Next Article in Journal
Quantitative Evaluation of the Integrity of Natural Ecosystems and Anthropogenic Impacts in Shennongjia National Park, China
Previous Article in Journal
A Re-Inventory after 12 Years—Increase in Red Wood Ant Nests and Woodpecker Cavities in Nests in the West Eifel Volcanic Field despite Climatic Changes
 
 
Search for Articles:
Title / Keyword
Author / Affiliation / Email
Journal
Article Type
 
 
Advanced Search
 
Section
Special Issue
Volume
Issue
Number
Page
 
Journals
Forests
Volume 14
Issue 5
10.3390/f14050986
forests-logo
Submit to this Journal Review for this Journal Propose a Special Issue
► Article Menu

Article Menu

share Share announcement Help format_quote Cite question_answer Discuss in SciProfiles thumb_up
...
Endorse textsms
...
Comment
first_page
settings
Order Article Reprints
Font Type:
Arial Georgia Verdana
Font Size:
Aa Aa Aa
Line Spacing:
Column Width:
Background:
Article

Improving Wildfire Danger Assessment Using Time Series Features of Weather and Fuel in the Great Xing’an Mountain Region, China

by
Zili Wang
,
Binbin He
*,
Rui Chen
and
Chunquan Fan
School of Resources and Environment, University of Electronic Science and Technology of China, Chengdu 611731, China
*
Author to whom correspondence should be addressed.
Forests 2023, 14(5), 986; https://doi.org/10.3390/f14050986
Submission received: 11 April 2023 / Revised: 27 April 2023 / Accepted: 28 April 2023 / Published: 10 May 2023
(This article belongs to the Section Forest Inventory, Modeling and Remote Sensing)
Browse Figures
Versions Notes

Abstract

:
Wildfires directly threaten the safety of life and property. Predicting wildfires with a model driven by wildfire danger factors can significantly reduce losses. Weather conditions continuously influence the drying rate of fuel as well as the occurrence probability and danger degree of wildfires. Previous studies have paid little attention to the continuous effects of weather and fuel on wildfires. This study improved the accuracy and effect of wildfire danger assessment using the time series features of weather and fuel. First, the time series features of weather and fuel factors within the 16 days before the fire were analyzed. Then, four feature groups were selected—feature group without time series values, feature group with time series values, feature group with Tsfresh transformation of time series values, and feature group with gradient and cumulative transformation of time series values—and three models were trained, respectively: random forest, balanced random forest, and extreme gradient boosting. The results showed that the f1-score of all feature groups with time series values (0.93) increased by 0.15, on average, compared with those without time series values (0.78) for the three models. The feature group with gradient and cumulative features had a more stable prediction accuracy and a more accurate wildfire danger map. The results suggest that using the appropriate time series features of weather and fuel can help improve the precision and effect of the wildfire danger assessment model.

1. Introduction

Although wildfire is an important forest regeneration agent and has several positive effects on the natural succession of forest ecosystems [1], it also poses risks to economic development, social security, and ecological protection. There are more than 100,000 wildfires around the world annually, burning more than 3.85 million hectares of forests each year [2]. The occurrence and spread of wildfire are largely dependent on weather and fuel factors. With global warming, the occurrence frequency and burning area of wildfires will continue to increase, especially in the global boreal coniferous forest [3]. It was predicted that the incidence and burning area of wildfires in northern China would increase sharply during the 21st century [4,5]. Therefore, wildfire danger assessment is imperative. Accurate quantification of wildfire occurrence probability is essential for wildfire danger assessment, which can provide information including time, location, and quantity of wildfire. In addition, accurate quantification can support wildfire management and prevention decisions [1,6].
To cope with the threat of wildfires, many countries, including Canada [7], the United States [8], Australia [9], Europe [10], and China [11,12], have built their wildfire danger assessment systems to improve the efficiency and effectiveness of wildfire management [13]. Among them, the Canadian Forest Fire Rating System (CFFDRS) and American National Fire Rating System (NFDRS) are the most established. The Canadian Fire Weather Index (FWI) system [7], which is a component of CFFDRS and has been widely used around the world [14,15,16], predicted three types of water content of fuel based on four weather variables collected at noon, and then simulated the wildfire danger index using those fuel water content codes and wind speed. Chinese scholars used the FWI system and topography to build statistical models, which have been successfully applied to forest fire prevention in the Great Xing’an Mountain region [14]. Australia has also developed its wildfire danger system based on the FWI system, the McArthur Forest Fire Danger Meter (FFDM) [9], with input elements including precipitation, air temperature, air relative humidity, wind speed, and long-term drought index. NFDRS [8] assessed the effect of topography and ignition source on wildfire danger besides weather and fuel. Therefore, NFDRS are essentially physical models developed based on combustion principles and laboratory experiments. These wildfire danger assessment systems all emphasize the importance of weather and fuel and employ several factors, including weather, fuel, topography, and ignition probability, to assess wildfire danger thoroughly [14,17].
According to the wildfire triangle [18], weather, fuel, and topography are the three most important factors in assessing wildfire danger. Weather and fuel play a driving role in the occurrence of wildfires. The continuous high temperatures, drought, and high wind speed accelerate the loss of water from the air, soil, and fuel [19] and then increase the vulnerability and danger of wildfires [20]. Weather variables, such as near surface temperature, relative humidity, wind speed, and precipitation, are often used to assess the wildfire danger [3,4,5,6,7,8,9,10,11,12] and can be collected by meteorological stations [21] or simulated by the climate reanalysis model [22,23]. Fuel variables, such as fuel moisture content (FMC), fuel load (FL), and fuel type (FT), can be obtained by weather simulation [7,24] or remote sensing retrieval [25,26,27] on a large scale and can be used as an assessment variable for wildfire danger [7,11,27,28]. Moreover, Nathan Phelps [29] investigated studies on wildfire danger assessment from the previous 20 years and proposed that a comprehensive consideration of terrain, ignition source, and fuel type is more helpful to improve the wildfire danger assessment effect.
Topography affects the distribution of wildfires in a fixed way. It affects the type and distribution of vegetation [30], solar radiation [31], and microclimate [32] and further affects the possibility of wildfires [33]. Studies have shown that the topographic roughness index (TRI) can effectively measure the effects of micro-topographic features on forest disturbance (vegetation growth, wind speed, ignition possibility, etc.) [34]. In addition, elevation, aspect, and slope are the most important topography variables and are often used in assessing wildfire danger [35,36].
Ignition sources can be roughly divided into human-caused ignition and lightning-caused ignition, randomly occurring and affecting the distribution of wildfires. The spatial frequency of human activities, such as the distance to roads, residential areas, rivers, and railways [11,36,37], have generally been used to account for the human-caused ignition when modeling wildfire danger. The lightning density, intensity, and the percentage of positive lightning can be collected [38,39,40] and used [41,42,43] to account for the lightning-caused ignition. However, because of the difficulties of acquiring and predicting lightning data, the lightning variables are still challenging to apply to predict future wildfire danger [44,45].
Wildfires are caused not only by current weather and fuel conditions but also by the historical trend of these driving factors. Since soil and plants have a certain lag in response to drought stress [29], drought and gale over a period of time are the real culprits of wildfire [46,47,48]. However, most studies on wildfire danger assessment have focused on the influences of weather and fuel conditions on the day of fire occurrence [3,4,5,6,7,8,9,10,15,16,17,18,19,20]. A few researchers have studied the effects of individual weather or fuel factors over a period of time on wildfire. Chen et al. [47] incorporated the number of days without rain (0–6 mm) into the model and explored the effects of a period of no rain on fire occurrence. Ryan Lagerquist et al. [49] used the drought code (DC) as an indicator to investigate the impact of short-term drought on major fires (burning area > 100 ha). Luo et al. [50] examined the changing moisture content of the forest canopy before the major wildfires and concluded that its cliff descent might be a signal of a major wildfire occurrence. However, these studies simply assessed wildfire danger using individual factors, with no comprehensive and complete consideration of the changes in weather and fuel conditions over a period of time before wildfire occurrence.
Therefore, this study aims to improve the accuracy and effect of wildfire danger assessment using the time series features of multiple weather and fuel factors. The specific objectives of this study are: (1) to explore the changing characteristics of weather and fuel factors before the fire, including time windows and time series features; (2) to explore whether the time series features can improve the prediction accuracy of models and the effect on wildfire danger assessment; and (3) to compare the effects of different time series feature groups and select the appropriate time series features. Therefore, our overarching goal is to understand the changing characteristics of weather and fuel factors before the fire, which can help us select the appropriate time series features and carry out wildfire danger assessment work.

