A deep learning approach with dense time series imagery of PlanetScope for alpine wetland plant diversity mapping

Dynamic monitoring of biodiversity in alpine wetlands is critical for addressing the threats posed by global climate change and species invasions. Comparing with expensive airborne hyperspectral measurement for limited spatial coverage, satellite multispectral data with high spatial and temporal resolutions (e.g., PlanetScope) provides an efficient alternative for monitoring wetland plant diversity (WPD). However, the capabilities of PlanetScope dense time series data for mapping plant diversity in alpine wetland landscapes remain unexplored. Here, with dense time-series PlanetScope data, we developed a novel network, called Self-Attention Wetland Plant Diversity Network (SAWPD-Net) for mapping plant diversity in Shennongjia Alpine Wetlands, one of global hotspots of wetland biodiversity. Additionally, the performances of a series of AI algorithms, including Self-Attention Wetland Plant Diversity Network (SAWPD-Net), Transformer, Long Short-Term Memory (LSTM), Recurrent Neural Network (RNN), and Random Forest (RF), were compared at different spatial and temporal resolutions with PlanetScope data. The results showed that: (1) Compared with other methods, our proposed SAWPD-Net achieved higher mapping accuracy at fine spatial resolution(9m × 9m ; R² = 0.57 ~ 0.72, RMSE = 0.24 ~ 0.19 ); RF achieved the highest mapping accuracy  with a temporal resolution of 1-day and a spatial resolution of 21m × 21m ( R² = 0.75, RMSE = 0.18 ); (2) WPD mapping accuracy is linearly correlated with the temporal resolution of the input data: when the temporal resolution increased from 120-day to 1-day, the R² of SAWPD-Net increased by 26.3%, while the RMSE decreased by 20.8%. This study uncovers the potential of high-resolution multispectral satellites and AI algorithms for tracking WPD dynamics, which can be vital for developing a new generation of global biodiversity monitoring networks.