Identification of the Most Important Environmental Variables for Predicting Sap Flow Using Machine Learning Techniques

Transpiration is an integral process of evapotranspiration and a primary driver of ecohydrology in forested systems. Sap flow, the upward movement of water from roots through the xylem system, is driven by transpiration. While sap flow measurement is a well-accepted technique for quantifying whole-plant transpiration, installing and maintaining a large number of sensors and data loggers in forest watersheds is costly and time-consuming. Using data measured in the field, we employed Random Forest (RF), a machine learning technique, to identify the most important environmental variables affecting the daily sap flow of eastern cottonwood (Populus deltoides (Bartr. ex Marsh.)) and to predict sap flow using only those variables. Among 11 environmental variables measured, three (daily solar radiation, photosynthetically active radiation, and air temperature) were most important to daily sap flow. Predicted sap flows generated with RF matched measured values closely, based on an R2 of 0.937, a root-mean-square-error (RMSE) of 96 g/m2/d (sapwood area), and a Nash-Sutcliff-efficiency (NSE) of 0.843. Compared with the Convolutional Neural Network–Bidirectional Long Short‑Term Memory–Gated Recurrent Unit (CNN‑BiLSTM‑GRU), a deep learning technique, RF more accurately predicted daily sap flow from the three most influential variables, particularly capturing peaks and valleys in the time series. Using these three variables, we further applied the RF model to predict daily sap flow for 2019, when no measurements were available. This study demonstrates an alternative approach for using RF to identify key environmental variables for cost‑effective and time‑saving daily sap flow prediction in forest plantations and watersheds.