2. Materials and Methods

2.1. Study Area

The study area is the Great Xing’an Mountain region (Figure 1), which is located in northeast China and is the highest latitude administrative region of China. This region spans the Heilongjiang Province and the Nei Monggol Autonomous Region, close to Russia in the north and bounded by the Heilongjiang River. It covers a total area of about 8.35 million ha between 121°12′~127°00′ E and 50°11′~53°33′ N. This region has a cold temperate monsoon climate, with little rain in spring and autumn and snow cover in winter. The average annual precipitation is about 460 mm, and the average annual temperature is −2.6 °C [51]. There are many mountains and hills in this region, with a latitude from 200 m to 1000 m above sea level and a slope of 0 to 7°. This region is suitable for plant growth, with a high forest coverage rate (79.83%). There is a lack of diversity of tree species in this region, mainly consisting of pine, white birch, and larch, among which larch is the most widely distributed, accounting for about 65% of the total area [52].
Wildfires are common in the Great Xing’an Mountain region. In this region, 846 wildfires were reported between 2000 and 2020, burning 1.02 million hectares of forest. Because the Great Xing’an Mountains prevent water vapor from entering the northern region, the burning area in the north (about 0.78 million ha) is much higher than the south (about 0.23 million ha). Lightning strikes are the primary source of wildfires, accounting for 96% of wildfire occurrences and 99% of burning area. Wildfires often occur in spring (May and June) and autumn (September and October), with the spring accounting for 66% of wildfire occurrences and 78% of burning area. The above information was obtained from the wildfire records of the Great Xing’an Mountain region from 2000 to 2020, which was provided by the Forest Fire Prevention Department of the Greater Xing’an Mountains Forestry Group Corporation. The wildfires in the study area all occurred in forest areas, so this study used wildfire to represent wildfires that occur in forests.

2.2. Data Preparation

2.2.1. Dependent Variables

The historical fire data contain 846 fire events in the Great Xing’an Mountain region from 2000 to 2020, which were extracted from the MODIS Burned Area (BA) product MCD64A1 Collection 6 (500 m). We confirmed the efficiency of each fire event by consulting the pertinent authorities. The information extracted included location, burning area, date of occurrence, date of extinction, and cause of fire. From 2015 to 2020, 293 fires with about 0.18 million ha burning area were used to assess the accuracy of model. The remaining 653 fires were used to train the model.
Non-fire points were selected to improve the model’s ability to distinguish between high and low wildfire danger areas. During the fire season, the difference between fire points and non-fire points is not significant, and using fire points alone as a dataset may not highlight this difference. The selection method of non-fire points is as follows.
(1)
First, a circular buffer zone was established with the fire location as the center and the burning area as the buffer area.
(2)
Then, the same buffer zone was established on each day between the 16th day before the fire occurrence and the 16th day after the fire extinction.
(3)
Third, non-fire points were chosen at random from the forest area outside the buffer zone, with a non-fire to fire ratio of 10:1. The forest area was extracted from the MODIS land cover product MCD12Q1 and classified into forest area according to categories 1–8 of the IGBP classification scheme.
Figure 2 shows the distribution of fire points, buffer zone of each fire point, and selected non-fire points from 31 May to 3 June 2018.

2.2.2. Explanatory Variables

Topography, weather, fuel, and ignition source are the most important factors in assessing wildfire danger. However, ignition source factors were not introduced in this study for the following reasons. (1) Wildfires in the study area are mostly caused by lightning (>90%), so we did not add human variables. (2) The ground lightning detection network in this study area was built and has provided lightning observation data since 2015 [53] but is unable to provide data over a sufficient time period (2000–2020) for this study. (3) The available remote sensing lightning data do not cover this study area [39] or are unable to provide a long enough time span [40]. (4) The uncertainty in location and intensity of lightning prediction data [44,45] makes it impossible to use it to forecast wildfire danger. In conclusion, the lightning data in the study area can only be obtained since 2015, which do not satisfy the requirements of this study. Studies [42,43,53,54] have shown that lightning wildfires almost always occur in high wildfire danger forest areas, where the weather and fuel are dry. It also makes sense to use the weather and fuel conditions before the fire to represent the high wildfire danger.
The remaining three types of data were used in this study: topography, weather, and fuel. All the data (Table 1) were used to extract the feature values of fire and non-fire points.

Topographic Variables

The topography variables selected in this study include elevation, slope, aspect, and TRI. TRI was calculated by the center elevation minus the mean of the surrounding elevation (Equation (1)). Aspect was divided into eight categories according to the standards in Table 2. The above topographic variables were extracted from GMTED2010 (the Global Multi-resolution Terrain Elevation Data 2010). GMTED2010 is a digital elevation model developed by the National Geospatial-Intelligence Agency and the United States Geological Survey, providing a maximum spatial resolution of 7.5 arcseconds [55].
TRI = e − ∑ i = 1 8 e i / 8
where e is the elevation of the center point of the 3 × 3 window, and e i is the elevation of the surrounding points.
Table
Table 1. Origin data.
Table 1. Origin data.
Variable TypeVariable NameVariable CodeSourceResolution
TopographyElevation (m)ElevationGMTED2010
[55]		0.001°, _
Slope (0–90°)Slope
AspectAspect
Topographic roughness indexTRI
WeatherDaily maximum temperature (°C)Tem_maxERA5-Land [23]0.1°, 1 day
Daily average relative humidity (0%–100%)RH_avg
Daily precipitation (cm)Pre
Daily average wind speed (m/s)Win_avg
FuelFine fuel moisture codeFFMCFire Danger Indices Historical Data [7,24]0.25°, 1 day
Duff moisture codeDMC
Drought codeDC
Initial buildup indexISI
Fire weather indexFWI
Dominant tree speciesTSthe third Forestry Survey of China0.001°, _
Leaf area indexLAIMODIS MOD15A2H500 m, 8 days
Table
Table 2. Aspect classification criteria.
Table 2. Aspect classification criteria.
AspectAzimuth (°)
North337.5°~360°, 0°~22.5°
Northeast22.5°~67.5°
East67.5°~112.5°
Southeast112.5°~157.5°
South157.5°~202.5°
Southwest202.5°~247.5°
West247.5°~292.5°
Northwest292.5°~337.5°

