Abstract Wetlands provide a wide variety of environmental services globally and detailed wetland inventory maps are always necessary to determine the conservation strategies and effectively monitor these productive ecosystems. During the last two decades, satellite remote sensing data have been extensively used for wetland mapping and monitoring worldwide. Polarimetric Synthetic Aperture Radar (PolSAR) imagery is a complex and multi-dimensional data, which has high potential to discriminate different land cover types. However, despite significant improvements to both information content in PolSAR imagery and advanced classification approaches, wetland classification using PolSAR data may not provide acceptable classification accuracy. This is because classification accuracy using PolSAR imagery strongly depends on the polarimetric features that are incorporated into the classification scheme. In this paper, a novel feature weighting method for PolSAR imagery is proposed to increase the classification accuracy of complex land cover. Specifically, a new coefficient is determined for each element of the coherency matrix by integration of Fisher Linear Discriminant Analysis (FLDA) and physical interpretation of the PolSAR data. The proposed methodology was applied to multi-temporal polarimetric C-band RADARSAT-2 data in the Avalon Peninsula, Deer Lake, and Gros Morne pilot sites in Newfoundland and Labrador, Canada. Different combinations of input features, including original PolSAR features, polarimetric decomposition features, and modified coherency matrix were used to evaluate the capacity of the proposed method for improving the classification accuracy using the Random Forest (RF) algorithm. The results demonstrated that the modified coherency matrix obtained by the proposed method, Van Zyl, and Freeman-Durden decomposition features were the most important features for wetland classification. The fine spatial resolution maps obtained in this study illustrate the distribution of terrestrial and aquatic habitats for the three wetland pilot sites in Newfoundland using the modified coherency matrix and other polarimetric features. The classified maps provide valuable baseline data for effectively monitoring climate and land cover changes, and support further scientific research in this area.

中文翻译：

---

使用 PolSAR 影像进行湿地分类的相干矩阵的 Fisher 线性判别分析

摘要 湿地在全球范围内提供各种各样的环境服务，详细的湿地清单地图对于确定保护策略和有效监测这些生产性生态系统始终是必要的。在过去的二十年中，卫星遥感数据已广泛用于全球湿地测绘和监测。极化合成孔径雷达 (PolSAR) 图像是一种复杂的多维数据，具有区分不同土地覆盖类型的潜力。然而，尽管 PolSAR 图像中的信息内容和高级分类方法都有显着改进，但使用 PolSAR 数据的湿地分类可能无法提供可接受的分类精度。这是因为使用 PolSAR 图像的分类精度在很大程度上取决于纳入分类方案的极化特征。在本文中，提出了一种新的 PolSAR 图像特征加权方法，以提高复杂土地覆盖的分类精度。具体而言，通过整合 Fisher 线性判别分析 (FLDA) 和 PolSAR 数据的物理解释，为相干矩阵的每个元素确定一个新系数。所提出的方法应用于加拿大纽芬兰和拉布拉多的阿瓦隆半岛、鹿湖和格罗斯莫讷试点站点的多时相极化 C 波段 RADARSAT-2 数据。输入特征的不同组合，包括原始PolSAR特征、极化分解特征、和修改的一致性矩阵被用来评估所提出的方法的能力，以使用随机森林（RF）算法提高分类精度。结果表明，该方法获得的修正相干矩阵、Van Zyl 和 Freeman-Durden 分解特征是湿地分类最重要的特征。本研究中获得的精细空间分辨率图使用修改后的相干矩阵和其他极化特征说明了纽芬兰三个湿地试验点的陆地和水生栖息地分布。机密地图为有效监测气候和土地覆盖变化提供了宝贵的基线数据，并支持该领域的进一步科学研究。