Abstract:

Objective: Due to the complexity and strong concealment of lightning-caused forest fire occurrence, it is difficult to monitor and early warning. For most available forest fire forecasting models, although main meteorological factors are taken into account the models, their adaptability and precision are still relatively low. In this study, based on the long-term multi-sources data, the driving factors of lightning-caused forest fires were analyzed, and a dynamic lightning-caused fire forecasting model with high spatial resolution was built by using machine learning method, to provide support for the fire prevention and control. Method: The spatial and temporal distribution of lightning-caused fires from 2010 to 2020 was analyzed. Multi-source data such as ground-based lightning observations, satellite data, meteorological reanalysis data and DEM data were used to extract 18 driving factors from 4 categories (i.e., lightning, meteorology, vegetation and terrain). The characteristics of each driving factor and the relationship with lightning-caused fires were studied. The driving factors of forest lightning-caused fire records and randomly generated non-lightning fire spots were extracted to establish our initial training sample dataset. Driving factors were selected by calculating feature importance and correlation matrix. Based on the optimized training sample dataset, three integrated learning models, namely, gradient ascending decision tree (GBDT), random forest (RF) and extreme random tree (ERT) were trained and evaluated, respectively. The model with the best performance was used to forecast forest lightning-caused fires in Daxing’ anling Mountains. Result: The results showed that from 2010 to 2020, the maximum and minimum number of lightning fires occurred in 2015 and 2012, respectively, mainly in May, June and July, and the occurrence period was mainly concentrated from 10:00 to 17:00. The areas with high lightning-caused fire density appeared in Mohe County, Tahe County, Xinlin District and Huzhong District. The spatial distribution of lightning was consistent with that of lightning-caused fires to a certain extent, but the more lightning number might not lead more lightning-caused fires. In 2011, the maximum number of lightning occurrences was 114 632, but only 11 lightning-caused fires. Under the following conditions: lightning intensity in –20– –40 kA, steepness ranged of – 4– –8 kA·μs^?1, relative humidity less than 40%, precipitation less than 4 mm, temperature more than 29 ℃, atmospheric pressure in 91–95 kPa and wind speed in 1–3 m?s^–1, lightning fires were more prone to occur. NDVI (normalized vegetation index), GPP (total primary productivity), Et (evapotranspiration), and NPP (net primary productivity) were positively correlated with the occurrence of lightning-caused fires. Lightning-caused fires occurred more frequently at altitude of 300–900 m and slope of 0–12°, and aspect had little effect on the occurrence of lightning fires. After feature selection, the remaining 13 features were involved in the model construction of the three ensemble learning algorithms, and the comparison showed that the ERT prediction effect was the best. The AUC value of ERT model reached 0.97, and the precision, recall and F1 score values were all higher than GBDT and RF models. The high-risk region of lighting fires predicted by ERT model had a good agreement with the location of the actual lightning fire spots. Conclusion: The multi-sources big data, especially satellite observation data, are used to obtain more potential driving factors related to the occurrence of lightning-caused fires. Combined with the advantages of machine learning method, our forecasting model of forest lightning-caused fires well performs on the high-risk region prediction of lightning-caused fires in Daxing’ anling Mountains, which has good generalization ability, good adaptability and high spatial resolution.

Key words: lightning-caused fire, fire forecasting, machine learning, remote sensing, lightning