Wetland inventory maps are essential information for the conservation and management of natural wetland areas. The classification framework is crucial for successful mapping of complex wetlands, including the model selection, input variables and training procedures. In this context, deep neural network (DNN) is a powerful technique for remote sensing image classification, but this model application for wetland mapping has not been discussed in the previous literature, especially using commercial WorldView-3 data. This study developed a new framework for wetland mapping using DNN algorithm and WorldView-3 image in the Millrace Flats Wildlife Management Area, Iowa, USA. The study area has several wetlands with a variety of shapes and sizes, and the minimum mapping unit was defined as 20 m² (0.002 ha). A set of potential variables was derived from WorldView-3 and auxiliary LiDAR data, and a feature selection procedure using principal components analysis (PCA) was used to identify the most important variables for wetland classification. Furthermore, traditional machine learning methods (support vector machine, random forest and k-nearest neighbor) were also implemented for the comparison of results. In general, the results show that DNN achieved satisfactory results in the study area (overall accuracy = 93.33 %), and we observed a high spatial overlap between reference and classified wetland polygons (Jaccard index ∼0.8). Our results confirm that PCA-based feature selection was effective in the optimization of DNN performance, and vegetation and textural indices were the most informative variables. In addition, the comparison of results indicated that DNN classification achieved relatively similar accuracies to other methods. The total classification errors vary from 0.104 to 0.111 among the methods, and the overlapped areas between reference and classified polygons range between 87.93 and 93.33 %. Finally, the findings of this study have three main implications. First, the integration of DNN model and WorldView-3 image is useful for wetland mapping at 1.2-m, but DNN results did not outperform other methods in this study area. Second, the feature selection was important for model performance, and the combination of most relevant input parameters contributes to the success of all tested models. Third, the spatial resolution of WorldView-3 is appropriate to preserve the shape and extent of small wetlands, while the application of medium resolution image (30-m) has a negative impact on the accurate delineation of these areas. Since commercial satellite data are becoming more affordable for remote sensing users, this study provides a framework that can be utilized to integrate very high-resolution imagery and deep learning in the classification of complex wetland areas.

湿地清单图是自然湿地保护和管理的重要信息。分类框架对于成功绘制复杂的湿地至关重要，包括模型选择，输入变量和培训程序。在这种情况下，深度神经网络（DNN）是一种用于遥感图像分类的强大技术，但是在以前的文献中，尤其是使用商业WorldView-3数据时，尚未讨论该模型在湿地制图中的应用。这项研究为美国爱荷华州Millrace Flats野生动物管理区的DNN算法和WorldView-3影像开发了湿地地图绘制的新框架。研究区域有数个形状和大小各异的湿地，最小测绘单位定义为20 m ²（0.002公顷）。从WorldView-3和辅助LiDAR数据中得出了一组潜在变量，并使用主成分分析（PCA）进行了特征选择程序来识别湿地分类中最重要的变量。此外，还采用了传统的机器学习方法（支持向量机，随机森林和k近邻）来比较结果。总体而言，结果表明，DNN在研究区域中取得了令人满意的结果（总精度= 93.33％），并且我们观察到参考湿地多边形和分类湿地多边形之间的空间重叠很大（Jaccard指数〜0.8）。我们的结果证实，基于PCA的特征选择对于DNN性能的优化是有效的，并且植被和质地指数是信息最多的变量。此外，结果比较表明，DNN分类获得了与其他方法相对相似的准确性。在这些方法中，总分类误差从0.104到0.111不等，参考多边形和分类多边形之间的重叠区域介于87.93％和93.33％之间。最后，这项研究的发现具有三个主要含义。首先，DNN模型和WorldView-3影像的集成对于在1.2米处进行湿地制图很有用，但是DNN的结果并没有优于该研究区域中的其他方法。其次，特征选择对于模型性能非常重要，并且最相关的输入参数的组合有助于所有测试模型的成功。第三，WorldView-3的空间分辨率适合保留小湿地的形状和范围，而中分辨率图像（30-m）的应用会对这些区域的精确描绘产生负面影响。由于商业卫星数据对于遥感用户而言变得越来越负担得起，因此本研究提供了一个框架，可用于将高分辨率的图像和深度学习集成到复杂湿地区域的分类中。