基于随机森林的高寒湿地地区土地覆盖遥感分类方法

高寒湿地是青藏高原典型独特的生态系统,是全球气候变化的敏感地带和预警区。利用遥感技术快速、准确地分类提取高寒湿地的土地覆盖信息,对当地生态安全监测和保护具有重要意义。本文以若尔盖湿地国家级自然保护区为研究区,首先,以高分一号(GF-1)遥感影像为数据源,融合光谱特征、水体指数、地形特征、植被指数和纹理信息等26个变量进行随机森林(Random forest,RF)分类实验;然后,根据袋外数据(Out of bag,OOB)的特征变量重要性得分和精度评价结果,选出高寒湿地地区土地覆盖类型的最优分类方案和特征;最后,对特征变量进行降维,并基于相同的变量,采用极大似然法(Maximum likelihood classification,MLC)、支持向量机(Support vector machine,SVM)、人工神经网络(Artificial neural network,ANN)和RF等方法进行分类,比较不同方法的优适性。结果表明:结合GF-1影像光谱、水体、植被、纹理特征和地形信息,使用26个变量的RF模型的分类精度最高,总体精度(Overall accuracy,OA)为90.07%,Kappa系数为0.86;通过RF模型的变量重要性分析可以有效选出重要的特征信息,在降低特征变量维度的同时,还能保证较高的分类精度; 4种分类方法中,RF算法是高寒湿地地区较合适的分类方法,OA比MLC基准方法高17.63个百分点,比SVM和ANN等机器学习算法分别高6.98、6.56个百分点。

Land Cover Remote Sensing Classification Method of Alpine Wetland Region Based on Random Forest Algorithm

Alpine wetland is a typical and unique ecosystem in the Qinghai-Tibet Plateau, which is considered as a sensitive zone and early warning area of global climate change. Using remote sensing technology to extract land cover information of alpine wetland more quickly and accurately is of great significance to the monitoring and protection of local ecological security. Firstly, taking Zoige Wetland National Nature Reserve as the study area and GF-1 remote sensing image as the data source, the random forest (RF) classification experiments were carried out based on 26 variables, including spectral characteristics, water index, topography feature, vegetation index and texture information. Then, through the out of bag (OOB) feature variable importance score and accuracy evaluation results, the optimal classification scheme and characteristics of land cover types in the alpine wetland region were selected. Finally, the feature variables were dimensionally reduced, and based on the same variables, the maximum likelihood classification (MLC), support vector machine (SVM), artificial neural network (ANN) and RF were used to classify, and the applicability of different methods was compared. The results showed that combining with the spectral characteristics, water and vegetation index, texture feature of GF-1 image and topography information, the RF model with 26 variables reached the highest classification accuracy, the overall accuracy (OA) was 90.07%, and the Kappa coefficient was 0.86. Using the variable importance analysis of RF model, important feature information could be effectively selected. Based on the importance analysis of RF model, the important feature information can be effectively selected, the dimension of feature variables can be reduced, and high classification accuracy was ensured. Among the four classification methods, RF algorithm was the most ideal one at present, OA was 17.63 percentage points higher than that of MLC, and 6.98 percentage points and 6.56 percentage points higher than those of SVM and ANN respectively. The RF classification method combined with multiple remote sensing information and feature selection can quickly and efficiently classify the land cover types of alpine wetland region, providing a quick and feasible technical means for the monitoring of local alpine wetland.