Distance to residence area Distance to road Slope Aspect Elevation Temperature Land use Wind speed Rainfall

0.659 0.281 0.218 0.194 0.132 0.103 0.071 0.059 0.044 0.013

0.008 0.009 0.011 0.015 0.017 0.008 0.008 0.018 0.014 0.009

1 2 3 4 5 6 7 8 9 10 11 12

True positive True negative False positive False negative Positive predictive value (%) Negative predictive value (%) Sensitivity (%) Specificity (%) Overall success rate (%) RMSE Kappa statistic AUC

363 312 66 15 84.6 95.4 96.0 82.5 89.3 0.322 0.786 0.932

345 308 70 33 83.1 90.3 91.3 81.5 86.4 0.327 0.728 0.908

346 306 72 32 82.8 90.5 91.5 81.0 86.2 0.342 0.725 0.894

(Benmiloud, 2012; Tien Bui et al., 2016e) was used to terminate the optimization process. For each generation in the optimization process, these particles fly in a constrained searching space and exchange their experiences to find the position that has the lowest RMSE. By this way, 500 combinations of antecedent and consequent parameters were explored of the forest fire model. The best parameters were derived based on the position with the lowest RMSE of the swarm of all iteration. 4.4. Stages 5: final forest fire susceptibility model Once the optimization process was terminated, the best antecedent and consequent parameters were extracted to build the final forest fire susceptibility model. The goodness-of-fit of the final model was evaluated using the training dataset and four statistical measures (RMSE, MAE, ROC curve, and AUC). In the next step, the prediction power of the model was assessed using the validation dataset and the same statistical measures. The final model was then used to estimate forest fire susceptibility values for all the pixels in the study area, and the forest fire susceptibility map was then generated. 5. Results and discussion In forest fire modeling, predictive ability of the ignition factors should be assessed and factor with non-predictive value should not incorporated into models. This will help to eliminate non-relevant factors that could improve the performance of prediction models (Tien Bui et al., 2016d). For this purpose, the Pearson correlation technique was adopted to evaluate correlations of each factor with the forest fires (Tien Bui et al., 2016b). The result is shown in Table 2. It could be seen that NDVI has the highest predictive ability value

(0.659), followed by distance to residence area (0.281), distance to road (0.218), slope (0.194), aspect (0.132), elevation (0.103), temperature (0.045), land use (0.059), wind speed (0.044), and rainfall (0.013). The finding is reasonable because NDVI is related to tree covers that influence the variability of fuel load, a main factor controlling the fire regime (Allan et al., 2003; Holsinger et al., 2016). The distance to residence area and the distance to road are related to anthropogenic factor that is the main cause of forest fires (Martínez et al., 2009). The finding is in agreement with Le et al. (2014) and Tien Bui et al. (2016b), who concluded that anthropogenic activities were the main cause of vegetation fires in Vietnam. Since no factor reveals non predictive ability value, all factors were incorporated in the modeling process. Fig. 6 shows the final structure of the PSO-NF model for tropical forest fire modeling in this study. It can be seen that 10 forest fire ignition factors were used as input variables whereas the output is forest fire susceptibility index. The model was built with 5 fuzzy rules that consist of 100 premise and 55 consequent parameters (Fig. 6). The best values for these parameters were determined in the training process using PSO. Training iterations versus RMSEs of the proposed PSO-NF model using PSO are shown in Fig. 7. It could be observed that RMSE started reducing from iteration 45 until iteration 483, and at this point RMSE was unchanged. RMSE of the final PSO-NF model is shown in Tables 3 and 4. It could be seen that RMSEs are 0.322 and 0.354 in the training and validation datasets, respectively. These values are lower than standard deviation of 0.5 indicating that the PSO-NF model was trained successfully. The performance of the PSO-NF model was evaluated using the training dataset and the result is shown in Table 3. It could be

40

D. Tien Bui et al. / Agricultural and Forest Meteorology 233 (2017) 32–44

Fig. 6. The structure of the proposed PSO-NF model for modeling of forest fires for the study area.

Fig. 7. Training iteration versus RMSE of the proposed PSO-NF model.

Fig. 8. ROC curves and AUCs of the PSO-NF model using: (a) the training dataset and (b) the validation dataset.

