Data modeling
Because of the complexity of the anthropogenic driving mechanisms, the relations among the 23 independent variables and ESV cannot be described by the conventional linear model. We used a feed-forward dense network as the deep learning model.
The network consists of 8 layers (4 dense layers), including 6 hidden layers in the network. Each layer has a certain number of neurons and activation functions (Table 1). Nonlinear activation functions such as the rectified linear unit (ReLU) were introduced in the 3rd and 5th hidden layers to learn the nonlinearity. The ReLU function was used to avoid vanishing gradient problems. Additionally, the dropout rate was set to 0.3 in all dropout layers to avoid overfitting problems.
We partitioned 70% of the 2191 units as training samples and 30% as testing samples. In the training phase, the optimizer and loss function were established based on adaptive moment estimation (ADAM) and the mean square error (MSE). After conventional model optimizations were performed, the above hyperparameters were determined. The corresponding model was trained and used in the study.
Table 1. The configuration of the model