Weather Variables

The weather variables selected in this study include temperature, relative humidity, wind speed, and precipitation. The weather data were extracted from the ERA5-Land reanalysis dataset, which is a global land atmosphere reanalysis dataset provided by the European Centre for Medium-Range Weather Forecasts (ECMWF) [23]. It has a temporal resolution of 1 h and a spatial resolution of 0.1°.
Weather variables include 2 m near surface temperature, 10 m near surface u-component of wind, 10 m near surface v-component of wind, 2 m near surface dewpoint temperature, and total precipitation, where the 2 m (or 10 m) near surface means the distance from the ground when the variable is measured. Although the ERA5-Land dataset does not provide surface relative humidity data directly, it can be fitted with surface temperature and surface dewpoint temperature. We used Tetens’s [56] empirical formula (Equation (2)) to calculate vapor pressure ( e 1 ) and saturation vapor pressure ( e 2 ) by surface temperature and surface dewpoint temperature. Then, surface relative humidity ( r h ) can be expressed by the ratio of e 1 and e 2 (Equation (3)).
e s = 6.1078 e [ 17.2693882 ( T − 237.16 ) T − 35.86 ]
r h = e 1 e 2 × 100 %

Fuel Variables

Five indexes of the FWI system [7] were selected to represent the moisture content and wildfire danger of forest fuel, including three fuel moisture codes that represent moisture contents of different fuel layers in mature coniferous forests: the Fine Fuel Moisture Code (FFMC), Duff Moisture Code (DMC), and Drought Code (DC), and two wildfire danger indexes that indicate fire behavior and danger degree: the Initial Spread Index (ISI) and Fire Weather Index (FWI). These index data were extracted from the Fire Danger Indices Historical Data [24], which were calculated from the historical simulation weather dataset provided by the ERA5 reanalysis data. The data have a spatial resolution of 0.25° and a temporal resolution of 1 day.
The leaf area index (LAI) was selected to represent the forest fuel load and was extracted from the MODIS product MOD15A2H, which provides a spatial resolution of 500 m and a temporal resolution of 8 days. A previous study [57] showed that the LAI is an excellent characteristic of vegetation growth in forest area. The dominant tree species were selected to represent the forest fuel type and extracted from the third Forestry Survey of China, which was conducted in 2018 and provides map spot data of 0.001°. The dominant tree species data were numerically encoded before input into the model.

2.3. Feature Groups

2.3.1. Time Series Features of Weather and Fuel

We extracted the weather and fuel values within the 16 days before the date of each fire and non-fire point and obtained the daily mean values. By comparing the dynamic changes within the 16 days, we determined the appropriate time window and time series features of each factor.
Weather factors include daily maximum temperature (Tem_max), daily mean relative humidity (RH_avg), daily precipitation (Pre), and daily mean wind speed (Win_avg). Figure 3 shows their changes within the 16 days before the fire and the non-fire points. We found that Tem_max and RH_avg started to change significantly on the 8th day before the fire. With the Tem_max increasing, RH_avg decreased continuously and reached the lowest on the fire occurrence day. The precipitation of the fire was lower than that of the non-fire within the 16 days, indicating that the vegetation and soil were continuously influenced by dry weather conditions before the fire. The Win_avg of the fire fluctuated considerably within the 16 days but was mostly higher than that of the non-fire.
Figure 4 shows the changes of fuel factors (FFMC, DMC, DC, ISI, and FWI) within the 16 days before the fire and the non-fire points. We found that all fuel factors of the fire were significantly higher than the non-fire, which was also consistent with the definition in the FWI system: the larger of the fire weather index, the higher the wildfire danger [7]. FFMC, ISI, and FWI had similar change patterns in that they started to increase significantly on the 8th day before the fire, and they represent the short-term effects of weather and fuel on wildfire danger. DC and DMC had similar change patterns and started to increase slowly on the 16th day before the fire. They represent the long-term wetting and drying process of deep fuel under the influences of weather conditions.
In conclusion, temperature, relative humidity, FFMC, ISI, and FWI had a relatively rapid impact on wildfire danger, and their changes started from the 8th day before the fire; therefore, their time windows were set to 8 days. Precipitation, wind speed, DMC, and DC required a longer period of time to influence wildfire danger, so their time windows were set to 16 days. The fire curve of each factor, in contrast to the non-fire curve, exhibited significant volatility, as demonstrated by the gradient rise or decrease and the general high or low of each factor. Therefore, we selected the gradient and cumulative feature types to show the changes of each factor. Among them, the average values of air temperature for the fire and non-fire were identical, so their cumulative features were deleted. Moreover, precipitation and wind speed showed unobvious trends, so their gradient features were deleted. As a result, we selected the time window and feature types of each factor (Table 3), which was the basis for generating the feature groups below.

2.3.2. Four Feature Groups

Tsfresh (https://tsfresh.readthedocs.io/en/latest/, accessed on 1 January 2023) is a temporal transformation tool that can be used to calculate a variety of time series features automatically. According to the above analysis, we selected one feature group without time series features and three feature groups with different time series features using the Tsfresh tool. They were shown in Figure 5 and labelled as follows: feature group without time series values (Group I), feature group with time series values (Group II), feature group with Tsfresh transformation of time series values (Group III), and feature group with gradient and cumulative transformation of time series values (Group IV). All four groups shared the same topography factors but differed in terms of weather and fuel factors. Group I contained the weather and fuel values on the day of the fire. Group II contained the daily weather and fuel values in the time windows (Table 3) before the fire. Group III maintained the Tsfresh transformation features based on the time series values. Based on the feature types provided in Table 3, Group IV selected the gradient and cumulative features from Group III.

2.4. Model

To explore effects of the four feature groups, we selected three machine learning models commonly used in wildfire danger assessment: random forest, balanced random forest, and extreme gradient boosting.

2.4.1. Random Forest

The random forest [58] model uses a bagging algorithm to integrate multiple decision trees, each of which is constructed by sampling datasets according to an entropy reduction strategy (Equation (4)). When predictions are required, random forest computes and votes on the predictions of each tree. The universality and robustness of random forest have made scholars inclined to use this model for wildfire danger assessment [29,59,60]. Wu et al. [60] used random forest to estimate the relative effects of factors on wildfire occurrence density in the Great Xing’an Mountain region.
G i n i ( D ) = 1 − ∑ i = 1 K ( | C i | | D | ) 2
where G i n i ( D ) refers to the entropy of a given set D , and the smaller of G i n i , the lower of the entropy. C i refers the ith subset of set D .
Training the random forest model, we first explored the number of trees in the forest (n_estimater), which has the greatest influence on the model. Then, we used grid search to determine other parameters, including the maximum tree depth (max_depth), minimum samples of leaf nodes (min_samples_leaf), and minimum samples of node splitting (min_samples_split).

2.4.2. Balanced Random Forest

The balanced random forest [61] model is comparable to random forest. The distinction is that random forest selects the training dataset of each tree at random, while balanced random forest selects positive and negative sample in a fixed proportion. Therefore, balanced random forest performs better with higher prediction accuracy for few classes. Studies have shown that models targeting imbalanced datasets are more conducive to improving the prediction accuracy of the fire class [29]. A previous study also showed that balanced random forest is more suitable for assessing wildfire danger in the Great Xing’an Mountain region [62].
The balanced random forest model needs to determine the parameter of a positive and negative sample ratio. When there are more positive samples, the recall will be higher, but the precision will be lower. Therefore, we tested six balanced random forest models with different parameters using Group II and determined the best parameter value (fire point: non-fire point = 1:4) by referring to the model with the highest f1-score value (Table 4). The training process of remaining parameters was completely consistent with the random forest model.