D. Tien Bui et al. / Agricultural and Forest Meteorology 233 (2017) 32–44

41

Fig. 9. Forest fire susceptibility map derived from the PSO-NF model.

Table 4 Prediction power of the proposed PSO-NF model, the RF model, and the SVM model using the validation dataset. No

Evaluation metrics

PSO-NF

RF

SVM

1 2 3 4 5 6 7 8 9 10 11 12

True positive True negative False positive False negative Positive predictive value (%) Negative predictive value (%) Sensitivity (%) Specificity (%) Overall success rate (%) RMSE Kappa statistic AUC

149 129 33 13 81.9 90.8 92.0 79.6 85.8 0.354 0.716 0.916

145 131 31 17 82.4 88.5 89.5 80.9 85.2 0.336 0.707 0.906

143 132 30 19 82.7 87.4 88.3 81.5 84.9 0.358 0.697 0.880

seen that the positive predictive value (PPV) is 84.6% indicating that the probability of classifying pixels to forest fires is 84.4%. The negative predictive value (NPV) is 95.4% indicating that the probability of classifying pixels to non-fire is 95.4%. Sensitivity is 96.0% indicating that 96.0% of the forest fire points are correctly classified, whereas only 82.5% of the non-forest fire points are correctly classified (specificity = 82.5%). Overall success rate is 89.3%, a high accuracy result. Kappa statistic is 0.786 indicating that the proposed model is better than random 78.6%. AUC of the PSO-NF model is 0.932 (Table 3 and Fig. 8a) indicating that the global fit of the model with the training data is 93.2%.

The prediction power of the PSO-NF model was assessed using the validation dataset that was not used during the training phase. The result is shown in Table 4 and Fig. 8b. It could be seen that the overall prediction capability is 91.6% (AUC = 0.916). PPV is 81.9% indicating that the probability of predicting pixels to forest fires is 81.9%. NPV is 90.8% indicating that the probability of predicting pixels to non-forest fires is 90.8%. Sensitivity and specificity of the PSO-NF model are 92.0% and 79.6% indicating that the model correctly predicted forest fires 92.0% and non-forest fires 79.6%. Overall success rate (85.8%) and Kappa statistic (0.716) show a satisfied result. 6. Model comparison and cartographic presentation of tropical forest fire susceptibility map Since this is the first time the PSO-NF model is proposed for tropical forest fire modeling, therefore the usability of the proposed model should be assessed and compared with those derived from benchmark methods. For this purpose, Random Forests (RF) and Support Vector Machine (SVM) that are state-of-the art soft computing methods were selected as the benchmark methods. RF was selected because it outperformed conventional methods for forest fire modeling (Oliveira et al., 2012), whereas SVM is widely accepted efficient method for nonlinear and complex problems (Tien Bui et al., 2016d). RF classifier is an ensemble of individual decision trees that has proven as a powerful supervised learning method for real-world problems (Kühnlein et al., 2014; Rodriguez-Galiano et al., 2012).

42

D. Tien Bui et al. / Agricultural and Forest Meteorology 233 (2017) 32–44

For forest fire modeling in this study, first, bootstrap subsets are generated from the training dataset, and then, each subset is used to build an individual decision tree. It is noted that numbers of features (ignition factors) in these subsets are randomly selected from ten fire ignition factors. In addition, 500 trees is used to ensure stable result as suggested by Ghimire et al. (2012) and Stevens et al. (2015). Finally, the RF model is formed by combining all decision tree classifiers. SVM is a powerful supervised learning method that has been shown to outperform various conventional methods in many fields (Sugumaran et al., 2007; Teodoro, 2015; Teodoro and Araujo, 2016; Tien Bui et al., 2013). For forest fire modeling using SVM in this study, radial basic function (RBF) kernel is used (Hoang and Tien Bui, 2016; Tien Bui et al., 2012a). Because the performance of the SVM model is influenced by the kernel width () and the regularization (C), therefore they should be carefully picked up. For this purpose, the grid-search method was used as suggested by Tien Bui et al. (2012a), with  of 0.395 and C of 12 were the best suited for the study area. The performance of the forest fire models derived from the RF and the SVM is shown in Table 3. It could be seen that the RF model and the SVM model have high performance in the training data. However, the performance of the two models is clearly lower those of the proposed PSO-NF model, in terms of overall success rate, RMSE, Kappa statistic, and AUC (Table 3). The prediction power of the RF model and the SVM model were assessed using the validation dataset. The results (Table 4) show that the two models have high prediction capability. However, although the overall success rate of the two models is almost equal to the PSO-NF model; the global prediction capability of the PSONF model has slightly better in terms of AUC, Kappa statistic, and RMSE (Table 4). From the aforementioned analysis, it could be concluded that the PSO-NF model is the best for the study area. The PSO-NF model was then used to calculate forest fire susceptibility index for all pixels in the study area. In the next step, these susceptibility indices were exported to a raster map using a C++ application developed by the authors. The map (Fig. 9) was cartographically represented by means of six classes (Tien Bui et al., 2016b) such as extremely high (10%, 980.5 km2 ), very high (15%, 1470.8 km2 ), high (15%, 1470.8 km2 ), moderate (20%, 1961.1 km2 ), low (20%, 1961.1 km2 ), and very low (20%, 1961.1 km2 ).

