Influence of terrain on lightning-ignited fires and prediction of firefighting factors based on NRBO-XGBoost in the Greater Khingan Mountains of Inner Mongolia

doi:10.11755/j.issn.1006-7639-2026-02-0273

Abstract

Abstract:

To deepen the understanding of the topographic influence mechanism of lightning-ignited fires in the Greater Khingan Range forest area of Inner Mongolia, identify the key driving factors of fire suppression demand, and provide a scientific basis for the monitoring and early warning of lightning-ignited fires as well as the optimal allocation of fire suppression resources in the forest area, based on historical lightning-ignited fire data and high-precision topographic data from 2015 to 2024 in the key state-owned forest area of the Greater Khingan Range of Inner Mongolia, this study systematically analyzed the spatial distribution characteristics of lightning-ignited fire points across multiple topographic factors including elevation, slope gradient, slope aspect and topographic relief. The XGBoost (eXtreme Gradient Boosting) algorithm optimized by the Newton-Raphson-Based Optimizer (NRBO) combined with the SHAP (SHapley Additive exPlanations) method was adopted to model and analyze fire suppression demand, and the contribution degrees of burned area and various topographic factors to the number of fire suppression personnel were revealed. The results show that lightning-ignited fire points in the study area present a gradient characteristic of “concentrated at medium-low elevations and sparse at high/low elevations” in terms of elevation; 83.06% of lightning-ignited fires occur in the medium-low elevation zone of 500-1 000 m. The regions with the highest frequency of lightning-ignited fire points are concentrated between 52.5°N-53.0°N and 120.5°E-122.5°E, while the absolute high-value areas of burned area appear at more southerly latitudes. Lightning-ignited fire points are mainly distributed in medium-slope areas of 5°-35°,and the slope range of 2° to 5° accounts for 37.83% of the total burned area. The medium-relief terrain with a relief value of 75 to 200 m serves as the “core interval” and “optimal” environment for lightning-ignited fires, contributing to 70.97% of such fires. Sunny slopes such as the south, southwest, and southeast slopes are high-risk areas for lightning-ignited fires, accounting for a relatively high proportion of the total fire occurrences. Burned area is the dominant factor in predicting the number of fire suppression personnel, and its contribution is significantly higher than that of topographic factors such as elevation, relief, slope gradient, and slope aspect. The distribution of the model’s prediction errors showed a significant leptokurtic pattern, with the peak closely around the zero error line. The R² value of the model’s prediction on the test set reached 0.723 9, and the prediction interval coverage probability reached 84.7%.

Key words: lightning-ignited fires, topography, fire suppression, contribution degree, the Greater Khingan Mountains

摘要：

为了深化对内蒙古大兴安岭林区雷击火地形影响机制的认识，明确扑火需求的关键驱动因素，基于内蒙古大兴安岭重点国有林区2015—2024年雷击火灾历史数据和高精度地形数据，分析雷击火点在高程、坡度、坡向及地形起伏度等多维地形因子上的空间分布特征，采用基于牛顿-拉夫逊优化器（Newton-Raphson-Based Optimizer，NRBO）优化的XGBoost（eXtreme Gradient Boosting）算法结合SHAP（SHapley Additive exPlanations）方法进行扑火需求建模和分析，揭示过火面积及各地形因子对扑火人数影响的贡献度。结果表明：内蒙古大兴安岭林区雷击火点在高程上呈现“中低海拔集中、高/低海拔稀疏”的梯度特征，83.06%的雷击火灾发生在500~1 000 m的中低海拔区，雷击火点频次最高的区域集中在52.5°N—53.0°N、120.5°E—122.5°E，而过火面积的绝对高值区则出现在更偏南的纬度；雷击火点集中在5°~35°的中坡度区域，[2°，5°）坡度区间过火面积占比高达37.83%；70.97%的雷击火灾发生在[75，200）m的中等起伏地形区域；南坡、西南坡、东南坡等阳坡是雷击火的高风险区；过火面积是预测扑火人数的主导因素，其贡献度明显高于高程、起伏度、坡度和坡向等地形因子，模型预测误差分布呈现显著的“尖峰”形态，且峰值围绕在零误差线附近，模型的测试集预测R²值达0.723 9，且预测区间覆盖概率达84.7%。

关键词: 雷击火, 地形, 扑火, 贡献度, 大兴安岭

CLC Number:

P446

LIU Xiaodong, HUANG Xingang, LI Gang, WANG Guosheng. Influence of terrain on lightning-ignited fires and prediction of firefighting factors based on NRBO-XGBoost in the Greater Khingan Mountains of Inner Mongolia[J]. Journal of Arid Meteorology, 2026, 44(2): 273-284.