2.4.3. Extreme Gradient Boosting

Extreme gradient boosting (XGBoost) [63] is an enhanced boosting algorithm that integrates cart regression trees into a more powerful classifier. XGBoost uses the number of leaf nodes as L1 regular term and the weight of leaf nodes as L2 regular term (Equation (6)), which can reduce the model variance, simplify the learned model, and prevent overfitting.
L ( φ ) = ∑ i L ( y ^ i , y i ) + ∑ i ( γ N + 1 2 λ ‖ ω ‖ 2 )
where L ( φ ) refers to the objective function of XGBoost, γ N refers to the L1 regular term, and λ ‖ ω ‖ 2 refers to the L2 regular term.
We used grid search to adjust the XGBoost parameters, including n_estimators, max_depth, min_samples_leaf, and min_samples_split.

2.4.4. Evaluation Indexes

To train each model, we adjusted the parameters to obtain the best f1-score value on the training set using the 10-fold cross-validation method and the f1-score as the evaluation index. We used metrics on the test set to measure the quality of the four feature groups, including the accuracy, precision, recall, f1-score, ROC-AUC, and PR-AUC. The f1-score is the primary metric used to assess the overall accuracy of each model, which combines accuracy and recall (Equation (6)). In wildfire danger assessment, we hope the model can maximize recall while maintaining high precision.
f 1 _ score = 2 ∗ P r e c i s i o n ∗ R e c a l l P r e c i s i o n + R e c a l l

3. Results

3.1. Evaluation Results

Table 5 shows the evaluation results for the four feature groups with the three models on the test set. Each row represents the evaluation result of a feature group and a model combination.
Comparing the evaluation results of all feature groups, all feature groups with time series features (Group II, III, and IV) had significantly higher evaluation values than those without time series features (Group I) on all models. In terms of f1-score, Group I received 0.779, on average, for the three models, which was significantly lower than that of Group II (0.928), Group III (0.924), and Group IV (0.935). On the XGBoost model, Group I obtained the highest score of the three models, but the score was also significantly lower than Group II (0.937), Group III (0.925), and Group IV (0.938).
We compared the evaluation results of Group II, III, and IV, which contained different time series features. Group III obtained similar evaluation values to Group II. The f1-score of Group II (0.928) was slightly lower than Group III (0.930) on the random forest model, while the f1-score of Group II (0.937) was slightly higher than Group III (0.925) on XGBoost. Moreover, the enormous number of features in Group III (1356 features) contributed to the high cost of finding the optimization model. Among the four feature groups, Group IV maintained the highest f1-score values on all three models (f1_score_rf = 0.938, f1_score_xgb = 0.938, f1_score_brf = 0.929).
Comparing the evaluation results of the three models, the random forest model had a relatively stable performance. In wildfire danger maps, random forest maintained a range that did not vary with the season, which was not presented by XGBoost and the balanced random forest model. Therefore, random forest was used in the following wildfire danger maps.

3.2. Wildfire Danger Map

In order to comprehensively compare performance of the four feature groups, we mapped wildfire danger maps using random forest and verified if fires broke out in high wildfire danger areas. Figure 6 shows the wildfire danger maps of the four feature groups in the Great Xing’an Mountain region on 31 May 2018.
According to the wildfire danger map of Group I, although there is a high wildfire danger on fire points in the northern, the wildfire danger is lower (low and very low) on fire points in most areas of the south and east. All the wildfire danger maps of Group II, III, and IV have higher wildfire danger (medium, high, and extremely high) on fire points. Especially in most of the southern and eastern regions, the corresponding situation of high wildfire danger and fire point are significantly better than the map of Group I.
Furthermore, we discovered that Group IV had extremely high wildfire danger (very high) areas in the western region, while Group II, III had high wildfire danger (median and high) in the same area. On 1 June 2018, this region, in the western part of the Great Xing’an Mountain region bordering the Nei Monggol Autonomous Region, was hit by the largest wildfire since 2015, burning 5100 ha of forest. Therefore, the extremely high wildfire danger expressed by Group IV is an excellent indication of the real wildfire danger situation on 31 May 2018.

4. Discussion

4.1. Changing Characteristics of Weather and Fuel

Weather and fuel are the primary driving factors determining wildfire occurrence [4,5,59,64,65], but the changing characteristics of these factors before the fire is incomplete. This study innovatively explored the changing characteristic of weather and fuel factors in a time window before the fire.
We analyzed the changes of weather and fuel factors within 16 days before the fire. The results showed that temperature, relative humidity, FFMC, ISI, and FWI had a rapid impact on wildfire danger, and they started to rapidly change on the eighth day before the fire. DMC and DC started to increase slowly on the 16th day before the fire, and precipitation remained below the average level within the 16 days. Among these, the temperature before the fire had a faster rising trend but was generally lower than that of the non-fire. The reason could be that fires in the Great Xing’an Mountain region are more common in spring when the ice and snow in forest have just melted and the temperature is low but rising rapidly. Studies have shown that the decrease of relative humidity is an important factor that stresses the water loss of surface fuels and the sudden rise of wildfire danger [64,66]. It was also observed in our results that the decreasing process of relative humidity was consistent with the increasing process of FFMC, ISI, and FWI. Moreover, the influences of these drivers on wildfire danger started from the eighth day before the fire, which has not been studied previously. Precipitation, DMC, and DC had slow effects on wildfire, and they require a long-term process (16 days and more) to increase the probability and severity of wildfire. This is consistent with the findings of Ryan Lagerquist et al. [49], who believed that the continuous reduction of deep fuel moisture content caused by a long-term lack of rain is a necessary condition for severe wildfire occurrence.

4.2. Using Timing Series Features to Assessing Wildfire Danger

At present, most studies have built wildfire danger assessment models using weather and fuel factors on the day of the fire [60,65,66,67]. However, we believe that it is partial to represent the natural increase of wildfire danger using values on the day of the fire. If only the value on the day of the fire is used, then the model will be unable to effectively identify the extreme wildfire danger induced by a sudden increase of temperature and a sudden decrease of air humidity and fuel moisture content. If the rainfall is low on the fire day because the preceding week’s plentiful rainfall has just finished, then the wildfire danger model will be unable to accurately represent the impact of continuous rainfall on wildfire danger. Some studies have used weather and fuel factor in a period for wildfire danger assessment and have achieved better effects [47,49]. Our results also showed that the evaluation results and wildfire danger maps of all groups containing time series features were significantly better than that of those without time series features. Therefore, it is essential to incorporate the time series features of weather and fuel factors into the wildfire danger assessment model.
In our study, we discovered that the feature group with gradient and cumulative transformation of time series values can better express the important information contained in the time series values. This could be because Tsfresh extracted a large number of invalid features which are involved in the model and play a negative function. The gradient and cumulative features generated from the naturally changing characteristics can emphasize the importance of effective features in the wildfire danger assessment model. This view is supported by a study [66], in which the authors believed that it was possible to avoid the reduction of important features to select features before building a wildfire danger assessment model. Only Group IV’s wildfire danger map showed the real situation of extremely high wildfire danger in the western Great Xing’an Mountain region before the major fire. The ability to analyze extremely high wildfire danger areas is essential for wildfire danger assessment work, which can improve people’s fire prevention vigilance in this area, provide advance warning, and limit the occurrence and spread of severe wildfire [68].
Combined with forecasted weather and fuel data, the wildfire danger assessment model can predict future wildfire danger and be used as a decision-making tool. Forecasted weather data can be obtained from the Global Forecast System (GFS), which is a weather forecast model produced by the National Centers for Environmental Prediction (NCEP) that provides hourly weather forecast data for the next 120 h. Forecasted fuel data can be calculated from GFS data because the code of FWI system only depends on the input of weather variables [69]. Studies [1,70] have shown that it is reliable to calculate FWI variables using weather raster data.