should be used to optimize the parameters for the neural fuzzy model. Since the performance of the proposed model outperforms those derived from the two benchmarks, the RF model and the SVM model, for the current study area, the new model is promising and should be used to assist decision makers in forest fire managements for other regions. In addition, the new model provided a list of five If-Then fuzzy rules; therefore, an early warning system for tropical forest fires could be established. One of the critical problems in forest fire modeling is the determination of the appropriate fire ignition factors. This study selected ten factors based on analysis of the historical forest fires, characteristics of the study area. All the ten factors have predictive values with the forest fires indicating that these factors have been selected and processed successfully. NDVI, the distance to residence area, and the distance to road have the highest predictive values indicating reasonable results. Using the proposed model, a forest fire susceptibility map for the study area was produced. Aerial interpretation of the map shows that areas in Dam Rong, Lam Ha, and Di Linh have high probability of forest fire; therefore these areas should be received higher priority regarding to fire prevention measures. Forest fire danger is generally low for areas in Da Teh and Cat Tien. In fact, these areas belong to the Cat Tien National Park that has been recognized as the world biodiversity conservation by the United Nations Educational, Scientific and Cultural Organization (UNESCO). Therefore, these areas have been protected carefully by the local authority. The main limitation of this research is that only Gaussian membership function was investigated for the PSO-NF model. Therefore, the prediction capability of the forest fire model may be enhanced if other membership functions, such as sigmoidal function and Bell-shaped function, are considered. In addition, the quality of the forest fire model could be improved when other new optimization algorithms are selected for searching the best parameters for the neural fuzzy model. Other possible improvements of this research consist of investigations of new machine learning models to improve the prediction capability of the forest fire model. As final conclusion, the result in this research is useful for forest planning and management in forest fire prone areas.

7. Concluding remarks

Acknowledgement

Forest fire modeling in regional scale is a typical complex and non-linear problem that is difficult to assess and predict. This study proposed and tested a novel hybrid intelligent model, named as Particle Swarm Optimized Neural Fuzzy (PSO-NF), for forest fire modeling at a case study of the Lam Dong province, Central Highland of Vietnam. This province has received many forest fires during the last five years, especially in 2013. The proposed model was established using the neural fuzzy inference system for classification of two classes, forest fire and non-fire. The probability belong to the forest fire class of each pixel is used as fire susceptibility index. In addition, the PSO is incorporated into the proposed model to find the best values for the premise and consequent parameters. Furthermore, a GIS databased for the study area was also constructed to train and validate the proposed model. Experimental results show that the performance of the proposed model was heavily dependent on the premise and consequent parameters. High performance of the model on both the training and validation datasets was obtained indicating that these parameters was successfully picked up by the PSO algorithm, therefore the algorithm is an efficient metaheuristic algorithm that

This research was supported by the Geographic Information System group, Department of Business Administration and Computer Science, University College of Southeast Norway, Bø i Telemark, Norway.