刘晓东, 黄薪钢, 李钢, 王国胜. 内蒙古大兴安岭地形对雷击火影响及基于NRBO-XGBoost的扑火因素预测[J]. 干旱气象, 2026, 44(2): 273-284.

Figures/Tables 23

Fig.1 Topography of the study area and distribution of surrounding meteorological observation stations

Tab.1 Classification of topographic factors in key state-owned forest areas of the Greater Khingan Mountains

地形因子	分级
地形因子	1	2	3	4	5	6
高程/m	[200，500）	[500，1 000）	[1 000，1 500）	[1 500，2 000）
坡度/（°）	[0，0.5）	[0.5，2）	[2，5）	[5，15）	[15，35）	[35，55）
地形起伏度/m	[0，20）	[20，75）	[75，200）	[200，600）

Fig.2 Spatial distribution of lightning-ignited fire points at different elevations

Tab.2 The proportion of lightning-ignited fire occurrences and burned areas in different elevation intervals

高程/m	面积占比	雷击火灾次数占比	过火面积占比
[200，500）	13.80	2.69	4.02
[500，1 000）	68.73	83.06	76.34
[1 000，1 500）	17.32	14.25	19.64
[1 500，2 000）	0.15	0	0

Fig.3 The variation of lightning-ignited fire occurrences and burned areas with elevation

Fig.4 The heat maps of lightning-ignited fire occurrences (a, Unit: times) and burned areas (b, Unit: hm2)

Fig.5 The spatial distribution of lightning-ignited fire points at different slopes

Tab.3 The proportion of lightning-ignited fire occurrences and burned areas in different slope intervals

坡度/（°）	面积占比	雷击火灾次数占比	过火面积占比
[0，0.5）	4.05	2.69	4.49
[0.5，2.0）	11.97	7.26	2.01
[2.0，5.0）	21.58	17.47	37.83
[5.0，15.0）	52.87	56.72	51.37
[15.0，35.0）	9.53	15.86	4.30
[35.0，55.0）	0	0	0

Fig.6 The change of lightning-ignited fire occurrences and burned areas with slope

Fig.7 The spatial distribution of lightning-ignited fire points on different terrain relief degrees

Tab.4 The proportion of lightning-ignited fire occurrences and burned areas across different topographic relief degree intervals

地形起伏度/m	面积占比	雷击火灾次数占比	过火面积占比
[0，20）	7.96	5.65	2.14
[20，75）	30.45	19.62	38.48
[75，200）	59.41	70.97	47.67
[200，600）	2.18	3.76	11.71

Fig.8 The variation of the number of lightning-ignited fires and burned area with terrain relief degree

Fig.9 The spatial distribution of lightning-ignited fire points on different slope aspects

Tab.5 The proportion of the number of lightning-ignited fires and burned area on different slope aspects

坡向	面积占比	雷击火灾次数占比	过火面积占比
平地	0.59	0	0
北坡	11.37	5.65	16.20
东北坡	13.06	11.56	5.34
东坡	13.49	11.83	11.35
东南坡	12.50	15.59	11.53
南坡	11.78	16.67	4.81
西南坡	12.48	15.86	6.99
西坡	12.71	13.17	30.08
西北坡	12.03	9.68	13.69

Fig.10 The principal component analysis of sample data

Fig.11 The optimization iteration curve

Fig.12 The prediction results of the model for the training set (a) and the test set (b)

Fig.13 The histogram of prediction errors on the test set

Fig.14 The R2plot of predictions on the test set

Fig.15 The kernel density estimation curve of interval prediction

Fig.16 The interval prediction results

Fig.17 The SHAP value distribution (a) and importance ranking (b) of wildfire suppression influencing factors

