基于无人机多光谱的沼泽植被识别方法

为了探究高分辨率无人机多光谱数据对沼泽植被群丛识别能力，在洪河国家级自然保护区的核心区、缓冲区和实验区分别建立典型样区，通过低空无人机搭载的RGB及多光谱相机获取研究区正射影像，构建多维数据集并确立4种分类方案.采用面向对象的随机森林（RF）算法，对输入的多维数据集进行变量选择和参数（mtry、ntree）调优，构建适合沼泽植被群丛尺度识别模型.结果表明：优化的面向对象的RF算法对沼泽湿地植被具有较高的识别能力，在95%的置信区间内，核心区方案四（结合了光谱波段、纹理特征、几何特征、位置特征、地表高程信息和植被指数）获得最高总体精度为87.12%，kappa系数为0.850，比方案二（结合了光谱波段、几何特征和位置特征）总体精度高12.27%，kappa系数高0.140；对于单一典型沼泽湿地植被识别精度中，芦苇获得最高的用户精度高于88%，生产者精度高于90%，小叶章的生产者精度高于85%，但是在核心区用户精度较低，仅为78%.该方法可以作为沼泽植被群丛识别的有效方法，为研究沼泽湿地生态环境变化提供更准确的数据支持.

Classification method of swamp vegetation using UAV multispectral data

This paper established machine learning models to classify swamp vegetation communities based on high-resolution UAV multispectral images. In Honghe National Nature Reserve, typical sample areas were selected in the core area, buffer zone and experimental area and ortho-images of these areas were acquired using low-altitude UAVs with RGB and multispectral cameras. Multidimensional datasets were then derived from multiresolution segmentation of ortho-images, and established four classification scenarios. The object-based random forest (RF) algorithm was used to classify vegetation communities after feature selection and parameters (mtry and ntree) optimization and tuning. This algorithm also could rank the importance of each feature in multidimensional datasets and eliminat data redundancy accordingly. The results showed that:The optimized object-based RF algorithm had a high recognition ability for swamp vegetation. The scenario 4 (combination of spectral bands, texture features, geometric features, location features, surface elevation information and vegetation indexes) in the core area obtained the highest overall accuracy (87.12%), and the kappa value was 0.850 at the 95% confidence interval, which was 12.27% higher than scenario 2 (combining spectral bands, geometric features and location features), and the kappa value improved 0.140; For an identification accuracy of typical swamp vegetation, the classification of the reed achieved the highest user's accuracy of above 88%, and its producer's accuracy was higher than 90%. The classification of Calamagrostis angustifolia also achieved over 85% of producer's accuracy, but its user's accuracy (78%) was lower in the core area. This method can be used as an effective method to identify swamp vegetation communities and provide more accurate data support for studying dynamic changes of wetland ecological environment.