采用全卷积神经网络与Stacking算法的湿地分类方法

高精度湿地制图对湿地生态保护与精细管理具有重要的支撑作用。针对传统湿地分类方法的精度不高和泛化能力弱等问题，提出了一种联合全卷积神经网路（fully convolutional neural network，FCN）与集成学习的湿地分类方法。首先利用全卷积神经网络（SegNet、UNet及RefineNet）对GF-6影像的语义特征进行提取与融合，然后利用Stacking集成算法对融合后的特征进行判别和分类。结果表明，联合全卷积神经网络与Stacking算法能有效提取湿地信息，总体分类精度为88.16%，Kappa系数为0.85。与联合全卷积神经网络与单一机器学习RF、SVM与kNN算法相比，该文提出的湿地分类方法在总体分类精度上分别提高了4.87%，5.31%和5.08%；与联合单一全卷积神经网络（RefineNet、SegNet、UNet）与Stacking算法下的湿地分类结果，该文提出的湿地分类方法在总体分类精度上分别提高了2.78%，4.48%与4.91%；该文方法一方面能通过卷积神经网络提取遥感影像深层的语义特征，另一方面通过集成学习根据各分类器的表征性能进行合理的选择并重组，从而提高分类精度及其泛化能力。该方法能为湿地信息提取及土地覆盖分类方法的研究提供参考。

Wetland classification method using fully convolutional neural network and Stacking algorithm

Abstract: High-precision wetland mapping plays an important supporting role in wetland ecological protection and fine management. To overcome the shortcomings of traditional wetland classification methods, this paper proposes a wetland classification method that couples fully convolutional neural network and ensemble learning. The convolutional layer and pooling layer in convolutional neural network constitute a highly versatile feature extractor, which can extract highly abstract deep features in remote sensing images. However, the input image size of many CNN models is 256 × 256 or 299 × 299 pixels, which cannot satisfy the extraction of large-scale wetland feature information. Fully Convolutional Neural Network replaces the fully connected layer with a convolutional layer, which can accept images of any size and can achieve pixel-by-pixel classification. Therefore, we employed fully convolutional neural networks (SegNet, UNet, and RefineNet) to extract and merge the deep semantic features of GF-6 images. A single machine learning algorithm is easy to fall into a local optimal solution, and the generalization ability of unknown samples is poor. Ensemble learning uses multiple base classifiers to predict the results. Therefore, it has a strong ability to apply to various scenarios and a high classification accuracy. Therefore, this paper selects the ensemble learning Stacking model which has better ensemble effect on the stable classifiers in the classification task for wetland classification research based on semantic features derived by fully convolution neural network. In some scenarios, the performance of the Stacking algorithm is better than other algorithms. However, in some applications, the performance of the integrated algorithm may degrade. Therefore, in order to further improve the stability and generalization ability of the Stacking algorithm, this paper proposes an adaptive Stacking algorithm based on the Stacking algorithm. In the adaptive Stacking algorithm, the meta-classifier was first determined. Pervious research demonstrated that using an ensemble classifier (e.g., random forest, RF) as the meta-classifier improves prediction performance. Thus, the RF algorithm was used as the meta-classifier in this study. Then, we let all the base-classifiers, including the support vector machine (SVM), RF, k-NearestNeighbor (kNN), logistic regression (LR), and na?ve Bayes (NB), combine freely and train the input dataset. The results show that the coupled fully convolution neural network and adaptive stacking algorithm can effectively extract wetland information, and the overall classification accuracy and kappa coefficient are 88.16% and 0.85, respectively. This method can effectively extract most types of wetlands. Among them, the producer accuracy and user accuracy of lakes/pools, mudflat, and sedge are all around 90%, but it is easy to form a misclassification of reed beaches and poplar forest beaches. The main reason is that the two have similar spectral characteristics during part of the growing season, and it is difficult to distinguish them with single-phase remote sensing images. Compared with coupling FCN and single classifier (SVM, RF and kNN), the overall accuracy of the proposed method is improved by 5.31%, 4.87%, and 5.08%, respectively. Compared with SVM, RF and kNN, the overall accuracy of the proposed method is improved by 10.28%, 7.90%, and 10.01%, respectively. Moreover, the proposed method has a higher classification accuracy than that of using SegNet, UNet or RefineNet. The results demonstrate that compared to traditional machine learning, convolutional neural networks can extract deep semantic features of remote sensing images. On the other hand, ensemble learning can make reasonable selection and reorganization according to the characterization performance of each classifier, thereby improving the classification accuracy and its generalization ability.