8. Conflict of interest The authors declare that there is no conflict of interest.

References Abraham, A., 2001. Neuro fuzzy systems: state-of-the-art modeling techniques. In: Connectionist Models of Neurons, Learning Processes, and Artificial Intelligence. Springer, pp. 269–276. Allan, G., Johnson, A., Cridland, S., Fitzgerald, N., 2003. Application of NDVI for predicting fuel curing at landscape scales in northern Australia: can remotely sensed data help schedule fire management operations? Int. J. Wildland Fire 12 (4), 299–308. Amatulli, G., Peréz-Cabello, F., de la Riva, J., 2007. Mapping lightning/human-caused wildfires occurrence under ignition point location uncertainty. Ecol. Modell. 200 (3), 321–333. ˜ Arganaraz, J.P., Pizarro, G.G., Zak, M., Landi, M.A., Bellis, L.M., 2015. Human and biophysical drivers of fires in Semiarid Chaco mountains of Central Argentina. Sci. Total Environ. 520, 1–12. Arndt, N., Vacik, H., Koch, V., Arpaci, A., Gossow, H., 2013. Modeling human-caused forest fire ignition for assessing forest fire danger in Austria. iFor.-Biogeosci. For. 6 (6), 315.

D. Tien Bui et al. / Agricultural and Forest Meteorology 233 (2017) 32–44 Arpaci, A., Malowerschnig, B., Sass, O., Vacik, H., 2014. Using multi variate data mining techniques for estimating fire susceptibility of Tyrolean forests. Appl. Geogr. 53, 258–270. Bajocco, S., et al., 2015. Mapping forest fuels through vegetation phenology: the role of coarse-Resolution satellite time-Series. PLoS One 10 (3), e0119811. Benmiloud, T., 2012. Improved adaptive neuro-fuzzy inference system. Neural Comput. Appl. 21 (3), 575–582. Benz, U.C., Hofmann, P., Willhauck, G., Lingenfelder, I., Heynen, M., 2004. Multi-resolution, object-oriented fuzzy analysis of remote sensing data for GIS-ready information. ISPRS J. Photogramm. Remote Sens. 58 (3), 239–258. Bermudez, P.d.Z., Mendes, J., Pereira, J., Turkman, K., Vasconcelos, M., 2009. Spatial and temporal extremes of wildfire sizes in Portugal (1984–2004). Int. J. Wildland Fire 18 (8), 983–991. Bonham-Carter, G.F., 2014. Geographic Information Systems for Geoscientists: Modelling with GIS, 13. Elsevier. Bui, K.-T.T., Tien Bui, D., Zou, J., Doan, C.V., Revhaug, I., 2016. 2016. A novel hybrid artificial intelligent approach based on neural fuzzy inference model and particle swarm optimization for horizontal displacement modeling of hydropower dam. Neural Comp. Appl., http://dx.doi.org/10.1007/s00521-0162666-0. Camp, A., Oliver, C., Hessburg, P., Everett, R., 1997. Predicting late-successional fire refugia pre-dating European settlement in the Wenatchee mountains. For. Ecol. Manage. 95 (1), 63–77. Cheng, M.-Y., Hoang, N.-D., 2015. A Swarm-optimized fuzzy instance-based learning approach for predicting slope collapses in mountain roads. Knowl.-Based Syst. 76, 256–263. Cheng, T., Wang, J., 2008. Integrated spatio-temporal data mining for forest fire prediction. Trans. GIS 12 (5), 591–611. Chuvieco, E., et al., 2010. Development of a framework for fire risk assessment using remote sensing and geographic information system technologies. Ecol. Modell. 221 (1), 46–58. Conedera, M., et al., 2011. Using Monte Carlo simulations to estimate relative fire ignition danger in a low-to-medium fire-prone region. For. Ecol. Manage. 