Ensemble algorithms for modeling forest live fuel loads and multivariate probability proportional to size sampling in Kunming, Yunnan, China

Abstract

Forest live fuel load (LFL) models of tree, shrub, and herbaceous layers were established using linear function (LF), power function (PF), K-nearest neighbor (KNN), gradient boosting regression tree (GBRT), and random forest (RF) algorithms to improve the fit and prediction accuracy and reduce the bias of forest LFL models. Multivariate probability proportional to size sampling (MPPS), probability proportional to size sampling (PPS), and probability proportional to frequency sampling (PPF) were used to obtain mean estimates of the forest LFLs, and the sampling accuracies were assessed. We utilized forest stock and biomass survey data (forest fire risk data and forest resource planning and design survey data) from Kunming, Yunnan, China. The results indicated that the RF model had the highest fit and prediction accuracy. The R2 values of the fitted models for the LFLs of Pinus yunnanensis, Quercus spp., Pinus armandi, Eucalyptus spp., other soft broadleaf species, other hard broadleaf species, other conifer species, shrubs, and herbaceous species were 0.79, 0.89, 0.86, 0.88, 0.91, 0.83, 0.88 0.88, and 0.92, respectively. The F-values were 4.652, 2.582, 2.565, 2.551, 2.738, 3.158, 16.156, 3.479, and 8.468, respectively. The sampling results of forest LFLs were much better for MPPS than for PPF and PPS. Moreover, the two-stage MPPS was the most efficient method for estimating tree, shrub, and herbaceous LFLs. The mean estimates of tree fuel loads, shrub fuel loads, and herbaceous fuel loads were 144.7760, 3.8094, and 0.6442 t/hm2, respectively. There were 555 overlapping subunits for the three types of forest LFLs. The proposed method can significantly reduce the cost of LFL surveys and improve estimation accuracy.