4.3. Limitations and Future Work

Some limitations remain in this study. First, different vegetation, geographical locations, and topography result in different soil and fuel water holding capacity, which, in turn, influences how wildfires react to weather and fuel. It has been demonstrated that it is possible to assess wildfire danger by dividing different geographical regions [50]. The uniformity of topography, weather, and vegetation in the Great Xing’an Mountain region enabled us to acquire similar wildfire danger change characteristics. However, different climatic and topographical conditions could be found over a considerable area of the world. It is necessary to establish their change characteristics of weather and fuel factors and carry out wildfire danger assessment work by dividing different geographical regions.
In addition, the effects of ignition sources on wildfire danger are not quantified in this study for several reasons. Ignition factors, such as density of population, GDP [11,36], lightning density [36,41,42], high tension lines [37], and graveyards [11], may play an essential role in the occurrence of wildfires. Figure 7 also demonstrates the high degree of coincidence between lightning strikes and fire points, which suggests that the addition of lightning data appears to effectively improve wildfire danger assessment. With the growing availability of lightning data in the Great Xing’an Mountain region, it is considerable to precisely quantify historical data of ignition factors and build more accurate wildfire danger models based on them. In addition, in forest areas with more human-caused wildfires, it is necessary to fully consider human activities as the ignition source.

5. Conclusions

This study discussed the effects of time series features of weather and fuel factors on wildfire danger assessment over the Great Xing’an Mountain region using all fire events records from 2000 to 2020. First, the time series features of weather factors (daily maximum temperature, daily average relative humidity, daily precipitation, daily average wind speed) and fuel factors (FFMC, DMC, DC, ISI, FWI) within 16 days before the fire were analyzed. Based on this, we established four feature groups and trained three models, respectively. The results showed that the evaluation values of all feature groups with time series features (f1-score > 0.9) were significantly higher than those of the feature group without time series values (F1-score < 0.8). The feature group with gradient and cumulative transformation of time series values had the best evaluation values across all models, and its wildfire danger map was the most accurate. Therefore, the main findings in this study are: (1) The weather and fuel factors showed significant gradient and cumulative characteristics over 8 or 16 days before the fire. (2) Adding the time series features of weather and fuel factors can effectively improve the accuracy of the wildfire danger assessment model. (3) Selecting appropriate time series features according to the changing characteristic can better reflect the high wildfire danger. Going forward, we expect to apply this method to forest areas with more complex climatic and topographic conditions in the global area and use ignition source variables to improve the effect of wildfire danger assessment.

Author Contributions

Conceptualization, Z.W. and B.H.; methodology, Z.W.; software, Z.W.; validation, Z.W.; formal analysis, Z.W.; investigation, Z.W.; resources, Z.W.; data curation, Z.W.; writing—original draft preparation, Z.W.; writing—review and editing, Z.W., R.C. and C.F.; visualization, Z.W.; supervision, B.H.; project administration, B.H.; funding acquisition, B.H. All authors have read and agreed to the published version of the manuscript.

Funding

This work was supported by the National Natural Science Foundation of China (Contract No. U20A2090).

Data Availability Statement

Publicly available datasets were analyzed in this study. This data can be found here: [ERA5-Land: https://cds.climate.copernicus.eu/cdsapp#!/dataset/reanalysis-era5-land?tab=overview, accessed on 1 January 2023; Fire Danger Indices Historical Data: https://cds.climate.copernicus.eu/cdsapp#!/dataset/cems-fire-historical?tab=overview, accessed on 1 January 2023].

Acknowledgments

The authors are grateful to the dataset providers: (1) the Land Processes Distributed Active Archive Center (LPDAAC) at the U.S. Geological Survey (USGS) Earth Resources Observation and Science Center (EROS, http://lpdaac.usgs.gov, accessed on 1 January 2023) for providing MODIS product MCD64A1, MCD12Q1, and MOD15A2H; (2) the United States Geological Survey, USGS National Geospatial-Intelligence Agency, NGA for providing the Global Multi-resolution Terrain Elevation Data 2010 and GMTED2010 digital elevation model (DEM); (3) the European Center for Medium-Range Weather Forecasts (ECMWF, http://apps.ecmwf.int, accessed on 1 January 2023) for providing ERA5-land dataset and Fire Danger Indices Historical Data dataset; (4) the Forest Fire Prevention Department of the Great Xing’an Mountains Forestry Group Corporation for providing suggestions.

Conflicts of Interest

The authors declare no conflict of interest.