261 (12), 2179–2187. Dan, N.L., Hieu, N.T., Lan, V.T.T., 2014. Drought, desertification Tay Nguyen territory associated with climate change scenarios. Vietnam J. Earth Sci. 35 (4), 310–317. Duarte, L., Teodoro, A.C., Gonc¸alves, J.A., Soares, D., Cunha, M., 2016. Assessing soil erosion risk using RUSLE through a GIS open source desktop and web application. Environ. Monit. Assess. 188 (6), 1–16. Eastaugh, C.S., Hasenauer, H., 2014. Deriving forest fire ignition risk with biogeochemical process modelling. Environ. Model. Softw. 55, 132–142. Eberhart, R.C., Kennedy, J., 1995. A new optimizer using particle swarm theory. In: Proceedings of the Sixth International Symposium on Micro Machine and Human Science, New York, NY, pp. 39–43. Ganteaume, A., et al., 2013. A review of the main driving factors of forest fire ignition over Europe. Environ. Manage. 51 (3), 651–662. Ghimire, B., Rogan, J., Galiano, V.R., Panday, P., Neeti, N., 2012. An evaluation of bagging, boosting, and random forests for land-cover classification in Cape Cod, Massachusetts, USA. GISci. Remote Sens. 49 (5), 623–643. Gillett, N., Weaver, A., Zwiers, F., Flannigan, M., 2004. Detecting the effect of climate change on Canadian forest fires. Geophys. Res. Lett. 31 (18). Gonzalez-Olabarria, J.R., Brotons, L., Gritten, D., Tudela, A., Teres, J.A., 2012. Identifying location and causality of fire ignition hotspots in a Mediterranean region. Int. J. Wildland Fire 21 (7), 905–914. Gordan, B., Armaghani, D.J., Hajihassani, M., Monjezi, M., 2016. Prediction of seismic slope stability through combination of particle swarm optimization and neural network. Eng. Comput. 32 (1), 85–97. Guo, F., et al., 2016. Wildfire ignition in the forests of southeast China: identifying drivers and spatial distribution to predict wildfire likelihood. Appl. Geogr. 66, 12–21. Guttman, N.B., 1998. Comparing the Palmer Drought Index and the Standardized Precipitation Index1. Wiley Online Library. Hargrove, W., 1994. Using EMBYR, a Large-scale Probabilistic Fire Model, to Re-create the Yellowstone Forest Lake Fire. Environmental Research News. Heimann, M., Reichstein, M., 2008. Terrestrial ecosystem carbon dynamics and climate feedbacks. Nature 451 (7176), 289–292. Higuera, P.E., Abatzoglou, J.T., Littell, J.S., Morgan, P., 2015. The changing strength and nature of fire-climate relationships in the northern rocky mountains, USA, 1902–2008. PLoS One 10 (6), e0127563. Hoang, N.-D., Tien Bui, D., 2016. A novel relevance vector machine classifier with cuckoo search optimization for spatial prediction of landslides. J. Comput. Civil Eng., http://dx.doi.org/10.1061/(ASCE)CP. 1943-5487.0000557. Holsinger, L., Parks, S.A., Miller, C., 2016. Weather, fuels, and topography impede wildland fire spread in western US landscapes. For. Ecol. Manage. 380, 59–69. Huebner, K., Lindo, Z., Lechowicz, M., 2012. Post-fire succession of collembolan communities in a northern hardwood forest. Eur. J. Soil Biol. 48, 59–65. Huesca, M., et al., 2009. Assessment of forest fire seasonality using MODIS fire potential: a time series approach. Agric. Forest Meteorol. 149 (11), 1946–1955. Jang, J.S.R., 1993. ANFIS: Adaptive-network-based fuzzy inference system. IEEE Trans. Syst. Man Cybern. 23 (3), 665–685. Jenks, G.F., Caspall, F.C., 1971. Error on choroplethic maps: definition, measurement, reduction. Ann. Assoc. Am. Geogr. 61 (2), 217–244. Kühnlein, M., Appelhans, T., Thies, B., Nauss, T., 2014. Improving the accuracy of rainfall rates from optical satellite sensors with machine learning — a random forests-based approach applied to MSG SEVIRI. Remote Sens. Environ. 141, 129–143.

