Coastline Extraction Method Based on Convolutional Neural Networks—A Case Study of Jiaozhou Bay in Qingdao, China

Abstract

The traditional edge detection-based shoreline extraction method is severely disturbed by noise, and it is difficult to obtain a continuous coastline. In response to the above problems, we propose a coastline extraction method based on convolutional neural networks. Firstly, we replace the standard convolution with the Mini-Inception structure in the backbone network to extract multi-scale features of the object, and all the multi-scale features are concatenated. Then, we use the leaky-ReLU activation function instead of the ReLU activation function to avoid the problem that ''dead'' neurons cannot learn the effective features of remote sensing images. Finally, the network fully exploits multi-level information of objects to perform the image-to-image prediction. We carried out experiments on the remote sensing images of Jiaozhou Bay in Qingdao. The experimental results showed that our method could effectively extract the coastline automatically, and the producer's accuracy and the user's accuracy were higher than the comparison methods. INDEX TERMS Coastline extraction, remote sensing images, convolutional neural networks, backbone networks, activation function.

Rotary Inverted Pendulum Identification for Control by Paraconsistent Neural Network

“…The literature presents different activation functions, such as rigid, linear limit, log-sigmoid (sigmoid), tangent hyperbolic sigmoid (tanh), positive linear (poslin) [1], [21], also called rectified linear unit (ReLU), Leaky-ReLU (LReLU), among others [1], [3], [20]. The sigmoid is one of the most used activation functions until recently [3].…”

“…ReLU is very similar to the identity function, making the neural net learning process, based on this activation function, faster than sigmoid [20]. The ReLU function and its derivative are expressed as:…”

“…A disadvantage of ReLU is that neurons tend to "die" during training, causing the neuron output to start producing only zeroes. A variation of ReLU, called Leaky-ReLU (LReLU), avoid this [20], whose function and derivative are presented in (8) and (9).…”

“…where α is a parameter introduced in LReLU, with values proposed between 0.01 and 0.2 [20]. LReLU allows the unit to give a small gradient when the unit is no active (n < 0), reducing the potential problem mentioned about ReLU [21].…”

“…In Section III, we describe the novel PAL2v neuron and the mathematical relative to the proposed method. The experimental results and comparison with other activation functions, including Leaky ReLU [20], are given in Section IV. Section V concludes with considerations about the paraconsistent neural network based on results presented in Section IV.…”