当前位置：
X-MOL 学术
›
Forest Ecol. Manag.
›
论文详情
Our official English website, www.x-mol.net, welcomes your feedback! (Note: you will need to create a separate account there.)
Forest fire susceptibility mapping based on precipitation-constrained cumulative dryness status information in Southeast China: A novel machine learning modeling approach
Forest Ecology and Management ( IF 3.7 ) Pub Date : 2024-02-29 , DOI: 10.1016/j.foreco.2024.121771 Longlong Zhao , Yuankai Ge , Shanxin Guo , Hongzhong Li , Xiaoli Li , Luyi Sun , Jinsong Chen
Forest Ecology and Management ( IF 3.7 ) Pub Date : 2024-02-29 , DOI: 10.1016/j.foreco.2024.121771 Longlong Zhao , Yuankai Ge , Shanxin Guo , Hongzhong Li , Xiaoli Li , Luyi Sun , Jinsong Chen
Frequent forest fires cause severe damage to the ecological service functions of forest ecosystems. Machine learning (ML) techniques have gained widespread use for forest fire susceptibility mapping due to their potent nonlinear learning capabilities. However, prior research has devoted insufficient attention to negative sample sampling, leading to overestimated fire susceptibility. Furthermore, the dynamic factors used for fire prediction are often extracted based on fixed time windows (FTW), making it challenging to accurately capture the spatial heterogeneity in forest cumulative dryness status (CDS). In this paper, we introduced a space-humidity-constrained (SHC-based) sampling method to generate negative samples with specific humidity. It was achieved by imposing distance and aridity constraints on randomly generated points, thus enhancing their representativeness. We also proposed a dynamic factor extraction method based on dynamic time windows (DTW), constrained by daily or cumulative precipitation thresholds (), to obtain fine-grained forest CDS information by limiting the statistical time windows of dynamic factors. Our results demonstrated that the forest fire susceptibility modeling approach, incorporating the above two methods for sample sets and feature sets construction, significantly enhanced the prediction performance of three ML models, namely random forest (RF), support vector machine (SVM), and multi-layer perceptron (MLP), underscoring its stability and robustness in improving model performance. The RF model outperformed the SVM and MLP models in terms of both performance improvement and prediction accuracy. The application of these two proposed methods improved the prediction accuracy of the RF model by more than 3% (FTW of at least 35 days) and 5% ( of at least 40 mm), respectively. As increased, the performance enhancement became even more pronounced (accuracy can increase by over 8% when is 60 mm). The novel modeling approach, rooted in an understanding of the mechanisms underlying forest fires and adeptly capturing the forest spatially heterogeneous CDS information, can significantly improve the mapping accuracy. Moreover, coupled with the adaptive acquisition advantage of dynamic factors, this study is able to generate near real-time fire susceptibility maps, which hold greater practical value for aiding decision-makers in formulating effective prevention and management strategies.
更新日期：2024-02-29
京公网安备 11010802027423号