An objective estimate of water stress - going beyond PDSI

Abstract

Drought detection is crucial for water resources management and global food security. Drought is typically detected using empirical indices, such as the Palmer Drought Severity Index (PDSI). However, those indices lack objectivity and are therefore suboptimal. Machine learning has been proven to be a powerful tool to define objective and optimal regressions, especially in hydrology. Here, we developed a machine learning (ML) model using Long Short-Term Memory (LSTM), which include memory effects, to predict the Evaporative Fraction (EF), an indicator of water stress at the surface, based on FLUXNET2015 Tier 1 eddy-covariance dataset. Compared to the widely used PDSI, EF is a more direct drought index to indicate water stress conditions. The results show that, firstly, with some routinely available variables, e.g., precipitation, net radiation, air temperature, relative humidity and other static variables like, Plant Functional Type (PFT) and soil property, the model can capture the EF dynamics, especially during the dry season. Secondly, we found there were different LSTM memory lengths across different Plant Functional Types. This indicates different rooting depth and different plant water use strategies that regulate the time scales of droughts. Our results have important implications for future water stress estimation, e.g., drought detection, in order to obtain a more direct and more accurate estimate of water stress.