43

Keane, R.E., et al., 1998. Development of Input Data Layers for the FARSITE Fire Growth Model for the Selway-Bitterroot Wilderness Complex, USA. Kumar, D., Verma, P., 2015. Comparative study of Mamdani & Takagi-Sugeno models for spectrum access in cognitive radio networks. In: 2015 IEEE International Conference on Computational Intelligence and Computing Research (ICCIC), IEEE, pp. 1–5. Le, T.H., et al., 2014. Vegetation fires and air pollution in Vietnam. Environ. Pollut. 195, 267–275. Lentile, L.B., Smith, F.W., Shepperd, W.D., 2006. Influence of topography and forest structure on patterns of mixed severity fire in ponderosa pine forests of the South Dakota Black Hills, USA. Int. J. Wildland Fire 15 (4), 557–566. Li, L., Chen, Y., Yu, X., Liu, R., Huang, C., 2015. Sub-pixel flood inundation mapping from multispectral remotely sensed images based on discrete particle swarm optimization. ISPRS J. Photogramm. Remote Sens. 101, 10–21. Lilly, J.H., 2010. Takagi–Sugeno fuzzy systems. In: Fuzzy Control and Identification., pp. 88–105. Linn, R., Reisner, J., Colman, J.J., Winterkamp, J., 2002. Studying wildfire behavior using FIRETEC. Int. J. Wildland Fire 11 (4), 233–246. Littell, J.S., Peterson, D.L., Riley, K.L., Liu, Y., Luce, C.H., 2016. A review of the relationships between drought and forest fire in the United States. In: Global Change Biology. Loboda, T.V., Csiszar, I.A., 2007. Assessing the risk of ignition in the Russian far east within a modeling framework of fire threat. Ecol. Appl. 17 (3), 791–805. Lopes, A.M.G., Cruz, M.G., Viegas, D.X., 2002. FireStation — an integrated software system for the numerical simulation of fire spread on complex topography. Environ. Model. Softw. 17 (3), 269–285. Martínez, J., Vega-Garcia, C., Chuvieco, E., 2009. Human-caused wildfire risk rating for prevention planning in Spain. J. Environ. Manage. 90 (2), 1241–1252. Massada, A.B., Syphard, A.D., Hawbaker, T.J., Stewart, S.I., Radeloff, V.C., 2011. Effects of ignition location models on the burn patterns of simulated wildfires. Environ. Model. Softw. 26 (5), 583–592. Massada, A.B., Syphard, A.D., Stewart, S.I., Radeloff, V.C., 2013. Wildfire ignition-distribution modelling: a comparative study in the Huron–Manistee National Forest, Michigan, USA. Int. J. Wildland Fire 22 (2), 174–183. Mavsar, R., Cabán, A.G., Varela, E., 2013. The state of development of fire management decision support systems in America and Europe. For. Policy Econ. 29, 45–55. McGrattan, K.B., et al., 2000. Fire Dynamics Simulator–Technical Reference Guide. National Institute of Standards and Technology. Building and Fire Research Laboratory. Ministry of Agriculture, Rural Development of Vietnam, 2016. The Vietnam’s FireWatch System for Online Monitoring and Management of Forest Fires. Ministry of Agriculture and Rural Development of Vietnam Hanoi http://www. kiemlam.org.vn/firewatchvn. Moritz, M.A., et al., 2012. Climate change and disruptions to global fire activity. Ecosphere 3 (6), art49. Nepstad, D.C., Stickler, C.M., Soares-Filho, B., Merry, F., 2008. Interactions among Amazon land use, forests and climate: prospects for a near-term forest tipping point. Philos. Trans. R. Soc. Lond.B: Biol. Sci.s 363 (1498), 1737–1746. North, M.A., 2009. A method for implementing a statistically significant number of data classes in the Jenks algorithm. In: Fuzzy Systems and Knowledge Discovery, 2009. FSKD’09. Sixth International Conference on, IEEE, pp. 35–38. Oliveira, S., Oehler, F., San-Miguel-Ayanz, J., Camia, A., Pereira, J.M.C., 2012. Modeling spatial patterns of fire occurrence in Mediterranean Europe using multiple regression and random forest. For. Ecol. Manage. 275, 117–129. Oliveira, S., Pereira, J.M., San-Miguel-Ayanz, J., Lourenc¸o, L., 2014. Exploring the spatial patterns of fire density in Southern Europe using geographically weighted regression. Appl. Geogr. 51, 143–157. Paiva, R.P., Dourado, A., 2004. Interpretability and learning in neuro-fuzzy systems. Fuzzy Sets Syst. 147 (1), 17–38. Pastor, E., Zárate, L., Planas, E., Arnaldos, J., 2003. Mathematical models and calculation systems for the study of wildland fire behaviour. Prog. Energy Combust. Sci. 29 (2), 139–153. Peterson, A.T., Papes¸, M., Soberón, J., 2008. Rethinking receiver operating characteristic analysis applications in ecological niche modeling. Ecol. Modell. 213 (1), 63–72. Pimont, F., et al., 2016. Modeling fuels and fire effects in 3D: model description and applications. Environ. Model. Softw. 80, 225–244. Podur, J., Martell, D.L., Csillag, F., 2003. Spatial patterns of lightning-caused forest fires in Ontario, 1976–1998. Ecol. Modell. 164 (1), 1–20. Poli, R., Kennedy, J., Blackwell, T., 2007. Particle swarm optimization. Swarm Intell. 1 (1), 33–57. Pourghasemi, H.R., 2015. GIS-based forest fire susceptibility mapping in Iran: a comparison between evidential belief function and binary logistic regression models. Scand. J. For. Res. 31 (3), 80–98. Pourtaghi, Z.S., Pourghasemi, H.R., Aretano, R., Semeraro, T., 2016. Investigation of general indicators influencing on forest fire and its susceptibility modeling using different data mining techniques. Ecol. Indic. 64, 72–84. Rodriguez-Galiano, V.F., Ghimire, B., Rogan, J., Chica-Olmo, M., Rigol-Sanchez, J.P., 2012. An assessment of the effectiveness of a random forest classifier for land-cover classification. ISPRS J. Photogramm. Remote Sens. 67, 93–104. Satir, O., Berberoglu, S., Donmez, C., 2015. Mapping regional forest fire probability using artificial neural network model in a Mediterranean forest ecosystem. In: Geomatics, Natural Hazards and Risk., pp. 1–14.

