基于ACRM模型与敏感波段的农作物LAI与LCC反演

Retrieving crop LAI and LCC based on their sensitive bands usingthe ACRM model

关键词:  叶面积指数  叶片叶绿素含量  冠层反射率模型  遥感反演  波段选择

Key words:  leaf area index  leaf chlorophyll content  canopy reflectance model  inversion  band selection

摘要:

【目的】面向现代农业生产和管理的数据需求，基于ACRM 冠层反射率模型，探索适 于冬小麦叶面积指数（LAI）和叶片叶绿素含量（LCC）反演的波段选择方案。【方法】文章 考虑高光谱数据降维和CR 模型模拟误差，选出覆盖蓝、绿、红与近红外的5 个波段（波段 选择方案B1），开展LAI 与LCC 同步反演。然后分别选择LAI 和LCC 的敏感波段，开展对 应参数的反演试验。【结果】（1）基于B1，能够在多数田块实现较为准确的LAI 与LCC 同 步反演（LAI 反演值与实测值间决定系数（R2）为0.860 4，均方根误差（RMSE）为0.963； LCC 反演的R2 为0.814 1，RMSE 为0.069）。（2）仅利用LAI 或LCC 敏感波段反演结果的R2 与RMSE 同时略有升高，但与基于B1 的反演结果相比，无明显差异。【结论】通过该研究与 利用相同数据的前期研究对比发现，旨在高光谱数据降维与限制CR 模型模拟误差的波段选 择，对LAI 反演精度改进作用较为显著。相较而言，仅选用单一目标参数（LAI 或LCC）的 敏感波段，对反演精度改进并不明显。由此，一方面证实了常规反演方法与面向对象反演法 不强调选用单一目标参数敏感波段的合理性；另一方面，并不否定多阶段目标决策（MSDT） 反演法以及一些相关研究提出的，仅采用单一目标参数敏感波段来开展反演的合理性。

Abstract:

[Purpose］Leaf area index（ LAI） and leaf chlorophyll content（ LCC） are promising variables for decision making in modern agriculture.Using remote sensing data，LAI and LCC can be retrieved simultaneously by inversing canopy reflectance（ CR） models.Such methodology is known for its better universality and less dependence on in-situ measurement.It has been stated by many studies that band selection is one of the key issues for retrieving crop variables based on a CR model.Aimed at monitoring LAI and LCC accurately for modern agriculture，we investigated the schemes of band selection for CR model inversion in this study，with particular attention on constraining the inversion by applying only the sensitive bands of LAI or LCC（ i.e.the spectral regions where LAI or LCC dominates the reflectance）.［Method］（1） a preliminary band selection was conducted for dimension reduction of hyperspectral data，and for eliminating the bands with significant discrepancies between the simulated and the remotely sensed spectra. This is realized by firstly assuming a combination of 5 bands，covering the spectral regions of blue，green，red and near-infrared. However，the exact bands were undetermined.Secondly， the bands in each spectral region，which achieved the optimum goodness of fitting between the simulated and the observed spectra，were selected.This scheme of band selection is denoted as B1. It was then tested for the simultaneous retrieval of LAI and LCC.（2） Based on B1，relevant studies，and a sensitivity evaluation on ACRM parameters using EFAST（ extended Fourier amplitude sensitivity test），the sensitive bands of LAI or LCC were selected respectively， denoted as B2-B5.And then，LAI or LCC was retrieved separately，using their sensitive bands only.［Result］Result shows that，（1） with B1，the LAI and LCC values in most（ 4 out of 5） fields can be retrieved simultaneously in reasonable accuracies（ R2=0.8604 and root-mean-square error（ RMSE）=0.963 for LAI，and R2=0.8141 and RMSE=0.0689 for LCC）.（2） The R2 and RMSE of the retrieved LAI or LCC based only on their sensitive bands are simultaneously higher than those based on B1. Nevertheless，their results showed no significant differences compared with the aforementioned results based on B1.［Conclusion］Comparing this study to our former studies using the same dataset，it can be found that band selection，which considering dimension reduction of hyperspectral data and avoiding errors of CR models，brings relevantly significant improvements on the retrieval accuracy of LAI. However，comparatively，the experiments in this study showed it was not so effective to constrain the inversion by using only the sensitive bands of a target variable. This study proved the rationality of conventional and object-based inversion approaches，in which constraining the inversion with only the sensitive bands of a target variable was not emphasized. Nevertheless，the potential of such constrain can neither be negated， according to the result of this study.