甘肃黑河流域上游森林地上生物量的多光谱遥感估测

Estimating Forest Above-Ground Biomass in the Upper Reaches of Heihe River Basin Using Multi-Spectral Remote Sensing

【目的】以黑河流域上游祁连山森林保护区为研究区,利用133个森林样地调查数据、Landsat-5 TM影像和ASTER GDEM产品为数据源,探讨地形对该流域森林地上生物量(above-ground biomass,AGB)估测的影响,以及选择合适的遥感估测方法反演该流域的森林AGB。【方法】首先利用青海云杉特殊的生境范围和绿色植被对比值植被指数(ratio vegetation index,RVI)的灵敏程度,及不同地物对纹理特征的不同响应,制定相应的决策树分类器,将研究区的土地覆盖类型分为两大类:森林(青海云杉)-非森林,并利用133个森林样地调查数据和Google Earth高分辨率影像的12 722个采样点对分类结果进行验证(总体分类精度达到90.39%,Kappa系数为0.81);然后运用多元线性逐步回归估测法,以及结合随机森林算法(random forest,RF)优化后的k最近邻分类法(k-nearest neighbors,k-NN)进行森林AGB的遥感估测,对比SCS+C地形校正前后青海云杉森林AGB的估测结果,同时比较2种不同估测方法的反演效果;最后利用得到的最优估测方法反演整个研究区的森林AGB,生成黑河流域上游祁连山森林保护区的森林AGB的等级分布图。【结果】SCS+C地形校正前多元线性逐步回归的估测精度为R2=0.31,RMSE=34.41 t·hm-2,地形校正后多元线性逐步回归的估测精度为R2=0.46,RMSE=30.51 t·hm-2;而基于SCS+C地形校正后的k-NN的交叉验证精度不仅明显高于地形校正前的精度,且显著优于多元线性逐步回归的估测结果,达到R2=0.54,RMSE=26.62 t·hm-2;另外基于最优的k-NN估测模型(窗口为7×7,采用马氏距离,k=3)反演的该流域青海云杉在2009年总的森林地上生物量为8.4×107t,平均森林地上生物量为96.20 t·hm-2。【结论】在地形复杂地区,运用SCS+C模型对地形进行适当校正,能够有效地消除太阳入射角变化引起的地表反射亮度的差异,使影像能够更准确地反映地表信息,提高森林AGB的遥感估测精度;在样本有限的情况下,相对于以大数定律作为理论基础的多元线性逐步回归估测法,k-NN能够避免发生过学习现象和样本不平衡问题,更适于该研究区青海云杉的森林AGB的估测。

【Objective】Forest biomass is the main source of energy and nutrients of the forest ecosystem operation.Qilian Mountain forest reserve at the upper reaches of Heihe River Basin was selected as the research area. The forest inventory data,Landsat-5 TM images and ASTER GDEM products were used as data sources. The purpose of this paper is to explore the effect of terrain on the estimation of forest above-ground biomass（ AGB） and select appropriate method for the inversion of forest AGB. 【Method】First,a decision-tree classifier was constructed by taking into account of the special habitat of Picea crassifolia and the sensitivity of the green vegetation for ratio vegetation index,and the different responses of various objects on the texture features. The land-cover types of the research area was divided into two categories： forest（ Picea crassifolia） —non-forest. The accuracy assessment of classification map was obtained by using field inventory data and high-resolution image of Google Earth（ The overall accuracy of the classification is 90. 39%,and the Kappa coefficient is 0. 81）. Then,the forest AGB was estimated using the multiple linear stepwise regression and k-NN. The k-NN was implemented by combining with RF algorithm. The change of the estimation accuracy before and after the topographic correction was analyzed. And the estimation accuracy of two different retrieval methods were compared with the forest survey data. Finally,the grade distribution of regional forest AGB was performed by the optimal estimation method.【Result】The estimation accuracy of multiple linear regression was R2= 0. 31,RMSE = 34. 41 t !hm- 2before SCS ＋ C topographic correction. But it was R2= 0. 46,RMSE = 30. 51 t !hm- 2after SCS ＋ C topographic correction. The optimal k-NN produced higher cross-validation accuracy（ R2= 0. 54,RMSE = 26. 62 t !hm- 2） by using the data after SCS ＋ C topographic correction than the outcome before SCS ＋ C topographic correction. At the same time,it performed better than the effect of the multiple linear stepwise Regression. The regional forest AGB which was performed by the optimized k-NN（ window sampling size was 7 × 7; distance measures was Mahalanobis Distance; k was 3） showed that the total of forest AGB of Picea crassifolia was 8. 4 × 107 t in this region,and the average was 96. 20 t !hm- 2. 【Conclusion 】The appropriate terrain correction with SCS ＋ C model could effectively eliminate the influence of the change of incident angle of the sun in complex terrain area. It could improve the estimation accuracy of the models. Compared with multiple linear stepwise regression,the optimal k-NN could avoid the phenomenon of learning and the problem of sample imbalances in the case of limited samples.