44

D. Tien Bui et al. / Agricultural and Forest Meteorology 233 (2017) 32–44

Schmidt, D.A., Taylor, A.H., Skinner, C.N., 2008. The influence of fuels treatment and landscape arrangement on simulated fire behavior, Southern Cascade range, California. For. Ecol. Manage. 255 (8), 3170–3184. Short, K., 2014. A spatial database of wildfires in the United States, 1992–2011. Earth Syst. Sci. Data 6 (1), 1–27. Stevens, F.R., Gaughan, A.E., Linard, C., Tatem, A.J., 2015. Disaggregating census data for population mapping using random forests with remotely-sensed and ancillary data. PLoS One 10 (2), e0107042. Sturtevant, B.R., Scheller, R.M., Miranda, B.R., Shinneman, D., Syphard, A., 2009. Simulating dynamic and mixed-severity fire regimes: a process-based fire extension for LANDIS-II. Ecol. Modell. 220 (23), 3380–3393. Sugumaran, V., Muralidharan, V., Ramachandran, K., 2007. Feature selection using decision tree and classification through proximal support vector machine for fault diagnostics of roller bearing. Mech. Syst. Sig. Process. 21 (2), 930–942. Taormina, R., Chau, K.-W., 2015. Data-driven input variable selection for rainfall–runoff modeling using binary-coded particle swarm optimization and extreme learning machines. J. Hydrol. 529, 1617–1632. Teodoro, A.C., Araujo, R., 2016. Comparison of performance of object-based image analysis techniques available in open source software (Spring and Orfeo toolbox/Monteverdi) considering very high spatial resolution data. J. Appl. Remote Sens. 10 (1), 016011. Teodoro, A.C., Duarte, L., 2013. Forest fire risk maps: a GIS open source application–a case study in Norwest of Portugal. Int. J. Geogr. Inf. Sci. 27 (4), 699–720. Teodoro, A., et al. An Integrated and Open Source GIS Environmental Management System for a Protected Area in the South of Portugal 2015; doi:10.1117/12.2193578 (pp. 96440U-96440U-12). Teodoro, A.C., 2015. Applicability of data mining algorithms in the identification of beach features/patterns on high-resolution satellite data. J. Appl. Remote Sens. 9 (1), 095095. The MathWorks, I, 2014. Fuzzy Logic Toolbox User’s Guide R2014b. Tien Bui, D., Pradhan, B., Lofman, O., Revhaug, I., Dick, O.B., 2012a. Application of support vector machines in landslide susceptibility assessment for the Hoa Binh province (Vietnam) with kernel functions analysis, iEMSs 2012 −managing resources of a limited planet. Proceedings of the 6th Biennial Meeting of the International Environmental Modelling and Software Society, 382–389. Tien Bui, D., Pradhan, B., Lofman, O., Revhaug, I., Dick, O.B., 2012b. Landslide susceptibility assessment in the Hoa Binh province of Vietnam: a comparison of the Levenberg-Marquardt and Bayesian regularized neural networks. Geomorphology 171–172 (0), 12–29. Tien Bui, D., Pradhan, B., Lofman, O., Revhaug, I., Dick, O.B., 2012c. Landslide susceptibility mapping at Hoa Binh province (Vietnam) using an adaptive neuro-fuzzy inference system and GIS. Comput. Geosci. 45 (0), 199–211. Tien Bui, D., Pradhan, B., Lofman, O., Revhaug, I., Dick, O., 2013. Regional prediction of landslide hazard using probability analysis of intense rainfall in the Hoa Binh province, Vietnam. Nat. Hazards 66 (2), 707–730.

