基于ZY-3卫星多光谱影像估算浙江省乔木林地上碳密度

目的基于覆盖浙江省的ZY-3卫星影像以及LULUCF碳汇监测样地数据，以浙江省乔木林地上碳密度为研究对象，尝试构建一个自动化提取浙江省乔木林地上碳密度的技术方法。方法分别在矢量标志建立、光谱信息提取、解译标志提纯、ZY-3卫星影像分类、自变量优选、建模方法优选、碳密度图制作等方面开展相关研究测试。结果本研究在解译标志提纯后对ZY-3影像进行分类的精度高于提纯前的影像分类精度；采用的kNN法对ZY-3影像进行分类的精度（平均总精度为80.31%，平均Kappa系数为0.69，乔木林平均用户精度为91.86%，乔木林平均生产者精度为80.85%）高于最大似然分类法（平均总精度为78.56%，平均Kappa系数为0.62，乔木林平均用户精度为89.68%，乔木林平均生产者精度为77.79%）；在选用的建模方法中，kNN法构建的模型精度（平均RMSE为15.64 t/hm2，平均RRMSE为23.53%）优于稳健估计法（平均RMSE为17.63 t/hm2，平均RRMSE为25.11%）。最后，生成了浙江省乔木林地上碳密度分布图。结论本研究可为省域或更大尺度范围的乔木林地上或森林碳密度估算提供一个新的路径，为实现自动化估算碳密度以及其他森林参数提供参考。 

Estimation of above-ground carbon density of arbor forest in Zhejiang Province of southern China based on ZY-3 satellite multispectral image

ObjectiveBased on the ZY-3 satellite imagery and the LULUCF carbon sink monitoring plot data covering Zhejiang Province of southern China, the study attempted to construct a technical method for automatically extracting the above-ground carbon density of arbor forest in this area.MethodTaking the carbon density of arbor forest in Zhejiang Province as the research object, relevant research tests were carried out in the aspects of vector sign constructing, extraction of spectral information, purification of interpretation sign, ZY-3 satellite image classification, optimization of independent variables, optimization of modeling methods, production of carbon density map, etc.ResultThe results showed that the accuracy of classification of ZY-3 imagery after purification of interpretation signs was higher than that of image classification before purification. The accuracy of classification of ZY-3 images by kNN method (average total accuracy was 80.31%, average Kappa coefficient was 0.69, average user accuracy of arbor forest was 91.86%, and the average producer accuracy of arbor forest was 80.85%), which was higher than the maximum likelihood classification method (average total accuracy was 78.56%, average Kappa coefficient was 0.62, average user accuracy of arbor forest was 89.68%, and the average producer accuracy of arbor forest was 77.79%). Among the selected modeling methods, the model accuracy constructed by the kNN method (average RMSE was 15.64 t/ha, average RRMSE was 23.53%) was better than the robust estimation method (average RMSE was 17.63 t/ha, average RRMSE was 25.11%). Finally, the above-mentioned carbon density distribution map of arbor forest in Zhejiang Province was generated.ConclusionThis study provides a new path for arbor forest or forest carbon density estimation at the provincial or larger scale, providing a reference for automated estimation of carbon density and other forest parameters.