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等问题，提出了一种采用全卷积神经（Fully Convolutional Neural，FCN）网络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，采用全卷积神经网络与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。与采用全卷积神经网络与单一机器学习的随机森林（Random Forest，RF）、支持向量机（Support Vector Machin，SVM）与k-近邻（Nearest Neighbor，kNN）算法相比，该研究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个百分点；与采用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个百分点；该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 is playing a critical role in wetland ecological protection and fine management. In order to improve traditional classification of wetland, a new wetland classification that coupled with the Fully Convolutional Neural (FCN) network and ensemble learning was proposed in this study. A highly versatile feature extractor, including the convolutional layers and pooling layers in Convolutional Neural Network (CNN), can be used to extract highly abstract deep features in remote sensing images. However, the input image size in many CNN models was 256 × 256 or 299 × 299 pixels, which cannot satisfy the features information extraction of wetland in a large scale. In the FCN, the fully connected layer can be replaced with a convolutional layer, where the images with any size can be accepted for the pixel-by-pixel classification. That was the reason that the FCN was chosen to extract image features. Here, the FCN (SegNet, UNet, and RefineNet) was employed to extract and merge the deep semantic features in the GF-6 images. Since single machine learning was easy to fall into a local optimal solution, with the relatively low generalization ability of unknown samples, ensemble learning can be utilized as multiple base classifiers to predict. A strong ability can be gained to apply for various scenarios with high classification accuracy. Therefore, a stacking ensemble learning model was selected for wetland classification, according to the semantic features derived by FCN, due mainly to a better ensemble effect on the stable classifiers in the task. In some scenarios, the performance of integrated Stacking was better than that of others. However, the performance of integrated stacking may be degraded in some applications. An adaptive stacking was proposed to further improve the stability and generalization ability of current Stacking. The meta-classifier was first determined in the adaptive stacking. The RF in an ensemble classifier was used as the meta-classifier to improve prediction performance. All the base-classifiers were combined freely to train the input dataset, including the Support Vector Machine (SVM), RF, k-Nearest Neighbor (kNN), Logistic Regression (LR), and Naive Bayes (NB). The results showed that the coupled FCN and adaptive Stacking can effectively extract the most types of wetland information, where the overall classification accuracy and Kappa coefficient were 88.16% and 0.85, respectively. The producer accuracy and user accuracy of lakes/pools, mudflat, and sedge were all around 90%, but it was easy to form a misclassification of reed beaches and poplar forest beaches. The main reason was that similar spectral characteristics during the growing season were difficult to distinguish with single-phase remote sensing images. Compared with coupling FCN and single classifier (SVM, RF and kNN), the overall accuracy was improved by 5.31, 4.87, and 5.08 percent points, respectively. Compared with the SVM, RF and kNN, the overall accuracy was improved by 9.68, 7.90, and 10.01 percent points, respectively. Moreover, a higher classification accuracy was achieved, compared with that of SegNet, UNet, or RefineNet. According to the characterization performance of each classifier, ensemble learning can make reasonable selection and reorganization, further to improve the classification accuracy and its generalization ability.