Deep learning-based identification of environmental memory effects on gross primary productivity of alpine grasslands in Qinghai-Tibetan Plateau

The extensively distributed grasslands of the Qinghai-Tibet Plateau (QTP) play a vital role in the global carbon cycle and climate regulation. Gross primary productivity (GPP), a crucial indicator of ecosystem carbon sequestration capacity, remains highly uncertain partly due to neglecting the memory effects of environmental conditions (i.e., the influence of past states on current GPP). Moreover, existing models have difficulty in simultaneously handle multidimensional spatio-temporal data and dynamic climate responses, leading to simulation deviations and exacerbating uncertainties. Here, we integrated climate and vegetation data with time series characteristics and spatial characteristics to simulate the GPP of alpine grassland on the QTP, by developing a deep learning model CNN-LSTM that combined Convolutional Neural Networks (CNNs) and Long Short-Term Memory networks (LSTM). The conclusions were as follows: (a) The CNN-LSTM model effectively captured spatial patterns using CNNs and temporal dependencies with LSTMs, incorporating memory effects to consider the impact of past environmental conditions. This integration enhanced GPP simulation accuracy and improved the model's ability to capture interannual variability. (b) The training and optimization of the CNN-LSTM models revealed that the comprehensive memory effect length of GPP on historical climate and vegetation dynamics operates in a 4-month timescale, with the memory effects of GPP varied across environmental variables in both duration and intensity. (c) During 2001–2021, The mean annual GPP of the alpine grassland in QTP was 332.29 g C m^-2 a^-1, with a growth rate of 1.84 g C m^-2 a^-2. (d) Precipitation exhibited relatively longer durations and higher intensities compared to other factors, and the interannual variability of GPP was mainly influenced by water conditions. This study highlights the importance of integrating environmental memory into GPP modeling, which would enhance our comprehension of the mechanisms driving GPP and the impacts of climate change on carbon cycling in terrestrial ecosystems.