基于物理模型训练神经网络的作物叶面积指数遥感反演研究

叶面积指数(LAI)是估算作物生长的关键参数。基于物理模型的LAI反演,被认为是当前最为可靠的方法,但其反演复杂。本文提出了将物理模型和神经网络相结合,从地表反射率反演叶面积指数的算法,利用MOD IS地表反射率和4-scale模型反演作物LAI。(1)利用4-scale模型模拟不同LAI与地表反射率的关系,生成训练数据;(2)利用模型模拟的LAI训练神经网络;(3)以MOD IS地表反射率输入训练后的神经网络,反演LAI。估算的LAI与其他LAI产品进行了比较,结果表明,估算的作物LAI和MOD IS及CYCLOPES LAI产品空间和时间分布一致,均方根误差分别为0.4994和0.6558。以2004年衡水的作物LAI地面观测数据进行了直接验证,估算的LAI与研究区地表植被分布一致,但是,三种卫星LAI产品都小于地表测量,故需针对华北平原浓密作物设计模型参数化方案。

Estimation of Crop LAI Based on the Neural Network Trained from Physical Model

Leaf area index（LAI） is an essential parameter for monitoring crop growth dynamic.In this paper,an algorithm,which is based on physical model and neural networks to derive leaf area index from land surface reflectance data,is presented.The algorithm utilizes MODIS land surface reflectances and 4-scale model to produce crop LAI.Firstly,the training dataset was created by running the 4-scale model to simulate crop LAI at different land surface reflectance and geometric situation.Then,the neural network was trained with the simulated LAI data set.After the neural network was trained,the LAI would be efficiently retrieved from MODIS reflectances and geometric data.This algorithm directly utilizes the directional reflectances instead of the BRDF normalized data in order to avoid complex BRDF normalization and the error from it.The estimated LAI is compared with existing LAI products.The results show that it is consistent with MODIS（RMSE=0.4994） and CYCLOPES（RMSE=0.6658） LAI products in temporal and spatial patterns.The algorithm is validated against ground measurements of annual crop LAI of 2004 in Hengshui,Hebei Province,China.The neural network derived LAI could represent the spatial pattern of the field LAI.However,all these LAI products are lower than field measurements.It would be suggested that the physical model should be modified to adapt to the dense crop in Northern China.