Fig.18 Global SHAP feature interaction matrix for wildfire suppression influencing factors

References 38

[1]	毕杰和, 2018. 内蒙古大兴安岭林区森林生态状况及保护建议[J]. 内蒙古林业调查设计, 41(6):6-8.
[2]	焦强英, 韩宗甫, 王炜烨, 等, 2023. 基于多源数据和机器学习方法的大兴安岭地区雷击火驱动因子及火险预测模型[J]. 林业科学, 59(6):74-87.
[3]	李婧, 2022. 内蒙古大兴安岭林火时空分布及火险区划[D]. 呼和浩特: 内蒙古农业大学.
[4]	李威, 舒立福, 王明玉, 等, 2023. 大兴安岭1980—2021年雷击火时空分布特征[J]. 林业科学, 59(10):22-31.
[5]	刘爱利, 2004. 基于1:100万DEM的我国地形地貌特征研究[D]. 西安: 西北大学.
[6]	刘晓东, 冯旭宇, 李庆君, 等, 2021. 内蒙古地形因素对雷电灾害的影响及其权重分析[J]. 科学技术与工程, 21(3):899-905.
[7]	刘晓东, 冯旭宇, 刘旭洋, 等, 2020. 内蒙古雷击人员伤亡特征及其致灾因素的熵权分析[J]. 中国安全生产科学技术, 16(4):182-188.
[8]	倪长虹, 邸雪颖, 2009. 黑龙江省大兴安岭雷击火发生规律[J]. 东北林业大学学报, 37(1):55-57.
[9]	聂卉, 吴晓燕, 2024. 结合梯度提升树算法与可解释机器学习模型SHAP的抑郁症影响因素研究[J]. 数据分析与知识发现, 8(3):41-52. DOI
[10]	田晓瑞, 舒立福, 王明玉, 等, 2009. 大兴安岭雷击火时空分布及预报模型[J]. 林业科学研究, 22(1):14-20.
[11]	田晓瑞, 舒立福, 赵凤君, 等, 2012. 大兴安岭雷击火发生条件分析[J]. 林业科学, 48(7):98-103.
[12]	万雨勤, 仝纪龙, 刘永乐, 等, 2026. 基于XGBoost算法量化气象要素对兰州市主城区夏季臭氧浓度的影响[J]. 环境科学, 47(1):223-232.
[13]	杨腾杰, 高新强, 杨志国, 等, 2025. 基于NRBO-XGBoost和ABKDE融合可解释模型的TBM掘进速度预测[J]. 河南科技大学学报:自然科学版, 46(4):73-87.
[14]	岳永杰, 韩军, 李玉柱, 等, 2013. 内蒙古大兴安岭森林涵养水源和保育土壤功能评估[J]. 中南林业科技大学学报, 33(12):91-95.
[15]	张冬有, 邓欧, 李亦秋, 等, 2012. 黑龙江省1980—2005年森林火灾时空特征[J]. 林业科学, 48(2):175-179.
[16]	赵凤君, 舒立福, 邸雪颖, 等, 2009. 气候变暖背景下内蒙古大兴安岭林区森林火灾发生日期的变化[J]. 林业科学, 45(6):166-172.
[17]	赵俊卉, 郭广猛, 亢新刚, 等, 2008. 基于LIS/OTD格点数据的中国东北地区雷击火时空分布原因分析[J]. 内蒙古农业大学学报:自然科学版, 29(3):49-54.
[18]	赵俊卉, 亢新刚, 郭广猛, 等, 2009. 基于LIS/OTD格点数据的中国东北地区雷击火时空分布[J]. 生态学杂志, 28(4): 715-720.
[19]	周梅, 2003. 大兴安岭落叶松林生态系统水文过程与规律研究[D]. 北京: 北京林业大学.
[20]	ABATZOGLOU J T, WILLIAMS A P, 2016. Impact of anthropogenic climate change on wildfire across western US forests[J]. Proceedings of the National Academy of Sciences of the United States of America, 113(42): 11 770-11 775.
[21]	BOWMAN D M J S, BALCH J K, ARTAXO P, et al, 2009. Fire in the Earth system[J]. Science, 324(5926): 481-484. DOI PMID
[22]	CHEN F, DU Y S, NIU S K, et al, 2015. Modeling forest lightning fire occurrence in the Daxinganling Mountains of northeastern China with MAXENT[J]. Forests, 6(5): 1 422-1 438.
[23]	COCHRANE M A, MORAN C J, WIMBERLY M C, et al, 2012. Estimation of wildfire size and risk changes due to fuels treatments[J]. International Journal of Wildland Fire, 21(4): 357-367. DOI URL
[24]	DU Y, LI S, FENG P, et al, 2021. Dynamic risk assessment of wildfire based on integrating multi-source data and a random forest algorithm: A case study of Chongqing, China[J]. Natural Hazards, 108(1): 1 071-1 090.
[25]	GUO F T, WANG G Y, INNES L, et al, 2016. Comparison of six generalized linear models for occurrence of lightning-induced fires in northern Daxing’an Mountains, China[J]. Journal of Forestry Research, 27(2): 379-388. DOI URL
