Wetland Restoration Prioritization Using Artificial Neural Networks

Abstract

Wetland destruction is currently one of the greatest environmental problems in the world. Despite the functions of wetlands, these valuable ecosystems have steadily decreased because of human activities and climate change. To protect these valuable ecosystems, wetland restoration and rehabilitation are important operations that have been conducted worldwide. Since a wetland is a complex ecosystem with a variety of phenomena, increasing the number of variables considered during a restoration project will further boost the success rate of a restoration project. However, the inclusion of more variables will increase the complexity of the analysis. Thus, a method that can analyze complex models using many input variables is valuable. In most scientific studies, artificial intelligence algorithms have been widely applied to complex projects. However, the main question is whether these algorithms can learn the ecological patterns of a restoration project. For this reason, a multilayer perceptron (MLP) neural network was applied in this paper to investigate the ability to use these algorithms for wetland restoration. An artificial neural network (ANN) with one hidden layer and 15 neurons was used to determine the best areas for wetland restoration. The neural network was trained using the Levenberg-Marquardt algorithm; then, the trained ANN was used to determine the best areas for wetland restoration. The root mean square error (RMSE) of the model that was trained to prioritize wetland restoration was 0.04 ha. Because of water limitations in the study area, it is not possible to restore entire wetlands. Therefore, areas for restoration are prioritized based on ecological objectives. The results of the ANN demonstrate its ability to learn the ecological patterns and illustrates the performance of using this method for wetland restoration. Neural networks can calculate the final weights mathematically, and these algorithms are able to analyze complex models using many input variables; thus, ANNs are practical for wetland restoration.

Appendix: Operating principle of Multi Layers Perceptron (MLP) neural network

Appendix Fig. 10 represents a MLP neural network with a single hidden layer.

Fig. 10

A MLP with a single hidden layer (Taravat et al. 2016)

A one-hidden-layer MLP can be described as a function f : R^D → R^L, where D is the size of input vector x and L is the size of the output vector f(x), such that, in matrix notation:

$$ f(x)=G\left({b}^{(2)}+{W}^{(2)}\left(s\left({b}^{(1)}+{W}^{(1)}x\right)\right)\right), $$

with bias vectors b⁽¹⁾, b⁽²⁾; weight matrices W⁽¹⁾, W⁽²⁾ and activation functions G and s (Taravat et al. 2016).

The vector h(x) = Φ(x) = s(b⁽¹⁾ + W⁽¹⁾x) constitutes the hidden layer.

${W}^{(1)}\in {R}^{D\times {D}_h}$ is the weight matrix connecting the input vector to the hidden layer.

Typical choices for s include tanh, with tanh(a) = (e^a − e^−a)/(e^a + e^−a), or the logistic sigmoid function, with sigmoid(a) = 1/(1 + e^−a).

The output vector is then obtained as: o(x) = G(b⁽²⁾ + W⁽²⁾h(x)). As before, class-membership probabilities can be obtained by choosing G as the softmax function (in the case of multi-class classification) (Taravat et al. 2016).

The set of parameters to learn the algorithm is the set θ = {W⁽²⁾, b⁽²⁾, W⁽¹⁾, b⁽¹⁾}. Obtaining the gradients ∂ℓ/∂θ can be achieved through the backpropagation algorithm (Taravat et al. 2016).

Tan-Sigmoid transfer function from the input layer to the hidden layer is used (Guo et al. 2011; Miao et al. 2018):

$$ f\ (x)=\frac{2}{1+{e}^{-2x}}-1 $$

For the transfer function from the hidden layer to the output layer, the pure-linearity expression is applied as follows (Ren et al. 2007):

$$ f\ (x)=k.x+b $$

where k is a non-zero real number and b is the bias term.

Appendix Fig. 11. illustrates the major transfer functions.

Fig. 11

The transfer function, which determines how a neuron’s firing rate varies with the input (Abbasi and Mahlooji 2012)