辐射传输模型模拟与深度学习结合的高分一号卫星植被光合有效辐射吸收比例产品反演算法

Fraction of absorbed photosynthetically active radiation inversion algorithm of GF-1 data combining radiative transfer model simulation and deep learning

植被光合有效辐射吸收比例FPAR (Fraction of absorbed Photosynthetically Active Radiation)是碳循环光能利用率模型中的关键参数之一。高分系列卫星的发射,为反演定量遥感产品提供了高时空分辨率的卫星遥感数据,基于高分数据的植被光合有效辐射吸收比例产品能够为生态系统碳循环的分析评估提供更加精细、精度更高的输入参数产品。本文发展了一种基于深度学习的光合有效辐射吸收比例反演方法。该方法利用SAIL(Scattering by Arbitrarily Inclined Leaves)模型模拟多种太阳入射角度、观测角度、大气条件下的植被冠层光合有效辐射吸收比例及冠层反射率,形成海量输入—输出模拟数据集,具有鲁棒性及更好的普适性;基于深度信念网络对数据集进行训练,得到高分一号(GF-1)卫星光合有效辐射吸收比例遥感反演模型。利用中国科学院怀来遥感综合试验站及黑河流域地表过程综合观测网FPAR地面站点连续观测数据对玉米作物、芦苇草地等下垫面反演的FPAR进行了对比验证,RMSE分别为0.15和0.17。本方法以辐射传输模型模拟的多维大气及地表输入—植被冠层输出作为深度学习的训练样本库,弥补了训练样本数量不足及观测数据不全带来的深度学习训练过程中的误差,从而使得模型兼具机理性和高效性。同时,反演的输入为具有太阳角度、观测角度信息的地表反射率产品,降低了输入参数获取的难度,减少输入参数误差传递的影响,有利于实现产品的业务化生产。

The Fraction of absorbed Photosynthetically Active Radiation(FPAR)is one of the key parameters in the light use efficiency model of the carbon cycle.High-spatiotemporal-resolution data have been provided for the inversion of quantitative remote sensing products since the launch of GF satellites.The FPAR products derived from GF satellite data provide precise and accurate input parameters for the analysis and evaluation of the ecosystem’s carbon cycle.In this study,a deep learning algorithm was developed to retrieve FPAR over China based on the simulated data of the radiative transfer model.The inputs are surface reflectance,cloud detection,and land cover products of GF-1 satellite data,whereas the output is FPAR.The FPAR product has a spatial resolution of 16 m and a temporal resolution of 10 days.This method uses the SAIL model to simulate output canopy FPAR and reflectance under various input variables,such as solar and observing angles and atmospheric conditions.The FPAR inversion model of GF-1 satellite data was obtained using a deep belief network.The long-term crop and grassland FPAR observation data in Huailai and Heihe were used to compare and validate the FPAR products,with a root mean square error of 0.15 and 0.17,respectively.The inversed FPAR is in good agreement with the measured FPAR in the low values,but lower than the measured FPAR in the high values.The radiative transfer model,the representativeness of the simulated data,and the preprocessing(calibration and geometric and atmospheric correction)of the GF-1 satellite data inevitably introduce some biases in the inversion process.This method uses the multidimensional atmospheric and surface variables as the input and the simulated vegetation canopy by the radiative transfer model parameters as the output.The simulated dataset,used as the training samples for deep learning,makes up for the errors in the deep learning training process caused by the insufficient number of training samples and incomplete observation data.The input of the inversion is only the surface reflectance product with the information of the sun angle and the observation angle.It lessens the difficulty of obtaining input parameters,reduces the influence of the error transmission of the input parameters,and is conducive to the realization of the commercial production of the product.The high FPAR is mainly distributed in the northeast,north,central,east,southwest,and south parts of China.The interannual variation of the FPAR time series,combined with the vegetation growing cycle and phenology,is high in spring and summer and low in autumn and winter.