References

  1. Johnston, L.M.; Wang, X.; Erni, S.; Taylor, S.W.; McFayden, C.B.; Oliver, J.A.; Stockdale, C.; Christianson, A.; Boulanger, Y.; Gauthier, S.; et al. Wildland fire risk research in Canada. Environ. Rev. 2020, 28, 164–186. [Google Scholar] [CrossRef]
  2. Schultz, M.G.; Heil, A.; Hoelzemann, J.J.; Spessa, A.; Thonicke, K.; Goldammer, J.G.; Held, A.C.; Pereira, J.M.C.; van het Bolscher, M. Global wildland fire emissions from 1960 to 2000. Glob. Biogeochem. Cycles 2008, 22, 198–215. [Google Scholar] [CrossRef]
  3. Jones, M.W.; Abatzoglou, J.T.; Veraverbeke, S.; Andela, N.; Lasslop, G.; Forkel, M.; Smith, A.J.P.; Burton, C.; Betts, R.A.; van der Werf, G.R.; et al. Global and Regional Trends and Drivers of Fire Under Climate Change. Rev. Geophys. 2022, 60, 108–129. [Google Scholar] [CrossRef]
  4. Tian, X.R.; Shu, L.; Zhao, F.; Wang, M.; McRae, D.J. Future impacts of climate change on forest fire danger in northeastern China. J. For. Res. 2011, 22, 437–446. [Google Scholar] [CrossRef]
  5. Liu, Z.; Yang, J.; Chang, Y.; Weisberg, P.J.; He, H.S. Spatial patterns and drivers of fire occurrence and its future trend under climate change in a boreal forest of Northeast China. Glob. Change Biol. 2012, 18, 2041–2056. [Google Scholar] [CrossRef]
  6. Costafreda-Aumedes, S.; Comas, C.; Vega-Garcia, C. Human-caused fire occurrence modelling in perspective: A review. Int. J. Wildland Fire 2017, 26, 983–998. [Google Scholar] [CrossRef]
  7. Wagner, C.E. Development and Structure of the Canadian Forest Fire Weather Index System; Canadian Forestry Service: Ottawa, ON, Canada, 1987.
  8. Cohen, J.D.; Deeming, J.E. The National Fire-Danger Rating System: Basic Equations; US Department of Agriculture, Forest Service, Pacific Southwest Forest and Range Experiment Station: Albany, CA, USA, 1985.
  9. McArthur, G.A. Fire Behavior Characteristics of the Longford Fire; Department of National Development, Forestry and Timber Bureau: Canberra, Australia, 1967.
  10. Pearce, H.; Majorhazi, K. Application of fire behaviour to fire danger and wildfire threat modelling in New Zealand. In Proceedings of the 3rd International Wildland Conference, Sydney, Australia, 3–6 October 2003. [Google Scholar]
  11. Chen, R.; He, B.; Quan, X.; Lai, X.; Fan, C. Improving Wildfire Probability Modeling by Integrating Dynamic-Step Weather Variables over Northwestern Sichuan, China. Int. J. Disaster Risk Sci. 2023, 14, 313–325. [Google Scholar] [CrossRef]
  12. Tian, X.R.; Douglas, J.; Mcrae; Zhang, Y.H. Review of forest fire risk classification forecasting systems. World For. Res. 2006, 19, 8. [Google Scholar]
  13. Martell, D.L. Forest fire management. In Handbook of Operations Research in Natural Resources; Springer: Boston, MA, USA, 2007; Volume 99. [Google Scholar]
  14. Tian, X.R.; Cui, W.; Shu, L.-F.; Zong, X. Effects of Climate Change on Burn Probability of Forests in Daxing’anling. Forests 2019, 10, 611. [Google Scholar] [CrossRef]
  15. Varela, V.; Vlachogiannis, D.; Sfetsos, A.; Politi, N.; Karozis, S. Methodology for the Study of Near-Future Changes of Fire Weather Patterns with Emphasis on Archaeological and Protected Touristic Areas in Greece. Forests 2020, 11, 1168. [Google Scholar] [CrossRef]
  16. Fernandes, P.M. Variation in the Canadian Fire Weather Index Thresholds for Increasingly Larger Fires in Portugal. Forests 2019, 10, 838. [Google Scholar] [CrossRef]
  17. Taylor Stephen, W.; Alexander Martin, E. Science, technology, and human factors in fire danger rating: The Canadian experience. Int. J. Wildland Fire 2006, 15, 121–135. [Google Scholar] [CrossRef]
  18. Pyne, S.J.; Andrews, P.L.; Laven, R. Introduction to Wildland Fire, 2nd ed.; John Wiley & Sons: New York, NY, USA, 1996. [Google Scholar]
  19. Hartley, I.P.; Hill, T.C.; Chadburn, S.E.; Hugelius, G. Temperature effects on carbon storage are controlled by soil stabilisation capacities. Nat. Commun. 2021, 12, 6713. [Google Scholar] [CrossRef] [PubMed]
  20. Hou, X.; Orth, R. Observational evidence of wildfire-promoting soil moisture anomalies. Sci. Rep. 2020, 10, 11008. [Google Scholar] [CrossRef]
  21. Abatzoglou, J.T. Development of gridded surface meteorological data for ecological applications and modelling. Int. J. Climatol. 2013, 27, 121–131. [Google Scholar] [CrossRef]
  22. Lawson, B.D.; Armitage, O.B. Weather Guide for the Canadian Forest Fire Danger Rating System; National Resources Canada: Edmonton, AB, Canada, 2008.
  23. Sabater, J.M. ERA5-Land Hourly Data from 1981 to Present; Copernicus Climate Change Service (C3S) Climate Data Store (CDS): The Hague, The Netherlands, 2019. [Google Scholar] [CrossRef]
  24. Vitolo, C.; Di Giuseppe, F.; Barnard, C.; Coughlan, R.; San-Miguel-Ayanz, J.; Liberta, G.; Krzeminski, B. ERA5-based global meteorological wildfire danger maps. Sci. Data 2020, 7, 216. [Google Scholar] [CrossRef]
  25. Quan, X.; He, B.; Yebra, M.; Yin, C.; Liao, Z.; Li, X. Retrieval of forest fuel moisture content using a coupled radiative transfer model. Environ. Model. Softw. 2017, 95, 290–302. [Google Scholar] [CrossRef]
  26. Quan, X.; Li, Y.; He, B.; Cary, G.J.; Lai, G. Application of Landsat ETM+ and OLI data for foliage fuel load monitoring using radiative transfer model and machine learning method. IEEE J. Sel. Top. Appl. Earth Obs. Remote Sens. 2021, 14, 5100–5110. [Google Scholar] [CrossRef]
  27. Quan, X.; Xie, Q.; He, B.; Luo, K.; Liu, X. Integrating remotely sensed fuel variables into wildfre danger assessment for China. Int. J. Wildland Fire 2021, 30, 807–821. [Google Scholar] [CrossRef]
  28. Trang, P.T.; Andrew, M.E.; Chu, T.; Enright, N.J. Forest fire and its key drivers in the tropical forests of northern Vietnam. Int. J. Wildland Fire 2022, 31, 213–229. [Google Scholar] [CrossRef]
  29. Phelps, N.; Woolford, D.G. Guidelines for effective evaluation and comparison of wildland fire occurrence prediction models. Int. J. Wildland Fire 2021, 30, 225–240. [Google Scholar] [CrossRef]
  30. De Frenne, P.; Lenoir, J.; Luoto, M.; Scheffers, B.R.; Zellweger, F.; Aalto, J.; Hylander, K. Topography and edge effects are more important than elevation as drivers of vegetation patterns in a neotropical montane forest. J. Veg. Sci. 2014, 25, 724–733. [Google Scholar] [CrossRef]
  31. Zhang, X.R.; Guan, D.J.; Mou, F.Y.; Dong, K. Calculation and Analysis of Annual Total Solar Radiation in Complicated Topography. Adv. Mater. Res. 2012, 512, 47–50. [Google Scholar] [CrossRef]
  32. De Frenne, P.; Lenoir, J.; Luoto, M.; Scheffers, B.R.; Zellweger, F.; Aalto, J.; Hylander, K. Forest microclimates and climate change: Importance, drivers and future research agenda. Glob. Change Biol. 2021, 5, 2279–2297. [Google Scholar] [CrossRef]
  33. Maingi, J.; Henry, M. Factors influencing wildfire occurrence and distribution in eastern Kentucky, USA. Int. J. Wildland Fire 2007, 16, 23–33. [Google Scholar] [CrossRef]
  34. Stambaugh, M.C.; Guyette, R.P. Predicting spatio-temporal variability in fire return intervals using a topographic roughness index. For. Ecol. Manag. 2008, 254, 463–473. [Google Scholar] [CrossRef]
