Convolutional Neural Networks for Comprehending Geographical Features of International Important Ramsar Wetland Ecological Habitat Scenes in China

摘要

In this paper, a novel architecture of a convolutional neural network for deeply comprehending geographical features of international important Ramsar wetland ecological habitat scenes in China was proposed. We explored convolutional neural network (CNN) which is consisted of 8 layers and receives a true color 227 × 227 pixels image as its input. Thereby, the networks have 51529 neurons at the input level. The following convolutional layer has the set of 96 filters. The subsampling layer contains the rectified linear units (RELU) layers and pooling layers. The final level is the fully connected layer with 4 neurons. The developed convolutional neural networks detector were able to perceive the instances of ecological habitat scenes at overall average accuracies of 91.7% and 84.6% in the training and test procedure. The experimental results indicate that the comprehending capability of well-trained CNN model is much better than the traditional machine learning strategy. The state-of-the-art proposal algorithms provide a superior alternative for accurately comprehending wetland ecological habitat scenes.