Prediction of Sap Flow in European Region Using Long Short-Term Memory (LSTM) Networks: The SAPFLUXNET Database

The development of a sap flow prediction model that covers the lack of transpiration measurement information is highly practical. However, a large-scale spatial model for forecasting plant sap flow has not yet emerged. The purpose of this thesis is to use the LSTM model to predict sap flow from the European datasets, using the open data SAPFLUXNET as a source. Our database was selected from 49 stands containing 618 plant data with 21 input features, including 6 dynamic variables and 15 static variables. In our results, it was argued that the treatment of missing data could be ignored for now, as it would have little interference with the model. In addition, the static variables played an important role in extending the model to a larger scale. Plant variables were fundamental to greatly improve the model, while site and stand variables also made the model robust. Lastly, it was found that ignoring the extreme values of the sap flow data was unhelpful and that higher sap flows might actually occur. The ultimate model was evaluated by the NSE value with an average of 0.779, at the “very good” level. This thesis is expected to be a pioneer on sap flow prediction using machine learning and open data within a regional or global scale.