Efficient estimation of plant species diversity in desert regions using UAV-based quadrats and advanced machine learning techniques

Abstract

Understanding the distribution of plant species diversity(PSD) along spatial and environmental gradients is essential for implementing effective conservation strategies. However, effective monitoring of large-scale PSD in desert regions remain challenging. In this study, traditional and unmanned aerial vehicle (UAV) quadrat surveys were employed to monitor the vegetation composition in the desert regions of the Junggar Basin, China. By combining multi-source data, two variable selection methods (elastic net regression and Boruta) and two machine learning algorithms (support vector machines and boosted regression trees) were used to develop PSD estimation models. This study aimed to investigate spatiotemporal variations in PSD and their driving factors. The results are as follows. (1) UAV method is more efficient and accurate than traditional methods in investigating PSD in desert areas. (2) The model combining variables selected by Elastic Net Regression and the Boosted Regression Trees algorithm is the optimal model for estimating PSD in desert areas(R² = 0.476–0.613, RMSE = 0.135–2.2, MAE = 0.1–1.72). (3) The central region of the basin exhibited lower PSD, whereas the peripheral regions demonstrated higher PSD but were more heavily impacted by external disturbances. Over the past 20 years, 5.99 %–13.87 % of the area has shown a significant decline in PSD. (4) Cumulative precipitation and soil organic carbon are the primary drivers of PSD's spatial patterns, while human disturbance dictates its temporal dynamics. This study introduced a novel method for estimating PSD, providing a theoretical foundation for ecological restoration, and biodiversity conservation in the study region.