Artificial intelligence classification of wetland vegetation morphology based on deep convolutional neural network
Ping Lin,Qun Lu,Du Li,Yongming Chen,Zhiyong Zou,Shanchao Jiang
DOI: https://doi.org/10.1111/nrm.12248
2020-01-01
Natural Resource Modeling
Abstract:In real-world wetland vegetation morphology (WVM) detection, large scene variations such as those due to landform, vegetation, sunlight, weather, and sky, as well as camera parameter settings such as focal length and shooting angle, require systematic and complicated artificial intelligence technology to accurately discriminate inter and intra-class wetland objections. To deal with these challenges, we introduced a deep-level discriminative model based on convolutional neural networks (CNN) for classifying the images of DongZhai Harbor intertidal, Lashi Lake alpine, Yancheng coastal and Zoige plateau wetlands in China. A 96-dimensional convolution operation with kernel sizes of 11x11x3 first applied to the resized 227x227 WVM input pictures to acquire the effective morphologic features. The perceptron layers of the rectified linear unit and the batch normalization were used in the middle layer to achieve the better gradient propagation property during the training process. The WVM features were down-sampled by the pooling networks to reduce the neuron dimensions. The fully connected layer linked to the output of the convolutional and pooling layers to obtain the high-level WVM species information for the final WVM classification purpose. The deep-level CNN-based method was compared with the traditional shallow-level feature-designed algorithms of conditional maximum entropy regression, multilayer perceptron and support vector machine. The deep-level algorithm showed the superior performance for detecting the WVM species, which provided a superior alternative routine for accurate artificial intelligence classification of WVM in the ecological engineering application. Recommendations for Resource Managers This proposed artificial intelligence model can be used to effectively help people understand the macroscopic scene changes of ecological elements from the wetland vegetation morphology images. The detection results show the robustness of the discriminative model, for it is not affected by the complex natural factors such as landform, vegetation, sunlight, weather, and sky, as well as human factors such as camera parameter settings of focal length and shooting angle to some extent. The deep-level model based on the cascade convolution framework can be used to extract more distinguishing detailed ecological features from the original wetland vegetation morphology images compared with the shallow-level algorithms using a single convolutional layer.