基于时间序列遥感影像及DTW算法的塞罕坝林场树种识别研究

Research on tree species identification of Saihanba Forest Farm based on time series remote sensing image and DTW algorithm

森林树种的识别和分布是森林资源监测的重要内容,是森林生态规划的基础。以塞罕坝机械林场为研究区域,利用14个时相的哨兵2号遥感影像组成时间序列数据,构建红、绿、蓝、近红外四基础波段特征时间序列和包括归一化植被指数(Normalized differential vegetation index,NDVI)在内的五特征时间序列,对这2个多特征时间序列分别采用动态时间规整(Dynamic time warping,DTW)算法计算的距离作为相似性度量标准,然后利用K均值算法完成研究区域树种聚类,并对树种识别精度分别进行评价。研究结果显示,NDVI特征时间序列非常有效体现植被的物候信息,加入NDVI特征时间序列的五特征时间序列运用DTW- K 均值方法可以达到现实中树种分类调查精度的需求,总体分类的精度为88.58%,Kappa系数为 0.84;落叶松分布最为广泛,面积达到35 892.23 hm^2,分类精度为94.14%,多集中分布于研究区域北部和东南部;桦树次之,面积为18 376.24 hm^2,其分类精度为76.30%,东部地区桦树分布明显高于西部地区;樟子松、云杉和柞树的分布面积较少,分别为8 749.30 hm^2、1 217.08 hm^2和3 814.40 hm^2,分类精度分别为83.82%、69.88%、72.22%。

Forest tree species identification and distribution is the basis of forest resource monitoring and forest ecological planning. Taking Saihanba Machinery Forest Farm as the study area, the four multi-characteristic time series of red, green, blue, near infrared and the five multi-characteristic time series of red, green, blue, near infrared and NDVI were constructed by the time series data consisting of 14 time-phase Sentinel 2 remote sensing images, and for these two multi-feature time series, the distance calculated by Dynamic time warping algorithm was used as the similarity measurement standard. Then the K-means algorithm was used to complete tree species clustering in the study areas, and the accuracy of tree species identification was evaluated. The results show that the five characteristic time series with NDVI feature time series can effectively reflect the phenological information of vegetation, and the DTW-K means method can meet the demand of tree species classification survey precision in reality. The overall classification accuracy is 88.58%,and the Kappa coefficient is 0.84;Larch is the most widely distributed, with an area of 35 892.23 hm^2 and a classification accuracy of 94.14%, most of which are distributed in the north and southeast of the study area. The area of birch is 18 376.24 hm^2, and the precision of classification is 76.30%. The distribution of birch in the eastern region is obviously higher than that in the western area. The distribution areas of scotch pine, spruce and oak were 8 749.30 hm^2, 1 217.08 hm^2 and 3 814.40 hm^2, respectively. The classification accuracy of scotch pine, spruce and oak was 83.82%, 69.88% and 72.22%, respectively.