Near-infrared spectroscopy and ensemble learning modeling for moisture detection in forest floor leaf litter

Abstract

The moisture content of forest floor litter is a critical indicator for assessing forest ecosystem stability and predicting wildfire risks. Traditional near-infrared (NIR) spectroscopy methods face limitations in species applicability and model accuracy. To enhance detection generalization capability and accuracy, this study proposes a moisture content detection model optimized by differential evolution (DE) algorithm and introduces an improved triangular kernel function (ITK) for least squares support vector machine (LSSVM) regression prediction, constructing a DE-LSSVM-ITK-based litter moisture content detection model. Using forest floor litter from Quercus mongolica, Fraxinus mandshurica, and Larix gmelinii as research subjects, the model employed a ten-fold cross-validation strategy to train and ensemble 10 optimal models, with the average prediction results on the test set serving as the final output. Experimental results demonstrate that the DE-LSSVM-ITK ensemble model achieves higher prediction accuracy and robustness, making it suitable for constructing moisture content detection models for different tree species. This provides a reliable technical approach for forest ecological monitoring and fire prevention.