Tien Bui, D., et al., 2016a. Spatial prediction of rainfall-induced landslides for the lao cai area (Vietnam) using a novel hybrid intelligent approach of least squares support vector machines inference model and artificial bee colony optimization. In: Landslides., http://dx.doi.org/10.1007/s10346-016-0711-9. Tien Bui, D., Le, K.-T., Nguyen, V., Le, H., Revhaug, I., 2016b. Tropical forest fire susceptibility mapping at the cat Ba national park area, hai phong city, vietnam, using GIS-Based kernel logistic regression. Remote Sens. 8 (4), 347. Tien Bui, D., Nguyen, Q.-P., Hoang, N.-D., Klempe, H., 2016c. A novel fuzzy K-Nearest neighbor inference model with differential evolution for spatial prediction of rainfall-Induced shallow landslides in a tropical hilly area using GIS. In: Landslides., http://dx.doi.org/10.1007/s10346-016-0708-4. Tien Bui, D., Pham, T.B., Nguyen, Q.-P., Hoang, N.-D., 2016d. Spatial prediction of rainfall-induced shallow landslides using hybrid integration approach of least squares support vector machines and differential evolution optimization: a case study in central vietnam. Int. J. Digit. Earth 9 (11), 1077–1097, http://dx. doi.org/10.1080/17538947.2016.1169561. Tien Bui, D., et al., 2016e. Hybrid artificial intelligence approach based on neural fuzzy inference model and metaheuristic optimization for flood susceptibility modelling in a high-frequency tropical cyclone area using GIS. J. Hydrol. 540, 317–330. Vasconcelos, M.P.d., Silva, S., Tome, M., Alvim, M., Pereira, J.C., 2001. Spatial prediction of fire ignition probabilities: comparing logistic regression and neural networks. Photogramm. Eng. Remote Sens. 67 (1), 73–81. Verde, J., Zêzere, J., 2010. Assessment and validation of wildfire susceptibility and hazard in Portugal. Nat. Hazards Earth Syst. Sci. 10 (3), 485–497. Vilar, L., Woolford, D.G., Martell, D.L., Martín, M.P., 2010. A model for predicting human-caused wildfire occurrence in the region of Madrid, Spain. Int. J. Wildland Fire 19 (3), 325–337. Vu, T.D., et al., 2013. Land Use, Forest Cover Change and Historical GHG Emission from 1990 to 2010, Lam Dong province, Vietnam, USAID, LEAF Hanoi. Wittenberg, L., Malkinson, D., 2009. Spatio-temporal perspectives of forest fires regimes in a maturing Mediterranean mixed pine landscape. Eur. J. For. Res. 128 (3), 297–304. Wotton, B., Martell, D., Logan, K., 2003. Climate change and people-caused forest fire occurrence in Ontario. Clim. Change 60 (3), 275–295. Yi, K., et al., 2013. Long-Term satellite detection of post-Fire vegetation trends in boreal forests of China. Remote Sens. 5 (12), 6938–6957. Zaitchik, B.F., Santanello, J.A., Kumar, S.V., Peters-Lidard, C.D., 2013. Representation of soil moisture feedbacks during drought in NASA unified WRF (NU-WRF). J. Hydrometeorol. 14 (1), 360–367. Zald, H.S., et al., 2014. Influence of lidar, Landsat imagery, disturbance history, plot location accuracy, and plot size on accuracy of imputation maps of forest composition and structure. Remote Sens. Environ. 143, 26–38. Zhou, C., Yin, K., Cao, Y., Ahmed, B., 2016. Application of time series analysis and PSO–SVM model in predicting the Bazimen landslide in the Three Gorges Reservoir, China. Eng. Geol. 204, 108–120.