中图分类号：S762.2 文献标识码：A DOI：10.7525/j.issn.1006-8023.2025.03.001

Abstract：Themoisturecontentofforestfloorliterisakeyfactorinforestfireoccurrences，nditsaccuratedetectionis crucial for fire prevention.Near-infrared spectroscopy（NIRS）can directly invert moisturecontent from spectral data， enabling rapid detectionof liter moisture content.However，spectral characteristics difer between fuel types due to variations inlightintensitydataatdiffrent wavelengths，requiringseparatedetectionmodelsforliterfrom dierenttree species to match specific light intensity-moisture content inversion relationships.Collcting and labeling spectral data across differentforeststands is time-consuming，limitingthepracticalapplicationofthespectralmethod.Toaddresthis issue，this study proposesa moisturecontent detection method for forest floor liter basedon Bi-LSTM（Bidirectional Long Short-Term Memory）transferlearning.By transfering the trained modelparameters to new models，we avoid training models from scratch，thereby improving model learning eficiencyand reducing the data required fortraining.The studydemonstrates thatthe Bi-LSTMmethodsurpases the traditional inversion approach using LSTMin termsof detection accuracy.Specifically，the mean absolute errr （MAE）for Quercus mongolicaand Larix gmelinii is reduced by

0 . 6 2 % and 0 . 8 7 % ，respectively，while the mean squared error（MSE）is reduced by 0 . 2 8 % and 0 . 7 0 % ，respectively. Moreover，the Bi-LSTM-based transfer learning approach significantly lessens the reliance on labeled NIR spectraldata. With a target domain sample size of 3OO and a source domain sample size of 1 0 0 0 ，the detection model record an MAE of 3 . 2 7 % ，an MSE of 1. 10 % ，and an of 0.918.When compared to models without source domain training，the MAE and MSE show reductions of 2. 3 6 % and 1 . 0 2 % ，respectively，and an increase in of 0.114.A comparative analysis before andafterimplementingtransferlearningreveals thathismethodologyoffersanovelstrategytodiminishthetimecostasociated withmodeling moisturecontentin spectralitterand to enhancethepracticalapplicationof spectraldetection. Keywords：Liter fall；moisture content;transfer learning；deep learning;near-infrared spectrum

0 引言

林内可燃物作为森林火灾产生的必要条件之二[1]，其中森林地表枯叶含水率（dead fuel forest mois-turecontent，DFFMC）作为一级引火要素是引起森林火灾的重要因素[2]。其含水率的大小与森林火灾发生的可能性、火灾强度以及蔓延速度具有直接关系，因此检测森林地表枯叶的含水率对于预测森林火灾风险和评估火险等级至关重要[3-4]。

目前DFFMC的研究方法主要有气象要素回归法、平衡含水率法以及遥感光谱法[5]，其中，气象要素回归法测量方便且建模简单，在各林区内应用广泛，但这种方法并未考虑到可燃物本身的种类差异，所以该方法测得的枯叶含水率在很大程度上受到可燃物自身特性和研究区域限制，并未将枯叶含水率按其自身种类细化[6]。平衡含水率法虽然在小规模范围内较为精确，但在扩展到更大范围时，其准确性可能会有所下降[7-8]。因此上述两种主流方法在一定程度上都会受到测量范围影响，当测量范围较大时，测量数值准确度不高。光谱法可以根据光谱数据直接反映水分含量从而实现枯叶含水率的快速检测，而且测量数值准，测量速度快，使用简便快捷[9-10]。

在以往的对近红外光谱检测枯叶含水率研究中，研究人员多采用线性回归算法对光谱数据进行含水率反演[1]。但目前没有研究表明枯叶含水率与叶片光谱的特征波段呈线性关系[12]。随着机器学习的发展，神经网络越来越多被用于枯叶含水率的检测研究中[13]，相比线性回归算法神经网络引入了对非线性关系的考虑。……