Scholarlink

<p>Litter moisture content (LMC) is a critical factor influencing forest fire ignition and spread, directly determining fire risk rating and fire behaviour, making it a core parameter in forest fire management. The temporal scale of the data analyzed and modelled is daily, as either LMC or weather data were collected daily. The daily minimum LMC is of significant importance for practical operations such as forest fire risk forecasting and prescribed burning. Currently, machine learning has been widely applied in forest fire management; however, research on daily predictions distinguishing between different LMC ranges, which are the original data with from 0% to maximum value, the subset of the data with LMC from 0 to 35%, and the sub-dataset with LMC grater than 35%, remains relatively scarce, and systematic comparisons with traditional methods are still insufficient, limiting the depth and breadth of model applications in this field. Therefore, this study focused on the surface litter of typical coniferous forests in southwestern China. Two machine learning (ML) methods, XGBoost and Random Forest (RF), were employed to predict daily LMC. Their performance was systematically compared with that of traditional semi-physical models (the Nelson and Simard mehtod) in order to evaluate the applicability of ML methods for operational prediction. The results indicated that when LMC ≤ 35%, relative humidity was the primary influencing factor, whereas wind speed and rainfall had significant effects on the full LMC dataset and the sub-dataset with LMC ≥ 35%. Both XGBoost and RF methods demonstrated excellent predictive capabilities across all LMC data, with a maximum mean absolute error (MAE) of only 9.23% and R<sup>2</sup> values consistently above 0.97. In contrast, the two traditional semi-physical methods, Nelson and Simard, exhibited higher prediction errors, with a minimum error still reaching 26.52%, failing to meet the practical management requirements for fire danger forecasting. This study confirms the significant advantages of the ML methods in daily LMC prediction, offering not only high predictive accuracy but also good interpretability in identifying key factors influencing LMC dynamics. It provides theoretical support and technical basis for forest fire prevention and fuel management in southwestern China.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

A study on machine learning-based prediction model for surface litter moisture content in coniferous forests

  • 期刊论文

  • 发表日期：

    2026年2月23日

查看全文
  • Yunlin Zhang,
  • Yanwei Zhang,
  • Jianfeng Li

摘要

Litter moisture content (LMC) is a critical factor influencing forest fire ignition and spread, directly determining fire risk rating and fire behaviour, making it a core parameter in forest fire management. The temporal scale of the data analyzed and modelled is daily, as either LMC or weather data were collected daily. The daily minimum LMC is of significant importance for practical operations such as forest fire risk forecasting and prescribed burning. Currently, machine learning has been widely applied in forest fire management; however, research on daily predictions distinguishing between different LMC ranges, which are the original data with from 0% to maximum value, the subset of the data with LMC from 0 to 35%, and the sub-dataset with LMC grater than 35%, remains relatively scarce, and systematic comparisons with traditional methods are still insufficient, limiting the depth and breadth of model applications in this field. Therefore, this study focused on the surface litter of typical coniferous forests in southwestern China. Two machine learning (ML) methods, XGBoost and Random Forest (RF), were employed to predict daily LMC. Their performance was systematically compared with that of traditional semi-physical models (the Nelson and Simard mehtod) in order to evaluate the applicability of ML methods for operational prediction. The results indicated that when LMC ≤ 35%, relative humidity was the primary influencing factor, whereas wind speed and rainfall had significant effects on the full LMC dataset and the sub-dataset with LMC ≥ 35%. Both XGBoost and RF methods demonstrated excellent predictive capabilities across all LMC data, with a maximum mean absolute error (MAE) of only 9.23% and R2 values consistently above 0.97. In contrast, the two traditional semi-physical methods, Nelson and Simard, exhibited higher prediction errors, with a minimum error still reaching 26.52%, failing to meet the practical management requirements for fire danger forecasting. This study confirms the significant advantages of the ML methods in daily LMC prediction, offering not only high predictive accuracy but also good interpretability in identifying key factors influencing LMC dynamics. It provides theoretical support and technical basis for forest fire prevention and fuel management in southwestern China.