Inland Marsh (IM) is a type of wetland characterized by the presence of non-woody plants as grasses, reeds or sedges, with a water surface smaller than 25% of the area. Historically, these areas have been suffering impacts related to pollution by urban, industrial and agrochemical waste, as well as drainage for agriculture. The IM delineation allows to understand the vegetation and hydrodynamic dynamics and also to monitor the degradation caused by human-induced activities. This work aimed to compare four machine learning algorithms (classification and regression tree (CART), artificial neural network (ANN), random forest (RF), and k-nearest neighbors (k-NN)) using active and passive remote sensing data in order to address the following questions: (1) which of the four machine learning methods has the greatest potential for inland marshes delineation? (2) are SAR features more important for inland marshes delineation than optical features? and (3) what are the most accurate classification parameters for inland marshes delineation? To address these questions, we used data from Sentinel 1A and Alos Palsar I (SAR) and Sentinel 2A (optical) sensors, in a geographic object-based image analysis (GEOBIA) approach. In addition, we performed a vectorization of a 1975 Brazilian Army topographic chart (first official document presenting marsh boundaries) in order to quantify the marsh area losses between 1975 and 2018 by comparing it with a Sentinel 2A image. Our results showed that the method with the highest overall accuracy was k-NN, with 98.5%. The accuracies for the RF, ANN, and CART methods were 98.3%, 96.0% and 95.5%, respectively. The four classifiers presented accuracies exceeding 95%, showing that all methods have potential for inland marsh delineation. However, we note that the classification results have a great dependence on the input layers. Regarding the importance of the features, SAR images were more important in RF and ANN models, especially in the HV, HV + VH and VH channels of the Alos Palsar I L-band satellite, while spectral indices from optical images were more important in the marshes delineation with the CART method. In addition, we found that the CART and ANN methods presented the largest variations of the overall accuracy (OA) in relation to the different parameters tested. The multi-sensor approach was critical for the high OA values found in the IM delineation (> 95%). The four machine learning methods can be accurately applied for IM delineation, acting as an important low-cost tool for monitoring and managing these environments, in the face of advances in agriculture, soil degradation and pollution of water resources due to agrochemical dumping.

中文翻译：

---

内陆沼泽划界数据挖掘技术与多传感器分析的比较

内陆沼泽（IM）是一种湿地，其特征是存在非木本植物，例如草，芦苇或莎草，水面小于面积的25％。从历史上看，这些地区一直受到与城市，工业和农业化学废物污染以及农业排水有关的影响。IM勾画不仅可以了解植被和水动力动力学，还可以监测人为活动引起的退化。这项工作旨在使用主动和被动遥感数据比较四种机器学习算法（分类和回归树（CART），人工神经网络（ANN），随机森林（RF）和k近邻（k-NN））。为了解决以下问题：（1）四种机器学习方法中哪一种最具内陆沼泽划界潜力？（2）SAR特征对内陆沼泽划界是否比光学特征更重要？（3）内陆沼泽划界最准确的分类参数是什么？为了解决这些问题，我们使用了基于地理对象的图像分析（GEOBIA）方法中的Sentinel 1A和Alos Palsar I（SAR）和Sentinel 2A（光学）传感器的数据。此外，我们对1975年巴西陆军地形图（显示沼泽边界的第一个官方文件）进行了矢量化处理，以便通过将其与Sentinel 2A图像进行比较来量化1975年至2018年之间的沼泽面积损失。我们的结果表明，整体精度最高的方法是k-NN，率为98.5％。RF，ANN，CART方法分别为98.3％，96.0％和95.5％。四个分类器的准确度均超过95％，表明所有方法都具有划定内陆沼泽的潜力。但是，我们注意到分类结果对输入层有很大的依赖性。关于这些功能的重要性，SAR图像在RF和ANN模型中更为重要，尤其是在Alos Palsar I L波段卫星的HV，HV + VH和VH通道中，而光学图像的光谱指数在用CART方法划定轮廓。此外，我们发现CART和ANN方法相对于所测试的不同参数，呈现出整体精度（OA）的最大变化。对于IM轮廓中发现的高OA值（> 95％），多传感器方法至关重要。