高级检索
    李春梅, 张王菲, 李增元, 陈尔学, 田昕. 基于多源数据的根河实验区生物量反演研究[J]. 北京林业大学学报, 2016, 38(3): 64-72. DOI: 10.13332/j.1000-1522.20150209
    引用本文: 李春梅, 张王菲, 李增元, 陈尔学, 田昕. 基于多源数据的根河实验区生物量反演研究[J]. 北京林业大学学报, 2016, 38(3): 64-72. DOI: 10.13332/j.1000-1522.20150209
    LI Chun-mei, ZHANG Wang-fei, LI Zeng-yuan, CHEN Er-xue, TIAN Xin. Retrieval of forest above-ground biomass using multi-source data in Genhe, Inner Mongolia[J]. Journal of Beijing Forestry University, 2016, 38(3): 64-72. DOI: 10.13332/j.1000-1522.20150209
    Citation: LI Chun-mei, ZHANG Wang-fei, LI Zeng-yuan, CHEN Er-xue, TIAN Xin. Retrieval of forest above-ground biomass using multi-source data in Genhe, Inner Mongolia[J]. Journal of Beijing Forestry University, 2016, 38(3): 64-72. DOI: 10.13332/j.1000-1522.20150209

    基于多源数据的根河实验区生物量反演研究

    Retrieval of forest above-ground biomass using multi-source data in Genhe, Inner Mongolia

    • 摘要: 森林是陆地生态系统的重要组成部分,精确估测森林地上生物量能够减少陆地生态系统碳储量的不确定性。本文以内蒙古大兴安岭根河实验区为研究区,基于森林样地调查数据、Landsat 8 OLI、机载P-波段PolSAR以及ASTER GDEM数据,分别采用多元线性逐步回归法和基于随机森林算法(Random Forest, RF)进行特征优化选择后的k-最近邻(k-nearest neighbors, k-NN)法对研究区森林地上生物量(above-ground biomass, AGB)进行估测,对比验证采用不同类型数据(单传感器数据和多传感器数据)时2种方法的反演结果来寻求森林AGB估测的最优方法和输入因子,最后利用最优的估测方法来反演整个研究区的森林AGB,生成根河实验区的森林AGB等级分布图。结果表明:对于多元线性逐步回归和k-NN 2种不同的方法,森林AGB的反演都表现出较为一致的结果,即采用多传感器遥感数据(Landsat 8 OLI和机载P-波段PolSAR数据)比采用单传感器遥感数据估算的森林AGB精度要高;而在同时采用多传感器遥感数据进行森林AGB的反演中,k-NN算法的估测结果(R2=0.65, RMSE=17.49 t/hm2)明显优于多元线性逐步回归算法(R2=0.36, RMSE=22.08 t/hm2)的估测结果。显然,多源数据协同反演森林AGB可以充分利用每种传感器的优点,提高遥感估测森林AGB的能力;与多元逐步回归方法相比,k-NN算法能够更多地考虑到森林参数同光谱值之间的非线性依赖关系,且能够避免发生过学习现象和样本不平衡问题。

       

      Abstract: Forest is an important component of terrestrial ecosystems; therefore, it is necessary to estimate the forest above-ground biomass (AGB) accurately in order to reduce the uncertainty of the carbon stock in forest ecosystem. We estimated forest AGB of the Genhe forest reserve which is located in Inner Mongolia using Landsat 8 OLI image, P-band PolSAR image and ASTER GDEM product based on the multiple linear stepwise regression model and k-nearest neighbors (k-NN) model. In particular, the Random Forest (RF) was applied to select the features for constructing the optimized k-NN. The results estimated by single-sensor and multi-sensor data were compared by the accuracy indicators of R2 and RMSE, aiming to understand the effects of data source on the estimation of forest AGB. Then regional forest AGB over the Genhe forest reserve was estimated by the optimal method. Validated against the field forest measurement, the estimation of forest AGB obtained from multi-sensor outperformed those obtained from single-sensor based on the multiple linear stepwise regression model and k-nearest neighbors (k-NN) model; the estimation of forest AGB obtained from k-NN (R2=0.65, RMSE=17.49 t/ha) agreed better with the field forest measurement than that obtained from the multiple linear stepwise regression model (R2=0.36, RMSE=22.08 t/ha) using multi-sensor data. The ability of estimating forest AGB using remote-sensing-based method was improved attributed to the integration of the advantages of multi sensor; The k-NN model is a more appropriate method to estimate the forest AGB over regional area than the multiple linear stepwise regression model, because the k-NN model focuses on the nonlinear dependence between forest parameters and spectral values, and it can avoid the problem of over learning and sample imbalances.

       

    /

    返回文章
    返回