基于最大熵模型的玉米冠层LAI升尺度方法

Upscaling method for corn canopy LAI using MaxEnt model

叶面积指数(leaf area index,LAI)是表达农作物冠层结构的关键参数之一,准确获取LAI对于农作物长势监测、估产等研究具有非常重要的意义。由于地物空间复杂性、数据源的不同以及遥感反演模型的非线性,LAI的反演结果会存在尺度效应,因此需要进行尺度转换研究。理想的升尺度转换应该只是数据空间分辨率的降低,而数据内在信息应保存到低分辨率中。最大熵(maximum entropy,Max Ent)模型是基于多种环境因子的广义学习模型,对分析因子的空间分布具有较高的估算精度,因此,该研究利用最大熵模型进行玉米冠层LAI升尺度方法研究,从而将野外实测的LAI点数据扩展到空间分辨率为30 m的面数据,所使用的数据源是Landsat8 OLI遥感影像、气象数据和野外样点上测量的LAI数据。研究结果表明:利用最大熵模型升尺度转换结果与实测LAI相比,R2为0.601、RMSE为0.898,说明两者的相关性较高;由于玉米冠层叶片之间的相互遮挡,导致整体结果偏低,但偏低误差在可接受范围内。因此,Max Ent模型可用于农作物LAI点数据到面数据的升尺度转换。

Leaf area index (LAI) is one of the key parameters to show canopy structure of crops. It is of great importance to obtain accurate LAI for monitoring and estimating the yield. Unfortunately, the LAI estimation results have scale effect resulting from feature space complexity, the difference of different remote sensing data source, and the nonlinear of remote sensing inversion model. So the scaling transformation is necessary when the multi-source remote sensing images are used. Scaling transformation is a process that extending the information and knowledge gained from a scale to other scales, including upscaling and downscaling. For a good upscaling method, the inherent information of high resolution image should be kept in low resolution image although its spatial resolution has been reduced. Statistical relationships between data for upscaling based on remote sensing products, and pixel decomposition method for upscaling the existed spatial continuously data from high-resolution to low resolution are commonly used. But most of agronomy parameters data are from site observation and sampling, these point data are informative and accurate. But the point locations are separated and dispersed, and space representatives are limited. So it is practical to up-scale these point data to spatial continuously data. There are regression analysis method, geo-statistical method, fractal method and others for point data’s upscaling, in which counter point sample size requires a high quality, and short of sufficient consideration on effect from various supplementary information related to the object to the sample point. MaxEnt is a general-purpose machine learning method with a simple and precise mathematical formulation. So the MaxEnt model was used in this study for upscaling maize canopy LAI from point data to spatial continuously data. The LAI point data measured in field work, Landsat8 OLI remote sensing image and the meteorological data were the data sources in this study. Firstly, classified the measured LAI data, getting meteorology temperature, surface temperature and relative humidity through spatial interpolation, spectral reflectance and spatial continuously data of vegetation index from remote sensing inversion. The data above could be used as environment variables, and distribution probabilities in different kinds of LAI data could be obtained through the MaxEnt Model. Secondly, we made a superposition for LAI probabilistic forecasting pictures, which converted LAI probability value to quantitative value. Lastly, we classified the research areas, and provided crops planting region, and then made a mask processing for the conversion results using the resulted pictures of classified fields so that we could get the results of LAI upscaling for maize canopy. The result showed that, comparing the upscaling conversion result with measured LAI data through the MaxEnt Model, we could found thatR2equaled 0.601 and RMSE was 0.898, indicating a high correlation between the two. The mutual occlusion between maize canopy leaves, resulting in an overall low result within an acceptable range. Therefore, the MaxEnt model can be used to upscaling crop canopy LAI from point data to spatial continuously data, and this method can be used to other crops as well.