Interannual variation of gross primary production detected from optimal convolutional neural network at multi‐timescale water stress

Abstract

AbstractSpatiotemporal patterns of water stress caused by global warming has significantly affected gross primary productivity (GPP). However, its impact is hard to capture as the water stress of different timescales simultaneously influence GPP through the effects of time lag and legacy. As a result, synthetically considering the roles of water stress in the current year and previous years in GPP is very important for accurate modeling of GPP. In this study, we introduced the water stress indicators of standardized precipitation evapotranspiration index (SPEI) at different timescales, GPP observations from global flux network and the remote sensing‐based indexes to models of convolutional neural network (CNN). We built the optimal CNN model which effectively simulated the spatial patterns of GPP and detected their trends in interannual variation. The findings were as follows: (1) The CNN model with introduction of multi‐timescale SPEIs that reflect both current and previous years' water conditions improved the accuracy of modeled GPP. This result indicated that although GPP is most sensitive to current water conditions, it is also clearly affected by previous years' water conditions. (2) The optimal CNN model that considered previous years' water conditions not only aptly simulated the spatial distribution of GPP, but also showed advantages in simulating the interannual fluctuations of GPP and its response to drought. Ignoring the previous water conditions will underestimate the interannual fluctuations of GPP and the impact of drought on GPP.