[26]	HEYERDAHL E K, BRUBAKER L B, AGEE J K, 2001. Spatial controls of historical fire regimes: A multiscale example from the interior west, USA[J]. Ecology, 82(3): 660-678. DOI URL
[27]	JAIN P, COOGAN S C P, SUBRAM-ANIAN S G, et al, 2020. A review of machine learning applications in wildfire science and management[J]. Environmental Reviews, 28(4): 478-505. DOI URL
[28]	JOLLY W M, COCHRANE M A, FREEBORN P H, et al, 2015. Climate-induced variations in global wildfire danger from 1979 to 2013[J]. Nature Communications, 6: 7537. DOI: 10.1038/ncomms7537. PMID
[29]	LUNDBERG S M, ERION G, CHEN H, et al, 2020. From local explanations to global understanding with explainable AI for trees[J]. Nature Machine Intelligence, 2(1): 56-67. DOI PMID
[30]	NARAYANARAJ G, WIMMERS M T, 2012. Modeling the influence of topographic roughness on fire spread[J]. Environmental Modelling & Software, 38: 290-302.
[31]	PARKS S A, PARISIEN M A, MILLER C, 2012. Spatial bottom-up controls on fire likelihood vary across western North America[J]. Ecosphere, 3(1): 1-20.
[32]	RODRIGUES M, DE LA RIVA J, FOTHERINGHAM S, 2014. Modeling the spatial variation of the explanatory factors of human-caused wildfires in Spain using geographically weighted logistic regression[J]. Applied Geography, 48: 52-63. DOI URL
[33]	ROTHERMEL R C, 1972. A mathematical model for predicting fire spread in wildland fuels[R]. Ogden: USDA Forest Service, Intermountain Forest and Range Experiment Station, 1-52.
[34]	RUDIN C, 2019. Stop explaining black box machine learning models for high stakes decisions and use interpretable models instead[J]. Nature Machine Intelligence, 1(5): 206-215. DOI PMID
[35]	SONG S D, ZHOU X, YUAN S B, et al, 2025. Interpretable artificial intelligence models for predicting lightning prone to inducing forest fires[J]. Journal of Atmospheric and Solar-Terrestrial Physics, 267: 106408. DOI:10.1016/j.jastp.2024.106408.
[36]	SOWMYA R, PREMKUMAR M, JANGIR P, 2024. Newton-Raphson-based optimizer: A new population-based metaheuristic algorithm for continuous optimization problems[J]. Engineering Applications of Artificial Intelligence, 128:107532. DOI:1016/j.engappai.2023.107532.
[37]	SYPHARD A D, RADELOFF V C, KEULER N S, et al, 2008. Predicting spatial patterns of fire on a southern California landscape[J]. International Journal of Wildland Fire, 17(5): 602-613. DOI URL
[38]	YUE W T, REN C, LIANG Y J, et al, 2023. Assessment of wildfire susceptibility and wildfire threats to ecological environment and urban development based on GIS and multi-source data: A case study of Guilin, China[J]. Remote Sensing, 15(10): 2659. DOI:10.3390/rs15102659.