联合GF-1和GF-3影像的森林地上生物量反演

Inversion of forest aboveground biomass from combined images of GF-1 and GF-3

【目的】探索高分(GF)光学、合成孔径雷达(SAR)数据及其联合数据在森林地上生物量(AGB)及其组成部分反演中的可行性。【方法】以云南省昆明市宜良县小哨林区的云南松为研究对象,结合实地调查数据,以GF-1光学数据和GF-3SAR数据作为数据源,提取光学数据常用的植被指数和纹理特征,SAR数据的各极化后向散射系数、纹理特征以及极化分解等参数,利用KNN-FIFS方法分别进行森林AGB及其分量的反演;然后采用留一交叉验证法对反演结果进行精度评价,并在此基础上绘制森林AGB及其分量空间分布图。【结果】联合GF-1和GF-3数据反演森林AGB及其分量的精度最高,R^(2)均超过了0.710,RMSEr的值在22%^(2)7%之间,其中树叶的反演精度最优,模型的R^(2)为0.714,RMSE为10.270 t/hm^(2),RMSEr为24.58%;除树叶生物量外,森林AGB和其他分量仅采用GF-1提取的特征进行反演时,精度均优于采用GF-3特征的反演结果。【结论】联合GF-1光学数据和GF-3全极化SAR可以实现一定程度的互补,提高森林AGB及其分量的反演精度,此外KNN-FIFS方法在低生物量水平的云南松纯林的AGB及其分量的反演中具有一定的鲁棒性,且KNN-FIFS优选的重要参数多为SAR和光学的纹理特征。

[Objective] The objective of this study was to explore the feasibility of GF-1, GF-3, and combination of GF-1 and GF-3 for total forest aboveground biomass(AGB) and its component inversion.[Method] In this study, the vegetation indices, texture characterizations extracted from GF-1,backscattering coefficients, texture characterizations and polarimetric decomposition features extracted from GF-3 were used independently and in combination to estimate total AGB and component AGB of a pure forest of Yunnan pine(Pinus yunnanensis), located in Xiaoshao Forest Region in Yiliang County, Yunnan Province of southwestern China. A fast iterative features selection method for k-NN method(KNN-FIFS)was applied in forest total AGB and component AGB inversion and the leave one out cross validation(LOOCV) method was used to evaluate the model and the inversion results and the results were mapped and analyzed. [Result] The joint GF-1 and GF-3 data had the highest accuracy for inversion of forest AGB and each component AGB, and the R^(2) of each exceeded 0.710, and the values of RMSEr were between 22% and 27%, among which the inversion accuracy of foliage was the best, with the model’s R^(2) of 0.714, RMSE of 10.270 t/ha, and RMSEr of 24.58%;except for foliage component AGB, forest AGB and other components of AGB had better accuracy than the inversion results with GF-3 features when only the features extracted from GF-1 were used for the inversion. [Conclusion] Combining GF-1 optical data and GF-3 fully polarized SAR data can achieve a certain degree of complementarity to improve the inversion accuracy of forest AGB and fractional AGB. In addition, the KNN-FIFS method is robust in the inversion of AGB and fractional AGB of pure Yunnan pine forests at low biomass levels, and the important parameters preferred by KNN-FIFS mostly are texture features extracted form SAR and optics data.