  35. Jaafari, A.; Gholami, D.M.; Zenner, E.K. A Bayesian modeling of wildfre probability in the Zagros Mountains, Iran. Ecol. Inform. 2017, 39, 32–44. [Google Scholar] [CrossRef]
  36. Nadeem, K.; Taylor, S.W.; Woolford, D.G.; Dean, C.B. Mesoscale spatiotemporal predictive models of daily human and lightning-caused wildland fire occurrence in British Columbia. Int. J. Wildland Fire 2020, 29, 11–27. [Google Scholar] [CrossRef]
  37. Vilar, L.; Woolford, D.G.; Martell, D.L.; Martin, M.P. A model for predicting human-caused wildfire occurrence in the region of Madrid, Spain. Int. J. Wildland Fire 2010, 19, 325–337. [Google Scholar] [CrossRef]
  38. Rodger, C.J.; Brundell, J.B.; Holzworth, R.H.; Lay, E.H. Growing Detection Efficiency of the Worldwide Lightning Location Network. Am. Inst. Phys. 2009, 1118, 15–20. [Google Scholar] [CrossRef]
  39. Lang, T. NRT Lightning Imaging Sensor (LIS) on International Space Station (ISS) Backgrounds [Indicate Subset Used]; NASA Global Hydrometeorology Resource Center DAAC: Huntsville, AL, USA, 2022. [CrossRef]
  40. Blakeslee, R.J. Lightning Imaging Sensor (LIS) on TRMM Backgrounds [Indicate Subset Used]; NASA Global Hydrometeorology Resource Center DAAC: Huntsville, AL, USA, 2010. [CrossRef]
  41. Burrows W, R.; King, P.; Lewis P, J.; Kochtubajda, B.; Snyder, V. Lightning occurrence patterns over Canada and adjacent United States from lightning detection network observations. Atmosphere 2002, 40, 59–80. [Google Scholar] [CrossRef]
  42. Wotton B, M.; Martell D, L. A lightning fire occurrence model for Ontario. Can. J. For. Res. 2005, 35, 1389–1401. [Google Scholar] [CrossRef]
  43. Krawchuk, M.A.; Cumming, S.G.; Flannigan, M.D. Predicted changes in fire weather suggest increases in lightning fire initiation and future area burned in the mixedwood boreal forest. Clim. Change 2009, 92, 83–97. [Google Scholar] [CrossRef]
  44. Coughlan, R.; Di Giuseppe, F.; Vitolo, C.; Barnard, C.; Lopez, P.; Drusch, M. Using machine learning to predict fire-ignition occurrences from lightning forecasts. Meteorol. Appl. 2021, 28, e1973. [Google Scholar] [CrossRef]
  45. Hessilt, T.D.; Abatzoglou, J.T.; Chen, Y.; Randerson, J.T.; Scholten, R.C.; van der Werf, G.; Veraverbeke, S. Future increases in lightning ignition efficiency and wildfire occurrence expected from drier fuels in boreal forest ecosystems of western North America. Environ. Res. Lett. 2022, 17, 054008. [Google Scholar] [CrossRef]
  46. Jolly, W.M.; Cochrane, M.A.; Freeborn, P.H.; Holden, Z.A.; Brown, T.J.; Williamson, G.J.; Bowman, D.M.J.S. Climate-induced variations in global wildfire danger from 1979 to 2013. Nat. Commun. 2015, 6, 7537. [Google Scholar] [CrossRef]
  47. Chen, F.; Fan, Z.; Niu, S.; Zheng, J. The Influence of Precipitation and Consecutive Dry Days on Burned Areas in Yunnan Province, Southwestern China. Adv. Meteorol. 2014, 2014, 748923. [Google Scholar] [CrossRef]
  48. Sharples, J.J.; McRae, R.H.D.; Weber, R.O. Wind characteristics over complex terrain with implications for bushfire risk management. Environ. Model. Softw. 2010, 25, 1099–1120. [Google Scholar] [CrossRef]
  49. Lagerquist, R.; Flannigan, M.D.; Wang, X.; Marshall, G.A. Automated prediction of extreme fire weather from synoptic patterns in northern Alberta, Canada. Can. J. For. Res. 2017, 47, 1175–1183. [Google Scholar] [CrossRef]
  50. Luo, K.W.; Quan, X.W.; He, B.B.; Yebra, M. Efects of live fuel moisture content on wildfre occurrence in fre-prone regions over southwest China. Forests 2019, 10, 887. [Google Scholar] [CrossRef]
  51. Du, Y.; Zhang, D. Spatial and temporal distribution of land surface temperature and its influencing factors in Greater Khingan Mountains from 2001 to 2019. For. Eng. 2020, 36, 10. (In Chinese) [Google Scholar]
  52. Xu, H. Forest in Great Xing’an Mountains of China; Science Press: Beijing, China, 1998. (In Chinese) [Google Scholar]
  53. Han, H.L. Design of Lightning Monitoring Location and Lightning Fire Warning System in the Northern Great Xingan’s Mountain; Beijing University of Posts and Telecommunications: Beijing, China, 2012. (In Chinese) [Google Scholar]
  54. Menezes, L.S.; Oliveira, A.M.D.; Santos, F.; Russo, A.; de Souza, R.A.; Roque, F.O.; Libonati, R. Lightning patterns in the Pantanal: Untangling natural and anthropogenic-induced wildfires. Sci. Total Environ. 2022, 820, 153021. [Google Scholar] [CrossRef] [PubMed]
  55. Danielson, J.J.; Gesch, D.B. Global Multi-Resolution Terrain Elevation Data 2010 (GMTED2010); U.S. Geological Survey Open: Reston, VA, USA, 2011.
  56. Coulson, K.L. Characteristics of the radiation emerging from the top of a rayleigh atmosphere—I: Intensity and polarization. Planet. Space Sci. 1959, 1, 265–276. [Google Scholar] [CrossRef]
  57. Fang, H.L.; Baret, F.; Plummer, S.; Schaepman-Strub, G. An Overview of Global Leaf Area Index (LAI): Methods, Products, Validation, and Applications. Rev. Geophys. 2019, 57, 739–799. [Google Scholar] [CrossRef]
  58. Cutler, A.; Cutler, D.; Stevens, J. Random Forests. Mach. Learn. 2001, 45, 157–176. [Google Scholar] [CrossRef]
  59. Jaafari, A.; Pourghasemi, H.R. Factors Influencing Regional-Scale Wildfire Probability in Iran. Spat. Model. GIS R Earth Environ. Sci. 2019, 28, 607–619. [Google Scholar] [CrossRef]
  60. Wu, Z.; He, H.S.; Yang, J.; Liu, Z.; Liang, Y. Relative effects of climatic and local factors on fire occurrence in boreal forest landscapes of northeastern China. Sci. Total Environ. 2014, 493, 472–480. [Google Scholar] [CrossRef]
  61. O’Brien, R.; Ishwaran, H. A random forests quantile classifier for class imbalanced data. Pattern Recognit 2019, 90, 232–249. [Google Scholar] [CrossRef]
  62. Wang, Z.; He, B.; Lai, X. Balanced Random Forest Model is More Suitable for Wildfire Risk Assessment. In Proceedings of the IGARSS 2022 IEEE International Geoscience and Remote Sensing Symposium, Kuala Lumpur, Malaysia, 17–22 July 2022; pp. 3596–3599. [Google Scholar] [CrossRef]
  63. Chen, T.; Guestrin, C. XGBoost: A Scalable Tree Boosting System. In Proceedings of the 22nd ACM SIGKDD International Conference on Knowledge Discovery and Data Mining, San Francisco, CA, USA, 13–17 August 2016. [Google Scholar] [CrossRef]
  64. Varela, V.; Vlachogiannis, D.; Sfetsos, A.; Karozis, S.; Politi, N.; Giroud, F. Projection of Forest Fire Danger due to Climate Change in the French Mediterranean Region. Sustainability 2019, 11, 4284. [Google Scholar] [CrossRef]
  65. Zong, X.; Tian, X.; Yin, Y. Impacts of Climate Change on Wildfires in Central Asia. Forests 2020, 11, 802. [Google Scholar] [CrossRef]
  66. Ma, W.; Feng, Z.; Cheng, Z.; Chen, S.; Wang, F. Identifying Forest Fire Driving Factors and Related Impacts in China Using Random Forest Algorithm. Forests 2020, 11, 507. [Google Scholar] [CrossRef]
  67. Taylor, S.W.; Woolford, D.G.; Dean, C.B.; Martell, D.L. Wildfire Prediction to Inform Fire Management: Statistical Science Challenges. Stat. Sci. 2013, 28, 586–615. [Google Scholar] [CrossRef]
  68. Zumbrunnen, T.; Pezzatti, G.B.; Menéndez, P.; Bugmann, H.; Bürgi, M.; Conedera, M. Weather and human impacts on forest fires: 100 years of fire history in two climatic regions of Switzerland. For. Ecol. Manag. 2011, 261, 2188–2199. [Google Scholar] [CrossRef]
