Landsat8光谱衍生数据分类体系下的牧草生物量反演

Grass Biomass Inversion Based on Landsat 8 Spectral Derived Data Classification System

牧草生物量的估算对于草地资源合理利用和载畜平衡监测具有重要的意义,是评价草地生态系统与草地资源可持续发展的关键指标。基于Landsat遥感技术快速、无损的大面积植被生物量估算研究已广泛应用,当前大多基于单一变量或几个常用植被指数构建反演模型,这些指数往往不能从多方面反映植被理化特征。归纳了不同Landsat8光谱衍生数据所反映的植被理化特征及它们间的关联方式,构建了Landsat8光谱衍生数据的分类体系;在此基础上提出了一种基于随机梯度Boosting(SGB)算法的多变量、非线性生物量估算模型,探讨不同类型光谱衍生数据组合对于牧草生物量反演结果的影响。以青海省海晏县为研究区进行方案可行性探讨。结果表明常用的Landsat8光谱衍生数据主要从植被的绿度、黄度、盖度、水分含量、纹理特征以及通过消除大气干扰和土壤背景干扰等7个方面反映植被的理化特征(7个小类),可归纳为直接因子(绿度、黄度、盖度、水分含量)、间接因子(消除大气干扰和消除土壤背景干扰)和空间因子(纹理特征)3大类型。在牧草生物量反演中,这些光谱衍生数据类型间具有较好的互补性,单一的直接因子模型估算结果最差,引入间接因子和空间因子均能提高模型的估算结果,而由直接因子(GNDVI, TCW, NDTI, NDSVI, TCD)、间接因子(SAVI, VARI)和空间因子(MeanB3, MeanB6, HomⅡ, DisB5)共同构建的SGB模型估算精度最优,R2达到了0.88;RMSE为141.00 g·m-2。与5种常用的生物量估算模型结果对比,该方法具有明显的优势。较单变量模型,R2提高了42%~60%,RMSE降低47%以上,R■提高了31%~53%, RMSEcv降低29%;较多变量模型,R2提高了29%~42%, RMSE降低35%以上,R■提高了2%~18%, RMSEcv降低2%以上。此外,所提出方法在消除反演模型过饱和方面也具一定成效。综上,利用Landsat8数据从反映植被不同理化特征角度构建反演模型实现了牧草生物量的精准估算,对于后期牧草生长状况实时监测以及草地资源可持续利用与管理具有重要的指导意义。研究结果还可以为今后进行大面积区域草地动态监测以及其他农业领域的研究提供参考和借鉴。

Estimation of forage biomass is of great significance for the rational use of grassland resources and monitoring of livestock load balance, and it is a key indicator for evaluating the sustainable development of grassland ecosystems and grassland resources. The rapid and non-destructive study of large-area vegetation biomass estimation based on Landsat remote sensing technology has been widely used. Most of the current researches are based on single variable or several commonly used vegetation indices to construct inversion models. These indices often cannot reflect the physical and chemical characteristics of vegetation inmany aspects. In this paper, the classification systems of different Landsat8-derived data were constructed by their corresponding physicochemical characteristics of vegetation andintersectional pattern with plants. A multivariable nonlinear biomass estimation model based on stochastic gradient boosting algorithm was proposed and the model estimation results were discussed with different combinations of derived data categories. The program feasibility study was carried out with Haiyan County in Qinghai Province as the study area. The results showed that the Landsat8-derived data reflected the physical and chemical characteristics of vegetation mainly from the aspects of vegetation greenness, yellowness, coverage, moisture content, texture characteristics and elimination of atmospheric disturbance and soil background interference(7 subcategories). On the other hand, these data can also be summarized into three categories: direct factors(greenness, yellowness, coverage, moisture content), indirect factors(eliminating atmospheric interference and eliminating soil background interference), and spatial factors(texture characteristics). The derived data categories have obvious complementarity. The direct factor(GNDVI, TCW, NDTI, NDSVI, TCD)-indirect factor(SAVI, VARI)-space factor(MeanB3, MeanB6, HomⅡ, DisB5) model had the best estimation accuracy, and R2 reached 0.88;the RMSE was 141.00 g·m-2, however the single direct factor model estimates result was the worst. Compared with the results of six typical biomass estimation models, the proposed method had obvious advantages. Compared with the univariate models, R2 increased by 42%~60%, RMSE decreased by more than 47%, R■ increased by 31%~53%, and RMSEcv decreased by more than 29%;Compared with the multivariate models, R2 increased by 29%~42%, RMSE decreased by more than 35%;and R■ increased by 2%~18%, RMSEcv decreased by more than 2%. In addition, the proposed model also had some effect in eliminatingoversaturation problem. In summary, this paper uses Landsat8 data to construct an inversion model from the perspective of reflecting different physical and chemical characteristics of vegetation to achieve accurate estimation of forage biomass, which has important guiding significance for the real-time monitoring of pastage growth and the sustainable use and management of grassland resources. The research results can also provide reference and reference for future large-area regional grassland dynamic monitoring and other agricultural research.