基于模拟退火算法的BP神经网络模型估算高分辨率叶面积指数

卫星遥感技术的快速发展使得获取全球大范围叶面积指数成为可能，但基于现有的算法和数据估算高分辨率LAI的精度还需要提高。针对农作物、草地和林地等3种典型地表类型，选取地面观测数据较多的4个研究区，包括3个各地类用于建模验证的研究区与一个用于适用性验证的独立研究区，针对4个研究区，分别获取地面测量数据以及对应的30 m空间分辨率地表反射率数据。在3个主要研究区建立并比较了NDVI植被指数经验模型、BP神经网络模型和基于模拟退火算法的BP神经网络模型，利用地面实测数据对模型进行验证。结果表明：在研究所选的3个主要研究区，基于模拟退火算法的BP神经网络模型的估算精度比BP神经网络模型和NDVI经验模型的估算精度高，农田、草地和林地站点估算结果的决定系数分别为0.899、0.858和0.863，BP神经网络模型的估算结果决定系数分别为：0.763、0.710和0.742，NDVI经验模型的精度最差，其估算结果的决定系数分别为0.622、0.536和0.637。为了验证SA-BP神经网络的适用性，选取独立研究区进行验证，结果显示验证精度较高，R²为0.842，RMSE为0.689 5，说明该模型外推能力较好。研究证明了基于模拟退火算法的BP神经网络模型提高了模型泛化能力，有效防止了BP神经网络模型滑入局部最小值，是提高高空间分辨率LAI估算精度的有效手段。

BP Neural Network based on Simulated Annealing Algorithm for High Resolution LAI Retrieval

The rapid development of satellite remote sensing technology makes it possible to obtain global large-scale Leaf Area Index （LAI）. However， it is difficult to estimate high-resolution LAI based on existing algorithms and data. In this paper， four typical research areas consist of three research areas for modeling validation and an independent research area for applicability validation including grassland， farmland and woodland were selected. Field data and correspondent satellite of the areas were then collected. The empirical model of NDVI vegetation index， the BP neural network model and the BP neural network based on simulated annealing algorithm were established and 30 m resolution LAI data were estimated with all models. Estimated results were validated with the Field data in three main research areas. The results indicated that NDVI empirical model has the worst accuracy in the three main research areas selected in this paper. The estimation accuracy of BP neural network model based on simulated annealing algorithm is higher than that of BP neural network model. The determinant coefficients of estimation results of farmland， grassland and woodland sites are 0.899， 0.858 and 0.863 respectively. The determinant coefficients of BP neural network model were 0.763， 0.710 and 0.742 respectively， while the determinant coefficients of NDVI empirical model were 0.622， 0.536 and 0.637 respectively. In order to verify the applicability of SA-BP neural network， an independent research area was selected for further verification. The results show that the validation accuracy is high，R² is 0.842 0， and RMSE is 0.689 5， which shows that the model has good extrapolation ability. This study proves that the BP neural network model based on simulated annealing algorithm improves the generalization ability of the model， effectively prevents the BP neural network model from sliding into the local minimum， and it is an effective method for LAI estimation in high spatial resolution.