军事知识服务平台
基于增量学习的森林地表细小死可燃物含水率预测方法
An Incremental Learning Approach for Predicting the Moisture Content of Forest Dead Fine Fuel
摘要 针对现有含水率预测模型在时间和跨树种预测条件下预测精度下降的问题,提出一种基于增量学习的长短期记忆网络(long short-term memory,LSTM)预测方法.以黑龙江省哈尔滨市东北林业大学实验林场的蒙古栎、胡桃楸和白桦3种典型林分为研究对象,基于2024年秋季高火险期采集20 min间隔的气象与含水率数据构建时序数据集,并从时间和林分类型2个维度对模型外推性能进行验证.结果表明,与传统LSTM模型相比,该方法在时间外推和跨树种预测条件下预测精度分别提高46.54%和47.47%,平均绝对误差(mean absolute error,MAE)分别为3.61%和3.97%.同时,补充设置的随机森林(random forest,RF)基线模型对比结果进一步支持了上述结论.该方法能够提高模型对环境变化的适应能力及跨林分预测能力.
Abstract:
To address the decline in predictive accuracy of existing fuel moisture content models under temporal and cross-species prediction scenarios,an incremental learning-based long short-term memory(LSTM)model is proposed.Taking three typical forest stands—Quercus mongolica,Juglans mandshurica,and Betula platyphylla—at the experimen-tal forest of Northeast Forestry University in Harbin,Heilongjiang Province as study objects,a time-series dataset was constructed using meteorological and moisture content data collected at 20-minute intervals during the high fire-risk pe-riod in autumn 2024.The extrapolation performance of the model was evaluated from both temporal and stand-type per-spectives.The results showed that,compared with the conventional LSTM model,the proposed method improved predic-tion accuracy by 46.54%and 47.47%under time extrapolation and cross-species prediction scenarios,respectively,with mean absolute errors(MAE)of 3.61%and 3.97%.In addition,comparisons with a supplementary random forest(RF)baseline model further supported the effectiveness of the proposed approach.The method demonstrates improved adaptability to environmental changes and enhanced predictive capability across different forest stands.
刊 名 森林工程 ISTICPKU
年,卷(期): 2026, 42(5)
分类号: S762.2
机标分类号: S762.1 TP391 TE341
在线出版日期: 2026年9月17日
基金项目: :中央财政林业科技推广示范项目,