森林地表枯叶含水率是森林火灾发生的重要因素，其精准检测对森林火灾的预防尤为重要。近红外光谱技术可以根据光谱数据直接反演水分含量，从而实现枯叶含水率的快速检测，但不同可燃物光谱光照强度数据的特征波段不尽相同，需要对不同树种的枯叶分别建立检测模型以匹配不同的光照强度与含水率反演关系，而对不同林分的光谱光照强度数据进行采集与标注也需要较高的时间成本，限制了光谱法的实际应用。为此，提出基于双向长短期记忆神经网络（Bi-LSTM）的森林地表枯叶含水率检测迁移学习方法，将训练好的模型参数迁移到新的模型中，避免重新训练模型，从而提高模型的学习效率、减小训练模型所需的数据量。结果表明，与经典反演方法长短期记忆网络（LSTM）相比，Bi-LSTM具有更好的检测性能，蒙古栎和落叶松的平均绝对误差（mean absolute error，MAE）分别减小了0.62%和0.87%，均方误差（mean square error，MSE）分别减小了0.28%和0.70%。且基于Bi-LSTM改进的迁移学习方法，大大降低了对标记近红外光谱数据的依赖。当目标域样本数为300个时，源域样本数为1 000个时，检测模型的MAE、MSE、决定系数（coefficient of determi nation，R ²）分别为3.27%、1.10%、0.918。MAE和MSE比没有源域训练的检测模型分别缩小了2.36%和1.02%，R ²提升了0.114。对比迁移前后说明迁移学习为降低光谱枯叶含水率建模时间成本、提高光谱检测实用性提供新的手段。

The moisture content of forest floor litter is a key factor in forest fire occurrences， and its accurate detection is crucial for fire prevention. Near-infrared spectroscopy （NIRS） can directly invert moisture content from spectral data， enabling rapid detection of litter moisture content. However， spectral characteristics differ between fuel types due to variations in light intensity data at different wavelengths， requiring separate detection models for litter from different tree species to match specific light intensity-moisture content inversion relationships. Collecting and labeling spectral data across different forest stands is time-consuming， limiting the practical application of the spectral method. To address this issue， this study proposes a moisture content detection method for forest floor litter based on Bi-LSTM （Bidirectional Long Short-Term Memory） transfer learning. By transferring the trained model parameters to new models， we avoid training models from scratch， thereby improving model learning efficiency and reducing the data required for training. The study demonstrates that the Bi-LSTM method surpasses the traditional inversion approach using LSTM in terms of detection accuracy. Specifically， the mean absolute error （MAE） for Quercus mongolica and Larix gmelinii is reduced by 0.62% and 0.87%， respectively， while the mean squared error （MSE） is reduced by 0.28% and 0.70%， respectively. Moreover， the Bi-LSTM-based transfer learning approach significantly lessens the reliance on labeled NIR spectral data. With a target domain sample size of 300 and a source domain sample size of 1 000， the detection model record an MAE of 3.27%， an MSE of 1.10%， and an R² of 0.918. When compared to models without source domain training， the MAE and MSE show reductions of 2.36% and 1.02%， respectively， and an increase in R² of 0.114. A comparative analysis before and after implementing transfer learning reveals that this methodology offers a novel strategy to diminish the time cost associated with modeling moisture content in spectral litter and to enhance the practical application of spectral detection.