Integrating rapid assessment, variable probability sampling, and machine learning to improve accuracy and consistency in mapping local spatial distribution of plant species richness

Abstract

Abstract Conserving plant diversity is integral to sustainable forest management. This study aims at diversifying tools to map spatial distribution of species richness. We develop a sampling strategy of using rapid assessments by local communities to gather prior information on species richness distribution to drive census cell selection by sampling with covariate designs. An artificial neural network model is built to predict the spatial patterns. Accuracy and consistency of rapid assessment factors, sample selection methods, and sampling intensity of census cells were tested in a simulation study with seven 25–50-ha census plots in the tropics and subtropics. Results showed that identifying more plant individuals in a rapid assessment improved accuracy and consistency, while transect was comparable to or slightly better than nearest-neighbor assessment, but knowing more species had little effects. Results of sampling with covariate designs depended on covariates. The covariate Ifreq, inverse of the frequency of the rapidly assessed species richness strata, was the best choice. List sampling and local pivotal method with Ifreq increased accuracy by 0.7%–1.6% and consistency by 7.6%–12.0% for 5% to 20% sampling intensity. This study recommends a rapid assessment method of selecting 20 individuals at every 20-m interval along a transect. Knowing at least half of the species in a forest that are abundant is sufficient. Local pivotal method is recommended at 5% sampling intensity or less. This study presents a methodology to directly involve local communities in probability-based forest resource assessment to support decision-making in forest management.