  69. Wang, Y. Updated Source Code for Calculating Fire Danger Indices in the Canadian Forest Fire Weather Index System; Natural Resources Canada, Canadian Forest Service, Northern Forestry Centre: Edmonton, AB, Canada, 2015; 26p. [Google Scholar]
  70. Tatli, H.; Türke, M. Climatological evaluation of Haines forest fire weather index over the Mediterranean Basin. Meteorol. Appl. 2014, 21, 545–552. [Google Scholar] [CrossRef]
Forests 14 00986 g001 550
Figure 1. Location of the study area and fire points from 2000 to 2020.
Figure 1. Location of the study area and fire points from 2000 to 2020.
Forests 14 00986 g001
Forests 14 00986 g002 550
Figure 2. The distribution of fire points, the buffer zone of each fire point, and selected non-fire points from 31 May to 3 June 2018. Forest area were extracted from the MODIS land cover product MCD12Q1 in 2010 and classified into forest areas according to categories 1–8 of the IGBP classification scheme.
Figure 2. The distribution of fire points, the buffer zone of each fire point, and selected non-fire points from 31 May to 3 June 2018. Forest area were extracted from the MODIS land cover product MCD12Q1 in 2010 and classified into forest areas according to categories 1–8 of the IGBP classification scheme.
Forests 14 00986 g002
Forests 14 00986 g003 550
Figure 3. Time series changes of weather factors within 16 days before the fire and the non-fire.
Figure 3. Time series changes of weather factors within 16 days before the fire and the non-fire.
Forests 14 00986 g003
Forests 14 00986 g004 550
Figure 4. Time series changes of fuel factors within 16 days before the fire and the non-fire.
Figure 4. Time series changes of fuel factors within 16 days before the fire and the non-fire.
Forests 14 00986 g004
Forests 14 00986 g005 550
Figure 5. Composition of the four feature groups.
Figure 5. Composition of the four feature groups.
Forests 14 00986 g005
Forests 14 00986 g006 550
Figure 6. Wildfire danger maps of the four feature groups with random forest on 31 May 2018. The wildfire danger map is rendered by five levels, with the probability of the fire predicted by the model were divided as follows: 0~0.2 is very low, 0.2~0.4 is low, 0.4~0.6 is medium, 0.6~0.8 is high, and 0.8~1.0 is very high. There were 33 wildfires with a total burning area of 22,160 ha from 31 May to 3 June 2018, which is the continuous date range with the largest total burning area and the highest frequency of wildfires in the test set (2015 to 2020).
Figure 6. Wildfire danger maps of the four feature groups with random forest on 31 May 2018. The wildfire danger map is rendered by five levels, with the probability of the fire predicted by the model were divided as follows: 0~0.2 is very low, 0.2~0.4 is low, 0.4~0.6 is medium, 0.6~0.8 is high, and 0.8~1.0 is very high. There were 33 wildfires with a total burning area of 22,160 ha from 31 May to 3 June 2018, which is the continuous date range with the largest total burning area and the highest frequency of wildfires in the test set (2015 to 2020).
Forests 14 00986 g006
Forests 14 00986 g007 550
Figure 7. Location of lightning points and wildfire points in the Great Xing’an Mountain region from 31 May to 3 June 2018. Each black dot represents a detected lightning strike.
Figure 7. Location of lightning points and wildfire points in the Great Xing’an Mountain region from 31 May to 3 June 2018. Each black dot represents a detected lightning strike.
Forests 14 00986 g007
Table
Table 3. Time windows and feature types of weather and fuel factors.
Table 3. Time windows and feature types of weather and fuel factors.
Variable TypeVariable CodeVariable ExplanationTime
Window		Selected Feature Type
WeatherTem_maxDaily maximum temperature extracted from ERA5-Land surface air temperature.8 daysGradient
RH_avgDaily average relative humidity extracted from ERA5-Land surface air temperature and surface dewpoint temperature.8 daysGradient, Cumulative
PreDaily precipitation extracted from ERA5-Land total precipitation.16 daysCumulative
Win_avgDaily average wind speed extracted from ERA5-Land surface u-wind and v-wind.16 daysCumulative
FuelFFMCFine fuel moisture code represents the moisture content of litter and other cured fine fuels.8 daysGradient, Cumulative
DMCDuff moisture code represents the average moisture content of loosely compacted organic layers of moderate depth.16 daysGradient, Cumulative
DCDrought code represents the average moisture content of deep, compact organic layers.16 daysGradient, Cumulative
ISIInitial buildup index represents the expected rate of fire spread.8 daysGradient, Cumulative
FWIFire weather index is a numeric rating of fire intensity.8 daysGradient, Cumulative
Table
Table 4. Choice of positive and negative sample ratio in the balanced random forest model.
Table 4. Choice of positive and negative sample ratio in the balanced random forest model.
Fire Points: Non-Fire PointsRecallPrecisionF1-Score
1:10.9330.7500.831
1:20.9150.8770.895
1:30.9040.9310.917
1:40.8960.9520.923
1:50.8870.9600.922
1:60.8680.9610.905
Table
Table 5. Evaluation results of the four feature groups on the three models.
Table 5. Evaluation results of the four feature groups on the three models.
ModelFeature GroupRecallPrecisionROC-AUCPR-AUCF1-Score
Random Forestgroup I0.7330.8310.8520.8040.779
group II0.8730.9900.9830.9600.928
group III0.8720.9960.9860.9570.930
group IV0.8860.9960.9860.9640.938
XGBoostgroup I0.7450.8530.8600.8170.795
group II0.8920.9870.9920.9670.937
group III0.8860.9680.9940.9710.925
group IV0.8890.9920.9980.9750.938
Balanced Random Forestgroup I0.8010.7310.9090.7980.764
group II0.8890.9480.9910.9560.918
group III0.8930.9260.9900.9590.909
group IV0.9110.9480.9960.9760.929
Disclaimer/Publisher’s Note: The statements, opinions and data contained in all publications are solely those of the individual author(s) and contributor(s) and not of MDPI and/or the editor(s). MDPI and/or the editor(s) disclaim responsibility for any injury to people or property resulting from any ideas, methods, instructions or products referred to in the content.

Share and Cite

MDPI and ACS Style

Wang, Z.; He, B.; Chen, R.; Fan, C. Improving Wildfire Danger Assessment Using Time Series Features of Weather and Fuel in the Great Xing’an Mountain Region, China. Forests 2023, 14, 986. https://doi.org/10.3390/f14050986

AMA Style

Wang Z, He B, Chen R, Fan C. Improving Wildfire Danger Assessment Using Time Series Features of Weather and Fuel in the Great Xing’an Mountain Region, China. Forests. 2023; 14(5):986. https://doi.org/10.3390/f14050986

Chicago/Turabian Style

Wang, Zili, Binbin He, Rui Chen, and Chunquan Fan. 2023. "Improving Wildfire Danger Assessment Using Time Series Features of Weather and Fuel in the Great Xing’an Mountain Region, China" Forests 14, no. 5: 986. https://doi.org/10.3390/f14050986

Note that from the first issue of 2016, this journal uses article numbers instead of page numbers. See further details here.

Article Metrics

Back to TopTop