Mechanisms of climate change impacts on vegetation and prediction of changes on the Loess Plateau, China

Monitoring and forecasting the spatiotemporal dynamics of vegetation across the Loess Plateau emerge as critical endeavors for environmental conservation, resource management, and strategic decision-making processes. Despite the swift advances in deep learning techniques for spatiotemporal prediction, their deployment for future vegetation forecasting remains underexplored. This investigation delves into vegetation alterations on the Loess Plateau from March 2000 to February 2023, employing fractional vegetation cover (FVC) as a metric, and scrutinizes its spatiotemporal interplay with precipitation and temperature. The introduction of a convolutional long short-term memory network enhanced by an attention mechanism (CBAM-ConvLSTM) aims to forecast vegetation dynamics on the Plateau over the ensuing 4 years, leveraging historical data on FVC, precipitation, and temperature. Findings revealed an ascending trajectory in the maximum annual FVC at a pace of 0.42% per annum, advancing from southeast to northwest, alongside a monthly average FVC increment at 0.02% per month. The principal driver behind FVC augmentation was identified as the growth season FVC surge in warm-temperate semi-arid and temperate semi-arid locales. Precipitation maintained a robust positive long-term association with FVC (Pearson coefficient > 0.7), whereas the temperature-FVC nexus displayed more variability, with periodic complementary trends. The CBAM-ConvLSTM framework, integrating FVC, precipitation, and temperature data, showcased commendable predictive accuracy. Future projections anticipate ongoing greening within the warm-temperate semi-arid region, contrasted by significant browning around the Loess Plateau's peripheries. This research lays the groundwork for employing deep learning in the simulation of vegetation's spatiotemporal dynamics.