Aboveground biomass retrieval of wetland vegetation at the species level using UAV hyperspectral imagery and machine learning

Abstract

The aboveground biomass (AGB) of wetland vegetation is a crucial indicator for assessing the health of wetland ecosystems. In the context of global biodiversity threats, biodiversity has become a focal point in ecological and remote sensing research. This study focuses on the Beiliuyao area on Chongming Island, and ground-based biomass data and unmanned aerial vehicle (UAV) hyperspectral imagery are employed for the regional-scale estimation of the AGB of wetland vegetation. Considering the significant differences in AGB between Spartina alterniflora and Phragmites australis during different phenological periods, and AGB retrieval models are constructed based on vegetation classification. Multivariate stepwise regression (MSR), BP neural network (BP), and random forest regression (RFR) models are used to estimate both the dry and wet weights of AGB at the species level. The research results are as follows: (1) Compared with the other three estimation models for the same period, the RFR model yields the highest accuracy, with an R2 reaching 0.82 and an RMSE of 116.14 g/m2. (2) The accuracy of the estimations in November is lower under the same model conditions than that in other months, with the lowest R2 of 0.57 and an RMSE of 228.42 g/m2. (3) The weight of the AGB gradually decreases from August to November, and the wet AGB density ranges from 6000 to 7000 g/m2 with an account of 4.2 % of the wet ABG falling within this range in August. The results of this study demonstrate that UAV hyperspectral imagery and the RFR model can be used to effectively estimate the biomass of dominant species in wetlands. This approach provides a theoretical basis for the large-scale, efficient and dynamic monitoring of the AGB of wetland vegetation.

Keywords

• Wetland vegetation
• Aboveground biomass
• UAV hyperspectral imagery
• Machine learning
• Vegetation classification