小麦生物量极化分解参数响应及反演

Inversion and Polarimetric Information Response Analysis of Wheat Biomass

生物量是反映农作物产量和生长健康状况的主要指标,然而直接进行农作物生物量估测不仅耗时耗力,而且具有破坏性。由于遥感技术在植被生物量反演中的方便、快捷及有效性,近年来不少研究者开始关注采用遥感技术进行农作物生物量的反演。小麦是包括中国在内的许多国家的主要粮食作物,寻找合适的小麦生物量精确估测方法在其产量估测中具有重要意义。以试验区小麦整个生长期内获取的5景全极化Radarsat^-2数据及同步的地面调查数据为基础,分析了Freeman-Durden和H/A/α两种极化分解的相关极化参数与其生物量的相关关系,并在此基础上构建了4种传统经验回归模型和随机森林(RF)非参数模型反演小麦生物量。另外,根据农作物散射的物理基础,结合Freeman-Durden分解参数,还构建了表面散射与整体散射能量比值参数、体散射与总体散射能量比值参数、体散射与表面散射比值参数和二次散射与整体散射能量比值参数(Odd/Span,Vol/Span,Vol/Odd and Dbl/Span)四个参数参与极化参数对小麦生物量的敏感性分析。研究中提取的各极化参数根据播后天数在各地块内取均值以降低随机性对提取的极化参数的影响,共计95块小麦地块用于研究。研究结果表明:Freeman-Durden分解的参数中,体散射分量、表面散射分量及这两者与总散射能量的比值、这两者的比值均表现出与小麦生物量的高相关性(R2>0.55);相比Freeman-Durden分解的参数,H/A/α分解参数中除反熵外,熵和散射角均表现出与生物量的高相关性(R2>0.70)。另外,小麦生物量反演模型的可行性研究结果表明:4种经验回归模型中,二次多项式和对数模型更适合采用极化参数对小麦生物量的反演。该类方法反演小麦生物量的最佳均方根误差为77.94g·m^-2,相对误差为29.05%。尽管采用随机森林的重要性参数排序中,H/A/α分解中的H和α参数也排序靠前,但是在单因子的经验回归中,其反演结果的误差较大,均方根误差为101.74g·m^-2,相对误差为37.91%。尽管如此,大多数非参数随机森林参数重要性排序结果与极化分解参数与生物量相关分析的结果基本一致,采用随机森林反演。小麦生物量的精度也有明显提高,均方根误差为54.53g·m^-2,相对误差为20.32%。

Biomass is one of the most useful indicators of crops vegetation development and health. It is also a good indicator of crop potential yield. However, direct biomass measurement is destructive and expensive. More recent estimates based on remote sensing technique have attracted the most interest because of its convenience, effectiveness and efficiency. Wheat is one of the main crops for both China and other countries, it is very important for its accurate yield estimation and finding the suitable approach of biomass inversion. In this study, 5 multi-temporal Radarsat^-2 quad polarimetric images were acquired during the entire growth cycle of wheat from sowing to near harvest. Meanwhile, 5 synchronous field campaigns were carried out at each satellite overpass with no lag more than one day to collect the information of wheat biomass. Then two popular quad polarimetric data decomposition methods like Freeman-Durden and H/A/α decomposition were chosen to extract polarimetric parameters. 6 decomposition parameters including surface scattering, dihedral scattering and volume scattering from Freeman-Durden decomposition and entropy, scattering alpha and anisotropy from H/A/α decomposition, the total power(Span) from Freeman-Durden decomposition and 4 ratios parameters including Odd/Span, Vol/Span, Vol/Odd and Dbl/Span were extracted from 5 multi-temporal Radarsat-2 images. The values of these extracted parameters were averaged according to day after sowing at each parcel in all95 wheat parcels. After these radar parameters were extracted, scatterplot of wheat biomass against day after sowing during the whole growth cycle was analyzed based on the comparison of the evolution of each polarimetric parameter against day after sowing with scatterplots. The coefficients of determination(R2) between wheat biomass and each polarimetric parameter was calculated with four empirical regression model to descript their relationship quantitatively. The five empirical regression models included linear function, quadratic function, power function, index function and logarithm function. The importance of these parameters was also computed by random forest(RF) parameter analyzed method. At last, the high fitting models with highest R2 and lowest root mean square error(RMSE) were selected for wheat biomass inversion. The results showed although several polarimetric parameters from both decomposition method revealed high sensitivity with wheat biomass, there are obvious distinct difference between wheat biomass and two kinds of decomposition parameters. When two decomposition methods were compared,the ratio parameters from Freeman-Durden decomposition showed better performance in wheat biomass inversion with empirical regression model. The RMSE of its results was 77.94 g·m^-2 and the relative error was 29.05%. Although H and α showed higher importance in RF parameter importance analysis, their performance for wheat biomass inversion with single parameter and empirical regression model was worse than many parameters from Freeman-Durden decomposition and their ratios. The RMSE of its results was 101.74 g·m^-2 and the relative error was 37.91%. Nonetheless, the inversion performance with RF, which combined polarimetric parameters from both decomposition methods, showed that the best inversion result with RMSE of 54.53 g·m^-2 and the relative error was 20.32%.